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 long hammer2_debug_inode;
  85 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
  86 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
  87 int hammer2_cluster_write = 0;          /* physical write clustering */
  88 int hammer2_dedup_enable = 1;
  89 int hammer2_always_compress = 0;        /* always try to compress */
  90 int hammer2_inval_enable = 0;
  91 int hammer2_flush_pipe = 100;
  92 int hammer2_dio_count;
  93 int hammer2_dio_limit = 256;
  94 int hammer2_bulkfree_tps = 5000;
  95 long hammer2_chain_allocs;
  96 long hammer2_chain_frees;
  97 long hammer2_limit_dirty_chains;
  98 long hammer2_limit_dirty_inodes;
  99 long hammer2_count_modified_chains;
 100 long hammer2_iod_invals;
 101 long hammer2_iod_file_read;
 102 long hammer2_iod_meta_read;
 103 long hammer2_iod_indr_read;
 104 long hammer2_iod_fmap_read;
 105 long hammer2_iod_volu_read;
 106 long hammer2_iod_file_write;
 107 long hammer2_iod_file_wembed;
 108 long hammer2_iod_file_wzero;
 109 long hammer2_iod_file_wdedup;
 110 long hammer2_iod_meta_write;
 111 long hammer2_iod_indr_write;
 112 long hammer2_iod_fmap_write;
 113 long hammer2_iod_volu_write;
 114 long hammer2_iod_inode_creates;
 115 long hammer2_iod_inode_deletes;
 116
 117 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 118 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 119                 "Buffer used for compression.");
 120
 121 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 122 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 123                 "Buffer used for decompression.");
 124
 125 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 126
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
 128            &hammer2_supported_version, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 130            &hammer2_debug, 0, "");
 131 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW,
 132            &hammer2_debug_inode, 0, "");
 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 134            &hammer2_cluster_meta_read, 0, "");
 135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 136            &hammer2_cluster_data_read, 0, "");
 137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 138            &hammer2_cluster_write, 0, "");
 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 140            &hammer2_dedup_enable, 0, "");
 141 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 142            &hammer2_always_compress, 0, "");
 143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 144            &hammer2_inval_enable, 0, "");
 145 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 146            &hammer2_flush_pipe, 0, "");
 147 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
 148            &hammer2_bulkfree_tps, 0, "");
 149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 150            &hammer2_chain_allocs, 0, "");
 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 152            &hammer2_chain_frees, 0, "");
 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 154            &hammer2_limit_dirty_chains, 0, "");
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
 156            &hammer2_limit_dirty_inodes, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 158            &hammer2_count_modified_chains, 0, "");
 159 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 160            &hammer2_dio_count, 0, "");
 161 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
 162            &hammer2_dio_limit, 0, "");
 163
 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 165            &hammer2_iod_invals, 0, "");
 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 167            &hammer2_iod_file_read, 0, "");
 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 169            &hammer2_iod_meta_read, 0, "");
 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 171            &hammer2_iod_indr_read, 0, "");
 172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 173            &hammer2_iod_fmap_read, 0, "");
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 175            &hammer2_iod_volu_read, 0, "");
 176
 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 178            &hammer2_iod_file_write, 0, "");
 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 180            &hammer2_iod_file_wembed, 0, "");
 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 182            &hammer2_iod_file_wzero, 0, "");
 183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 184            &hammer2_iod_file_wdedup, 0, "");
 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 186            &hammer2_iod_meta_write, 0, "");
 187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 188            &hammer2_iod_indr_write, 0, "");
 189 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 190            &hammer2_iod_fmap_write, 0, "");
 191 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 192            &hammer2_iod_volu_write, 0, "");
 193 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW,
 194            &hammer2_iod_inode_creates, 0, "");
 195 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW,
 196            &hammer2_iod_inode_deletes, 0, "");
 197
 198 long hammer2_process_icrc32;
 199 long hammer2_process_xxhash64;
 200 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
 201            &hammer2_process_icrc32, 0, "");
 202 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
 203            &hammer2_process_xxhash64, 0, "");
 204
 205 static int hammer2_vfs_init(struct vfsconf *conf);
 206 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 207 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 208                                 struct ucred *cred);
 209 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 210                                 struct vnode *, struct ucred *);
 211 static int hammer2_recovery(hammer2_dev_t *hmp);
 212 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 213 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 214 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 215                                 struct ucred *cred);
 216 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 217                                 struct ucred *cred);
 218 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 219                                 struct fid *fhp, struct vnode **vpp);
 220 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 221 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 222                                 int *exflagsp, struct ucred **credanonp);
 223 static void hammer2_vfs_modifying(struct mount *mp);
 224
 225 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 226 #if 0
 227 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 228 #endif
 229
 230 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 231
 232 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 233 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 234                                 hammer2_dev_t *hmp);
 235 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
 236
 237 /*
 238  * HAMMER2 vfs operations.
 239  */
 240 static struct vfsops hammer2_vfsops = {
 241         .vfs_init       = hammer2_vfs_init,
 242         .vfs_uninit     = hammer2_vfs_uninit,
 243         .vfs_sync       = hammer2_vfs_sync,
 244         .vfs_mount      = hammer2_vfs_mount,
 245         .vfs_unmount    = hammer2_vfs_unmount,
 246         .vfs_root       = hammer2_vfs_root,
 247         .vfs_statfs     = hammer2_vfs_statfs,
 248         .vfs_statvfs    = hammer2_vfs_statvfs,
 249         .vfs_vget       = hammer2_vfs_vget,
 250         .vfs_vptofh     = hammer2_vfs_vptofh,
 251         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 252         .vfs_checkexp   = hammer2_vfs_checkexp,
 253         .vfs_modifying  = hammer2_vfs_modifying
 254 };
 255
 256 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 257
 258 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 259 MODULE_VERSION(hammer2, 1);
 260
 261 static
 262 int
 263 hammer2_vfs_init(struct vfsconf *conf)
 264 {
 265         static struct objcache_malloc_args margs_read;
 266         static struct objcache_malloc_args margs_write;
 267         static struct objcache_malloc_args margs_vop;
 268
 269         int error;
 270
 271         error = 0;
 272         kmalloc_raise_limit(M_HAMMER2, 0);      /* unlimited */
 273
 274         /*
 275          * A large DIO cache is needed to retain dedup enablement masks.
 276          * The bulkfree code clears related masks as part of the disk block
 277          * recycling algorithm, preventing it from being used for a later
 278          * dedup.
 279          *
 280          * NOTE: A large buffer cache can actually interfere with dedup
 281          *       operation because we dedup based on media physical buffers
 282          *       and not logical buffers.  Try to make the DIO case large
 283          *       enough to avoid this problem, but also cap it.
 284          */
 285         hammer2_dio_limit = nbuf * 2;
 286         if (hammer2_dio_limit > 100000)
 287                 hammer2_dio_limit = 100000;
 288
 289         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 290                 error = EINVAL;
 291         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 292                 error = EINVAL;
 293         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 294                 error = EINVAL;
 295
 296         if (error)
 297                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 298
 299         margs_read.objsize = 65536;
 300         margs_read.mtype = M_HAMMER2_DEBUFFER;
 301
 302         margs_write.objsize = 32768;
 303         margs_write.mtype = M_HAMMER2_CBUFFER;
 304
 305         margs_vop.objsize = sizeof(hammer2_xop_t);
 306         margs_vop.mtype = M_HAMMER2;
 307
 308         /*
 309          * Note thaht for the XOPS cache we want backing store allocations
 310          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 311          * confusion), so use the backing store function that does it.  This
 312          * means that initial XOPS objects are zerod but REUSED objects are
 313          * not.  So we are responsible for cleaning the object up sufficiently
 314          * for our needs before objcache_put()ing it back (typically just the
 315          * FIFO indices).
 316          */
 317         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 318                                 0, 1, NULL, NULL, NULL,
 319                                 objcache_malloc_alloc,
 320                                 objcache_malloc_free,
 321                                 &margs_read);
 322         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 323                                 0, 1, NULL, NULL, NULL,
 324                                 objcache_malloc_alloc,
 325                                 objcache_malloc_free,
 326                                 &margs_write);
 327         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 328                                 0, 1, NULL, NULL, NULL,
 329                                 objcache_malloc_alloc_zero,
 330                                 objcache_malloc_free,
 331                                 &margs_vop);
 332
 333
 334         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 335         TAILQ_INIT(&hammer2_mntlist);
 336         TAILQ_INIT(&hammer2_pfslist);
 337         TAILQ_INIT(&hammer2_spmplist);
 338
 339         hammer2_limit_dirty_chains = maxvnodes / 10;
 340         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 341                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 342         if (hammer2_limit_dirty_chains < 1000)
 343                 hammer2_limit_dirty_chains = 1000;
 344
 345         hammer2_limit_dirty_inodes = maxvnodes / 25;
 346         if (hammer2_limit_dirty_inodes < 100)
 347                 hammer2_limit_dirty_inodes = 100;
 348         if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
 349                 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
 350
 351         return (error);
 352 }
 353
 354 static
 355 int
 356 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 357 {
 358         objcache_destroy(cache_buffer_read);
 359         objcache_destroy(cache_buffer_write);
 360         objcache_destroy(cache_xops);
 361         return 0;
 362 }
 363
 364 /*
 365  * Core PFS allocator.  Used to allocate or reference the pmp structure
 366  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 367  * The pmp can be passed in or loaded by this function using the chain and
 368  * inode data.
 369  *
 370  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 371  * transactions.  Note that synchronization does not use this field.
 372  * (typically frontend operations and synchronization cannot run on the
 373  * same PFS node at the same time).
 374  *
 375  * XXX check locking
 376  */
 377 hammer2_pfs_t *
 378 hammer2_pfsalloc(hammer2_chain_t *chain,
 379                  const hammer2_inode_data_t *ripdata,
 380                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 381 {
 382         hammer2_pfs_t *pmp;
 383         hammer2_inode_t *iroot;
 384         int count;
 385         int i;
 386         int j;
 387
 388         pmp = NULL;
 389
 390         /*
 391          * Locate or create the PFS based on the cluster id.  If ripdata
 392          * is NULL this is a spmp which is unique and is always allocated.
 393          *
 394          * If the device is mounted in local mode all PFSs are considered
 395          * independent and not part of any cluster (for debugging only).
 396          */
 397         if (ripdata) {
 398                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 399                         if (force_local != pmp->force_local)
 400                                 continue;
 401                         if (force_local == NULL &&
 402                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 403                                  sizeof(pmp->pfs_clid)) == 0) {
 404                                         break;
 405                         } else if (force_local && pmp->pfs_names[0] &&
 406                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 407                                         break;
 408                         }
 409                 }
 410         }
 411
 412         if (pmp == NULL) {
 413                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 414                 pmp->force_local = force_local;
 415                 hammer2_trans_manage_init(pmp);
 416                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 417                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 418                 lockinit(&pmp->lock, "pfslk", 0, 0);
 419                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 420                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 421                 spin_init(&pmp->xop_spin, "h2xop");
 422                 spin_init(&pmp->lru_spin, "h2lru");
 423                 RB_INIT(&pmp->inum_tree);
 424                 TAILQ_INIT(&pmp->syncq);
 425                 TAILQ_INIT(&pmp->depq);
 426                 TAILQ_INIT(&pmp->lru_list);
 427                 spin_init(&pmp->list_spin, "h2pfsalloc_list");
 428
 429                 /*
 430                  * Distribute backend operations to threads
 431                  */
 432                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 433                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 434
 435                 /*
 436                  * Save the last media transaction id for the flusher.  Set
 437                  * initial
 438                  */
 439                 if (ripdata) {
 440                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 441                         TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 442                 } else {
 443                         pmp->flags |= HAMMER2_PMPF_SPMP;
 444                         TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
 445                 }
 446
 447                 /*
 448                  * The synchronization thread may start too early, make
 449                  * sure it stays frozen until we are ready to let it go.
 450                  * XXX
 451                  */
 452                 /*
 453                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 454                                          HAMMER2_THREAD_REMASTER;
 455                 */
 456         }
 457
 458         /*
 459          * Create the PFS's root inode and any missing XOP helper threads.
 460          */
 461         if ((iroot = pmp->iroot) == NULL) {
 462                 iroot = hammer2_inode_get(pmp, NULL, 1, -1);
 463                 if (ripdata)
 464                         iroot->meta = ripdata->meta;
 465                 pmp->iroot = iroot;
 466                 hammer2_inode_ref(iroot);
 467                 hammer2_inode_unlock(iroot);
 468         }
 469
 470         /*
 471          * Stop here if no chain is passed in.
