sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2018 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         int pass;
  73 };
  74
  75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  76 static struct hammer2_mntlist hammer2_mntlist;
  77
  78 struct hammer2_pfslist hammer2_pfslist;
  79 struct hammer2_pfslist hammer2_spmplist;
  80 struct lock hammer2_mntlk;
  81
  82 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
  83 int hammer2_debug;
  84 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
  85 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
  86 int hammer2_dedup_enable = 1;
  87 int hammer2_always_compress = 0;        /* always try to compress */
  88 int hammer2_inval_enable = 0;
  89 int hammer2_flush_pipe = 100;
  90 int hammer2_dio_count;
  91 int hammer2_dio_limit = 256;
  92 int hammer2_bulkfree_tps = 5000;
  93 long hammer2_chain_allocs;
  94 long hammer2_chain_frees;
  95 long hammer2_limit_dirty_chains;
  96 long hammer2_count_modified_chains;
  97 long hammer2_iod_invals;
  98 long hammer2_iod_file_read;
  99 long hammer2_iod_meta_read;
 100 long hammer2_iod_indr_read;
 101 long hammer2_iod_fmap_read;
 102 long hammer2_iod_volu_read;
 103 long hammer2_iod_file_write;
 104 long hammer2_iod_file_wembed;
 105 long hammer2_iod_file_wzero;
 106 long hammer2_iod_file_wdedup;
 107 long hammer2_iod_meta_write;
 108 long hammer2_iod_indr_write;
 109 long hammer2_iod_fmap_write;
 110 long hammer2_iod_volu_write;
 111
 112 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 113 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 114                 "Buffer used for compression.");
 115
 116 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 117 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 118                 "Buffer used for decompression.");
 119
 120 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 121
 122 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
 123            &hammer2_supported_version, 0, "");
 124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 125            &hammer2_debug, 0, "");
 126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 127            &hammer2_cluster_meta_read, 0, "");
 128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 129            &hammer2_cluster_data_read, 0, "");
 130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 131            &hammer2_dedup_enable, 0, "");
 132 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 133            &hammer2_always_compress, 0, "");
 134 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 135            &hammer2_inval_enable, 0, "");
 136 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 137            &hammer2_flush_pipe, 0, "");
 138 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
 139            &hammer2_bulkfree_tps, 0, "");
 140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 141            &hammer2_chain_allocs, 0, "");
 142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 143            &hammer2_chain_frees, 0, "");
 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 145            &hammer2_limit_dirty_chains, 0, "");
 146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 147            &hammer2_count_modified_chains, 0, "");
 148 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 149            &hammer2_dio_count, 0, "");
 150 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
 151            &hammer2_dio_limit, 0, "");
 152
 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 154            &hammer2_iod_invals, 0, "");
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 156            &hammer2_iod_file_read, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 158            &hammer2_iod_meta_read, 0, "");
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 160            &hammer2_iod_indr_read, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 162            &hammer2_iod_fmap_read, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 164            &hammer2_iod_volu_read, 0, "");
 165
 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 167            &hammer2_iod_file_write, 0, "");
 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 169            &hammer2_iod_file_wembed, 0, "");
 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 171            &hammer2_iod_file_wzero, 0, "");
 172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 173            &hammer2_iod_file_wdedup, 0, "");
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 175            &hammer2_iod_meta_write, 0, "");
 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 177            &hammer2_iod_indr_write, 0, "");
 178 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 179            &hammer2_iod_fmap_write, 0, "");
 180 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 181            &hammer2_iod_volu_write, 0, "");
 182
 183 long hammer2_process_icrc32;
 184 long hammer2_process_xxhash64;
 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
 186            &hammer2_process_icrc32, 0, "");
 187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
 188            &hammer2_process_xxhash64, 0, "");
 189
 190 static int hammer2_vfs_init(struct vfsconf *conf);
 191 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 192 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 193                                 struct ucred *cred);
 194 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 195                                 struct vnode *, struct ucred *);
 196 static int hammer2_recovery(hammer2_dev_t *hmp);
 197 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 198 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 199 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 200                                 struct ucred *cred);
 201 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 202                                 struct ucred *cred);
 203 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 204                                 struct fid *fhp, struct vnode **vpp);
 205 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 206 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 207                                 int *exflagsp, struct ucred **credanonp);
 208
 209 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 210 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 211
 212 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 213
 214 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 215 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 216                                 hammer2_dev_t *hmp);
 217
 218 /*
 219  * HAMMER2 vfs operations.
 220  */
 221 static struct vfsops hammer2_vfsops = {
 222         .vfs_init       = hammer2_vfs_init,
 223         .vfs_uninit     = hammer2_vfs_uninit,
 224         .vfs_sync       = hammer2_vfs_sync,
 225         .vfs_mount      = hammer2_vfs_mount,
 226         .vfs_unmount    = hammer2_vfs_unmount,
 227         .vfs_root       = hammer2_vfs_root,
 228         .vfs_statfs     = hammer2_vfs_statfs,
 229         .vfs_statvfs    = hammer2_vfs_statvfs,
 230         .vfs_vget       = hammer2_vfs_vget,
 231         .vfs_vptofh     = hammer2_vfs_vptofh,
 232         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 233         .vfs_checkexp   = hammer2_vfs_checkexp
 234 };
 235
 236 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 237
 238 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 239 MODULE_VERSION(hammer2, 1);
 240
 241 static
 242 int
 243 hammer2_vfs_init(struct vfsconf *conf)
 244 {
 245         static struct objcache_malloc_args margs_read;
 246         static struct objcache_malloc_args margs_write;
 247         static struct objcache_malloc_args margs_vop;
 248
 249         int error;
 250
 251         error = 0;
 252
 253         /*
 254          * A large DIO cache is needed to retain dedup enablement masks.
 255          * The bulkfree code clears related masks as part of the disk block
 256          * recycling algorithm, preventing it from being used for a later
 257          * dedup.
 258          *
 259          * NOTE: A large buffer cache can actually interfere with dedup
 260          *       operation because we dedup based on media physical buffers
 261          *       and not logical buffers.  Try to make the DIO case large
 262          *       enough to avoid this problem, but also cap it.
 263          */
 264         hammer2_dio_limit = nbuf * 2;
 265         if (hammer2_dio_limit > 100000)
 266                 hammer2_dio_limit = 100000;
 267
 268         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 269                 error = EINVAL;
 270         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 271                 error = EINVAL;
 272         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 273                 error = EINVAL;
 274
 275         if (error)
 276                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 277
 278         margs_read.objsize = 65536;
 279         margs_read.mtype = M_HAMMER2_DEBUFFER;
 280
 281         margs_write.objsize = 32768;
 282         margs_write.mtype = M_HAMMER2_CBUFFER;
 283
 284         margs_vop.objsize = sizeof(hammer2_xop_t);
 285         margs_vop.mtype = M_HAMMER2;
 286
 287         /*
 288          * Note thaht for the XOPS cache we want backing store allocations
 289          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 290          * confusion), so use the backing store function that does it.  This
 291          * means that initial XOPS objects are zerod but REUSED objects are
 292          * not.  So we are responsible for cleaning the object up sufficiently
 293          * for our needs before objcache_put()ing it back (typically just the
 294          * FIFO indices).
 295          */
 296         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 297                                 0, 1, NULL, NULL, NULL,
 298                                 objcache_malloc_alloc,
 299                                 objcache_malloc_free,
 300                                 &margs_read);
 301         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 302                                 0, 1, NULL, NULL, NULL,
 303                                 objcache_malloc_alloc,
 304                                 objcache_malloc_free,
 305                                 &margs_write);
 306         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 307                                 0, 1, NULL, NULL, NULL,
 308                                 objcache_malloc_alloc_zero,
 309                                 objcache_malloc_free,
 310                                 &margs_vop);
 311
 312
 313         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 314         TAILQ_INIT(&hammer2_mntlist);
 315         TAILQ_INIT(&hammer2_pfslist);
 316         TAILQ_INIT(&hammer2_spmplist);
 317
 318         hammer2_limit_dirty_chains = maxvnodes / 10;
 319         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 320                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 321
 322         return (error);
 323 }
 324
 325 static
 326 int
 327 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 328 {
 329         objcache_destroy(cache_buffer_read);
 330         objcache_destroy(cache_buffer_write);
 331         objcache_destroy(cache_xops);
 332         return 0;
 333 }
 334
 335 /*
 336  * Core PFS allocator.  Used to allocate or reference the pmp structure
 337  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 338  * The pmp can be passed in or loaded by this function using the chain and
 339  * inode data.
 340  *
 341  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 342  * transactions.  Note that synchronization does not use this field.
 343  * (typically frontend operations and synchronization cannot run on the
 344  * same PFS node at the same time).
 345  *
 346  * XXX check locking
 347  */
 348 hammer2_pfs_t *
 349 hammer2_pfsalloc(hammer2_chain_t *chain,
 350                  const hammer2_inode_data_t *ripdata,
 351                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 352 {
 353         hammer2_pfs_t *pmp;
 354         hammer2_inode_t *iroot;
 355         int count;
 356         int i;
 357         int j;
 358
 359         pmp = NULL;
 360
 361         /*
 362          * Locate or create the PFS based on the cluster id.  If ripdata
 363          * is NULL this is a spmp which is unique and is always allocated.
 364          *
 365          * If the device is mounted in local mode all PFSs are considered
 366          * independent and not part of any cluster (for debugging only).
 367          */
 368         if (ripdata) {
 369                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 370                         if (force_local != pmp->force_local)
 371                                 continue;
 372                         if (force_local == NULL &&
 373                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 374                                  sizeof(pmp->pfs_clid)) == 0) {
 375                                         break;
 376                         } else if (force_local && pmp->pfs_names[0] &&
 377                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 378                                         break;
 379                         }
 380                 }
 381         }
 382
 383         if (pmp == NULL) {
 384                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 385                 pmp->force_local = force_local;
 386                 hammer2_trans_manage_init(pmp);
 387                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 388                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 389                 lockinit(&pmp->lock, "pfslk", 0, 0);
 390                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 391                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 392                 spin_init(&pmp->xop_spin, "h2xop");
 393                 spin_init(&pmp->lru_spin, "h2lru");
 394                 RB_INIT(&pmp->inum_tree);
 395                 TAILQ_INIT(&pmp->sideq);
 396                 TAILQ_INIT(&pmp->lru_list);
 397                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 398
 399                 /*
 400                  * Distribute backend operations to threads
 401                  */
 402                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 403                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 404
 405                 /*
 406                  * Save the last media transaction id for the flusher.  Set
 407                  * initial
 408                  */
 409                 if (ripdata) {
 410                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 411                         TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 412                 } else {
 413                         pmp->flags |= HAMMER2_PMPF_SPMP;
 414                         TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
 415                 }
 416
 417                 /*
 418                  * The synchronization thread may start too early, make
 419                  * sure it stays frozen until we are ready to let it go.
