sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2018 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72         int pass;
  73 };
  74
  75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  76 static struct hammer2_mntlist hammer2_mntlist;
  77
  78 struct hammer2_pfslist hammer2_pfslist;
  79 struct hammer2_pfslist hammer2_spmplist;
  80 struct lock hammer2_mntlk;
  81
  82 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
  83 int hammer2_debug;
  84 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
  85 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
  86 int hammer2_cluster_write = 0;          /* physical write clustering */
  87 int hammer2_dedup_enable = 1;
  88 int hammer2_always_compress = 0;        /* always try to compress */
  89 int hammer2_inval_enable = 0;
  90 int hammer2_flush_pipe = 100;
  91 int hammer2_dio_count;
  92 int hammer2_dio_limit = 256;
  93 int hammer2_bulkfree_tps = 5000;
  94 long hammer2_chain_allocs;
  95 long hammer2_chain_frees;
  96 long hammer2_limit_dirty_chains;
  97 long hammer2_limit_dirty_inodes;
  98 long hammer2_count_modified_chains;
  99 long hammer2_iod_invals;
 100 long hammer2_iod_file_read;
 101 long hammer2_iod_meta_read;
 102 long hammer2_iod_indr_read;
 103 long hammer2_iod_fmap_read;
 104 long hammer2_iod_volu_read;
 105 long hammer2_iod_file_write;
 106 long hammer2_iod_file_wembed;
 107 long hammer2_iod_file_wzero;
 108 long hammer2_iod_file_wdedup;
 109 long hammer2_iod_meta_write;
 110 long hammer2_iod_indr_write;
 111 long hammer2_iod_fmap_write;
 112 long hammer2_iod_volu_write;
 113 long hammer2_iod_inode_creates;
 114 long hammer2_iod_inode_deletes;
 115
 116 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 117 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 118                 "Buffer used for compression.");
 119
 120 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 121 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 122                 "Buffer used for decompression.");
 123
 124 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 125
 126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
 127            &hammer2_supported_version, 0, "");
 128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 129            &hammer2_debug, 0, "");
 130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 131            &hammer2_cluster_meta_read, 0, "");
 132 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 133            &hammer2_cluster_data_read, 0, "");
 134 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 135            &hammer2_cluster_write, 0, "");
 136 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 137            &hammer2_dedup_enable, 0, "");
 138 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 139            &hammer2_always_compress, 0, "");
 140 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 141            &hammer2_inval_enable, 0, "");
 142 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 143            &hammer2_flush_pipe, 0, "");
 144 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
 145            &hammer2_bulkfree_tps, 0, "");
 146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 147            &hammer2_chain_allocs, 0, "");
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 149            &hammer2_chain_frees, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 151            &hammer2_limit_dirty_chains, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
 153            &hammer2_limit_dirty_inodes, 0, "");
 154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 155            &hammer2_count_modified_chains, 0, "");
 156 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 157            &hammer2_dio_count, 0, "");
 158 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
 159            &hammer2_dio_limit, 0, "");
 160
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 162            &hammer2_iod_invals, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 164            &hammer2_iod_file_read, 0, "");
 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 166            &hammer2_iod_meta_read, 0, "");
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 168            &hammer2_iod_indr_read, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 170            &hammer2_iod_fmap_read, 0, "");
 171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 172            &hammer2_iod_volu_read, 0, "");
 173
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 175            &hammer2_iod_file_write, 0, "");
 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 177            &hammer2_iod_file_wembed, 0, "");
 178 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 179            &hammer2_iod_file_wzero, 0, "");
 180 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 181            &hammer2_iod_file_wdedup, 0, "");
 182 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 183            &hammer2_iod_meta_write, 0, "");
 184 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 185            &hammer2_iod_indr_write, 0, "");
 186 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 187            &hammer2_iod_fmap_write, 0, "");
 188 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 189            &hammer2_iod_volu_write, 0, "");
 190 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW,
 191            &hammer2_iod_inode_creates, 0, "");
 192 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW,
 193            &hammer2_iod_inode_deletes, 0, "");
 194
 195 long hammer2_process_icrc32;
 196 long hammer2_process_xxhash64;
 197 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
 198            &hammer2_process_icrc32, 0, "");
 199 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
 200            &hammer2_process_xxhash64, 0, "");
 201
 202 static int hammer2_vfs_init(struct vfsconf *conf);
 203 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 204 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 205                                 struct ucred *cred);
 206 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 207                                 struct vnode *, struct ucred *);
 208 static int hammer2_recovery(hammer2_dev_t *hmp);
 209 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 210 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 211 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 212                                 struct ucred *cred);
 213 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 214                                 struct ucred *cred);
 215 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 216                                 struct fid *fhp, struct vnode **vpp);
 217 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 218 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 219                                 int *exflagsp, struct ucred **credanonp);
 220 static void hammer2_vfs_modifying(struct mount *mp);
 221
 222 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 223 #if 0
 224 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 225 #endif
 226
 227 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 228
 229 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 230 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 231                                 hammer2_dev_t *hmp);
 232 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
 233
 234 /*
 235  * HAMMER2 vfs operations.