 472          */
 473         if (chain == NULL)
 474                 goto done;
 475
 476         /*
 477          * When a chain is passed in we must add it to the PFS's root
 478          * inode, update pmp->pfs_types[], and update the syncronization
 479          * threads.
 480          *
 481          * When forcing local mode, mark the PFS as a MASTER regardless.
 482          *
 483          * At the moment empty spots can develop due to removals or failures.
 484          * Ultimately we want to re-fill these spots but doing so might
 485          * confused running code. XXX
 486          */
 487         hammer2_inode_ref(iroot);
 488         hammer2_mtx_ex(&iroot->lock);
 489         j = iroot->cluster.nchains;
 490
 491         if (j == HAMMER2_MAXCLUSTER) {
 492                 kprintf("hammer2_mount: cluster full!\n");
 493                 /* XXX fatal error? */
 494         } else {
 495                 KKASSERT(chain->pmp == NULL);
 496                 chain->pmp = pmp;
 497                 hammer2_chain_ref(chain);
 498                 iroot->cluster.array[j].chain = chain;
 499                 if (force_local)
 500                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 501                 else
 502                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 503                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 504                 pmp->pfs_hmps[j] = chain->hmp;
 505                 hammer2_spin_ex(&pmp->inum_spin);
 506                 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset;
 507                 hammer2_spin_unex(&pmp->inum_spin);
 508
 509                 /*
 510                  * If the PFS is already mounted we must account
 511                  * for the mount_count here.
 512                  */
 513                 if (pmp->mp)
 514                         ++chain->hmp->mount_count;
 515
 516                 /*
 517                  * May have to fixup dirty chain tracking.  Previous
 518                  * pmp was NULL so nothing to undo.
 519                  */
 520                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 521                         hammer2_pfs_memory_inc(pmp);
 522                 ++j;
 523         }
 524         iroot->cluster.nchains = j;
 525
 526         /*
 527          * Update nmasters from any PFS inode which is part of the cluster.
 528          * It is possible that this will result in a value which is too
 529          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 530          * override this value later on.
 531          *
 532          * (This informs us of masters that might not currently be
 533          *  discoverable by this mount).
 534          */
 535         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 536                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 537         }
 538
 539         /*
 540          * Count visible masters.  Masters are usually added with
 541          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 542          * are more (XXX and must update the master inodes).
 543          */
 544         count = 0;
 545         for (i = 0; i < iroot->cluster.nchains; ++i) {
 546                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 547                         ++count;
 548         }
 549         if (pmp->pfs_nmasters < count)
 550                 pmp->pfs_nmasters = count;
 551
 552         /*
 553          * Create missing synchronization and support threads.
 554          *
 555          * Single-node masters (including snapshots) have nothing to
 556          * synchronize and do not require this thread.
 557          *
 558          * Multi-node masters or any number of soft masters, slaves, copy,
 559          * or other PFS types need the thread.
 560          *
 561          * Each thread is responsible for its particular cluster index.
 562          * We use independent threads so stalls or mismatches related to
 563          * any given target do not affect other targets.
 564          */
 565         for (i = 0; i < iroot->cluster.nchains; ++i) {
 566                 /*
 567                  * Single-node masters (including snapshots) have nothing
 568                  * to synchronize and will make direct xops support calls,
 569                  * thus they do not require this thread.
 570                  *
 571                  * Note that there can be thousands of snapshots.  We do not
 572                  * want to create thousands of threads.
 573                  */
 574                 if (pmp->pfs_nmasters <= 1 &&
 575                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 576                         continue;
 577                 }
 578
 579                 /*
 580                  * Sync support thread
 581                  */
 582                 if (pmp->sync_thrs[i].td == NULL) {
 583                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 584                                            "h2nod", i, -1,
 585                                            hammer2_primary_sync_thread);
 586                 }
 587         }
 588
 589         /*
 590          * Create missing Xop threads
 591          *
 592          * NOTE: We create helper threads for all mounted PFSs or any
 593          *       PFSs with 2+ nodes (so the sync thread can update them,
 594          *       even if not mounted).
 595          */
 596         if (pmp->mp || iroot->cluster.nchains >= 2)
 597                 hammer2_xop_helper_create(pmp);
 598
 599         hammer2_mtx_unlock(&iroot->lock);
 600         hammer2_inode_drop(iroot);
 601 done:
 602         return pmp;
 603 }
 604
 605 /*
 606  * Deallocate an element of a probed PFS.  If destroying and this is a
 607  * MASTER, adjust nmasters.
 608  *
 609  * This function does not physically destroy the PFS element in its device
 610  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 611  */
 612 void
 613 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 614 {
 615         hammer2_inode_t *iroot;
 616         hammer2_chain_t *chain;
 617         int j;
 618
 619         /*
 620          * Cleanup our reference on iroot.  iroot is (should) not be needed
 621          * by the flush code.
 622          */
 623         iroot = pmp->iroot;
 624         if (iroot) {
 625                 /*
 626                  * Stop synchronizing
 627                  *
 628                  * XXX flush after acquiring the iroot lock.
 629                  * XXX clean out the cluster index from all inode structures.
 630                  */
 631                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 632
 633                 /*
 634                  * Remove the cluster index from the group.  If destroying
 635                  * the PFS and this is a master, adjust pfs_nmasters.
 636                  */
 637                 hammer2_mtx_ex(&iroot->lock);
 638                 chain = iroot->cluster.array[clindex].chain;
 639                 iroot->cluster.array[clindex].chain = NULL;
 640
 641                 switch(pmp->pfs_types[clindex]) {
 642                 case HAMMER2_PFSTYPE_MASTER:
 643                         if (destroying && pmp->pfs_nmasters > 0)
 644                                 --pmp->pfs_nmasters;
 645                         /* XXX adjust ripdata->meta.pfs_nmasters */
 646                         break;
 647                 default:
 648                         break;
 649                 }
 650                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 651
 652                 hammer2_mtx_unlock(&iroot->lock);
 653
 654                 /*
 655                  * Release the chain.
 656                  */
 657                 if (chain) {
 658                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 659                         hammer2_chain_drop(chain);
 660                 }
 661
 662                 /*
 663                  * Terminate all XOP threads for the cluster index.
 664                  */
 665                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 666                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 667         }
 668 }
 669
 670 /*
 671  * Destroy a PFS, typically only occurs after the last mount on a device
 672  * has gone away.
 673  */
 674 static void
 675 hammer2_pfsfree(hammer2_pfs_t *pmp)
 676 {
 677         hammer2_inode_t *iroot;
 678         hammer2_chain_t *chain;
 679         int chains_still_present = 0;
 680         int i;
 681         int j;
 682
 683         /*
 684          * Cleanup our reference on iroot.  iroot is (should) not be needed
 685          * by the flush code.
 686          */
 687         if (pmp->flags & HAMMER2_PMPF_SPMP)
 688                 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
 689         else
 690                 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 691
 692         /*
 693          * Cleanup chains remaining on LRU list.
 694          */
 695         hammer2_spin_ex(&pmp->lru_spin);
 696         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 697                 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
 698                 atomic_add_int(&pmp->lru_count, -1);
 699                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
 700                 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
 701                 hammer2_chain_ref(chain);
 702                 hammer2_spin_unex(&pmp->lru_spin);
 703                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 704                 hammer2_chain_drop(chain);
 705                 hammer2_spin_ex(&pmp->lru_spin);
 706         }
 707         hammer2_spin_unex(&pmp->lru_spin);
 708
 709         /*
 710          * Clean up iroot
 711          */
 712         iroot = pmp->iroot;
 713         if (iroot) {
 714                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 715                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 716                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 717                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 718                         chain = iroot->cluster.array[i].chain;
 719                         if (chain && !RB_EMPTY(&chain->core.rbtree)) {
 720                                 kprintf("hammer2: Warning pmp %p still "
 721                                         "has active chains\n", pmp);
 722                                 chains_still_present = 1;
 723                         }
 724                 }
 725 #if REPORT_REFS_ERRORS
 726                 if (iroot->refs != 1)
 727                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 728                                 iroot, iroot->refs);
 729 #else
 730                 KKASSERT(iroot->refs == 1);
 731 #endif
 732                 /* ref for iroot */
 733                 hammer2_inode_drop(iroot);
 734                 pmp->iroot = NULL;
 735         }
 736
 737         /*
 738          * Free remaining pmp resources
 739          */
 740         if (chains_still_present) {
 741                 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp);
 742         } else {
 743                 kmalloc_destroy(&pmp->mmsg);
 744                 kmalloc_destroy(&pmp->minode);
 745                 kfree(pmp, M_HAMMER2);
 746         }
 747 }
 748
 749 /*
 750  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 751  * empty is terminated and freed.
 752  *
 753  * XXX inefficient.
 754  */
 755 static void
 756 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
 757 {
 758         hammer2_pfs_t *pmp;
 759         hammer2_inode_t *iroot;
 760         hammer2_chain_t *rchain;
 761         int i;
 762         int j;
 763         struct hammer2_pfslist *wlist;
 764
 765         if (which == 0)
 766                 wlist = &hammer2_pfslist;
 767         else
 768                 wlist = &hammer2_spmplist;
 769 again:
 770         TAILQ_FOREACH(pmp, wlist, mntentry) {
 771                 if ((iroot = pmp->iroot) == NULL)
 772                         continue;
 773
 774                 /*
 775                  * Determine if this PFS is affected.  If it is we must
 776                  * freeze all management threads and lock its iroot.
 777                  *
 778                  * Freezing a management thread forces it idle, operations
 779                  * in-progress will be aborted and it will have to start
 780                  * over again when unfrozen, or exit if told to exit.
 781                  */
 782                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 783                         if (pmp->pfs_hmps[i] == hmp)
 784                                 break;
 785                 }
 786                 if (i == HAMMER2_MAXCLUSTER)
 787                         continue;
 788
 789                 hammer2_vfs_sync_pmp(pmp, MNT_WAIT);
 790
 791                 /*
 792                  * Make sure all synchronization threads are locked
 793                  * down.
 794                  */
 795                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 796                         if (pmp->pfs_hmps[i] == NULL)
 797                                 continue;
 798                         hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 799                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 800                                 hammer2_thr_freeze_async(
 801                                         &pmp->xop_groups[j].thrs[i]);
 802                         }
 803                 }
 804                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 805                         if (pmp->pfs_hmps[i] == NULL)
 806                                 continue;
 807                         hammer2_thr_freeze(&pmp->sync_thrs[i]);
 808                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 809                                 hammer2_thr_freeze(
 810                                         &pmp->xop_groups[j].thrs[i]);
 811                         }
 812                 }
 813
 814                 /*
 815                  * Lock the inode and clean out matching chains.
 816                  * Note that we cannot use hammer2_inode_lock_*()
 817                  * here because that would attempt to validate the
 818                  * cluster that we are in the middle of ripping
 819                  * apart.
 820                  *
 821                  * WARNING! We are working directly on the inodes
 822                  *          embedded cluster.
 823                  */
 824                 hammer2_mtx_ex(&iroot->lock);
 825
 826                 /*
 827                  * Remove the chain from matching elements of the PFS.
 828                  */
 829                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 830                         if (pmp->pfs_hmps[i] != hmp)
 831                                 continue;
 832                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 833                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 834                                 hammer2_thr_delete(
 835                                         &pmp->xop_groups[j].thrs[i]);
 836                         }
 837                         rchain = iroot->cluster.array[i].chain;
 838                         iroot->cluster.array[i].chain = NULL;
 839                         pmp->pfs_types[i] = 0;
 840                         if (pmp->pfs_names[i]) {
 841                                 kfree(pmp->pfs_names[i], M_HAMMER2);
 842                                 pmp->pfs_names[i] = NULL;
 843                         }
 844                         if (rchain) {
 845                                 hammer2_chain_drop(rchain);
 846                                 /* focus hint */
 847                                 if (iroot->cluster.focus == rchain)
 848                                         iroot->cluster.focus = NULL;
 849                         }
 850                         pmp->pfs_hmps[i] = NULL;
 851                 }
 852                 hammer2_mtx_unlock(&iroot->lock);
 853
 854                 /*
 855                  * Cleanup trailing chains.  Gaps may remain.
 856                  */
 857                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 858                         if (pmp->pfs_hmps[i])
 859                                 break;
 860                 }
 861                 iroot->cluster.nchains = i + 1;
 862
 863                 /*
 864                  * If the PMP has no elements remaining we can destroy it.
 865                  * (this will transition management threads from frozen->exit).
 866                  */
 867                 if (iroot->cluster.nchains == 0) {
 868                         /*
 869                          * If this was the hmp's spmp, we need to clean
 870                          * a little more stuff out.