 420                  * XXX
 421                  */
 422                 /*
 423                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 424                                          HAMMER2_THREAD_REMASTER;
 425                 */
 426         }
 427
 428         /*
 429          * Create the PFS's root inode and any missing XOP helper threads.
 430          */
 431         if ((iroot = pmp->iroot) == NULL) {
 432                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 433                 if (ripdata)
 434                         iroot->meta = ripdata->meta;
 435                 pmp->iroot = iroot;
 436                 hammer2_inode_ref(iroot);
 437                 hammer2_inode_unlock(iroot);
 438         }
 439
 440         /*
 441          * Stop here if no chain is passed in.
 442          */
 443         if (chain == NULL)
 444                 goto done;
 445
 446         /*
 447          * When a chain is passed in we must add it to the PFS's root
 448          * inode, update pmp->pfs_types[], and update the syncronization
 449          * threads.
 450          *
 451          * When forcing local mode, mark the PFS as a MASTER regardless.
 452          *
 453          * At the moment empty spots can develop due to removals or failures.
 454          * Ultimately we want to re-fill these spots but doing so might
 455          * confused running code. XXX
 456          */
 457         hammer2_inode_ref(iroot);
 458         hammer2_mtx_ex(&iroot->lock);
 459         j = iroot->cluster.nchains;
 460
 461         if (j == HAMMER2_MAXCLUSTER) {
 462                 kprintf("hammer2_mount: cluster full!\n");
 463                 /* XXX fatal error? */
 464         } else {
 465                 KKASSERT(chain->pmp == NULL);
 466                 chain->pmp = pmp;
 467                 hammer2_chain_ref(chain);
 468                 iroot->cluster.array[j].chain = chain;
 469                 if (force_local)
 470                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 471                 else
 472                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 473                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 474                 pmp->pfs_hmps[j] = chain->hmp;
 475
 476                 /*
 477                  * If the PFS is already mounted we must account
 478                  * for the mount_count here.
 479                  */
 480                 if (pmp->mp)
 481                         ++chain->hmp->mount_count;
 482
 483                 /*
 484                  * May have to fixup dirty chain tracking.  Previous
 485                  * pmp was NULL so nothing to undo.
 486                  */
 487                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 488                         hammer2_pfs_memory_inc(pmp);
 489                 ++j;
 490         }
 491         iroot->cluster.nchains = j;
 492
 493         /*
 494          * Update nmasters from any PFS inode which is part of the cluster.
 495          * It is possible that this will result in a value which is too
 496          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 497          * override this value later on.
 498          *
 499          * (This informs us of masters that might not currently be
 500          *  discoverable by this mount).
 501          */
 502         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 503                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 504         }
 505
 506         /*
 507          * Count visible masters.  Masters are usually added with
 508          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 509          * are more (XXX and must update the master inodes).
 510          */
 511         count = 0;
 512         for (i = 0; i < iroot->cluster.nchains; ++i) {
 513                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 514                         ++count;
 515         }
 516         if (pmp->pfs_nmasters < count)
 517                 pmp->pfs_nmasters = count;
 518
 519         /*
 520          * Create missing synchronization and support threads.
 521          *
 522          * Single-node masters (including snapshots) have nothing to
 523          * synchronize and do not require this thread.
 524          *
 525          * Multi-node masters or any number of soft masters, slaves, copy,
 526          * or other PFS types need the thread.
 527          *
 528          * Each thread is responsible for its particular cluster index.
 529          * We use independent threads so stalls or mismatches related to
 530          * any given target do not affect other targets.
 531          */
 532         for (i = 0; i < iroot->cluster.nchains; ++i) {
 533                 /*
 534                  * Single-node masters (including snapshots) have nothing
 535                  * to synchronize and will make direct xops support calls,
 536                  * thus they do not require this thread.
 537                  *
 538                  * Note that there can be thousands of snapshots.  We do not
 539                  * want to create thousands of threads.
 540                  */
 541                 if (pmp->pfs_nmasters <= 1 &&
 542                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 543                         continue;
 544                 }
 545
 546                 /*
 547                  * Sync support thread
 548                  */
 549                 if (pmp->sync_thrs[i].td == NULL) {
 550                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 551                                            "h2nod", i, -1,
 552                                            hammer2_primary_sync_thread);
 553                 }
 554         }
 555
 556         /*
 557          * Create missing Xop threads
 558          *
 559          * NOTE: We create helper threads for all mounted PFSs or any
 560          *       PFSs with 2+ nodes (so the sync thread can update them,
 561          *       even if not mounted).
 562          */
 563         if (pmp->mp || iroot->cluster.nchains >= 2)
 564                 hammer2_xop_helper_create(pmp);
 565
 566         hammer2_mtx_unlock(&iroot->lock);
 567         hammer2_inode_drop(iroot);
 568 done:
 569         return pmp;
 570 }
 571
 572 /*
 573  * Deallocate an element of a probed PFS.  If destroying and this is a
 574  * MASTER, adjust nmasters.
 575  *
 576  * This function does not physically destroy the PFS element in its device
 577  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 578  */
 579 void
 580 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 581 {
 582         hammer2_inode_t *iroot;
 583         hammer2_chain_t *chain;
 584         int j;
 585
 586         /*
 587          * Cleanup our reference on iroot.  iroot is (should) not be needed
 588          * by the flush code.
 589          */
 590         iroot = pmp->iroot;
 591         if (iroot) {
 592                 /*
 593                  * Stop synchronizing
 594                  *
 595                  * XXX flush after acquiring the iroot lock.
 596                  * XXX clean out the cluster index from all inode structures.
 597                  */
 598                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 599
 600                 /*
 601                  * Remove the cluster index from the group.  If destroying
 602                  * the PFS and this is a master, adjust pfs_nmasters.
 603                  */
 604                 hammer2_mtx_ex(&iroot->lock);
 605                 chain = iroot->cluster.array[clindex].chain;
 606                 iroot->cluster.array[clindex].chain = NULL;
 607
 608                 switch(pmp->pfs_types[clindex]) {
 609                 case HAMMER2_PFSTYPE_MASTER:
 610                         if (destroying && pmp->pfs_nmasters > 0)
 611                                 --pmp->pfs_nmasters;
 612                         /* XXX adjust ripdata->meta.pfs_nmasters */
 613                         break;
 614                 default:
 615                         break;
 616                 }
 617                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 618
 619                 hammer2_mtx_unlock(&iroot->lock);
 620
 621                 /*
 622                  * Release the chain.
 623                  */
 624                 if (chain) {
 625                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 626                         hammer2_chain_drop(chain);
 627                 }
 628
 629                 /*
 630                  * Terminate all XOP threads for the cluster index.
 631                  */
 632                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 633                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 634         }
 635 }
 636
 637 /*
 638  * Destroy a PFS, typically only occurs after the last mount on a device
 639  * has gone away.
 640  */
 641 static void
 642 hammer2_pfsfree(hammer2_pfs_t *pmp)
 643 {
 644         hammer2_inode_t *iroot;
 645         hammer2_chain_t *chain;
 646         int i;
 647         int j;
 648
 649         /*
 650          * Cleanup our reference on iroot.  iroot is (should) not be needed
 651          * by the flush code.
 652          */
 653         if (pmp->flags & HAMMER2_PMPF_SPMP)
 654                 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
 655         else
 656                 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 657
 658         iroot = pmp->iroot;
 659         if (iroot) {
 660                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 661                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 662                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 663                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 664                 }
 665 #if REPORT_REFS_ERRORS
 666                 if (pmp->iroot->refs != 1)
 667                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 668                                 pmp->iroot, pmp->iroot->refs);
 669 #else
 670                 KKASSERT(pmp->iroot->refs == 1);
 671 #endif
 672                 /* ref for pmp->iroot */
 673                 hammer2_inode_drop(pmp->iroot);
 674                 pmp->iroot = NULL;
 675         }
 676
 677         /*
 678          * Cleanup chains remaining on LRU list.
 679          */
 680         hammer2_spin_ex(&pmp->lru_spin);
 681         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 682                 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
 683                 atomic_add_int(&pmp->lru_count, -1);
 684                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
 685                 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
 686                 hammer2_chain_ref(chain);
 687                 hammer2_spin_unex(&pmp->lru_spin);
 688                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 689                 hammer2_chain_drop(chain);
 690                 hammer2_spin_ex(&pmp->lru_spin);
 691         }
 692         hammer2_spin_unex(&pmp->lru_spin);
 693
 694         /*
 695          * Free remaining pmp resources
 696          */
 697         kmalloc_destroy(&pmp->mmsg);
 698         kmalloc_destroy(&pmp->minode);
 699
 700         kfree(pmp, M_HAMMER2);
 701 }
 702
 703 /*
 704  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 705  * empty is terminated and freed.
 706  *
 707  * XXX inefficient.
 708  */
 709 static void
 710 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
 711 {
 712         hammer2_pfs_t *pmp;
 713         hammer2_inode_t *iroot;
 714         hammer2_chain_t *rchain;
 715         int didfreeze;
 716         int i;
 717         int j;
 718         struct hammer2_pfslist *wlist;
 719
 720         if (which == 0)
 721                 wlist = &hammer2_pfslist;
 722         else
 723                 wlist = &hammer2_spmplist;
 724 again:
 725         TAILQ_FOREACH(pmp, wlist, mntentry) {
 726                 if ((iroot = pmp->iroot) == NULL)
 727                         continue;
 728                 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
 729                 hammer2_inode_run_sideq(pmp, 1);
 730                 hammer2_bioq_sync(pmp);
 731                 hammer2_trans_done(pmp);
 732
 733                 /*
 734                  * Determine if this PFS is affected.  If it is we must
 735                  * freeze all management threads and lock its iroot.
 736                  *
 737                  * Freezing a management thread forces it idle, operations
 738                  * in-progress will be aborted and it will have to start
 739                  * over again when unfrozen, or exit if told to exit.
 740                  */
 741                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 742                         if (pmp->pfs_hmps[i] == hmp)
 743                                 break;
 744                 }
 745                 if (i != HAMMER2_MAXCLUSTER) {
 746                         /*
 747                          * Make sure all synchronization threads are locked
 748                          * down.
 749                          */
 750                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 751                                 if (pmp->pfs_hmps[i] == NULL)
 752                                         continue;
 753                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 754                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 755                                         hammer2_thr_freeze_async(
 756                                                 &pmp->xop_groups[j].thrs[i]);
 757                                 }
 758                         }
 759                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 760                                 if (pmp->pfs_hmps[i] == NULL)
 761                                         continue;
 762                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 763                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 764                                         hammer2_thr_freeze(
 765                                                 &pmp->xop_groups[j].thrs[i]);
 766                                 }
 767                         }
 768
 769                         /*
 770                          * Lock the inode and clean out matching chains.
 771                          * Note that we cannot use hammer2_inode_lock_*()
 772                          * here because that would attempt to validate the
 773                          * cluster that we are in the middle of ripping
 774                          * apart.