 236  */
 237 static struct vfsops hammer2_vfsops = {
 238         .vfs_init       = hammer2_vfs_init,
 239         .vfs_uninit     = hammer2_vfs_uninit,
 240         .vfs_sync       = hammer2_vfs_sync,
 241         .vfs_mount      = hammer2_vfs_mount,
 242         .vfs_unmount    = hammer2_vfs_unmount,
 243         .vfs_root       = hammer2_vfs_root,
 244         .vfs_statfs     = hammer2_vfs_statfs,
 245         .vfs_statvfs    = hammer2_vfs_statvfs,
 246         .vfs_vget       = hammer2_vfs_vget,
 247         .vfs_vptofh     = hammer2_vfs_vptofh,
 248         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 249         .vfs_checkexp   = hammer2_vfs_checkexp,
 250         .vfs_modifying  = hammer2_vfs_modifying
 251 };
 252
 253 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 254
 255 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 256 MODULE_VERSION(hammer2, 1);
 257
 258 static
 259 int
 260 hammer2_vfs_init(struct vfsconf *conf)
 261 {
 262         static struct objcache_malloc_args margs_read;
 263         static struct objcache_malloc_args margs_write;
 264         static struct objcache_malloc_args margs_vop;
 265
 266         int error;
 267
 268         error = 0;
 269
 270         /*
 271          * A large DIO cache is needed to retain dedup enablement masks.
 272          * The bulkfree code clears related masks as part of the disk block
 273          * recycling algorithm, preventing it from being used for a later
 274          * dedup.
 275          *
 276          * NOTE: A large buffer cache can actually interfere with dedup
 277          *       operation because we dedup based on media physical buffers
 278          *       and not logical buffers.  Try to make the DIO case large
 279          *       enough to avoid this problem, but also cap it.
 280          */
 281         hammer2_dio_limit = nbuf * 2;
 282         if (hammer2_dio_limit > 100000)
 283                 hammer2_dio_limit = 100000;
 284
 285         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 286                 error = EINVAL;
 287         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 288                 error = EINVAL;
 289         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 290                 error = EINVAL;
 291
 292         if (error)
 293                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 294
 295         margs_read.objsize = 65536;
 296         margs_read.mtype = M_HAMMER2_DEBUFFER;
 297
 298         margs_write.objsize = 32768;
 299         margs_write.mtype = M_HAMMER2_CBUFFER;
 300
 301         margs_vop.objsize = sizeof(hammer2_xop_t);
 302         margs_vop.mtype = M_HAMMER2;
 303
 304         /*
 305          * Note thaht for the XOPS cache we want backing store allocations
 306          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 307          * confusion), so use the backing store function that does it.  This
 308          * means that initial XOPS objects are zerod but REUSED objects are
 309          * not.  So we are responsible for cleaning the object up sufficiently
 310          * for our needs before objcache_put()ing it back (typically just the
 311          * FIFO indices).
 312          */
 313         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 314                                 0, 1, NULL, NULL, NULL,
 315                                 objcache_malloc_alloc,
 316                                 objcache_malloc_free,
 317                                 &margs_read);
 318         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 319                                 0, 1, NULL, NULL, NULL,
 320                                 objcache_malloc_alloc,
 321                                 objcache_malloc_free,
 322                                 &margs_write);
 323         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 324                                 0, 1, NULL, NULL, NULL,
 325                                 objcache_malloc_alloc_zero,
 326                                 objcache_malloc_free,
 327                                 &margs_vop);
 328
 329
 330         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 331         TAILQ_INIT(&hammer2_mntlist);
 332         TAILQ_INIT(&hammer2_pfslist);
 333         TAILQ_INIT(&hammer2_spmplist);
 334
 335         hammer2_limit_dirty_chains = maxvnodes / 10;
 336         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 337                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 338         if (hammer2_limit_dirty_chains < 1000)
 339                 hammer2_limit_dirty_chains = 1000;
 340
 341         hammer2_limit_dirty_inodes = maxvnodes / 25;
 342         if (hammer2_limit_dirty_inodes < 100)
 343                 hammer2_limit_dirty_inodes = 100;
 344         if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
 345                 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
 346
 347         return (error);
 348 }
 349
 350 static
 351 int
 352 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 353 {
 354         objcache_destroy(cache_buffer_read);
 355         objcache_destroy(cache_buffer_write);
 356         objcache_destroy(cache_xops);
 357         return 0;
 358 }
 359
 360 /*
 361  * Core PFS allocator.  Used to allocate or reference the pmp structure
 362  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 363  * The pmp can be passed in or loaded by this function using the chain and
 364  * inode data.