 871                          */
 872                         if (hmp->spmp == pmp) {
 873                                 hmp->spmp = NULL;
 874                                 hmp->vchain.pmp = NULL;
 875                                 hmp->fchain.pmp = NULL;
 876                         }
 877
 878                         /*
 879                          * Free the pmp and restart the loop
 880                          */
 881                         KKASSERT(TAILQ_EMPTY(&pmp->syncq));
 882                         KKASSERT(TAILQ_EMPTY(&pmp->depq));
 883                         hammer2_pfsfree(pmp);
 884                         goto again;
 885                 }
 886
 887                 /*
 888                  * If elements still remain we need to set the REMASTER
 889                  * flag and unfreeze it.
 890                  */
 891                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 892                         if (pmp->pfs_hmps[i] == NULL)
 893                                 continue;
 894                         hammer2_thr_remaster(&pmp->sync_thrs[i]);
 895                         hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 896                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 897                                 hammer2_thr_remaster(
 898                                         &pmp->xop_groups[j].thrs[i]);
 899                                 hammer2_thr_unfreeze(
 900                                         &pmp->xop_groups[j].thrs[i]);
 901                         }
 902                 }
 903         }
 904 }
 905
 906 /*
 907  * Mount or remount HAMMER2 fileystem from physical media
 908  *
 909  *      mountroot
 910  *              mp              mount point structure
 911  *              path            NULL
 912  *              data            <unused>
 913  *              cred            <unused>
 914  *
 915  *      mount
 916  *              mp              mount point structure
 917  *              path            path to mount point
 918  *              data            pointer to argument structure in user space
 919  *                      volume  volume path (device@LABEL form)
 920  *                      hflags  user mount flags
 921  *              cred            user credentials
 922  *
 923  * RETURNS:     0       Success
 924  *              !0      error number
 925  */
 926 static
 927 int
 928 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 929                   struct ucred *cred)
 930 {
 931         struct hammer2_mount_info info;
 932         hammer2_pfs_t *pmp;
 933         hammer2_pfs_t *spmp;
 934         hammer2_dev_t *hmp;
 935         hammer2_dev_t *force_local;
 936         hammer2_key_t key_next;
 937         hammer2_key_t key_dummy;
 938         hammer2_key_t lhc;
 939         struct vnode *devvp;
 940         struct nlookupdata nd;
 941         hammer2_chain_t *parent;
 942         hammer2_chain_t *chain;
 943         const hammer2_inode_data_t *ripdata;
 944         hammer2_blockref_t bref;
 945         struct file *fp;
 946         char devstr[MNAMELEN];
 947         size_t size;
 948         size_t done;
 949         char *dev;
 950         char *label;
 951         int ronly = 1;
 952         int error;
 953         int i;
 954
 955         hmp = NULL;
 956         pmp = NULL;
 957         dev = NULL;
 958         label = NULL;
 959         devvp = NULL;
 960
 961         if (path == NULL) {
 962                 /*
 963                  * Root mount
 964                  */
 965                 bzero(&info, sizeof(info));
 966                 info.cluster_fd = -1;
 967                 ksnprintf(devstr, sizeof(devstr), "%s",
 968                           mp->mnt_stat.f_mntfromname);
 969                 kprintf("hammer2_mount: root '%s'\n", devstr);
 970                 done = strlen(devstr) + 1;
 971         } else {
 972                 /*
 973                  * Non-root mount or updating a mount
 974                  */
 975                 error = copyin(data, &info, sizeof(info));
 976                 if (error)
 977                         return (error);
 978
 979                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 980                 if (error)
 981                         return (error);
 982                 kprintf("hammer2_mount: '%s'\n", devstr);
 983         }
 984
 985         /*
 986          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 987          * if no label specified, based on the partition id.  Error out if no
 988          * label or device (with partition id) is specified.  This is strictly
 989          * a convenience to match the default label created by newfs_hammer2,
 990          * our preference is that a label always be specified.
 991          *
 992          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
 993          *       that does not specify a device, as long as some H2 label
 994          *       has already been mounted from that device.  This makes
 995          *       mounting snapshots a lot easier.
 996          */
 997         dev = devstr;
 998         label = strchr(devstr, '@');
 999         if (label && ((label + 1) - dev) > done) {
1000                 kprintf("hammer2: mount: bad label %s/%zd\n",
1001                         devstr, done);
1002                 return (EINVAL);
1003         }
1004         if (label == NULL || label[1] == 0) {
1005                 char slice;
1006
1007                 if (label == NULL)
1008                         label = devstr + strlen(devstr);
1009                 else
1010                         *label = '\0';          /* clean up trailing @ */
1011
1012                 slice = label[-1];
1013                 switch(slice) {
1014                 case 'a':
1015                         label = "BOOT";
1016                         break;
1017                 case 'd':
1018                         label = "ROOT";
1019                         break;
1020                 default:
1021                         label = "DATA";
1022                         break;
1023                 }
1024         } else {
1025                 *label = '\0';
1026                 label++;
1027         }
1028
1029         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
1030                 dev, label, (mp->mnt_flag & MNT_RDONLY));
1031
1032         if (mp->mnt_flag & MNT_UPDATE) {
1033                 /*
1034                  * Update mount.  Note that pmp->iroot->cluster is
1035                  * an inode-embedded cluster and thus cannot be
1036                  * directly locked.
1037                  *
1038                  * XXX HAMMER2 needs to implement NFS export via
1039                  *     mountctl.
1040                  */
1041                 hammer2_cluster_t *cluster;
1042
1043                 pmp = MPTOPMP(mp);
1044                 pmp->hflags = info.hflags;
1045                 cluster = &pmp->iroot->cluster;
1046                 for (i = 0; i < cluster->nchains; ++i) {
1047                         if (cluster->array[i].chain == NULL)
1048                                 continue;
1049                         hmp = cluster->array[i].chain->hmp;
1050                         devvp = hmp->devvp;
1051                         error = hammer2_remount(hmp, mp, path,
1052                                                 devvp, cred);
1053                         if (error)
1054                                 break;
1055                 }
1056
1057                 return error;
1058         }
1059
1060         /*
1061          * HMP device mount
1062          *
1063          * If a path is specified and dev is not an empty string, lookup the
1064          * name and verify that it referes to a block device.
1065          *
1066          * If a path is specified and dev is an empty string we fall through
1067          * and locate the label in the hmp search.
1068          */
1069         if (path && *dev != 0) {
1070                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1071                 if (error == 0)
1072                         error = nlookup(&nd);
1073                 if (error == 0)
1074                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1075                 nlookup_done(&nd);
1076         } else if (path == NULL) {
1077                 /* root mount */
1078                 cdev_t cdev = kgetdiskbyname(dev);
1079                 error = bdevvp(cdev, &devvp);
1080                 if (error)
1081                         kprintf("hammer2: cannot find '%s'\n", dev);
1082         } else {
1083                 /*
1084                  * We will locate the hmp using the label in the hmp loop.
1085                  */
1086                 error = 0;
1087         }
1088
1089         /*
1090          * Make sure its a block device.  Do not check to see if it is
1091          * already mounted until we determine that its a fresh H2 device.
1092          */
1093         if (error == 0 && devvp) {
1094                 vn_isdisk(devvp, &error);
1095         }
1096
1097         /*
1098          * Determine if the device has already been mounted.  After this
1099          * check hmp will be non-NULL if we are doing the second or more
1100          * hammer2 mounts from the same device.
1101          */
1102         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1103         if (devvp) {
1104                 /*
1105                  * Match the device.  Due to the way devfs works,
1106                  * we may not be able to directly match the vnode pointer,
1107                  * so also check to see if the underlying device matches.
1108                  */
1109                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1110                         if (hmp->devvp == devvp)
1111                                 break;
1112                         if (devvp->v_rdev &&
1113                             hmp->devvp->v_rdev == devvp->v_rdev) {
1114                                 break;
1115                         }
1116                 }
1117
1118                 /*
1119                  * If no match this may be a fresh H2 mount, make sure
1120                  * the device is not mounted on anything else.
1121                  */
1122                 if (hmp == NULL)
1123                         error = vfs_mountedon(devvp);
1124         } else if (error == 0) {
1125                 /*
1126                  * Match the label to a pmp already probed.
1127                  */
1128                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1129                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1130                                 if (pmp->pfs_names[i] &&
1131                                     strcmp(pmp->pfs_names[i], label) == 0) {
1132                                         hmp = pmp->pfs_hmps[i];
1133                                         break;
1134                                 }
1135                         }
1136                         if (hmp)
1137                                 break;
1138                 }
1139                 if (hmp == NULL)
1140                         error = ENOENT;
1141         }
1142
1143         /*
1144          * Open the device if this isn't a secondary mount and construct
1145          * the H2 device mount (hmp).
1146          */
1147         if (hmp == NULL) {
1148                 hammer2_chain_t *schain;
1149                 hammer2_xid_t xid;
1150                 hammer2_xop_head_t xop;
1151
1152                 if (error == 0 && vcount(devvp) > 0) {
1153                         kprintf("Primary device already has references\n");
1154                         error = EBUSY;
1155                 }
1156
1157                 /*
1158                  * Now open the device
1159                  */
1160                 if (error == 0) {
1161                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1162                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1163                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1164                         if (error == 0) {
1165                                 error = VOP_OPEN(devvp,
1166                                              (ronly ? FREAD : FREAD | FWRITE),
1167                                              FSCRED, NULL);
1168                         }
1169                         vn_unlock(devvp);
1170                 }
1171                 if (error && devvp) {
1172                         vrele(devvp);
1173                         devvp = NULL;
1174                 }
1175                 if (error) {
1176                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1177                         return error;
1178                 }
1179                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1180                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1181                 hmp->ronly = ronly;
1182                 hmp->devvp = devvp;
1183                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1184                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1185                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1186                 RB_INIT(&hmp->iotree);
1187                 spin_init(&hmp->io_spin, "h2mount_io");
1188                 spin_init(&hmp->list_spin, "h2mount_list");
1189
1190                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1191                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1192                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1193
1194                 /*
1195                  * vchain setup. vchain.data is embedded.
1196                  * vchain.refs is initialized and will never drop to 0.
1197                  *
1198                  * NOTE! voldata is not yet loaded.
1199                  */
1200                 hmp->vchain.hmp = hmp;
1201                 hmp->vchain.refs = 1;
1202                 hmp->vchain.data = (void *)&hmp->voldata;
1203                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1204                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1205                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1206
1207                 hammer2_chain_core_init(&hmp->vchain);
1208                 /* hmp->vchain.u.xxx is left NULL */
1209
1210                 /*
1211                  * fchain setup.  fchain.data is embedded.
1212                  * fchain.refs is initialized and will never drop to 0.
1213                  *
1214                  * The data is not used but needs to be initialized to
1215                  * pass assertion muster.  We use this chain primarily
1216                  * as a placeholder for the freemap's top-level RBTREE
1217                  * so it does not interfere with the volume's topology
1218                  * RBTREE.
1219                  */
1220                 hmp->fchain.hmp = hmp;
1221                 hmp->fchain.refs = 1;
1222                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1223                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1224                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1225                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1226                 hmp->fchain.bref.methods =
1227                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1228                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1229
1230                 hammer2_chain_core_init(&hmp->fchain);
1231                 /* hmp->fchain.u.xxx is left NULL */
1232
1233                 /*
1234                  * Install the volume header and initialize fields from
1235                  * voldata.
1236                  */
1237                 error = hammer2_install_volume_header(hmp);
1238                 if (error) {
1239                         hammer2_unmount_helper(mp, NULL, hmp);
1240                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1241                         hammer2_vfs_unmount(mp, MNT_FORCE);
1242                         return error;
1243                 }
1244
1245                 /*
1246                  * Really important to get these right or the flush and
1247                  * teardown code will get confused.
1248                  */
1249                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1250                 spmp = hmp->spmp;
1251                 spmp->pfs_hmps[0] = hmp;
1252
1253                 /*
1254                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1255                  * is inherited from the volume header.
1256                  */
1257                 xid = 0;
1258                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1259                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1260                 hmp->vchain.pmp = spmp;
1261                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1262                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1263                 hmp->fchain.pmp = spmp;
1264
1265                 /*
1266                  * First locate the super-root inode, which is key 0
1267                  * relative to the volume header's blockset.
1268                  *
1269                  * Then locate the root inode by scanning the directory keyspace
1270                  * represented by the label.