 775                          *
 776                          * WARNING! We are working directly on the inodes
 777                          *          embedded cluster.
 778                          */
 779                         hammer2_mtx_ex(&iroot->lock);
 780
 781                         /*
 782                          * Remove the chain from matching elements of the PFS.
 783                          */
 784                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 785                                 if (pmp->pfs_hmps[i] != hmp)
 786                                         continue;
 787                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 788                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 789                                         hammer2_thr_delete(
 790                                                 &pmp->xop_groups[j].thrs[i]);
 791                                 }
 792                                 rchain = iroot->cluster.array[i].chain;
 793                                 iroot->cluster.array[i].chain = NULL;
 794                                 pmp->pfs_types[i] = 0;
 795                                 if (pmp->pfs_names[i]) {
 796                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 797                                         pmp->pfs_names[i] = NULL;
 798                                 }
 799                                 if (rchain) {
 800                                         hammer2_chain_drop(rchain);
 801                                         /* focus hint */
 802                                         if (iroot->cluster.focus == rchain)
 803                                                 iroot->cluster.focus = NULL;
 804                                 }
 805                                 pmp->pfs_hmps[i] = NULL;
 806                         }
 807                         hammer2_mtx_unlock(&iroot->lock);
 808                         didfreeze = 1;  /* remaster, unfreeze down below */
 809                 } else {
 810                         didfreeze = 0;
 811                 }
 812
 813                 /*
 814                  * Cleanup trailing chains.  Gaps may remain.
 815                  */
 816                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 817                         if (pmp->pfs_hmps[i])
 818                                 break;
 819                 }
 820                 iroot->cluster.nchains = i + 1;
 821
 822                 /*
 823                  * If the PMP has no elements remaining we can destroy it.
 824                  * (this will transition management threads from frozen->exit).
 825                  */
 826                 if (iroot->cluster.nchains == 0) {
 827                         /*
 828                          * If this was the hmp's spmp, we need to clean
 829                          * a little more stuff out.
 830                          */
 831                         if (hmp->spmp == pmp) {
 832                                 hmp->spmp = NULL;
 833                                 hmp->vchain.pmp = NULL;
 834                                 hmp->fchain.pmp = NULL;
 835                         }
 836
 837                         /*
 838                          * Free the pmp and restart the loop
 839                          */
 840                         hammer2_pfsfree(pmp);
 841                         goto again;
 842                 }
 843
 844                 /*
 845                  * If elements still remain we need to set the REMASTER
 846                  * flag and unfreeze it.
 847                  */
 848                 if (didfreeze) {
 849                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 850                                 if (pmp->pfs_hmps[i] == NULL)
 851                                         continue;
 852                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 853                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 854                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 855                                         hammer2_thr_remaster(
 856                                                 &pmp->xop_groups[j].thrs[i]);
 857                                         hammer2_thr_unfreeze(
 858                                                 &pmp->xop_groups[j].thrs[i]);
 859                                 }
 860                         }
 861                 }
 862         }
 863 }
 864
 865 /*
 866  * Mount or remount HAMMER2 fileystem from physical media
 867  *
 868  *      mountroot
 869  *              mp              mount point structure
 870  *              path            NULL
 871  *              data            <unused>
 872  *              cred            <unused>
 873  *
 874  *      mount
 875  *              mp              mount point structure
 876  *              path            path to mount point
 877  *              data            pointer to argument structure in user space
 878  *                      volume  volume path (device@LABEL form)
 879  *                      hflags  user mount flags
 880  *              cred            user credentials
 881  *
 882  * RETURNS:     0       Success
 883  *              !0      error number
 884  */
 885 static
 886 int
 887 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 888                   struct ucred *cred)
 889 {
 890         struct hammer2_mount_info info;
 891         hammer2_pfs_t *pmp;
 892         hammer2_pfs_t *spmp;
 893         hammer2_dev_t *hmp;
 894         hammer2_dev_t *force_local;
 895         hammer2_key_t key_next;
 896         hammer2_key_t key_dummy;
 897         hammer2_key_t lhc;
 898         struct vnode *devvp;
 899         struct nlookupdata nd;
 900         hammer2_chain_t *parent;
 901         hammer2_chain_t *chain;
 902         hammer2_cluster_t *cluster;
 903         const hammer2_inode_data_t *ripdata;
 904         hammer2_blockref_t bref;
 905         struct file *fp;
 906         char devstr[MNAMELEN];
 907         size_t size;
 908         size_t done;
 909         char *dev;
 910         char *label;
 911         int ronly = 1;
 912         int error;
 913         int i;
 914
 915         hmp = NULL;
 916         pmp = NULL;
 917         dev = NULL;
 918         label = NULL;
 919         devvp = NULL;
 920
 921         kprintf("hammer2_mount\n");
 922
 923         if (path == NULL) {
 924                 /*
 925                  * Root mount
 926                  */
 927                 bzero(&info, sizeof(info));
 928                 info.cluster_fd = -1;
 929                 ksnprintf(devstr, sizeof(devstr), "%s",
 930                           mp->mnt_stat.f_mntfromname);
 931                 kprintf("hammer2_mount: root '%s'\n", devstr);
 932         } else {
 933                 /*
 934                  * Non-root mount or updating a mount
 935                  */
 936                 error = copyin(data, &info, sizeof(info));
 937                 if (error)
 938                         return (error);
 939
 940                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 941                 if (error)
 942                         return (error);
 943         }
 944
 945         /*
 946          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 947          * if no label specified, based on the partition id.  Error out if no
 948          * label or device (with partition id) is specified.  This is strictly
 949          * a convenience to match the default label created by newfs_hammer2,
 950          * our preference is that a label always be specified.
 951          *
 952          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
 953          *       that does not specify a device, as long as some H2 label
 954          *       has already been mounted from that device.  This makes
 955          *       mounting snapshots a lot easier.
 956          */
 957         dev = devstr;
 958         label = strchr(devstr, '@');
 959         if (label && ((label + 1) - dev) > done)
 960                 return (EINVAL);
 961         if (label == NULL || label[1] == 0) {
 962                 char slice;
 963
 964                 if (label == NULL)
 965                         label = devstr + strlen(devstr);
 966                 slice = label[-1];
 967                 switch(slice) {
 968                 case 'a':
 969                         label = "BOOT";
 970                         break;
 971                 case 'd':
 972                         label = "ROOT";
 973                         break;
 974                 default:
 975                         label = "DATA";
 976                         break;
 977                 }
 978         } else {
 979                 *label = '\0';
 980                 label++;
 981         }
 982
 983         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
 984                 dev, label, (mp->mnt_flag & MNT_RDONLY));
 985
 986         if (mp->mnt_flag & MNT_UPDATE) {
 987                 /*
 988                  * Update mount.  Note that pmp->iroot->cluster is
 989                  * an inode-embedded cluster and thus cannot be
 990                  * directly locked.
 991                  *
 992                  * XXX HAMMER2 needs to implement NFS export via
 993                  *     mountctl.
 994                  */
 995                 pmp = MPTOPMP(mp);
 996                 pmp->hflags = info.hflags;
 997                 cluster = &pmp->iroot->cluster;
 998                 for (i = 0; i < cluster->nchains; ++i) {
 999                         if (cluster->array[i].chain == NULL)
1000                                 continue;
1001                         hmp = cluster->array[i].chain->hmp;
1002                         devvp = hmp->devvp;
1003                         error = hammer2_remount(hmp, mp, path,
1004                                                 devvp, cred);
1005                         if (error)
1006                                 break;
1007                 }
1008
1009                 return error;
1010         }
1011
1012         /*
1013          * HMP device mount
1014          *
1015          * If a path is specified and dev is not an empty string, lookup the
1016          * name and verify that it referes to a block device.
1017          *
1018          * If a path is specified and dev is an empty string we fall through
1019          * and locate the label in the hmp search.
1020          */
1021         if (path && *dev != 0) {
1022                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1023                 if (error == 0)
1024                         error = nlookup(&nd);
1025                 if (error == 0)
1026                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1027                 nlookup_done(&nd);
1028         } else if (path == NULL) {
1029                 /* root mount */
1030                 cdev_t cdev = kgetdiskbyname(dev);
1031                 error = bdevvp(cdev, &devvp);
1032                 if (error)
1033                         kprintf("hammer2: cannot find '%s'\n", dev);
1034         } else {
1035                 /*
1036                  * We will locate the hmp using the label in the hmp loop.
1037                  */
1038                 error = 0;
1039         }
1040
1041         /*
1042          * Make sure its a block device.  Do not check to see if it is
1043          * already mounted until we determine that its a fresh H2 device.
1044          */
1045         if (error == 0 && devvp) {
1046                 vn_isdisk(devvp, &error);
1047         }
1048
1049         /*
1050          * Determine if the device has already been mounted.  After this
1051          * check hmp will be non-NULL if we are doing the second or more
1052          * hammer2 mounts from the same device.
1053          */
1054         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1055         if (devvp) {
1056                 /*
1057                  * Match the device.  Due to the way devfs works,
1058                  * we may not be able to directly match the vnode pointer,
1059                  * so also check to see if the underlying device matches.
1060                  */
1061                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1062                         if (hmp->devvp == devvp)
1063                                 break;
1064                         if (devvp->v_rdev &&
1065                             hmp->devvp->v_rdev == devvp->v_rdev) {
1066                                 break;
1067                         }
1068                 }
1069
1070                 /*
1071                  * If no match this may be a fresh H2 mount, make sure
1072                  * the device is not mounted on anything else.
1073                  */
1074                 if (hmp == NULL)
1075                         error = vfs_mountedon(devvp);
1076         } else if (error == 0) {
1077                 /*
1078                  * Match the label to a pmp already probed.
1079                  */
1080                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1081                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1082                                 if (pmp->pfs_names[i] &&
1083                                     strcmp(pmp->pfs_names[i], label) == 0) {
1084                                         hmp = pmp->pfs_hmps[i];
1085                                         break;
1086                                 }
1087                         }
1088                         if (hmp)
1089                                 break;
1090                 }
1091                 if (hmp == NULL)
1092                         error = ENOENT;
1093         }
1094
1095         /*
1096          * Open the device if this isn't a secondary mount and construct
1097          * the H2 device mount (hmp).
1098          */
1099         if (hmp == NULL) {
1100                 hammer2_chain_t *schain;
1101                 hammer2_xid_t xid;
1102
1103                 if (error == 0 && vcount(devvp) > 0) {
1104                         kprintf("Primary device already has references\n");
1105                         error = EBUSY;
1106                 }
1107
1108                 /*
1109                  * Now open the device
1110                  */
1111                 if (error == 0) {
1112                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1113                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1114                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1115                         if (error == 0) {
1116                                 error = VOP_OPEN(devvp,
1117                                              (ronly ? FREAD : FREAD | FWRITE),
1118                                              FSCRED, NULL);
1119                         }
1120                         vn_unlock(devvp);
1121                 }
1122                 if (error && devvp) {
1123                         vrele(devvp);
1124                         devvp = NULL;
1125                 }
1126                 if (error) {
1127                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1128                         return error;
1129                 }
1130                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1131                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1132                 hmp->ronly = ronly;
1133                 hmp->devvp = devvp;
1134                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1135                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1136                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1137                 RB_INIT(&hmp->iotree);
1138                 spin_init(&hmp->io_spin, "hm2mount_io");
1139                 spin_init(&hmp->list_spin, "hm2mount_list");
1140                 TAILQ_INIT(&hmp->flushq);
1141
1142                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1143                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1144                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1145
1146                 /*
1147                  * vchain setup. vchain.data is embedded.