 365  *
 366  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 367  * transactions.  Note that synchronization does not use this field.
 368  * (typically frontend operations and synchronization cannot run on the
 369  * same PFS node at the same time).
 370  *
 371  * XXX check locking
 372  */
 373 hammer2_pfs_t *
 374 hammer2_pfsalloc(hammer2_chain_t *chain,
 375                  const hammer2_inode_data_t *ripdata,
 376                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 377 {
 378         hammer2_pfs_t *pmp;
 379         hammer2_inode_t *iroot;
 380         int count;
 381         int i;
 382         int j;
 383
 384         pmp = NULL;
 385
 386         /*
 387          * Locate or create the PFS based on the cluster id.  If ripdata
 388          * is NULL this is a spmp which is unique and is always allocated.
 389          *
 390          * If the device is mounted in local mode all PFSs are considered
 391          * independent and not part of any cluster (for debugging only).
 392          */
 393         if (ripdata) {
 394                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 395                         if (force_local != pmp->force_local)
 396                                 continue;
 397                         if (force_local == NULL &&
 398                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 399                                  sizeof(pmp->pfs_clid)) == 0) {
 400                                         break;
 401                         } else if (force_local && pmp->pfs_names[0] &&
 402                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 403                                         break;
 404                         }
 405                 }
 406         }
 407
 408         if (pmp == NULL) {
 409                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 410                 pmp->force_local = force_local;
 411                 hammer2_trans_manage_init(pmp);
 412                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 413                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 414                 lockinit(&pmp->lock, "pfslk", 0, 0);
 415                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 416                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 417                 spin_init(&pmp->xop_spin, "h2xop");
 418                 spin_init(&pmp->lru_spin, "h2lru");
 419                 RB_INIT(&pmp->inum_tree);
 420                 TAILQ_INIT(&pmp->syncq);
 421                 TAILQ_INIT(&pmp->depq);
 422                 TAILQ_INIT(&pmp->lru_list);
 423                 spin_init(&pmp->list_spin, "h2pfsalloc_list");
 424
 425                 /*
 426                  * Distribute backend operations to threads
 427                  */
 428                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 429                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 430
 431                 /*
 432                  * Save the last media transaction id for the flusher.  Set
 433                  * initial
 434                  */
 435                 if (ripdata) {
 436                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 437                         TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 438                 } else {
 439                         pmp->flags |= HAMMER2_PMPF_SPMP;
 440                         TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
 441                 }
 442
 443                 /*
 444                  * The synchronization thread may start too early, make
 445                  * sure it stays frozen until we are ready to let it go.
 446                  * XXX
 447                  */
 448                 /*
 449                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 450                                          HAMMER2_THREAD_REMASTER;
 451                 */
 452         }
 453
 454         /*
 455          * Create the PFS's root inode and any missing XOP helper threads.
 456          */
 457         if ((iroot = pmp->iroot) == NULL) {
 458                 iroot = hammer2_inode_get(pmp, NULL, 1, -1);
 459                 if (ripdata)
 460                         iroot->meta = ripdata->meta;
 461                 pmp->iroot = iroot;
 462                 hammer2_inode_ref(iroot);
 463                 hammer2_inode_unlock(iroot);
 464         }
 465
 466         /*
 467          * Stop here if no chain is passed in.
 468          */
 469         if (chain == NULL)
 470                 goto done;
 471
 472         /*
 473          * When a chain is passed in we must add it to the PFS's root
 474          * inode, update pmp->pfs_types[], and update the syncronization
 475          * threads.
 476          *
 477          * When forcing local mode, mark the PFS as a MASTER regardless.
 478          *
 479          * At the moment empty spots can develop due to removals or failures.