1271                  */
1272                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1273                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1274                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1275                                       &error, 0);
1276                 hammer2_chain_lookup_done(parent);
1277                 if (schain == NULL) {
1278                         kprintf("hammer2_mount: invalid super-root\n");
1279                         hammer2_unmount_helper(mp, NULL, hmp);
1280                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1281                         hammer2_vfs_unmount(mp, MNT_FORCE);
1282                         return EINVAL;
1283                 }
1284                 if (schain->error) {
1285                         kprintf("hammer2_mount: error %s reading super-root\n",
1286                                 hammer2_error_str(schain->error));
1287                         hammer2_chain_unlock(schain);
1288                         hammer2_chain_drop(schain);
1289                         schain = NULL;
1290                         hammer2_unmount_helper(mp, NULL, hmp);
1291                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1292                         hammer2_vfs_unmount(mp, MNT_FORCE);
1293                         return EINVAL;
1294                 }
1295
1296                 /*
1297                  * The super-root always uses an inode_tid of 1 when
1298                  * creating PFSs.
1299                  */
1300                 spmp->inode_tid = 1;
1301                 spmp->modify_tid = schain->bref.modify_tid + 1;
1302
1303                 /*
1304                  * Sanity-check schain's pmp and finish initialization.
1305                  * Any chain belonging to the super-root topology should
1306                  * have a NULL pmp (not even set to spmp).
1307                  */
1308                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1309                 KKASSERT(schain->pmp == NULL);
1310                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1311
1312                 /*
1313                  * Replace the dummy spmp->iroot with a real one.  It's
1314                  * easier to just do a wholesale replacement than to try
1315                  * to update the chain and fixup the iroot fields.
1316                  *
1317                  * The returned inode is locked with the supplied cluster.
1318                  */
1319                 hammer2_dummy_xop_from_chain(&xop, schain);
1320                 hammer2_inode_drop(spmp->iroot);
1321                 spmp->iroot = NULL;
1322                 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1);
1323                 spmp->spmp_hmp = hmp;
1324                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1325                 spmp->pfs_hmps[0] = hmp;
1326                 hammer2_inode_ref(spmp->iroot);
1327                 hammer2_inode_unlock(spmp->iroot);
1328                 hammer2_cluster_unlock(&xop.cluster);
1329                 hammer2_chain_drop(schain);
1330                 /* do not call hammer2_cluster_drop() on an embedded cluster */
1331                 schain = NULL;  /* now invalid */
1332                 /* leave spmp->iroot with one ref */
1333
1334                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1335                         error = hammer2_recovery(hmp);
1336                         if (error == 0)
1337                                 error |= hammer2_fixup_pfses(hmp);
1338                         /* XXX do something with error */
1339                 }
1340                 hammer2_update_pmps(hmp);
1341                 hammer2_iocom_init(hmp);
1342                 hammer2_bulkfree_init(hmp);
1343
1344                 /*
1345                  * Ref the cluster management messaging descriptor.  The mount
1346                  * program deals with the other end of the communications pipe.
1347                  *
1348                  * Root mounts typically do not supply one.
1349                  */
1350                 if (info.cluster_fd >= 0) {
1351                         fp = holdfp(curthread, info.cluster_fd, -1);
1352                         if (fp) {
1353                                 hammer2_cluster_reconnect(hmp, fp);
1354                         } else {
1355                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1356                         }
1357                 }
1358         } else {
1359                 spmp = hmp->spmp;
1360                 if (info.hflags & HMNT2_DEVFLAGS) {
1361                         kprintf("hammer2: Warning: mount flags pertaining "
1362                                 "to the whole device may only be specified "
1363                                 "on the first mount of the device: %08x\n",
1364                                 info.hflags & HMNT2_DEVFLAGS);
1365                 }
1366         }
1367
1368         /*
1369          * Force local mount (disassociate all PFSs from their clusters).
1370          * Used primarily for debugging.
1371          */
1372         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1373
1374         /*
1375          * Lookup the mount point under the media-localized super-root.
1376          * Scanning hammer2_pfslist doesn't help us because it represents
1377          * PFS cluster ids which can aggregate several named PFSs together.
1378          *
1379          * cluster->pmp will incorrectly point to spmp and must be fixed
1380          * up later on.
1381          */
1382         hammer2_inode_lock(spmp->iroot, 0);
1383         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1384         lhc = hammer2_dirhash(label, strlen(label));
1385         chain = hammer2_chain_lookup(&parent, &key_next,
1386                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1387                                      &error, 0);
1388         while (chain) {
1389                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1390                     strcmp(label, chain->data->ipdata.filename) == 0) {
1391                         break;
1392                 }
1393                 chain = hammer2_chain_next(&parent, chain, &key_next,
1394                                             key_next,
1395                                             lhc + HAMMER2_DIRHASH_LOMASK,
1396                                             &error, 0);
1397         }
1398         if (parent) {
1399                 hammer2_chain_unlock(parent);
1400                 hammer2_chain_drop(parent);
1401         }
1402         hammer2_inode_unlock(spmp->iroot);
1403
1404         /*
1405          * PFS could not be found?
1406          */
1407         if (chain == NULL) {
1408                 if (error)
1409                         kprintf("hammer2_mount: PFS label I/O error\n");
1410                 else
1411                         kprintf("hammer2_mount: PFS label not found\n");
1412                 hammer2_unmount_helper(mp, NULL, hmp);
1413                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1414                 hammer2_vfs_unmount(mp, MNT_FORCE);
1415
1416                 return EINVAL;
1417         }
1418
1419         /*
1420          * Acquire the pmp structure (it should have already been allocated
1421          * via hammer2_update_pmps() so do not pass cluster in to add to
1422          * available chains).
1423          *
1424          * Check if the cluster has already been mounted.  A cluster can
1425          * only be mounted once, use null mounts to mount additional copies.
1426          */
1427         if (chain->error) {
1428                 kprintf("hammer2_mount: PFS label I/O error\n");
1429         } else {
1430                 ripdata = &chain->data->ipdata;
1431                 bref = chain->bref;
1432                 pmp = hammer2_pfsalloc(NULL, ripdata,
1433                                        bref.modify_tid, force_local);
1434         }
1435         hammer2_chain_unlock(chain);
1436         hammer2_chain_drop(chain);
1437
1438         /*
1439          * Finish the mount
1440          */
1441         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1442
1443         if (pmp->mp) {
1444                 kprintf("hammer2_mount: PFS already mounted!\n");
1445                 hammer2_unmount_helper(mp, NULL, hmp);
1446                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1447                 hammer2_vfs_unmount(mp, MNT_FORCE);
1448
1449                 return EBUSY;
1450         }
1451
1452         pmp->hflags = info.hflags;
1453         mp->mnt_flag |= MNT_LOCAL;
1454         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1455         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1456
1457         /*
1458          * required mount structure initializations
1459          */
1460         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1461         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1462
1463         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1464         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1465
1466         /*
1467          * Optional fields
1468          */
1469         mp->mnt_iosize_max = MAXPHYS;
1470
1471         /*
1472          * Connect up mount pointers.
1473          */
1474         hammer2_mount_helper(mp, pmp);
1475
1476         lockmgr(&hammer2_mntlk, LK_RELEASE);
1477
1478         /*
1479          * Finish setup
1480          */
1481         vfs_getnewfsid(mp);
1482         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1483         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1484         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1485
1486         if (path) {
1487                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1488                           MNAMELEN - 1, &size);
1489                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1490         } /* else root mount, already in there */
1491
1492         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1493         if (path) {
1494                 copyinstr(path, mp->mnt_stat.f_mntonname,
1495                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1496                           &size);
1497         } else {
1498                 /* root mount */
1499                 mp->mnt_stat.f_mntonname[0] = '/';
1500         }
1501
1502         /*
1503          * Initial statfs to prime mnt_stat.
1504          */
1505         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1506
1507         return 0;
1508 }
1509
1510 /*
1511  * Scan PFSs under the super-root and create hammer2_pfs structures.
1512  */
1513 static
1514 void
1515 hammer2_update_pmps(hammer2_dev_t *hmp)
1516 {
1517         const hammer2_inode_data_t *ripdata;
1518         hammer2_chain_t *parent;
1519         hammer2_chain_t *chain;
1520         hammer2_blockref_t bref;
1521         hammer2_dev_t *force_local;
1522         hammer2_pfs_t *spmp;
1523         hammer2_pfs_t *pmp;
1524         hammer2_key_t key_next;
1525         int error;
1526
1527         /*
1528          * Force local mount (disassociate all PFSs from their clusters).
1529          * Used primarily for debugging.
1530          */
1531         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1532
1533         /*
1534          * Lookup mount point under the media-localized super-root.
1535          *
1536          * cluster->pmp will incorrectly point to spmp and must be fixed
1537          * up later on.
1538          */
1539         spmp = hmp->spmp;
1540         hammer2_inode_lock(spmp->iroot, 0);
1541         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1542         chain = hammer2_chain_lookup(&parent, &key_next,
1543                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1544                                          &error, 0);
1545         while (chain) {
1546                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1547                         continue;
1548                 if (chain->error) {
1549                         kprintf("I/O error scanning PFS labels\n");
1550                 } else {
1551                         ripdata = &chain->data->ipdata;
1552                         bref = chain->bref;
1553
1554                         pmp = hammer2_pfsalloc(chain, ripdata,
1555                                                bref.modify_tid, force_local);
1556                 }
1557                 chain = hammer2_chain_next(&parent, chain, &key_next,
1558                                            key_next, HAMMER2_KEY_MAX,
1559                                            &error, 0);
1560         }
1561         if (parent) {
1562                 hammer2_chain_unlock(parent);
1563                 hammer2_chain_drop(parent);
1564         }
1565         hammer2_inode_unlock(spmp->iroot);
1566 }
1567
1568 static
1569 int
1570 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1571                 struct vnode *devvp, struct ucred *cred)
1572 {
1573         int error;
1574
1575         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1576                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1577                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1578                 vn_unlock(devvp);
1579                 error = hammer2_recovery(hmp);
1580                 if (error == 0)
1581                         error |= hammer2_fixup_pfses(hmp);
1582                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1583                 if (error == 0) {
1584                         VOP_CLOSE(devvp, FREAD, NULL);
1585                         hmp->ronly = 0;
1586                 } else {
1587                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1588                 }
1589                 vn_unlock(devvp);
1590         } else {
1591                 error = 0;
1592         }
1593         return error;
1594 }
1595
1596 static
1597 int
1598 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1599 {
1600         hammer2_pfs_t *pmp;
1601         int flags;
1602         int error = 0;
1603
1604         pmp = MPTOPMP(mp);
1605
1606         if (pmp == NULL)
1607                 return(0);
1608
1609         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1610
1611         /*
1612          * If mount initialization proceeded far enough we must flush
1613          * its vnodes and sync the underlying mount points.  Three syncs
1614          * are required to fully flush the filesystem (freemap updates lag
1615          * by one flush, and one extra for safety).
1616          */
1617         if (mntflags & MNT_FORCE)
1618                 flags = FORCECLOSE;
1619         else
1620                 flags = 0;
1621         if (pmp->iroot) {
1622                 error = vflush(mp, 0, flags);
1623                 if (error)
1624                         goto failed;
1625                 hammer2_vfs_sync(mp, MNT_WAIT);
1626                 hammer2_vfs_sync(mp, MNT_WAIT);
1627                 hammer2_vfs_sync(mp, MNT_WAIT);
1628         }
1629
1630         /*
1631          * Cleanup the frontend support XOPS threads
1632          */
1633         hammer2_xop_helper_cleanup(pmp);
1634
1635         if (pmp->mp)
1636                 hammer2_unmount_helper(mp, pmp, NULL);
1637
1638         error = 0;
1639 failed:
1640         lockmgr(&hammer2_mntlk, LK_RELEASE);
1641
1642         return (error);
1643 }
1644
1645 /*
1646  * Mount helper, hook the system mount into our PFS.
1647  * The mount lock is held.
1648  *
1649  * We must bump the mount_count on related devices for any
1650  * mounted PFSs.
1651  */
1652 static
1653 void
1654 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1655 {
1656         hammer2_cluster_t *cluster;
1657         hammer2_chain_t *rchain;
1658         int i;
1659
1660         mp->mnt_data = (qaddr_t)pmp;
1661         pmp->mp = mp;
1662
1663         /*
1664          * After pmp->mp is set we have to adjust hmp->mount_count.
1665          */
1666         cluster = &pmp->iroot->cluster;
1667         for (i = 0; i < cluster->nchains; ++i) {
1668                 rchain = cluster->array[i].chain;
1669                 if (rchain == NULL)
1670                         continue;
1671                 ++rchain->hmp->mount_count;
1672         }
1673
1674         /*
1675          * Create missing Xop threads
1676          */
1677         hammer2_xop_helper_create(pmp);
1678 }
1679
1680 /*
1681  * Mount helper, unhook the system mount from our PFS.