1148                  * vchain.refs is initialized and will never drop to 0.
1149                  *
1150                  * NOTE! voldata is not yet loaded.
1151                  */
1152                 hmp->vchain.hmp = hmp;
1153                 hmp->vchain.refs = 1;
1154                 hmp->vchain.data = (void *)&hmp->voldata;
1155                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1156                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1157                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1158
1159                 hammer2_chain_core_init(&hmp->vchain);
1160                 /* hmp->vchain.u.xxx is left NULL */
1161
1162                 /*
1163                  * fchain setup.  fchain.data is embedded.
1164                  * fchain.refs is initialized and will never drop to 0.
1165                  *
1166                  * The data is not used but needs to be initialized to
1167                  * pass assertion muster.  We use this chain primarily
1168                  * as a placeholder for the freemap's top-level RBTREE
1169                  * so it does not interfere with the volume's topology
1170                  * RBTREE.
1171                  */
1172                 hmp->fchain.hmp = hmp;
1173                 hmp->fchain.refs = 1;
1174                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1175                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1176                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1177                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1178                 hmp->fchain.bref.methods =
1179                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1180                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1181
1182                 hammer2_chain_core_init(&hmp->fchain);
1183                 /* hmp->fchain.u.xxx is left NULL */
1184
1185                 /*
1186                  * Install the volume header and initialize fields from
1187                  * voldata.
1188                  */
1189                 error = hammer2_install_volume_header(hmp);
1190                 if (error) {
1191                         hammer2_unmount_helper(mp, NULL, hmp);
1192                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1193                         hammer2_vfs_unmount(mp, MNT_FORCE);
1194                         return error;
1195                 }
1196
1197                 /*
1198                  * Really important to get these right or flush will get
1199                  * confused.
1200                  */
1201                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1202                 kprintf("alloc spmp %p tid %016jx\n",
1203                         hmp->spmp, hmp->voldata.mirror_tid);
1204                 spmp = hmp->spmp;
1205
1206                 /*
1207                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1208                  * is inherited from the volume header.
1209                  */
1210                 xid = 0;
1211                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1212                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1213                 hmp->vchain.pmp = spmp;
1214                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1215                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1216                 hmp->fchain.pmp = spmp;
1217
1218                 /*
1219                  * First locate the super-root inode, which is key 0
1220                  * relative to the volume header's blockset.
1221                  *
1222                  * Then locate the root inode by scanning the directory keyspace
1223                  * represented by the label.
1224                  */
1225                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1226                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1227                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1228                                       &error, 0);
1229                 hammer2_chain_lookup_done(parent);
1230                 if (schain == NULL) {
1231                         kprintf("hammer2_mount: invalid super-root\n");
1232                         hammer2_unmount_helper(mp, NULL, hmp);
1233                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1234                         hammer2_vfs_unmount(mp, MNT_FORCE);
1235                         return EINVAL;
1236                 }
1237                 if (schain->error) {
1238                         kprintf("hammer2_mount: error %s reading super-root\n",
1239                                 hammer2_error_str(schain->error));
1240                         hammer2_chain_unlock(schain);
1241                         hammer2_chain_drop(schain);
1242                         schain = NULL;
1243                         hammer2_unmount_helper(mp, NULL, hmp);
1244                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1245                         hammer2_vfs_unmount(mp, MNT_FORCE);
1246                         return EINVAL;
1247                 }
1248
1249                 /*
1250                  * The super-root always uses an inode_tid of 1 when
1251                  * creating PFSs.
1252                  */
1253                 spmp->inode_tid = 1;
1254                 spmp->modify_tid = schain->bref.modify_tid + 1;
1255
1256                 /*
1257                  * Sanity-check schain's pmp and finish initialization.
1258                  * Any chain belonging to the super-root topology should
1259                  * have a NULL pmp (not even set to spmp).
1260                  */
1261                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1262                 KKASSERT(schain->pmp == NULL);
1263                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1264
1265                 /*
1266                  * Replace the dummy spmp->iroot with a real one.  It's
1267                  * easier to just do a wholesale replacement than to try
1268                  * to update the chain and fixup the iroot fields.
1269                  *
1270                  * The returned inode is locked with the supplied cluster.
1271                  */
1272                 cluster = hammer2_cluster_from_chain(schain);
1273                 hammer2_inode_drop(spmp->iroot);
1274                 spmp->iroot = NULL;
1275                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1276                 spmp->spmp_hmp = hmp;
1277                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1278                 spmp->pfs_hmps[0] = hmp;
1279                 hammer2_inode_ref(spmp->iroot);
1280                 hammer2_inode_unlock(spmp->iroot);
1281                 hammer2_cluster_unlock(cluster);
1282                 hammer2_cluster_drop(cluster);
1283                 schain = NULL;
1284                 /* leave spmp->iroot with one ref */
1285
1286                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1287                         error = hammer2_recovery(hmp);
1288                         /* XXX do something with error */
1289                 }
1290                 hammer2_update_pmps(hmp);
1291                 hammer2_iocom_init(hmp);
1292                 hammer2_bulkfree_init(hmp);
1293
1294                 /*
1295                  * Ref the cluster management messaging descriptor.  The mount
1296                  * program deals with the other end of the communications pipe.
1297                  *
1298                  * Root mounts typically do not supply one.
1299                  */
1300                 if (info.cluster_fd >= 0) {
1301                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1302                         if (fp) {
1303                                 hammer2_cluster_reconnect(hmp, fp);
1304                         } else {
1305                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1306                         }
1307                 }
1308         } else {
1309                 spmp = hmp->spmp;
1310                 if (info.hflags & HMNT2_DEVFLAGS) {
1311                         kprintf("hammer2: Warning: mount flags pertaining "
1312                                 "to the whole device may only be specified "
1313                                 "on the first mount of the device: %08x\n",
1314                                 info.hflags & HMNT2_DEVFLAGS);
1315                 }
1316         }
1317
1318         /*
1319          * Force local mount (disassociate all PFSs from their clusters).
1320          * Used primarily for debugging.
1321          */
1322         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1323
1324         /*
1325          * Lookup the mount point under the media-localized super-root.
1326          * Scanning hammer2_pfslist doesn't help us because it represents
1327          * PFS cluster ids which can aggregate several named PFSs together.
1328          *
1329          * cluster->pmp will incorrectly point to spmp and must be fixed
1330          * up later on.
1331          */
1332         hammer2_inode_lock(spmp->iroot, 0);
1333         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1334         lhc = hammer2_dirhash(label, strlen(label));
1335         chain = hammer2_chain_lookup(&parent, &key_next,
1336                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1337                                      &error, 0);
1338         while (chain) {
1339                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1340                     strcmp(label, chain->data->ipdata.filename) == 0) {
1341                         break;
1342                 }
1343                 chain = hammer2_chain_next(&parent, chain, &key_next,
1344                                             key_next,
1345                                             lhc + HAMMER2_DIRHASH_LOMASK,
1346                                             &error, 0);
1347         }
1348         if (parent) {
1349                 hammer2_chain_unlock(parent);
1350                 hammer2_chain_drop(parent);
1351         }
1352         hammer2_inode_unlock(spmp->iroot);
1353
1354         /*
1355          * PFS could not be found?
1356          */
1357         if (chain == NULL) {
1358                 if (error)
1359                         kprintf("hammer2_mount: PFS label I/O error\n");
1360                 else
1361                         kprintf("hammer2_mount: PFS label not found\n");
1362                 hammer2_unmount_helper(mp, NULL, hmp);
1363                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1364                 hammer2_vfs_unmount(mp, MNT_FORCE);
1365
1366                 return EINVAL;
1367         }
1368
1369         /*
1370          * Acquire the pmp structure (it should have already been allocated
1371          * via hammer2_update_pmps() so do not pass cluster in to add to
1372          * available chains).
1373          *
1374          * Check if the cluster has already been mounted.  A cluster can
1375          * only be mounted once, use null mounts to mount additional copies.
1376          */
1377         if (chain->error) {
1378                 kprintf("hammer2_mount: PFS label I/O error\n");
1379         } else {
1380                 ripdata = &chain->data->ipdata;
1381                 bref = chain->bref;
1382                 pmp = hammer2_pfsalloc(NULL, ripdata,
1383                                        bref.modify_tid, force_local);
1384         }
1385         hammer2_chain_unlock(chain);
1386         hammer2_chain_drop(chain);
1387
1388         /*
1389          * Finish the mount
1390          */
1391         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1392
1393         if (pmp->mp) {
1394                 kprintf("hammer2_mount: PFS already mounted!\n");
1395                 hammer2_unmount_helper(mp, NULL, hmp);
1396                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1397                 hammer2_vfs_unmount(mp, MNT_FORCE);
1398
1399                 return EBUSY;
1400         }
1401
1402         pmp->hflags = info.hflags;
1403         mp->mnt_flag |= MNT_LOCAL;
1404         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1405         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1406
1407         /*
1408          * required mount structure initializations
1409          */
1410         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1411         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1412
1413         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1414         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1415
1416         /*
1417          * Optional fields
1418          */
1419         mp->mnt_iosize_max = MAXPHYS;
1420
1421         /*
1422          * Connect up mount pointers.
1423          */
1424         hammer2_mount_helper(mp, pmp);
1425
1426         lockmgr(&hammer2_mntlk, LK_RELEASE);
1427
1428         /*
1429          * Finish setup
1430          */
1431         vfs_getnewfsid(mp);
1432         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1433         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1434         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1435
1436         if (path) {
1437                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1438                           MNAMELEN - 1, &size);
1439                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1440         } /* else root mount, already in there */
1441
1442         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1443         if (path) {
1444                 copyinstr(path, mp->mnt_stat.f_mntonname,
1445                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1446                           &size);
1447         } else {
1448                 /* root mount */
1449                 mp->mnt_stat.f_mntonname[0] = '/';
1450         }
1451
1452         /*
1453          * Initial statfs to prime mnt_stat.
1454          */
1455         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1456
1457         return 0;
1458 }
1459
1460 /*
1461  * Scan PFSs under the super-root and create hammer2_pfs structures.