 480          * Ultimately we want to re-fill these spots but doing so might
 481          * confused running code. XXX
 482          */
 483         hammer2_inode_ref(iroot);
 484         hammer2_mtx_ex(&iroot->lock);
 485         j = iroot->cluster.nchains;
 486
 487         if (j == HAMMER2_MAXCLUSTER) {
 488                 kprintf("hammer2_mount: cluster full!\n");
 489                 /* XXX fatal error? */
 490         } else {
 491                 KKASSERT(chain->pmp == NULL);
 492                 chain->pmp = pmp;
 493                 hammer2_chain_ref(chain);
 494                 iroot->cluster.array[j].chain = chain;
 495                 if (force_local)
 496                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 497                 else
 498                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 499                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 500                 pmp->pfs_hmps[j] = chain->hmp;
 501                 hammer2_spin_ex(&pmp->inum_spin);
 502                 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset;
 503                 hammer2_spin_unex(&pmp->inum_spin);
 504
 505                 /*
 506                  * If the PFS is already mounted we must account
 507                  * for the mount_count here.
 508                  */
 509                 if (pmp->mp)
 510                         ++chain->hmp->mount_count;
 511
 512                 /*
 513                  * May have to fixup dirty chain tracking.  Previous
 514                  * pmp was NULL so nothing to undo.
 515                  */
 516                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 517                         hammer2_pfs_memory_inc(pmp);
 518                 ++j;
 519         }
 520         iroot->cluster.nchains = j;
 521
 522         /*
 523          * Update nmasters from any PFS inode which is part of the cluster.
 524          * It is possible that this will result in a value which is too
 525          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 526          * override this value later on.
 527          *
 528          * (This informs us of masters that might not currently be
 529          *  discoverable by this mount).
 530          */
 531         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 532                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 533         }
 534
 535         /*
 536          * Count visible masters.  Masters are usually added with
 537          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 538          * are more (XXX and must update the master inodes).
 539          */
 540         count = 0;
 541         for (i = 0; i < iroot->cluster.nchains; ++i) {
 542                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 543                         ++count;
 544         }
 545         if (pmp->pfs_nmasters < count)
 546                 pmp->pfs_nmasters = count;
 547
 548         /*
 549          * Create missing synchronization and support threads.
 550          *
 551          * Single-node masters (including snapshots) have nothing to
 552          * synchronize and do not require this thread.
 553          *
 554          * Multi-node masters or any number of soft masters, slaves, copy,
 555          * or other PFS types need the thread.
 556          *
 557          * Each thread is responsible for its particular cluster index.
 558          * We use independent threads so stalls or mismatches related to
 559          * any given target do not affect other targets.
 560          */
 561         for (i = 0; i < iroot->cluster.nchains; ++i) {
 562                 /*
 563                  * Single-node masters (including snapshots) have nothing
 564                  * to synchronize and will make direct xops support calls,
 565                  * thus they do not require this thread.
 566                  *
 567                  * Note that there can be thousands of snapshots.  We do not
 568                  * want to create thousands of threads.
 569                  */
 570                 if (pmp->pfs_nmasters <= 1 &&
 571                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 572                         continue;
 573                 }
 574
 575                 /*
 576                  * Sync support thread
 577                  */
 578                 if (pmp->sync_thrs[i].td == NULL) {
 579                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 580                                            "h2nod", i, -1,
 581                                            hammer2_primary_sync_thread);
 582                 }
 583         }
 584
 585         /*
 586          * Create missing Xop threads
 587          *
 588          * NOTE: We create helper threads for all mounted PFSs or any
 589          *       PFSs with 2+ nodes (so the sync thread can update them,
 590          *       even if not mounted).
 591          */
 592         if (pmp->mp || iroot->cluster.nchains >= 2)
 593                 hammer2_xop_helper_create(pmp);
 594
 595         hammer2_mtx_unlock(&iroot->lock);
 596         hammer2_inode_drop(iroot);
 597 done:
 598         return pmp;
 599 }
 600
 601 /*
 602  * Deallocate an element of a probed PFS.  If destroying and this is a
 603  * MASTER, adjust nmasters.
 604  *
 605  * This function does not physically destroy the PFS element in its device
 606  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 607  */
 608 void
 609 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 610 {
 611         hammer2_inode_t *iroot;
 612         hammer2_chain_t *chain;
 613         int j;
 614
 615         /*
 616          * Cleanup our reference on iroot.  iroot is (should) not be needed
 617          * by the flush code.