1682  * The mount lock is held.
1683  *
1684  * If hmp is supplied a mount responsible for being the first to open
1685  * the block device failed and the block device and all PFSs using the
1686  * block device must be cleaned up.
1687  *
1688  * If pmp is supplied multiple devices might be backing the PFS and each
1689  * must be disconnected.  This might not be the last PFS using some of the
1690  * underlying devices.  Also, we have to adjust our hmp->mount_count
1691  * accounting for the devices backing the pmp which is now undergoing an
1692  * unmount.
1693  */
1694 static
1695 void
1696 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1697 {
1698         hammer2_cluster_t *cluster;
1699         hammer2_chain_t *rchain;
1700         struct vnode *devvp;
1701         int dumpcnt;
1702         int ronly;
1703         int i;
1704
1705         /*
1706          * If no device supplied this is a high-level unmount and we have to
1707          * to disconnect the mount, adjust mount_count, and locate devices
1708          * that might now have no mounts.
1709          */
1710         if (pmp) {
1711                 KKASSERT(hmp == NULL);
1712                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1713                 pmp->mp = NULL;
1714                 mp->mnt_data = NULL;
1715
1716                 /*
1717                  * After pmp->mp is cleared we have to account for
1718                  * mount_count.
1719                  */
1720                 cluster = &pmp->iroot->cluster;
1721                 for (i = 0; i < cluster->nchains; ++i) {
1722                         rchain = cluster->array[i].chain;
1723                         if (rchain == NULL)
1724                                 continue;
1725                         --rchain->hmp->mount_count;
1726                         /* scrapping hmp now may invalidate the pmp */
1727                 }
1728 again:
1729                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1730                         if (hmp->mount_count == 0) {
1731                                 hammer2_unmount_helper(NULL, NULL, hmp);
1732                                 goto again;
1733                         }
1734                 }
1735                 return;
1736         }
1737
1738         /*
1739          * Try to terminate the block device.  We can't terminate it if
1740          * there are still PFSs referencing it.
1741          */
1742         if (hmp->mount_count)
1743                 return;
1744
1745         /*
1746          * Decomission the network before we start messing with the
1747          * device and PFS.
1748          */
1749         hammer2_iocom_uninit(hmp);
1750
1751         hammer2_bulkfree_uninit(hmp);
1752         hammer2_pfsfree_scan(hmp, 0);
1753 #if 0
1754         hammer2_dev_exlock(hmp);        /* XXX order */
1755 #endif
1756
1757         /*
1758          * Cycle the volume data lock as a safety (probably not needed any
1759          * more).  To ensure everything is out we need to flush at least
1760          * three times.  (1) The running of the sideq can dirty the
1761          * filesystem, (2) A normal flush can dirty the freemap, and
1762          * (3) ensure that the freemap is fully synchronized.
1763          *
1764          * The next mount's recovery scan can clean everything up but we want
1765          * to leave the filesystem in a 100% clean state on a normal unmount.
1766          */
1767 #if 0
1768         hammer2_voldata_lock(hmp);
1769         hammer2_voldata_unlock(hmp);
1770 #endif
1771
1772         /*
1773          * Flush whatever is left.  Unmounted but modified PFS's might still
1774          * have some dirty chains on them.
1775          */
1776         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1777         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1778
1779         if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1780                 hammer2_voldata_modify(hmp);
1781                 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1782                                             HAMMER2_FLUSH_ALL);
1783         }
1784         hammer2_chain_unlock(&hmp->fchain);
1785
1786         if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1787                 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1788                                             HAMMER2_FLUSH_ALL);
1789         }
1790         hammer2_chain_unlock(&hmp->vchain);
1791
1792         if ((hmp->vchain.flags | hmp->fchain.flags) &
1793             HAMMER2_CHAIN_FLUSH_MASK) {
1794                 kprintf("hammer2_unmount: chains left over "
1795                         "after final sync\n");
1796                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1797                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1798
1799                 if (hammer2_debug & 0x0010)
1800                         Debugger("entered debugger");
1801         }
1802
1803         hammer2_pfsfree_scan(hmp, 1);
1804
1805         KKASSERT(hmp->spmp == NULL);
1806
1807         /*
1808          * Finish up with the device vnode
1809          */
1810         if ((devvp = hmp->devvp) != NULL) {
1811                 ronly = hmp->ronly;
1812                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1813                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1814                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1815                         ronly);
1816                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1817                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1818                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1819                 hmp->devvp = NULL;
1820                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1821                 vn_unlock(devvp);
1822                 vrele(devvp);
1823                 devvp = NULL;
1824         }
1825
1826         /*
1827          * Clear vchain/fchain flags that might prevent final cleanup
1828          * of these chains.
1829          */
1830         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1831                 atomic_add_long(&hammer2_count_modified_chains, -1);
1832                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1833                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1834         }
1835         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1836                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1837         }
1838
1839         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1840                 atomic_add_long(&hammer2_count_modified_chains, -1);
1841                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1842                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1843         }
1844         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1845                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1846         }
1847
1848         /*
1849          * Final drop of embedded freemap root chain to
1850          * clean up fchain.core (fchain structure is not
1851          * flagged ALLOCATED so it is cleaned out and then
1852          * left to rot).
1853          */
1854         hammer2_chain_drop(&hmp->fchain);
1855
1856         /*
1857          * Final drop of embedded volume root chain to clean
1858          * up vchain.core (vchain structure is not flagged
1859          * ALLOCATED so it is cleaned out and then left to
1860          * rot).
1861          */
1862         dumpcnt = 50;
1863         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1864         dumpcnt = 50;
1865         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1866 #if 0
1867         hammer2_dev_unlock(hmp);
1868 #endif
1869         hammer2_chain_drop(&hmp->vchain);
1870
1871         hammer2_io_cleanup(hmp, &hmp->iotree);
1872         if (hmp->iofree_count) {
1873                 kprintf("io_cleanup: %d I/O's left hanging\n",
1874                         hmp->iofree_count);
1875         }
1876
1877         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1878         kmalloc_destroy(&hmp->mchain);
1879         kfree(hmp, M_HAMMER2);
1880 }
1881
1882 int
1883 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1884                  ino_t ino, struct vnode **vpp)
1885 {
1886         hammer2_xop_lookup_t *xop;
1887         hammer2_pfs_t *pmp;
1888         hammer2_inode_t *ip;
1889         hammer2_tid_t inum;
1890         int error;
1891
1892         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1893
1894         error = 0;
1895         pmp = MPTOPMP(mp);
1896
1897         /*
1898          * Easy if we already have it cached
1899          */
1900         ip = hammer2_inode_lookup(pmp, inum);
1901         if (ip) {
1902                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1903                 *vpp = hammer2_igetv(ip, &error);
1904                 hammer2_inode_unlock(ip);
1905                 hammer2_inode_drop(ip);         /* from lookup */
1906
1907                 return error;
1908         }
1909
1910         /*
1911          * Otherwise we have to find the inode
1912          */
1913         xop = hammer2_xop_alloc(pmp->iroot, 0);
1914         xop->lhc = inum;
1915         hammer2_xop_start(&xop->head, &hammer2_lookup_desc);
1916         error = hammer2_xop_collect(&xop->head, 0);
1917
1918         if (error == 0)
1919                 ip = hammer2_inode_get(pmp, &xop->head, -1, -1);
1920         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1921
1922         if (ip) {
1923                 *vpp = hammer2_igetv(ip, &error);
1924                 hammer2_inode_unlock(ip);
1925         } else {
1926                 *vpp = NULL;
1927                 error = ENOENT;
1928         }
1929         return (error);
1930 }
1931
1932 static
1933 int
1934 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1935 {
1936         hammer2_pfs_t *pmp;
1937         struct vnode *vp;
1938         int error;
1939
1940         pmp = MPTOPMP(mp);
1941         if (pmp->iroot == NULL) {
1942                 kprintf("hammer2 (%s): no root inode\n",
1943                         mp->mnt_stat.f_mntfromname);
1944                 *vpp = NULL;
1945                 return EINVAL;
1946         }
1947
1948         error = 0;
1949         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1950
1951         while (pmp->inode_tid == 0) {
1952                 hammer2_xop_ipcluster_t *xop;
1953                 const hammer2_inode_meta_t *meta;
1954
1955                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1956                 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc);
1957                 error = hammer2_xop_collect(&xop->head, 0);
1958
1959                 if (error == 0) {
1960                         meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1961                         pmp->iroot->meta = *meta;
1962                         pmp->inode_tid = meta->pfs_inum + 1;
1963                         hammer2_xop_pdata(&xop->head);
1964                         /* meta invalid */
1965
1966                         if (pmp->inode_tid < HAMMER2_INODE_START)
1967                                 pmp->inode_tid = HAMMER2_INODE_START;
1968                         pmp->modify_tid =
1969                                 xop->head.cluster.focus->bref.modify_tid + 1;
1970 #if 0
1971                         kprintf("PFS: Starting inode %jd\n",
1972                                 (intmax_t)pmp->inode_tid);
1973                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1974                                 pmp->inode_tid, pmp->modify_tid);
1975 #endif
1976                         wakeup(&pmp->iroot);
1977
1978                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1979
1980                         /*
1981                          * Prime the mount info.
1982                          */
1983                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1984                         break;
1985                 }
1986
1987                 /*
1988                  * Loop, try again
1989                  */
1990                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1991                 hammer2_inode_unlock(pmp->iroot);
1992                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1993                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1994                 if (error == EINTR)
1995                         break;
1996         }
1997
1998         if (error) {
1999                 hammer2_inode_unlock(pmp->iroot);
2000                 *vpp = NULL;
2001         } else {
2002                 vp = hammer2_igetv(pmp->iroot, &error);
2003                 hammer2_inode_unlock(pmp->iroot);
2004                 *vpp = vp;
2005         }
2006
2007         return (error);
2008 }
2009
2010 /*
2011  * Filesystem status
2012  *
2013  * XXX incorporate ipdata->meta.inode_quota and data_quota
2014  */
2015 static
2016 int
2017 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
2018 {
2019         hammer2_pfs_t *pmp;
2020         hammer2_dev_t *hmp;
2021         hammer2_blockref_t bref;
2022         struct statfs tmp;
2023         int i;
2024
2025         /*
2026          * NOTE: iroot might not have validated the cluster yet.
2027          */
2028         pmp = MPTOPMP(mp);
2029
2030         bzero(&tmp, sizeof(tmp));
2031
2032         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2033                 hmp = pmp->pfs_hmps[i];
2034                 if (hmp == NULL)
2035                         continue;
2036                 if (pmp->iroot->cluster.array[i].chain)
2037                         bref = pmp->iroot->cluster.array[i].chain->bref;
2038                 else
2039                         bzero(&bref, sizeof(bref));
2040
2041                 tmp.f_files = bref.embed.stats.inode_count;
2042                 tmp.f_ffree = 0;
2043                 tmp.f_blocks = hmp->voldata.allocator_size /
2044                                mp->mnt_vstat.f_bsize;
2045                 tmp.f_bfree = hmp->voldata.allocator_free /
2046                               mp->mnt_vstat.f_bsize;
2047                 tmp.f_bavail = tmp.f_bfree;
2048
2049                 if (cred && cred->cr_uid != 0) {
2050                         uint64_t adj;
2051
2052                         /* 5% */
2053                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2054                         tmp.f_blocks -= adj;
2055                         tmp.f_bfree -= adj;
2056                         tmp.f_bavail -= adj;
2057                 }
2058
2059                 mp->mnt_stat.f_blocks = tmp.f_blocks;
2060                 mp->mnt_stat.f_bfree = tmp.f_bfree;
2061                 mp->mnt_stat.f_bavail = tmp.f_bavail;
2062                 mp->mnt_stat.f_files = tmp.f_files;
2063                 mp->mnt_stat.f_ffree = tmp.f_ffree;
2064
2065                 *sbp = mp->mnt_stat;
2066         }
2067         return (0);
2068 }
2069
2070 static
2071 int
2072 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2073 {
2074         hammer2_pfs_t *pmp;
2075         hammer2_dev_t *hmp;
2076         hammer2_blockref_t bref;
2077         struct statvfs tmp;
2078         int i;
2079
2080         /*
2081          * NOTE: iroot might not have validated the cluster yet.