1462  */
1463 static
1464 void
1465 hammer2_update_pmps(hammer2_dev_t *hmp)
1466 {
1467         const hammer2_inode_data_t *ripdata;
1468         hammer2_chain_t *parent;
1469         hammer2_chain_t *chain;
1470         hammer2_blockref_t bref;
1471         hammer2_dev_t *force_local;
1472         hammer2_pfs_t *spmp;
1473         hammer2_pfs_t *pmp;
1474         hammer2_key_t key_next;
1475         int error;
1476
1477         /*
1478          * Force local mount (disassociate all PFSs from their clusters).
1479          * Used primarily for debugging.
1480          */
1481         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1482
1483         /*
1484          * Lookup mount point under the media-localized super-root.
1485          *
1486          * cluster->pmp will incorrectly point to spmp and must be fixed
1487          * up later on.
1488          */
1489         spmp = hmp->spmp;
1490         hammer2_inode_lock(spmp->iroot, 0);
1491         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1492         chain = hammer2_chain_lookup(&parent, &key_next,
1493                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1494                                          &error, 0);
1495         while (chain) {
1496                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1497                         continue;
1498                 if (chain->error) {
1499                         kprintf("I/O error scanning PFS labels\n");
1500                 } else {
1501                         ripdata = &chain->data->ipdata;
1502                         bref = chain->bref;
1503
1504                         pmp = hammer2_pfsalloc(chain, ripdata,
1505                                                bref.modify_tid, force_local);
1506                 }
1507                 chain = hammer2_chain_next(&parent, chain, &key_next,
1508                                            key_next, HAMMER2_KEY_MAX,
1509                                            &error, 0);
1510         }
1511         if (parent) {
1512                 hammer2_chain_unlock(parent);
1513                 hammer2_chain_drop(parent);
1514         }
1515         hammer2_inode_unlock(spmp->iroot);
1516 }
1517
1518 static
1519 int
1520 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1521                 struct vnode *devvp, struct ucred *cred)
1522 {
1523         int error;
1524
1525         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1526                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1527                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1528                 vn_unlock(devvp);
1529                 error = hammer2_recovery(hmp);
1530                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1531                 if (error == 0) {
1532                         VOP_CLOSE(devvp, FREAD, NULL);
1533                         hmp->ronly = 0;
1534                 } else {
1535                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1536                 }
1537                 vn_unlock(devvp);
1538         } else {
1539                 error = 0;
1540         }
1541         return error;
1542 }
1543
1544 static
1545 int
1546 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1547 {
1548         hammer2_pfs_t *pmp;
1549         int flags;
1550         int error = 0;
1551
1552         pmp = MPTOPMP(mp);
1553
1554         if (pmp == NULL)
1555                 return(0);
1556
1557         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1558
1559         /*
1560          * If mount initialization proceeded far enough we must flush
1561          * its vnodes and sync the underlying mount points.  Three syncs
1562          * are required to fully flush the filesystem (freemap updates lag
1563          * by one flush, and one extra for safety).
1564          */
1565         if (mntflags & MNT_FORCE)
1566                 flags = FORCECLOSE;
1567         else
1568                 flags = 0;
1569         if (pmp->iroot) {
1570                 error = vflush(mp, 0, flags);
1571                 if (error)
1572                         goto failed;
1573                 hammer2_vfs_sync(mp, MNT_WAIT);
1574                 hammer2_vfs_sync(mp, MNT_WAIT);
1575                 hammer2_vfs_sync(mp, MNT_WAIT);
1576         }
1577
1578         /*
1579          * Cleanup the frontend support XOPS threads
1580          */
1581         hammer2_xop_helper_cleanup(pmp);
1582
1583         if (pmp->mp)
1584                 hammer2_unmount_helper(mp, pmp, NULL);
1585
1586         error = 0;
1587 failed:
1588         lockmgr(&hammer2_mntlk, LK_RELEASE);
1589
1590         return (error);
1591 }
1592
1593 /*
1594  * Mount helper, hook the system mount into our PFS.
1595  * The mount lock is held.
1596  *
1597  * We must bump the mount_count on related devices for any
1598  * mounted PFSs.
1599  */
1600 static
1601 void
1602 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1603 {
1604         hammer2_cluster_t *cluster;
1605         hammer2_chain_t *rchain;
1606         int i;
1607
1608         mp->mnt_data = (qaddr_t)pmp;
1609         pmp->mp = mp;
1610
1611         /*
1612          * After pmp->mp is set we have to adjust hmp->mount_count.
1613          */
1614         cluster = &pmp->iroot->cluster;
1615         for (i = 0; i < cluster->nchains; ++i) {
1616                 rchain = cluster->array[i].chain;
1617                 if (rchain == NULL)
1618                         continue;
1619                 ++rchain->hmp->mount_count;
1620         }
1621
1622         /*
1623          * Create missing Xop threads
1624          */
1625         hammer2_xop_helper_create(pmp);
1626 }
1627
1628 /*
1629  * Mount helper, unhook the system mount from our PFS.
1630  * The mount lock is held.
1631  *
1632  * If hmp is supplied a mount responsible for being the first to open
1633  * the block device failed and the block device and all PFSs using the
1634  * block device must be cleaned up.
1635  *
1636  * If pmp is supplied multiple devices might be backing the PFS and each
1637  * must be disconnected.  This might not be the last PFS using some of the
1638  * underlying devices.  Also, we have to adjust our hmp->mount_count
1639  * accounting for the devices backing the pmp which is now undergoing an
1640  * unmount.
1641  */
1642 static
1643 void
1644 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1645 {
1646         hammer2_cluster_t *cluster;
1647         hammer2_chain_t *rchain;
1648         struct vnode *devvp;
1649         int dumpcnt;
1650         int ronly;
1651         int i;
1652
1653         /*
1654          * If no device supplied this is a high-level unmount and we have to
1655          * to disconnect the mount, adjust mount_count, and locate devices
1656          * that might now have no mounts.
1657          */
1658         if (pmp) {
1659                 KKASSERT(hmp == NULL);
1660                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1661                 pmp->mp = NULL;
1662                 mp->mnt_data = NULL;
1663
1664                 /*
1665                  * After pmp->mp is cleared we have to account for
1666                  * mount_count.
1667                  */
1668                 cluster = &pmp->iroot->cluster;
1669                 for (i = 0; i < cluster->nchains; ++i) {
1670                         rchain = cluster->array[i].chain;
1671                         if (rchain == NULL)
1672                                 continue;
1673                         --rchain->hmp->mount_count;
1674                         /* scrapping hmp now may invalidate the pmp */
1675                 }
1676 again:
1677                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1678                         if (hmp->mount_count == 0) {
1679                                 hammer2_unmount_helper(NULL, NULL, hmp);
1680                                 goto again;
1681                         }
1682                 }
1683                 return;
1684         }
1685
1686         /*
1687          * Try to terminate the block device.  We can't terminate it if
1688          * there are still PFSs referencing it.
1689          */
1690         if (hmp->mount_count)
1691                 return;
1692
1693         /*
1694          * Decomission the network before we start messing with the
1695          * device and PFS.
1696          */
1697         hammer2_iocom_uninit(hmp);
1698
1699         hammer2_bulkfree_uninit(hmp);
1700         hammer2_pfsfree_scan(hmp, 0);
1701         hammer2_dev_exlock(hmp);        /* XXX order */
1702
1703         /*
1704          * Cycle the volume data lock as a safety (probably not needed any
1705          * more).  To ensure everything is out we need to flush at least
1706          * three times.  (1) The running of the sideq can dirty the
1707          * filesystem, (2) A normal flush can dirty the freemap, and
1708          * (3) ensure that the freemap is fully synchronized.
1709          *
1710          * The next mount's recovery scan can clean everything up but we want
1711          * to leave the filesystem in a 100% clean state on a normal unmount.
1712          */
1713 #if 0
1714         hammer2_voldata_lock(hmp);
1715         hammer2_voldata_unlock(hmp);
1716 #endif
1717
1718         /*
1719          * Flush whatever is left.  Unmounted but modified PFS's might still
1720          * have some dirty chains on them.
1721          */
1722         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1723         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1724
1725         if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1726                 hammer2_voldata_modify(hmp);
1727                 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1728                                             HAMMER2_FLUSH_ALL);
1729         }
1730         hammer2_chain_unlock(&hmp->fchain);
1731
1732         if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1733                 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1734                                             HAMMER2_FLUSH_ALL);
1735         }
1736         hammer2_chain_unlock(&hmp->vchain);
1737
1738         if ((hmp->vchain.flags | hmp->fchain.flags) &
1739             HAMMER2_CHAIN_FLUSH_MASK) {
1740                 kprintf("hammer2_unmount: chains left over "
1741                         "after final sync\n");
1742                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1743                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1744
1745                 if (hammer2_debug & 0x0010)
1746                         Debugger("entered debugger");
1747         }
1748
1749         hammer2_pfsfree_scan(hmp, 1);
1750
1751         KKASSERT(hmp->spmp == NULL);
1752
1753         /*
1754          * Finish up with the device vnode
1755          */
1756         if ((devvp = hmp->devvp) != NULL) {
1757                 ronly = hmp->ronly;
1758                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1759                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1760                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1761                         ronly);
1762                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1763                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1764                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1765                 hmp->devvp = NULL;
1766                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1767                 vn_unlock(devvp);
1768                 vrele(devvp);
1769                 devvp = NULL;
1770         }
1771
1772         /*
1773          * Clear vchain/fchain flags that might prevent final cleanup
1774          * of these chains.
1775          */
1776         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1777                 atomic_add_long(&hammer2_count_modified_chains, -1);
1778                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1779                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1780         }
1781         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1782                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1783         }
1784
1785         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1786                 atomic_add_long(&hammer2_count_modified_chains, -1);
1787                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1788                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1789         }
1790         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1791                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1792         }
1793
1794         /*
1795          * Final drop of embedded freemap root chain to
1796          * clean up fchain.core (fchain structure is not
1797          * flagged ALLOCATED so it is cleaned out and then
1798          * left to rot).
1799          */
1800         hammer2_chain_drop(&hmp->fchain);
1801
1802         /*
1803          * Final drop of embedded volume root chain to clean
1804          * up vchain.core (vchain structure is not flagged
1805          * ALLOCATED so it is cleaned out and then left to
1806          * rot).
1807          */
1808         dumpcnt = 50;
1809         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1810         dumpcnt = 50;
1811         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1812         hammer2_dev_unlock(hmp);
1813         hammer2_chain_drop(&hmp->vchain);
1814
1815         hammer2_io_cleanup(hmp, &hmp->iotree);
1816         if (hmp->iofree_count) {
1817                 kprintf("io_cleanup: %d I/O's left hanging\n",
1818                         hmp->iofree_count);
1819         }
1820
1821         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1822         kmalloc_destroy(&hmp->mchain);
1823         kfree(hmp, M_HAMMER2);
1824 }
1825
1826 int
1827 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1828                  ino_t ino, struct vnode **vpp)
1829 {
1830         hammer2_xop_lookup_t *xop;
1831         hammer2_pfs_t *pmp;
1832         hammer2_inode_t *ip;
1833         hammer2_tid_t inum;
1834         int error;
1835
1836         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1837
1838         error = 0;
1839         pmp = MPTOPMP(mp);
1840
1841         /*
1842          * Easy if we already have it cached
1843          */
1844         ip = hammer2_inode_lookup(pmp, inum);
1845         if (ip) {
1846                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1847                 *vpp = hammer2_igetv(ip, &error);
1848                 hammer2_inode_unlock(ip);
1849                 hammer2_inode_drop(ip);         /* from lookup */
1850
1851                 return error;
1852         }
1853
1854         /*
1855          * Otherwise we have to find the inode
1856          */
1857         xop = hammer2_xop_alloc(pmp->iroot, 0);
1858         xop->lhc = inum;
1859         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1860         error = hammer2_xop_collect(&xop->head, 0);
1861
1862         if (error == 0) {
1863                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1864                         kprintf("vget: no collect error but also no rdata\n");
1865                         kprintf("xop %p\n", xop);
1866                         while ((hammer2_debug & 0x80000) == 0) {
1867                                 tsleep(xop, PCATCH, "wait", hz * 10);
1868                         }
1869                         ip = NULL;
1870                 } else {
1871                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1872                 }
1873         }
1874         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1875
1876         if (ip) {
1877                 *vpp = hammer2_igetv(ip, &error);
1878                 hammer2_inode_unlock(ip);
1879         } else {
1880                 *vpp = NULL;
1881                 error = ENOENT;
1882         }
1883         return (error);
1884 }
1885
1886 static
1887 int
1888 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1889 {
1890         hammer2_pfs_t *pmp;
1891         struct vnode *vp;
1892         int error;
1893
1894         pmp = MPTOPMP(mp);
1895         if (pmp->iroot == NULL) {
1896                 *vpp = NULL;
1897                 return EINVAL;
1898         }
1899
1900         error = 0;
1901         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1902
1903         while (pmp->inode_tid == 0) {
1904                 hammer2_xop_ipcluster_t *xop;
1905                 hammer2_inode_meta_t *meta;
1906
1907                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1908                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1909                 error = hammer2_xop_collect(&xop->head, 0);
1910
1911                 if (error == 0) {
1912                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1913                         pmp->iroot->meta = *meta;
1914                         pmp->inode_tid = meta->pfs_inum + 1;
1915                         if (pmp->inode_tid < HAMMER2_INODE_START)
1916                                 pmp->inode_tid = HAMMER2_INODE_START;
1917                         pmp->modify_tid =
1918                                 xop->head.cluster.focus->bref.modify_tid + 1;
1919 #if 0
1920                         kprintf("PFS: Starting inode %jd\n",
1921                                 (intmax_t)pmp->inode_tid);
1922                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1923                                 pmp->inode_tid, pmp->modify_tid);
1924 #endif
1925                         wakeup(&pmp->iroot);
1926
1927                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1928
1929                         /*
1930                          * Prime the mount info.
1931                          */
1932                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1933                         break;
1934                 }
1935
1936                 /*
1937                  * Loop, try again
1938                  */
1939                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1940                 hammer2_inode_unlock(pmp->iroot);
1941                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1942                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1943                 if (error == EINTR)
1944                         break;
1945         }
1946
1947         if (error) {
1948                 hammer2_inode_unlock(pmp->iroot);
1949                 *vpp = NULL;
1950         } else {
1951                 vp = hammer2_igetv(pmp->iroot, &error);
1952                 hammer2_inode_unlock(pmp->iroot);
1953                 *vpp = vp;
1954         }
1955
1956         return (error);
1957 }
1958
1959 /*
1960  * Filesystem status
1961  *
1962  * XXX incorporate ipdata->meta.inode_quota and data_quota
1963  */
1964 static
1965 int
1966 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1967 {
1968         hammer2_pfs_t *pmp;
1969         hammer2_dev_t *hmp;
1970         hammer2_blockref_t bref;
1971         struct statfs tmp;
1972         int i;
1973
1974         /*
1975          * NOTE: iroot might not have validated the cluster yet.
1976          */
1977         pmp = MPTOPMP(mp);
1978
1979         bzero(&tmp, sizeof(tmp));
1980
1981         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1982                 hmp = pmp->pfs_hmps[i];
1983                 if (hmp == NULL)
1984                         continue;
1985                 if (pmp->iroot->cluster.array[i].chain)
1986                         bref = pmp->iroot->cluster.array[i].chain->bref;
1987                 else
1988                         bzero(&bref, sizeof(bref));
1989
1990                 tmp.f_files = bref.embed.stats.inode_count;
1991                 tmp.f_ffree = 0;
1992                 tmp.f_blocks = hmp->voldata.allocator_size /
1993                                mp->mnt_vstat.f_bsize;
1994                 tmp.f_bfree = hmp->voldata.allocator_free /
1995                               mp->mnt_vstat.f_bsize;
1996                 tmp.f_bavail = tmp.f_bfree;
1997
1998                 if (cred && cred->cr_uid != 0) {
1999                         uint64_t adj;
2000
2001                         /* 5% */
2002                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2003                         tmp.f_blocks -= adj;
2004                         tmp.f_bfree -= adj;
2005                         tmp.f_bavail -= adj;
2006                 }
2007
2008                 mp->mnt_stat.f_blocks = tmp.f_blocks;
2009                 mp->mnt_stat.f_bfree = tmp.f_bfree;
2010                 mp->mnt_stat.f_bavail = tmp.f_bavail;
2011                 mp->mnt_stat.f_files = tmp.f_files;
2012                 mp->mnt_stat.f_ffree = tmp.f_ffree;
2013
2014                 *sbp = mp->mnt_stat;
2015         }
2016         return (0);
2017 }
2018
2019 static
2020 int
2021 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2022 {
2023         hammer2_pfs_t *pmp;
2024         hammer2_dev_t *hmp;
2025         hammer2_blockref_t bref;
2026         struct statvfs tmp;
2027         int i;
2028
2029         /*
2030          * NOTE: iroot might not have validated the cluster yet.
2031          */
2032         pmp = MPTOPMP(mp);
2033         bzero(&tmp, sizeof(tmp));
2034
2035         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2036                 hmp = pmp->pfs_hmps[i];
2037                 if (hmp == NULL)
2038                         continue;
2039                 if (pmp->iroot->cluster.array[i].chain)
2040                         bref = pmp->iroot->cluster.array[i].chain->bref;
2041                 else
2042                         bzero(&bref, sizeof(bref));
2043
2044                 tmp.f_files = bref.embed.stats.inode_count;
2045                 tmp.f_ffree = 0;
2046                 tmp.f_blocks = hmp->voldata.allocator_size /
2047                                mp->mnt_vstat.f_bsize;
2048                 tmp.f_bfree = hmp->voldata.allocator_free /
2049                               mp->mnt_vstat.f_bsize;
2050                 tmp.f_bavail = tmp.f_bfree;
2051
2052                 if (cred && cred->cr_uid != 0) {
2053                         uint64_t adj;
2054
2055                         /* 5% */
2056                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2057                         tmp.f_blocks -= adj;
2058                         tmp.f_bfree -= adj;
2059                         tmp.f_bavail -= adj;
2060                 }
2061
2062                 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2063                 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2064                 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2065                 mp->mnt_vstat.f_files = tmp.f_files;
2066                 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2067
2068                 *sbp = mp->mnt_vstat;
2069         }
2070         return (0);
2071 }
2072
2073 /*
2074  * Mount-time recovery (RW mounts)
2075  *
2076  * Updates to the free block table are allowed to lag flushes by one
2077  * transaction.  In case of a crash, then on a fresh mount we must do an
2078  * incremental scan of the last committed transaction id and make sure that
2079  * all related blocks have been marked allocated.
2080  *
2081  * The super-root topology and each PFS has its own transaction id domain,
2082  * so we must track PFS boundary transitions.
2083  */
2084 struct hammer2_recovery_elm {
2085         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2086         hammer2_chain_t *chain;
2087         hammer2_tid_t sync_tid;
2088 };
2089
2090 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2091
2092 struct hammer2_recovery_info {
2093         struct hammer2_recovery_list list;
2094         hammer2_tid_t   mtid;
2095         int     depth;
2096 };
2097
2098 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2099                         hammer2_chain_t *parent,
2100                         struct hammer2_recovery_info *info,
2101                         hammer2_tid_t sync_tid);
2102
2103 #define HAMMER2_RECOVERY_MAXDEPTH       10
2104
2105 static
2106 int
2107 hammer2_recovery(hammer2_dev_t *hmp)
2108 {
2109         struct hammer2_recovery_info info;
2110         struct hammer2_recovery_elm *elm;
2111         hammer2_chain_t *parent;
2112         hammer2_tid_t sync_tid;
2113         hammer2_tid_t mirror_tid;
2114         int error;
2115
2116         hammer2_trans_init(hmp->spmp, 0);
2117
2118         sync_tid = hmp->voldata.freemap_tid;
2119         mirror_tid = hmp->voldata.mirror_tid;
2120
2121         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2122         if (sync_tid >= mirror_tid) {
2123                 kprintf(" no recovery needed\n");
2124         } else {
2125                 kprintf(" freemap recovery %016jx-%016jx\n",
2126                         sync_tid + 1, mirror_tid);
2127         }
2128
2129         TAILQ_INIT(&info.list);
2130         info.depth = 0;
2131         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2132         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2133         hammer2_chain_lookup_done(parent);
2134
2135         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2136                 TAILQ_REMOVE(&info.list, elm, entry);
2137                 parent = elm->chain;
2138                 sync_tid = elm->sync_tid;
2139                 kfree(elm, M_HAMMER2);
2140
2141                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2142                 error |= hammer2_recovery_scan(hmp, parent, &info,
2143                                               hmp->voldata.freemap_tid);
2144                 hammer2_chain_unlock(parent);
2145                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2146         }
2147         hammer2_trans_done(hmp->spmp);
2148
2149         return error;
2150 }
2151
2152 static
2153 int
2154 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2155                       struct hammer2_recovery_info *info,
2156                       hammer2_tid_t sync_tid)
2157 {
2158         const hammer2_inode_data_t *ripdata;
2159         hammer2_chain_t *chain;
2160         hammer2_blockref_t bref;
2161         int tmp_error;
2162         int rup_error;
2163         int error;
2164         int first;
2165
2166         /*
2167          * Adjust freemap to ensure that the block(s) are marked allocated.
2168          */
2169         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2170                 hammer2_freemap_adjust(hmp, &parent->bref,
2171                                        HAMMER2_FREEMAP_DORECOVER);
2172         }
2173
2174         /*
2175          * Check type for recursive scan
2176          */
2177         switch(parent->bref.type) {
2178         case HAMMER2_BREF_TYPE_VOLUME:
2179                 /* data already instantiated */
2180                 break;
2181         case HAMMER2_BREF_TYPE_INODE:
2182                 /*
2183                  * Must instantiate data for DIRECTDATA test and also
2184                  * for recursion.