2082          */
2083         pmp = MPTOPMP(mp);
2084         bzero(&tmp, sizeof(tmp));
2085
2086         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2087                 hmp = pmp->pfs_hmps[i];
2088                 if (hmp == NULL)
2089                         continue;
2090                 if (pmp->iroot->cluster.array[i].chain)
2091                         bref = pmp->iroot->cluster.array[i].chain->bref;
2092                 else
2093                         bzero(&bref, sizeof(bref));
2094
2095                 tmp.f_files = bref.embed.stats.inode_count;
2096                 tmp.f_ffree = 0;
2097                 tmp.f_blocks = hmp->voldata.allocator_size /
2098                                mp->mnt_vstat.f_bsize;
2099                 tmp.f_bfree = hmp->voldata.allocator_free /
2100                               mp->mnt_vstat.f_bsize;
2101                 tmp.f_bavail = tmp.f_bfree;
2102
2103                 if (cred && cred->cr_uid != 0) {
2104                         uint64_t adj;
2105
2106                         /* 5% */
2107                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2108                         tmp.f_blocks -= adj;
2109                         tmp.f_bfree -= adj;
2110                         tmp.f_bavail -= adj;
2111                 }
2112
2113                 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2114                 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2115                 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2116                 mp->mnt_vstat.f_files = tmp.f_files;
2117                 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2118
2119                 *sbp = mp->mnt_vstat;
2120         }
2121         return (0);
2122 }
2123
2124 /*
2125  * Mount-time recovery (RW mounts)
2126  *
2127  * Updates to the free block table are allowed to lag flushes by one
2128  * transaction.  In case of a crash, then on a fresh mount we must do an
2129  * incremental scan of the last committed transaction id and make sure that
2130  * all related blocks have been marked allocated.
2131  *
2132  * The super-root topology and each PFS has its own transaction id domain,
2133  * so we must track PFS boundary transitions.
2134  */
2135 struct hammer2_recovery_elm {
2136         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2137         hammer2_chain_t *chain;
2138         hammer2_tid_t sync_tid;
2139 };
2140
2141 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2142
2143 struct hammer2_recovery_info {
2144         struct hammer2_recovery_list list;
2145         hammer2_tid_t   mtid;
2146         int     depth;
2147 };
2148
2149 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2150                         hammer2_chain_t *parent,
2151                         struct hammer2_recovery_info *info,
2152                         hammer2_tid_t sync_tid);
2153
2154 #define HAMMER2_RECOVERY_MAXDEPTH       10
2155
2156 static
2157 int
2158 hammer2_recovery(hammer2_dev_t *hmp)
2159 {
2160         struct hammer2_recovery_info info;
2161         struct hammer2_recovery_elm *elm;
2162         hammer2_chain_t *parent;
2163         hammer2_tid_t sync_tid;
2164         hammer2_tid_t mirror_tid;
2165         int error;
2166
2167         hammer2_trans_init(hmp->spmp, 0);
2168
2169         sync_tid = hmp->voldata.freemap_tid;
2170         mirror_tid = hmp->voldata.mirror_tid;
2171
2172         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2173         if (sync_tid >= mirror_tid) {
2174                 kprintf(" no recovery needed\n");
2175         } else {
2176                 kprintf(" freemap recovery %016jx-%016jx\n",
2177                         sync_tid + 1, mirror_tid);
2178         }
2179
2180         TAILQ_INIT(&info.list);
2181         info.depth = 0;
2182         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2183         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2184         hammer2_chain_lookup_done(parent);
2185
2186         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2187                 TAILQ_REMOVE(&info.list, elm, entry);
2188                 parent = elm->chain;
2189                 sync_tid = elm->sync_tid;
2190                 kfree(elm, M_HAMMER2);
2191
2192                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2193                 error |= hammer2_recovery_scan(hmp, parent, &info,
2194                                               hmp->voldata.freemap_tid);
2195                 hammer2_chain_unlock(parent);
2196                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2197         }
2198
2199         hammer2_trans_done(hmp->spmp, 0);
2200
2201         return error;
2202 }
2203
2204 static
2205 int
2206 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2207                       struct hammer2_recovery_info *info,
2208                       hammer2_tid_t sync_tid)
2209 {
2210         const hammer2_inode_data_t *ripdata;
2211         hammer2_chain_t *chain;
2212         hammer2_blockref_t bref;
2213         int tmp_error;
2214         int rup_error;
2215         int error;
2216         int first;
2217
2218         /*
2219          * Adjust freemap to ensure that the block(s) are marked allocated.
2220          */
2221         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2222                 hammer2_freemap_adjust(hmp, &parent->bref,
2223                                        HAMMER2_FREEMAP_DORECOVER);
2224         }
2225
2226         /*
2227          * Check type for recursive scan
2228          */
2229         switch(parent->bref.type) {
2230         case HAMMER2_BREF_TYPE_VOLUME:
2231                 /* data already instantiated */
2232                 break;
2233         case HAMMER2_BREF_TYPE_INODE:
2234                 /*
2235                  * Must instantiate data for DIRECTDATA test and also
2236                  * for recursion.
2237                  */
2238                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2239                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2240                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2241                         /* not applicable to recovery scan */
2242                         hammer2_chain_unlock(parent);
2243                         return 0;
2244                 }
2245                 hammer2_chain_unlock(parent);
2246                 break;
2247         case HAMMER2_BREF_TYPE_INDIRECT:
2248                 /*
2249                  * Must instantiate data for recursion
2250                  */
2251                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2252                 hammer2_chain_unlock(parent);
2253                 break;
2254         case HAMMER2_BREF_TYPE_DIRENT:
2255         case HAMMER2_BREF_TYPE_DATA:
2256         case HAMMER2_BREF_TYPE_FREEMAP:
2257         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2258         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2259                 /* not applicable to recovery scan */
2260                 return 0;
2261                 break;
2262         default:
2263                 return HAMMER2_ERROR_BADBREF;
2264         }
2265
2266         /*
2267          * Defer operation if depth limit reached or if we are crossing a
2268          * PFS boundary.
2269          */
2270         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2271                 struct hammer2_recovery_elm *elm;
2272
2273                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2274                 elm->chain = parent;
2275                 elm->sync_tid = sync_tid;
2276                 hammer2_chain_ref(parent);
2277                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2278                 /* unlocked by caller */
2279
2280                 return(0);
2281         }
2282
2283
2284         /*
2285          * Recursive scan of the last flushed transaction only.  We are
2286          * doing this without pmp assignments so don't leave the chains
2287          * hanging around after we are done with them.
2288          *
2289          * error        Cumulative error this level only
2290          * rup_error    Cumulative error for recursion
2291          * tmp_error    Specific non-cumulative recursion error
2292          */
2293         chain = NULL;
2294         first = 1;
2295         rup_error = 0;
2296         error = 0;
2297
2298         for (;;) {
2299                 error |= hammer2_chain_scan(parent, &chain, &bref,
2300                                             &first,
2301                                             HAMMER2_LOOKUP_NODATA);
2302
2303                 /*
2304                  * Problem during scan or EOF
2305                  */
2306                 if (error)
2307                         break;
2308
2309                 /*
2310                  * If this is a leaf
2311                  */
2312                 if (chain == NULL) {
2313                         if (bref.mirror_tid > sync_tid) {
2314                                 hammer2_freemap_adjust(hmp, &bref,
2315                                                      HAMMER2_FREEMAP_DORECOVER);
2316                         }
2317                         continue;
2318                 }
2319
2320                 /*
2321                  * This may or may not be a recursive node.
2322                  */
2323                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2324                 if (bref.mirror_tid > sync_tid) {
2325                         ++info->depth;
2326                         tmp_error = hammer2_recovery_scan(hmp, chain,
2327                                                            info, sync_tid);
2328                         --info->depth;
2329                 } else {
2330                         tmp_error = 0;
2331                 }
2332
2333                 /*
2334                  * Flush the recovery at the PFS boundary to stage it for
2335                  * the final flush of the super-root topology.
2336                  */
2337                 if (tmp_error == 0 &&
2338                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2339                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2340                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2341                                              HAMMER2_FLUSH_ALL);
2342                 }
2343                 rup_error |= tmp_error;
2344         }
2345         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2346 }
2347
2348 /*
2349  * This fixes up an error introduced in earlier H2 implementations where
2350  * moving a PFS inode into an indirect block wound up causing the
2351  * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2352  */
2353 static
2354 int
2355 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2356 {
2357         const hammer2_inode_data_t *ripdata;
2358         hammer2_chain_t *parent;
2359         hammer2_chain_t *chain;
2360         hammer2_key_t key_next;
2361         hammer2_pfs_t *spmp;
2362         int error;
2363
2364         error = 0;
2365
2366         /*
2367          * Lookup mount point under the media-localized super-root.
2368          *
2369          * cluster->pmp will incorrectly point to spmp and must be fixed
2370          * up later on.
2371          */
2372         spmp = hmp->spmp;
2373         hammer2_inode_lock(spmp->iroot, 0);
2374         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2375         chain = hammer2_chain_lookup(&parent, &key_next,
2376                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2377                                          &error, 0);
2378         while (chain) {
2379                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2380                         continue;
2381                 if (chain->error) {
2382                         kprintf("I/O error scanning PFS labels\n");
2383                         error |= chain->error;
2384                 } else if ((chain->bref.flags &
2385                             HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2386                         int error2;
2387
2388                         ripdata = &chain->data->ipdata;
2389                         hammer2_trans_init(hmp->spmp, 0);
2390                         error2 = hammer2_chain_modify(chain,
2391                                                       chain->bref.modify_tid,
2392                                                       0, 0);
2393                         if (error2 == 0) {
2394                                 kprintf("hammer2: Correct mis-flagged PFS %s\n",
2395                                         ripdata->filename);
2396                                 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2397                         } else {
2398                                 error |= error2;
2399                         }
2400                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2401                                              HAMMER2_FLUSH_ALL);
2402                         hammer2_trans_done(hmp->spmp, 0);
2403                 }
2404                 chain = hammer2_chain_next(&parent, chain, &key_next,
2405                                            key_next, HAMMER2_KEY_MAX,
2406                                            &error, 0);
2407         }
2408         if (parent) {
2409                 hammer2_chain_unlock(parent);
2410                 hammer2_chain_drop(parent);
2411         }
2412         hammer2_inode_unlock(spmp->iroot);
2413
2414         return error;
2415 }
2416
2417 /*
2418  * Sync a mount point; this is called periodically on a per-mount basis from
2419  * the filesystem syncer, and whenever a user issues a sync.
2420  */
2421 int
2422 hammer2_vfs_sync(struct mount *mp, int waitfor)
2423 {
2424         int error;
2425
2426         error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor);
2427
2428         return error;
2429 }
2430
2431 /*
2432  * Because frontend operations lock vnodes before we get a chance to
2433  * lock the related inode, we can't just acquire a vnode lock without
2434  * risking a deadlock.  The frontend may be holding a vnode lock while
2435  * also blocked on our SYNCQ flag while trying to get the inode lock.
2436  *
2437  * To deal with this situation we can check the vnode lock situation
2438  * after locking the inode and perform a work-around.
2439  */
2440 int
2441 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor)
2442 {
2443         struct mount *mp;
2444         /*hammer2_xop_flush_t *xop;*/
2445         /*struct hammer2_sync_info info;*/
2446         hammer2_inode_t *ip;
2447         hammer2_depend_t *depend;
2448         hammer2_depend_t *depend_next;
2449         struct vnode *vp;
2450         uint32_t pass2;
2451         int error;
2452         int dorestart;
2453
2454         mp = pmp->mp;
2455
2456         /*
2457          * Move all inodes on sideq to syncq.  This will clear sideq.
2458          * This should represent all flushable inodes.  These inodes
2459          * will already have refs due to being on syncq or sideq.  We
2460          * must do this all at once with the spinlock held to ensure that
2461          * all inode dependencies are part of the same flush.
2462          *
2463          * We should be able to do this asynchronously from frontend
2464          * operations because we will be locking the inodes later on
2465          * to actually flush them, and that will partition any frontend
2466          * op using the same inode.  Either it has already locked the
2467          * inode and we will block, or it has not yet locked the inode
2468          * and it will block until we are finished flushing that inode.
2469          *
2470          * When restarting, only move the inodes flagged as PASS2 from
2471          * SIDEQ to SYNCQ.  PASS2 propagation by inode_lock4() and
2472          * inode_depend() are atomic with the spin-lock.
2473          */
2474         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2475 #ifdef HAMMER2_DEBUG_SYNC
2476         kprintf("FILESYSTEM SYNC BOUNDARY\n");
2477 #endif
2478         dorestart = 0;
2479
2480         /*
2481          * Move inodes from depq to syncq, releasing the related
2482          * depend structures.
2483          */
2484 restart:
2485 #ifdef HAMMER2_DEBUG_SYNC
2486         kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart);
2487 #endif
2488         hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/);
2489         hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN);
2490
2491         /*
2492          * Move inodes from depq to syncq.  When restarting, only depq's
2493          * marked pass2 are moved.