2185                  */
2186                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2187                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2188                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2189                         /* not applicable to recovery scan */
2190                         hammer2_chain_unlock(parent);
2191                         return 0;
2192                 }
2193                 hammer2_chain_unlock(parent);
2194                 break;
2195         case HAMMER2_BREF_TYPE_INDIRECT:
2196                 /*
2197                  * Must instantiate data for recursion
2198                  */
2199                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2200                 hammer2_chain_unlock(parent);
2201                 break;
2202         case HAMMER2_BREF_TYPE_DIRENT:
2203         case HAMMER2_BREF_TYPE_DATA:
2204         case HAMMER2_BREF_TYPE_FREEMAP:
2205         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2206         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2207                 /* not applicable to recovery scan */
2208                 return 0;
2209                 break;
2210         default:
2211                 return HAMMER2_ERROR_BADBREF;
2212         }
2213
2214         /*
2215          * Defer operation if depth limit reached or if we are crossing a
2216          * PFS boundary.
2217          */
2218         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2219                 struct hammer2_recovery_elm *elm;
2220
2221                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2222                 elm->chain = parent;
2223                 elm->sync_tid = sync_tid;
2224                 hammer2_chain_ref(parent);
2225                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2226                 /* unlocked by caller */
2227
2228                 return(0);
2229         }
2230
2231
2232         /*
2233          * Recursive scan of the last flushed transaction only.  We are
2234          * doing this without pmp assignments so don't leave the chains
2235          * hanging around after we are done with them.
2236          *
2237          * error        Cumulative error this level only
2238          * rup_error    Cumulative error for recursion
2239          * tmp_error    Specific non-cumulative recursion error
2240          */
2241         chain = NULL;
2242         first = 1;
2243         rup_error = 0;
2244         error = 0;
2245
2246         for (;;) {
2247                 error |= hammer2_chain_scan(parent, &chain, &bref,
2248                                             &first,
2249                                             HAMMER2_LOOKUP_NODATA);
2250
2251                 /*
2252                  * Problem during scan or EOF
2253                  */
2254                 if (error)
2255                         break;
2256
2257                 /*
2258                  * If this is a leaf
2259                  */
2260                 if (chain == NULL) {
2261                         if (bref.mirror_tid > sync_tid) {
2262                                 hammer2_freemap_adjust(hmp, &bref,
2263                                                      HAMMER2_FREEMAP_DORECOVER);
2264                         }
2265                         continue;
2266                 }
2267
2268                 /*
2269                  * This may or may not be a recursive node.
2270                  */
2271                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2272                 if (bref.mirror_tid > sync_tid) {
2273                         ++info->depth;
2274                         tmp_error = hammer2_recovery_scan(hmp, chain,
2275                                                            info, sync_tid);
2276                         --info->depth;
2277                 } else {
2278                         tmp_error = 0;
2279                 }
2280
2281                 /*
2282                  * Flush the recovery at the PFS boundary to stage it for
2283                  * the final flush of the super-root topology.
2284                  */
2285                 if (tmp_error == 0 &&
2286                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2287                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2288                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2289                                              HAMMER2_FLUSH_ALL);
2290                 }
2291                 rup_error |= tmp_error;
2292         }
2293         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2294 }
2295
2296 /*
2297  * Sync a mount point; this is called periodically on a per-mount basis from
2298  * the filesystem syncer, and whenever a user issues a sync.
2299  */
2300 int
2301 hammer2_vfs_sync(struct mount *mp, int waitfor)
2302 {
2303         hammer2_xop_flush_t *xop;
2304         struct hammer2_sync_info info;
2305         hammer2_inode_t *iroot;
2306         hammer2_pfs_t *pmp;
2307         int flags;
2308         int error;
2309
2310         pmp = MPTOPMP(mp);
2311         iroot = pmp->iroot;
2312         KKASSERT(iroot);
2313         KKASSERT(iroot->pmp == pmp);
2314
2315         /*
2316          * We can't acquire locks on existing vnodes while in a transaction
2317          * without risking a deadlock.  This assumes that vfsync() can be
2318          * called without the vnode locked (which it can in DragonFly).
2319          * Otherwise we'd have to implement a multi-pass or flag the lock
2320          * failures and retry.
2321          *
2322          * The reclamation code interlocks with the sync list's token
2323          * (by removing the vnode from the scan list) before unlocking
2324          * the inode, giving us time to ref the inode.
2325          */
2326         /*flags = VMSC_GETVP;*/
2327         flags = 0;
2328         if (waitfor & MNT_LAZY)
2329                 flags |= VMSC_ONEPASS;
2330
2331         /*
2332          * Flush vnodes individually using a normal transaction to avoid
2333          * stalling any concurrent operations.  This will flush the related
2334          * buffer cache buffers and inodes to the media.
2335          *
2336          * For efficiency do an async pass before making sure with a
2337          * synchronous pass on all related buffer cache buffers.
2338          */
2339         hammer2_trans_init(pmp, 0);
2340
2341         info.error = 0;
2342
2343         info.waitfor = MNT_NOWAIT;
2344         info.pass = 1;
2345         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2346
2347         /*
2348          * Now do two passes making sure we get everything.  The first pass
2349          * vfsync()s dirty vnodes.  The second pass waits for their I/O's
2350          * to finish and cleans up the dirty flag on the vnode.
2351          */
2352         info.pass = 1;
2353         info.waitfor = MNT_WAIT;
2354         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2355
2356         info.pass = 2;
2357         info.waitfor = MNT_WAIT;
2358         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2359
2360         /*
2361          * We must also run the sideq to handle any disconnected inodes
2362          * as the vnode scan will not see these.
2363          */
2364         hammer2_inode_run_sideq(pmp, 1);
2365         hammer2_trans_done(pmp);
2366
2367         /*
2368          * Start our flush transaction and flush the root topology down to
2369          * the inodes, but not the inodes themselves (which we already flushed
2370          * above).  Any concurrent activity effecting inode contents will not
2371          *
2372          * The flush sequence will
2373          *
2374          * NOTE! It is still possible for the paging code to push pages
2375          *       out via a UIO_NOCOPY hammer2_vop_write() during the main
2376          *       flush.
2377          */
2378         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2379
2380         /*
2381          * sync dirty vnodes again while in the flush transaction.  This is
2382          * currently an expensive shim to makre sure the logical topology is
2383          * completely consistent before we flush the volume header.
2384          */
2385         info.pass = 1;
2386         info.waitfor = MNT_WAIT;
2387         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2388
2389         info.pass = 2;
2390         info.waitfor = MNT_WAIT;
2391         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2392
2393         /*
2394          * Use the XOP interface to concurrently flush all nodes to
2395          * synchronize the PFSROOT subtopology to the media.  A standard
2396          * end-of-scan ENOENT error indicates cluster sufficiency.
2397          *
2398          * Note that this flush will not be visible on crash recovery until
2399          * we flush the super-root topology in the next loop.
2400          *
2401          * XXX For now wait for all flushes to complete.
2402          */
2403         if (iroot) {
2404                 /*
2405                  * If unmounting try to flush everything including any
2406                  * sub-trees under inodes, just in case there is dangling
2407                  * modified data, as a safety.  Otherwise just flush up to
2408                  * the inodes in this stage.
2409                  */
2410                 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2411                         xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2412                                                        HAMMER2_XOP_VOLHDR);
2413                 } else {
2414                         xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2415                                                        HAMMER2_XOP_INODE_STOP |
2416                                                        HAMMER2_XOP_VOLHDR);
2417                 }
2418                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2419                 error = hammer2_xop_collect(&xop->head,
2420                                             HAMMER2_XOP_COLLECT_WAITALL);
2421                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2422                 if (error == HAMMER2_ERROR_ENOENT)
2423                         error = 0;
2424                 else
2425                         error = hammer2_error_to_errno(error);
2426         } else {
2427                 error = 0;
2428         }
2429         hammer2_trans_done(pmp);
2430
2431         return (error);
2432 }
2433
2434 /*
2435  * Sync passes.
2436  *
2437  * Note that we ignore the tranasction mtid we got above.  Instead,
2438  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2439  * transactions.
2440  *
2441  * WARNING! The frontend might be waiting on chnmem (limit_dirty_chains)
2442  * while holding a vnode locked.  When this situation occurs we cannot
2443  * safely test whether it is ok to clear the dirty bit on the vnode.
2444  * However, we can still flush the inode's topology.
2445  */
2446 static int
2447 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2448 {
2449         struct hammer2_sync_info *info = data;
2450         hammer2_inode_t *ip;
2451         int error;
2452
2453         /*
2454          * Degenerate cases.  Note that ip == NULL typically means the
2455          * syncer vnode itself and we don't want to vclrisdirty() in that
2456          * situation.
2457          */
2458         ip = VTOI(vp);
2459         if (ip == NULL) {
2460                 return(0);
2461         }
2462         if (vp->v_type == VNON || vp->v_type == VBAD) {
2463                 vclrisdirty(vp);
2464                 return(0);
2465         }
2466
2467         /*
2468          * Synchronize the buffer cche and inode meta-data to the backing
2469          * chain topology.
2470          *
2471          * vfsync is not necessarily synchronous, so it is best NOT to try
2472          * to flush the backing topology to media at this point.
2473          */
2474         hammer2_inode_ref(ip);
2475         if ((ip->flags & (HAMMER2_INODE_RESIZED|HAMMER2_INODE_MODIFIED)) ||
2476             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2477                 if (info->pass == 1)
2478                         vfsync(vp, info->waitfor, 1, NULL, NULL);
2479                 else
2480                         bio_track_wait(&vp->v_track_write, 0, 0);
2481         }
2482         if (info->pass == 2 && (vp->v_flag & VISDIRTY)) {
2483                 /*
2484                  * v_token is needed to interlock v_rbdirty_tree.