2494          */
2495         hammer2_spin_ex(&pmp->list_spin);
2496         depend_next = TAILQ_FIRST(&pmp->depq);
2497
2498         while ((depend = depend_next) != NULL) {
2499                 depend_next = TAILQ_NEXT(depend, entry);
2500                 if (dorestart && depend->pass2 == 0)
2501                         continue;
2502                 TAILQ_FOREACH(ip, &depend->sideq, entry) {
2503                         KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ);
2504                         atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ);
2505                         atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ);
2506                         ip->depend = NULL;
2507                 }
2508                 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry);
2509                 pmp->sideq_count -= depend->count;
2510                 depend->count = 0;
2511                 depend->pass2 = 0;
2512                 TAILQ_REMOVE(&pmp->depq, depend, entry);
2513         }
2514
2515         hammer2_spin_unex(&pmp->list_spin);
2516         hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/
2517                                       HAMMER2_TRANS_WAITING);
2518         dorestart = 0;
2519
2520         /*
2521          * sideq_count may have dropped enough to allow us to unstall
2522          * the frontend.
2523          */
2524         hammer2_pfs_memory_inc(pmp);
2525         hammer2_pfs_memory_wakeup(pmp);
2526
2527         /*
2528          * Now run through all inodes on syncq.
2529          *
2530          * Flush transactions only interlock with other flush transactions.
2531          * Any conflicting frontend operations will block on the inode, but
2532          * may hold a vnode lock while doing so.
2533          */
2534         hammer2_spin_ex(&pmp->list_spin);
2535         while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) {
2536                 /*
2537                  * Remove the inode from the SYNCQ, transfer the syncq ref
2538                  * to us.  We must clear SYNCQ to allow any potential
2539                  * front-end deadlock to proceed.  We must set PASS2 so
2540                  * the dependency code knows what to do.
2541                  */
2542                 pass2 = ip->flags;
2543                 cpu_ccfence();
2544                 if (atomic_cmpset_int(&ip->flags,
2545                               pass2,
2546                               (pass2 & ~(HAMMER2_INODE_SYNCQ |
2547                                          HAMMER2_INODE_SYNCQ_WAKEUP)) |
2548                               HAMMER2_INODE_SYNCQ_PASS2) == 0) {
2549                         continue;
2550                 }
2551                 TAILQ_REMOVE(&pmp->syncq, ip, entry);
2552                 hammer2_spin_unex(&pmp->list_spin);
2553                 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP)
2554                         wakeup(&ip->flags);
2555
2556                 /*
2557                  * Relock the inode, and we inherit a ref from the above.
2558                  * We will check for a race after we acquire the vnode.
2559                  */
2560                 hammer2_mtx_ex(&ip->lock);
2561
2562                 /*
2563                  * We need the vp in order to vfsync() dirty buffers, so if
2564                  * one isn't attached we can skip it.
2565                  *
2566                  * Ordering the inode lock and then the vnode lock has the
2567                  * potential to deadlock.  If we had left SYNCQ set that could
2568                  * also deadlock us against the frontend even if we don't hold
2569                  * any locks, but the latter is not a problem now since we
2570                  * cleared it.  igetv will temporarily release the inode lock
2571                  * in a safe manner to work-around the deadlock.
2572                  *
2573                  * Unfortunately it is still possible to deadlock when the
2574                  * frontend obtains multiple inode locks, because all the
2575                  * related vnodes are already locked (nor can the vnode locks
2576                  * be released and reacquired without messing up RECLAIM and
2577                  * INACTIVE sequencing).
2578                  *
2579                  * The solution for now is to move the vp back onto SIDEQ
2580                  * and set dorestart, which will restart the flush after we
2581                  * exhaust the current SYNCQ.  Note that additional
2582                  * dependencies may build up, so we definitely need to move
2583                  * the whole SIDEQ back to SYNCQ when we restart.
2584                  */
2585                 vp = ip->vp;
2586                 if (vp) {
2587                         if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) {
2588                                 /*
2589                                  * Failed to get the vnode, requeue the inode
2590                                  * (PASS2 is already set so it will be found
2591                                  * again on the restart).
2592                                  *
2593                                  * Then unlock, possibly sleep, and retry
2594                                  * later.  We sleep if PASS2 was *previously*
2595                                  * set, before we set it again above.
2596                                  */
2597                                 vp = NULL;
2598                                 dorestart = 1;
2599 #ifdef HAMMER2_DEBUG_SYNC
2600                                 kprintf("inum %ld (sync delayed by vnode)\n",
2601                                         (long)ip->meta.inum);
2602 #endif
2603                                 hammer2_inode_delayed_sideq(ip);
2604
2605                                 hammer2_mtx_unlock(&ip->lock);
2606                                 hammer2_inode_drop(ip);
2607
2608                                 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) {
2609                                         tsleep(&dorestart, 0, "h2syndel", 2);
2610                                 }
2611                                 hammer2_spin_ex(&pmp->list_spin);
2612                                 continue;
2613                         }
2614                 } else {
2615                         vp = NULL;
2616                 }
2617
2618                 /*
2619                  * If the inode wound up on a SIDEQ again it will already be
2620                  * prepped for another PASS2.  In this situation if we flush
2621                  * it now we will just wind up flushing it again in the same
2622                  * syncer run, so we might as well not flush it now.
2623                  */
2624                 if (ip->flags & HAMMER2_INODE_SIDEQ) {
2625                         hammer2_mtx_unlock(&ip->lock);
2626                         hammer2_inode_drop(ip);
2627                         if (vp)
2628                                 vput(vp);
2629                         dorestart = 1;
2630                         hammer2_spin_ex(&pmp->list_spin);
2631                         continue;
2632                 }
2633
2634                 /*
2635                  * Ok we have the inode exclusively locked and if vp is
2636                  * not NULL that will also be exclusively locked.  Do the
2637                  * meat of the flush.
2638                  *
2639                  * vp token needed for v_rbdirty_tree check / vclrisdirty
2640                  * sequencing.  Though we hold the vnode exclusively so
2641                  * we shouldn't need to hold the token also in this case.
2642                  */
2643                 if (vp) {
2644                         vfsync(vp, MNT_WAIT, 1, NULL, NULL);
2645                         bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */
2646                 }
2647
2648                 /*
2649                  * If the inode has not yet been inserted into the tree
2650                  * we must do so.  Then sync and flush it.  The flush should
2651                  * update the parent.
2652                  */
2653                 if (ip->flags & HAMMER2_INODE_DELETING) {
2654 #ifdef HAMMER2_DEBUG_SYNC
2655                         kprintf("inum %ld destroy\n", (long)ip->meta.inum);
2656 #endif
2657                         hammer2_inode_chain_des(ip);
2658                         atomic_add_long(&hammer2_iod_inode_deletes, 1);
2659                 } else if (ip->flags & HAMMER2_INODE_CREATING) {
2660 #ifdef HAMMER2_DEBUG_SYNC
2661                         kprintf("inum %ld insert\n", (long)ip->meta.inum);
2662 #endif
2663                         hammer2_inode_chain_ins(ip);
2664                         atomic_add_long(&hammer2_iod_inode_creates, 1);
2665                 }
2666 #ifdef HAMMER2_DEBUG_SYNC
2667                 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum);
2668 #endif
2669
2670                 /*
2671                  * Because I kinda messed up the design and index the inodes
2672                  * under the root inode, along side the directory entries,
2673                  * we can't flush the inode index under the iroot until the
2674                  * end.  If we do it now we might miss effects created by
2675                  * other inodes on the SYNCQ.
2676                  *
2677                  * Do a normal (non-FSSYNC) flush instead, which allows the
2678                  * vnode code to work the same.  We don't want to force iroot
2679                  * back onto the SIDEQ, and we also don't want the flush code
2680                  * to update pfs_iroot_blocksets until the final flush later.
2681                  *
2682                  * XXX at the moment this will likely result in a double-flush
2683                  * of the iroot chain.
2684                  */
2685                 hammer2_inode_chain_sync(ip);
2686                 if (ip == pmp->iroot) {
2687                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP);
2688                 } else {
2689                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2690                                                       HAMMER2_XOP_FSSYNC);
2691                 }
2692                 if (vp) {
2693                         lwkt_gettoken(&vp->v_token);
2694                         if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2695                                           HAMMER2_INODE_RESIZED |
2696                                           HAMMER2_INODE_DIRTYDATA)) == 0 &&
2697                             RB_EMPTY(&vp->v_rbdirty_tree) &&
2698                             !bio_track_active(&vp->v_track_write)) {
2699                                 vclrisdirty(vp);
2700                         } else {
2701                                 hammer2_inode_delayed_sideq(ip);
2702                         }
2703                         lwkt_reltoken(&vp->v_token);
2704                         vput(vp);
2705                         vp = NULL;      /* safety */
2706                 }
2707                 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2);
2708                 hammer2_inode_unlock(ip);       /* unlock+drop */
2709                 /* ip pointer invalid */
2710
2711                 /*
2712                  * If the inode got dirted after we dropped our locks,
2713                  * it will have already been moved back to the SIDEQ.
2714                  */
2715                 hammer2_spin_ex(&pmp->list_spin);
2716         }
2717         hammer2_spin_unex(&pmp->list_spin);
2718         if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) {
2719 #ifdef HAMMER2_DEBUG_SYNC
2720                 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n");
2721                 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/
2722 #endif
2723                 dorestart = 1;
2724                 goto restart;
2725         }
2726 #ifdef HAMMER2_DEBUG_SYNC
2727         kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n");
2728         /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/
2729 #endif
2730
2731         /*
2732          * We have to flush the PFS root last, even if it does not appear to
2733          * be dirty, because all the inodes in the PFS are indexed under it.
2734          * The normal flushing of iroot above would only occur if directory
2735          * entries under the root were changed.
2736          *
2737          * Specifying VOLHDR will cause an additionl flush of hmp->spmp
2738          * for the media making up the cluster.
2739          */
2740         if ((ip = pmp->iroot) != NULL) {
2741                 hammer2_inode_ref(ip);
2742                 hammer2_mtx_ex(&ip->lock);
2743                 hammer2_inode_chain_sync(ip);
2744                 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2745                                               HAMMER2_XOP_FSSYNC |
2746                                               HAMMER2_XOP_VOLHDR);
2747                 hammer2_inode_unlock(ip);       /* unlock+drop */
2748         }
2749 #ifdef HAMMER2_DEBUG_SYNC
2750         kprintf("FILESYSTEM SYNC STAGE 2 DONE\n");
2751 #endif
2752
2753         /*
2754          * device bioq sync
2755          */
2756         hammer2_bioq_sync(pmp);
2757
2758 #if 0
2759         info.pass = 1;
2760         info.waitfor = MNT_WAIT;
2761         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2762
2763         info.pass = 2;
2764         info.waitfor = MNT_WAIT;
2765         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2766 #endif
2767 #if 0
2768         /*
2769          * Generally speaking we now want to flush the media topology from
2770          * the iroot through to the inodes.  The flush stops at any inode
2771          * boundary, which allows the frontend to continue running concurrent
2772          * modifying operations on inodes (including kernel flushes of
2773          * buffers) without interfering with the main sync.
2774          *
2775          * Use the XOP interface to concurrently flush all nodes to
2776          * synchronize the PFSROOT subtopology to the media.  A standard
2777          * end-of-scan ENOENT error indicates cluster sufficiency.
2778          *
2779          * Note that this flush will not be visible on crash recovery until
2780          * we flush the super-root topology in the next loop.
2781          *
2782          * XXX For now wait for all flushes to complete.
2783          */
2784         if (mp && (ip = pmp->iroot) != NULL) {
2785                 /*
2786                  * If unmounting try to flush everything including any
2787                  * sub-trees under inodes, just in case there is dangling
2788                  * modified data, as a safety.  Otherwise just flush up to
2789                  * the inodes in this stage.
2790                  */
2791                 kprintf("MP & IROOT\n");
2792 #ifdef HAMMER2_DEBUG_SYNC
2793                 kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n");
2794 #endif
2795                 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2796                         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2797                                                     HAMMER2_XOP_VOLHDR |
2798                                                     HAMMER2_XOP_FSSYNC |
2799                                                     HAMMER2_XOP_INODE_STOP);
2800                 } else {
2801                         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2802                                                     HAMMER2_XOP_INODE_STOP |
2803                                                     HAMMER2_XOP_VOLHDR |
2804                                                     HAMMER2_XOP_FSSYNC |
2805                                                     HAMMER2_XOP_INODE_STOP);
2806                 }
2807                 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc);
2808                 error = hammer2_xop_collect(&xop->head,
2809                                             HAMMER2_XOP_COLLECT_WAITALL);
2810                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2811 #ifdef HAMMER2_DEBUG_SYNC
2812                 kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n");
2813 #endif
2814                 if (error == HAMMER2_ERROR_ENOENT)
2815                         error = 0;
2816                 else
2817                         error = hammer2_error_to_errno(error);
2818         } else {
2819                 error = 0;
2820         }
2821 #endif
2822         error = 0;      /* XXX */
2823         hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH);
2824
2825         return (error);
2826 }
2827
2828 static
2829 int
2830 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2831 {
2832         hammer2_inode_t *ip;
2833
2834         KKASSERT(MAXFIDSZ >= 16);
2835         ip = VTOI(vp);
2836         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2837         fhp->fid_ext = 0;
2838         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2839         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2840
2841         return 0;
2842 }
2843
2844 static
2845 int
2846 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2847                struct fid *fhp, struct vnode **vpp)
2848 {
2849         hammer2_pfs_t *pmp;
2850         hammer2_tid_t inum;
2851         int error;
2852
2853         pmp = MPTOPMP(mp);
2854         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2855         if (vpp) {
2856                 if (inum == 1)
2857                         error = hammer2_vfs_root(mp, vpp);
2858                 else
2859                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2860         } else {
2861                 error = 0;
2862         }
2863         if (error)
2864                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2865         return error;
2866 }
2867
2868 static
2869 int
2870 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2871                  int *exflagsp, struct ucred **credanonp)
2872 {
2873         hammer2_pfs_t *pmp;
2874         struct netcred *np;
2875         int error;
2876
2877         pmp = MPTOPMP(mp);
2878         np = vfs_export_lookup(mp, &pmp->export, nam);
2879         if (np) {
2880                 *exflagsp = np->netc_exflags;
2881                 *credanonp = &np->netc_anon;
2882                 error = 0;
2883         } else {
2884                 error = EACCES;
2885         }
2886         return error;
2887 }
2888
2889 /*
2890  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2891  * header into the HMP
2892  *
2893  * XXX read four volhdrs and use the one with the highest TID whos CRC
2894  *     matches.
2895  *
2896  * XXX check iCRCs.
2897  *
2898  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2899  *     nonexistant locations.
2900  *
2901  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2902  */
2903 static
2904 int
2905 hammer2_install_volume_header(hammer2_dev_t *hmp)
2906 {
2907         hammer2_volume_data_t *vd;
2908         struct buf *bp;
2909         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2910         int error_reported;
2911         int error;
2912         int valid;
2913         int i;
2914
2915         error_reported = 0;
2916         error = 0;
2917         valid = 0;
2918         bp = NULL;
2919
2920         /*
2921          * There are up to 4 copies of the volume header (syncs iterate
2922          * between them so there is no single master).  We don't trust the
2923          * volu_size field so we don't know precisely how large the filesystem
2924          * is, so depend on the OS to return an error if we go beyond the
2925          * block device's EOF.
2926          */
2927         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2928                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2929                               HAMMER2_VOLUME_BYTES, &bp);
2930                 if (error) {
2931                         brelse(bp);
2932                         bp = NULL;
2933                         continue;
2934                 }
2935
2936                 vd = (struct hammer2_volume_data *) bp->b_data;
2937                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2938                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2939                         brelse(bp);
2940                         bp = NULL;
2941                         continue;
2942                 }
2943
2944                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2945                         /* XXX: Reversed-endianness filesystem */
2946                         kprintf("hammer2: reverse-endian filesystem detected");
2947                         brelse(bp);
2948                         bp = NULL;
2949                         continue;
2950                 }
2951
2952                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2953                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2954                                       HAMMER2_VOLUME_ICRC0_SIZE);
2955                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2956                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2957                                        HAMMER2_VOLUME_ICRC1_SIZE);
2958                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2959                         kprintf("hammer2 volume header crc "
2960                                 "mismatch copy #%d %08x/%08x\n",
2961                                 i, crc0, crc);
2962                         error_reported = 1;
2963                         brelse(bp);
2964                         bp = NULL;
2965                         continue;
2966                 }
2967                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2968                         valid = 1;
2969                         hmp->voldata = *vd;
2970                         hmp->volhdrno = i;
2971                 }
2972                 brelse(bp);
2973                 bp = NULL;
2974         }
2975         if (valid) {
2976                 hmp->volsync = hmp->voldata;
2977                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2978                 error = 0;
2979                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2980                         kprintf("hammer2: using volume header #%d\n",
2981                                 hmp->volhdrno);
2982                 }
2983         } else {
2984                 error = EINVAL;
2985                 kprintf("hammer2: no valid volume headers found!\n");
2986         }
2987         return (error);
2988 }
2989
2990 /*
2991  * This handles hysteresis on regular file flushes.  Because the BIOs are
2992  * routed to a thread it is possible for an excessive number to build up
2993  * and cause long front-end stalls long before the runningbuffspace limit
2994  * is hit, so we implement hammer2_flush_pipe to control the
2995  * hysteresis.
2996  *
2997  * This is a particular problem when compression is used.
2998  */
2999 void
3000 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
3001 {
3002         atomic_add_int(&pmp->count_lwinprog, 1);
3003 }
3004
3005 void
3006 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
3007 {
3008         int lwinprog;
3009
3010         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
3011         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
3012             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
3013                 atomic_clear_int(&pmp->count_lwinprog,
3014                                  HAMMER2_LWINPROG_WAITING);
3015                 wakeup(&pmp->count_lwinprog);
3016         }
3017         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
3018             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
3019                 atomic_clear_int(&pmp->count_lwinprog,
3020                                  HAMMER2_LWINPROG_WAITING0);
3021                 wakeup(&pmp->count_lwinprog);
3022         }
3023 }
3024
3025 void
3026 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
3027 {
3028         int lwinprog;
3029         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
3030                                     HAMMER2_LWINPROG_WAITING0;
3031
3032         for (;;) {
3033                 lwinprog = pmp->count_lwinprog;
3034                 cpu_ccfence();
3035                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3036                         break;
3037                 tsleep_interlock(&pmp->count_lwinprog, 0);
3038                 atomic_set_int(&pmp->count_lwinprog, lwflag);
3039                 lwinprog = pmp->count_lwinprog;
3040                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3041                         break;
3042                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
3043         }
3044 }
3045
3046 /*
3047  * It is possible for an excessive number of dirty chains or dirty inodes
3048  * to build up.  When this occurs we start an asynchronous filesystem sync.
3049  * If the level continues to build up, we stall, waiting for it to drop,
3050  * with some hysteresis.
3051  *
3052  * We limit the stall to two seconds per call.
3053  *
3054  * This relies on the kernel calling hammer2_vfs_modifying() prior to
3055  * obtaining any vnode locks before making a modifying VOP call.
3056  */
3057 static void
3058 hammer2_vfs_modifying(struct mount *mp)
3059 {
3060         hammer2_pfs_memory_wait(MPTOPMP(mp));
3061 }
3062
3063 /*
3064  * Initiate an asynchronous filesystem sync and, with hysteresis,
3065  * stall if the internal data structure count becomes too bloated.
3066  */
3067 void
3068 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
3069 {
3070         uint32_t waiting;
3071         int loops;
3072
3073         if (pmp == NULL || pmp->mp == NULL)
3074                 return;
3075
3076         for (loops = 0; loops < 2; ++loops) {
3077                 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK;
3078                 cpu_ccfence();
3079
3080                 /*
3081                  * Start the syncer running at 1/2 the limit
3082                  */
3083                 if (waiting > hammer2_limit_dirty_chains / 2 ||
3084                     pmp->sideq_count > hammer2_limit_dirty_inodes / 2) {
3085                         trigger_syncer(pmp->mp);
3086                 }
3087
3088                 /*
3089                  * Stall at the limit waiting for the counts to drop.
3090                  * This code will typically be woken up once the count
3091                  * drops below 3/4 the limit, or in one second.
3092                  */
3093                 if (waiting < hammer2_limit_dirty_chains &&
3094                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
3095                         break;
3096                 }
3097                 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
3098                 atomic_set_int(&pmp->inmem_dirty_chains,
3099                                HAMMER2_DIRTYCHAIN_WAITING);
3100                 if (waiting < hammer2_limit_dirty_chains &&
3101                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
3102                         break;
3103                 }
3104                 trigger_syncer(pmp->mp);
3105                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED, "h2memw", hz);
3106 #if 0
3107                 limit = pmp->mp->mnt_nvnodelistsize / 10;
3108                 if (limit < hammer2_limit_dirty_chains)
3109                         limit = hammer2_limit_dirty_chains;
3110                 if (limit < 1000)
3111                         limit = 1000;
3112 #endif
3113         }
3114 }
3115
3116 void
3117 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
3118 {
3119         if (pmp) {
3120                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
3121         }
3122 }
3123
3124 void
3125 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
3126 {
3127         uint32_t waiting;
3128
3129         if (pmp) {
3130                 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
3131                 /* don't need --waiting to test flag */
3132
3133                 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) &&
3134                     (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <=
3135                     hammer2_limit_dirty_chains * 2 / 3 &&
3136                     pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) {
3137                         atomic_clear_int(&pmp->inmem_dirty_chains,
3138                                          HAMMER2_DIRTYCHAIN_WAITING);
3139                         wakeup(&pmp->inmem_dirty_chains);
3140                 }
3141         }
3142 }
3143
3144 /*
3145  * Returns 0 if the filesystem has tons of free space
3146  * Returns 1 if the filesystem has less than 10% remaining
3147  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
3148  */
3149 int
3150 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
3151 {
3152         hammer2_pfs_t *pmp;
3153         hammer2_dev_t *hmp;
3154         hammer2_off_t free_reserved;
3155         hammer2_off_t free_nominal;
3156         int i;
3157
3158         pmp = ip->pmp;
3159
3160         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
3161                 free_reserved = HAMMER2_SEGSIZE;
3162                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
3163                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
3164                         hmp = pmp->pfs_hmps[i];
3165                         if (hmp == NULL)
3166                                 continue;
3167                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
3168                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
3169                                 continue;
3170
3171                         if (free_nominal > hmp->voldata.allocator_free)
3172                                 free_nominal = hmp->voldata.allocator_free;
3173                         if (free_reserved < hmp->free_reserved)
3174                                 free_reserved = hmp->free_reserved;
3175                 }
3176
3177                 /*
3178                  * SMP races ok
3179                  */
3180                 pmp->free_reserved = free_reserved;
3181                 pmp->free_nominal = free_nominal;
3182                 pmp->free_ticks = ticks;
3183         } else {
3184                 free_reserved = pmp->free_reserved;
3185                 free_nominal = pmp->free_nominal;
3186         }
3187         if (cred && cred->cr_uid != 0) {
3188                 if ((int64_t)(free_nominal - bytes) <
3189                     (int64_t)free_reserved) {
3190                         return 2;
3191                 }
3192         } else {
3193                 if ((int64_t)(free_nominal - bytes) <
3194                     (int64_t)free_reserved / 2) {
3195                         return 2;
3196                 }
3197         }
3198         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
3199                 return 1;
3200         return 0;
3201 }
3202
3203 /*
3204  * Debugging
3205  */
3206 void
3207 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
3208                    u_int flags)
3209 {
3210         hammer2_chain_t *scan;
3211         hammer2_chain_t *parent;
3212
3213         --*countp;
3214         if (*countp == 0) {
3215                 kprintf("%*.*s...\n", tab, tab, "");
3216                 return;
3217         }
3218         if (*countp < 0)
3219                 return;
3220         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
3221                 tab, tab, "", pfx,
3222                 chain, chain->bref.type,
3223                 chain->bref.key, chain->bref.keybits,
3224                 chain->bref.mirror_tid);
3225
3226         kprintf("%*.*s      [%08x] (%s) refs=%d",
3227                 tab, tab, "",
3228                 chain->flags,
3229                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3230                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
3231                 chain->refs);
3232
3233         parent = chain->parent;
3234         if (parent)
3235                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
3236                         tab, tab, "",
3237                         parent, parent->flags, parent->refs);
3238         if (RB_EMPTY(&chain->core.rbtree)) {
3239                 kprintf("\n");
3240         } else {
3241                 kprintf(" {\n");
3242                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3243                         if ((scan->flags & flags) || flags == (u_int)-1) {
3244                                 hammer2_dump_chain(scan, tab + 4, countp, 'a',
3245                                                    flags);
3246                         }
3247                 }
3248                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3249                         kprintf("%*.*s}(%s)\n", tab, tab, "",
3250                                 chain->data->ipdata.filename);
3251                 else
3252                         kprintf("%*.*s}\n", tab, tab, "");
3253         }
3254 }