2485                  */
2486                 lwkt_gettoken(&vp->v_token);
2487                 hammer2_inode_lock(ip, 0);
2488                 hammer2_inode_chain_sync(ip);
2489                 hammer2_inode_chain_flush(ip);
2490                 if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2491                                   HAMMER2_INODE_RESIZED |
2492                                   HAMMER2_INODE_DIRTYDATA)) == 0 &&
2493                     RB_EMPTY(&vp->v_rbdirty_tree) &&
2494                     !bio_track_active(&vp->v_track_write)) {
2495                         vclrisdirty(vp);
2496                 }
2497                 hammer2_inode_unlock(ip);
2498                 lwkt_reltoken(&vp->v_token);
2499         }
2500         hammer2_inode_drop(ip);
2501 #if 1
2502         error = 0;
2503         if (error)
2504                 info->error = error;
2505 #endif
2506         return(0);
2507 }
2508
2509 static
2510 int
2511 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2512 {
2513         hammer2_inode_t *ip;
2514
2515         KKASSERT(MAXFIDSZ >= 16);
2516         ip = VTOI(vp);
2517         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2518         fhp->fid_ext = 0;
2519         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2520         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2521
2522         return 0;
2523 }
2524
2525 static
2526 int
2527 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2528                struct fid *fhp, struct vnode **vpp)
2529 {
2530         hammer2_pfs_t *pmp;
2531         hammer2_tid_t inum;
2532         int error;
2533
2534         pmp = MPTOPMP(mp);
2535         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2536         if (vpp) {
2537                 if (inum == 1)
2538                         error = hammer2_vfs_root(mp, vpp);
2539                 else
2540                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2541         } else {
2542                 error = 0;
2543         }
2544         if (error)
2545                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2546         return error;
2547 }
2548
2549 static
2550 int
2551 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2552                  int *exflagsp, struct ucred **credanonp)
2553 {
2554         hammer2_pfs_t *pmp;
2555         struct netcred *np;
2556         int error;
2557
2558         pmp = MPTOPMP(mp);
2559         np = vfs_export_lookup(mp, &pmp->export, nam);
2560         if (np) {
2561                 *exflagsp = np->netc_exflags;
2562                 *credanonp = &np->netc_anon;
2563                 error = 0;
2564         } else {
2565                 error = EACCES;
2566         }
2567         return error;
2568 }
2569
2570 /*
2571  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2572  * header into the HMP
2573  *
2574  * XXX read four volhdrs and use the one with the highest TID whos CRC
2575  *     matches.
2576  *
2577  * XXX check iCRCs.
2578  *
2579  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2580  *     nonexistant locations.
2581  *
2582  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2583  */
2584 static
2585 int
2586 hammer2_install_volume_header(hammer2_dev_t *hmp)
2587 {
2588         hammer2_volume_data_t *vd;
2589         struct buf *bp;
2590         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2591         int error_reported;
2592         int error;
2593         int valid;
2594         int i;
2595
2596         error_reported = 0;
2597         error = 0;
2598         valid = 0;
2599         bp = NULL;
2600
2601         /*
2602          * There are up to 4 copies of the volume header (syncs iterate
2603          * between them so there is no single master).  We don't trust the
2604          * volu_size field so we don't know precisely how large the filesystem
2605          * is, so depend on the OS to return an error if we go beyond the
2606          * block device's EOF.
2607          */
2608         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2609                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2610                               HAMMER2_VOLUME_BYTES, &bp);
2611                 if (error) {
2612                         brelse(bp);
2613                         bp = NULL;
2614                         continue;
2615                 }
2616
2617                 vd = (struct hammer2_volume_data *) bp->b_data;
2618                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2619                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2620                         brelse(bp);
2621                         bp = NULL;
2622                         continue;
2623                 }
2624
2625                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2626                         /* XXX: Reversed-endianness filesystem */
2627                         kprintf("hammer2: reverse-endian filesystem detected");
2628                         brelse(bp);
2629                         bp = NULL;
2630                         continue;
2631                 }
2632
2633                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2634                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2635                                       HAMMER2_VOLUME_ICRC0_SIZE);
2636                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2637                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2638                                        HAMMER2_VOLUME_ICRC1_SIZE);
2639                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2640                         kprintf("hammer2 volume header crc "
2641                                 "mismatch copy #%d %08x/%08x\n",
2642                                 i, crc0, crc);
2643                         error_reported = 1;
2644                         brelse(bp);
2645                         bp = NULL;
2646                         continue;
2647                 }
2648                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2649                         valid = 1;
2650                         hmp->voldata = *vd;
2651                         hmp->volhdrno = i;
2652                 }
2653                 brelse(bp);
2654                 bp = NULL;
2655         }
2656         if (valid) {
2657                 hmp->volsync = hmp->voldata;
2658                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2659                 error = 0;
2660                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2661                         kprintf("hammer2: using volume header #%d\n",
2662                                 hmp->volhdrno);
2663                 }
2664         } else {
2665                 error = EINVAL;
2666                 kprintf("hammer2: no valid volume headers found!\n");
2667         }
2668         return (error);
2669 }
2670
2671 /*
2672  * This handles hysteresis on regular file flushes.  Because the BIOs are
2673  * routed to a thread it is possible for an excessive number to build up
2674  * and cause long front-end stalls long before the runningbuffspace limit
2675  * is hit, so we implement hammer2_flush_pipe to control the
2676  * hysteresis.
2677  *
2678  * This is a particular problem when compression is used.
2679  */
2680 void
2681 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2682 {
2683         atomic_add_int(&pmp->count_lwinprog, 1);
2684 }
2685
2686 void
2687 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2688 {
2689         int lwinprog;
2690
2691         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2692         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2693             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2694                 atomic_clear_int(&pmp->count_lwinprog,
2695                                  HAMMER2_LWINPROG_WAITING);
2696                 wakeup(&pmp->count_lwinprog);
2697         }
2698         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2699             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2700                 atomic_clear_int(&pmp->count_lwinprog,
2701                                  HAMMER2_LWINPROG_WAITING0);
2702                 wakeup(&pmp->count_lwinprog);
2703         }
2704 }
2705
2706 void
2707 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2708 {
2709         int lwinprog;
2710         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2711                                     HAMMER2_LWINPROG_WAITING0;
2712
2713         for (;;) {
2714                 lwinprog = pmp->count_lwinprog;
2715                 cpu_ccfence();
2716                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2717                         break;
2718                 tsleep_interlock(&pmp->count_lwinprog, 0);
2719                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2720                 lwinprog = pmp->count_lwinprog;
2721                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2722                         break;
2723                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2724         }
2725 }
2726
2727 /*
2728  * Manage excessive memory resource use for chain and related
2729  * structures.
2730  */
2731 void
2732 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2733 {
2734         uint32_t waiting;
2735         uint32_t count;
2736         uint32_t limit;
2737 #if 0
2738         static int zzticks;
2739 #endif
2740
2741         /*
2742          * Atomic check condition and wait.  Also do an early speedup of
2743          * the syncer to try to avoid hitting the wait.
2744          */
2745         for (;;) {
2746                 waiting = pmp->inmem_dirty_chains;
2747                 cpu_ccfence();
2748                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2749
2750                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2751                 if (limit < hammer2_limit_dirty_chains)
2752                         limit = hammer2_limit_dirty_chains;
2753                 if (limit < 1000)
2754                         limit = 1000;
2755
2756 #if 0
2757                 if ((int)(ticks - zzticks) > hz) {
2758                         zzticks = ticks;
2759                         kprintf("count %ld %ld\n", count, limit);
2760                 }
2761 #endif
2762
2763                 /*
2764                  * Block if there are too many dirty chains present, wait
2765                  * for the flush to clean some out.
2766                  */
2767                 if (count > limit) {
2768                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2769                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2770                                                waiting,
2771                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2772                                 speedup_syncer(pmp->mp);
2773                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2774                                        "chnmem", hz);
2775                         }
2776                         continue;       /* loop on success or fail */
2777                 }
2778
2779                 /*
2780                  * Try to start an early flush before we are forced to block.
2781                  */
2782                 if (count > limit * 5 / 10)
2783                         speedup_syncer(pmp->mp);
2784                 break;
2785         }
2786 }
2787
2788 void
2789 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2790 {
2791         if (pmp) {
2792                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2793         }
2794 }
2795
2796 void
2797 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2798 {
2799         uint32_t waiting;
2800
2801         if (pmp) {
2802                 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
2803                 /* don't need --waiting to test flag */
2804                 if (waiting & HAMMER2_DIRTYCHAIN_WAITING) {
2805                         atomic_clear_int(&pmp->inmem_dirty_chains,
2806                                          HAMMER2_DIRTYCHAIN_WAITING);
2807                         wakeup(&pmp->inmem_dirty_chains);
2808                 }
2809         }
2810 }
2811
2812 /*
2813  * Returns 0 if the filesystem has tons of free space
2814  * Returns 1 if the filesystem has less than 10% remaining
2815  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2816  */
2817 int
2818 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2819 {
2820         hammer2_pfs_t *pmp;
2821         hammer2_dev_t *hmp;
2822         hammer2_off_t free_reserved;
2823         hammer2_off_t free_nominal;
2824         int i;
2825
2826         pmp = ip->pmp;
2827
2828         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2829                 free_reserved = HAMMER2_SEGSIZE;
2830                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2831                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2832                         hmp = pmp->pfs_hmps[i];
2833                         if (hmp == NULL)
2834                                 continue;
2835                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2836                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2837                                 continue;
2838
2839                         if (free_nominal > hmp->voldata.allocator_free)
2840                                 free_nominal = hmp->voldata.allocator_free;
2841                         if (free_reserved < hmp->free_reserved)
2842                                 free_reserved = hmp->free_reserved;
2843                 }
2844
2845                 /*
2846                  * SMP races ok
2847                  */
2848                 pmp->free_reserved = free_reserved;
2849                 pmp->free_nominal = free_nominal;
2850                 pmp->free_ticks = ticks;
2851         } else {
2852                 free_reserved = pmp->free_reserved;
2853                 free_nominal = pmp->free_nominal;
2854         }
2855         if (cred && cred->cr_uid != 0) {
2856                 if ((int64_t)(free_nominal - bytes) <
2857                     (int64_t)free_reserved) {
2858                         return 2;
2859                 }
2860         } else {
2861                 if ((int64_t)(free_nominal - bytes) <
2862                     (int64_t)free_reserved / 2) {
2863                         return 2;
2864                 }
2865         }
2866         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2867                 return 1;
2868         return 0;
2869 }
2870
2871 /*
2872  * Debugging
2873  */
2874 void
2875 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
2876                    u_int flags)
2877 {
2878         hammer2_chain_t *scan;
2879         hammer2_chain_t *parent;
2880
2881         --*countp;
2882         if (*countp == 0) {
2883                 kprintf("%*.*s...\n", tab, tab, "");
2884                 return;
2885         }
2886         if (*countp < 0)
2887                 return;
2888         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2889                 tab, tab, "", pfx,
2890                 chain, chain->bref.type,
2891                 chain->bref.key, chain->bref.keybits,
2892                 chain->bref.mirror_tid);
2893
2894         kprintf("%*.*s      [%08x] (%s) refs=%d",
2895                 tab, tab, "",
2896                 chain->flags,
2897                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2898                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2899                 chain->refs);
2900
2901         parent = chain->parent;
2902         if (parent)
2903                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2904                         tab, tab, "",
2905                         parent, parent->flags, parent->refs);
2906         if (RB_EMPTY(&chain->core.rbtree)) {
2907                 kprintf("\n");
2908         } else {
2909                 kprintf(" {\n");
2910                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
2911                         if ((scan->flags & flags) || flags == (u_int)-1) {
2912                                 hammer2_dump_chain(scan, tab + 4, countp, 'a',
2913                                                    flags);
2914                         }
2915                 }
2916                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2917                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2918                                 chain->data->ipdata.filename);
2919                 else
2920                         kprintf("%*.*s}\n", tab, tab, "");
2921         }
2922 }