 618          */
 619         iroot = pmp->iroot;
 620         if (iroot) {
 621                 /*
 622                  * Stop synchronizing
 623                  *
 624                  * XXX flush after acquiring the iroot lock.
 625                  * XXX clean out the cluster index from all inode structures.
 626                  */
 627                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 628
 629                 /*
 630                  * Remove the cluster index from the group.  If destroying
 631                  * the PFS and this is a master, adjust pfs_nmasters.
 632                  */
 633                 hammer2_mtx_ex(&iroot->lock);
 634                 chain = iroot->cluster.array[clindex].chain;
 635                 iroot->cluster.array[clindex].chain = NULL;
 636
 637                 switch(pmp->pfs_types[clindex]) {
 638                 case HAMMER2_PFSTYPE_MASTER:
 639                         if (destroying && pmp->pfs_nmasters > 0)
 640                                 --pmp->pfs_nmasters;
 641                         /* XXX adjust ripdata->meta.pfs_nmasters */
 642                         break;
 643                 default:
 644                         break;
 645                 }
 646                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 647
 648                 hammer2_mtx_unlock(&iroot->lock);
 649
 650                 /*
 651                  * Release the chain.
 652                  */
 653                 if (chain) {
 654                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 655                         hammer2_chain_drop(chain);
 656                 }
 657
 658                 /*
 659                  * Terminate all XOP threads for the cluster index.
 660                  */
 661                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 662                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 663         }
 664 }
 665
 666 /*
 667  * Destroy a PFS, typically only occurs after the last mount on a device
 668  * has gone away.
 669  */
 670 static void
 671 hammer2_pfsfree(hammer2_pfs_t *pmp)
 672 {
 673         hammer2_inode_t *iroot;
 674         hammer2_chain_t *chain;
 675         int chains_still_present = 0;
 676         int i;
 677         int j;
 678
 679         /*
 680          * Cleanup our reference on iroot.  iroot is (should) not be needed
 681          * by the flush code.
 682          */
 683         if (pmp->flags & HAMMER2_PMPF_SPMP)
 684                 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
 685         else
 686                 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 687
 688         /*
 689          * Cleanup chains remaining on LRU list.
 690          */
 691         hammer2_spin_ex(&pmp->lru_spin);
 692         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 693                 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
 694                 atomic_add_int(&pmp->lru_count, -1);
 695                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
 696                 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
 697                 hammer2_chain_ref(chain);
 698                 hammer2_spin_unex(&pmp->lru_spin);
 699                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 700                 hammer2_chain_drop(chain);
 701                 hammer2_spin_ex(&pmp->lru_spin);
 702         }
 703         hammer2_spin_unex(&pmp->lru_spin);
 704
 705         /*
 706          * Clean up iroot
 707          */
 708         iroot = pmp->iroot;
 709         if (iroot) {
 710                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 711                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 712                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 713                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 714                         chain = iroot->cluster.array[i].chain;
 715                         if (chain && !RB_EMPTY(&chain->core.rbtree)) {
 716                                 kprintf("hammer2: Warning pmp %p still "
 717                                         "has active chains\n", pmp);
 718                                 chains_still_present = 1;
 719                         }
 720                 }
 721 #if REPORT_REFS_ERRORS
 722                 if (iroot->refs != 1)
 723                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 724                                 iroot, iroot->refs);
 725 #else
 726                 KKASSERT(iroot->refs == 1);
 727 #endif
 728                 /* ref for iroot */
 729                 hammer2_inode_drop(iroot);
 730                 pmp->iroot = NULL;
 731         }
 732
 733         /*
 734          * Free remaining pmp resources
 735          */
 736         if (chains_still_present) {
 737                 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp);
 738         } else {
 739                 kmalloc_destroy(&pmp->mmsg);
 740                 kmalloc_destroy(&pmp->minode);
 741                 kfree(pmp, M_HAMMER2);
 742         }
 743 }
 744
 745 /*
 746  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 747  * empty is terminated and freed.
 748  *
 749  * XXX inefficient.
 750  */
 751 static void
 752 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
 753 {
 754         hammer2_pfs_t *pmp;
 755         hammer2_inode_t *iroot;
 756         hammer2_chain_t *rchain;
 757         int i;
 758         int j;
 759         struct hammer2_pfslist *wlist;
 760
 761         if (which == 0)
 762                 wlist = &hammer2_pfslist;
 763         else
 764                 wlist = &hammer2_spmplist;
 765 again:
 766         TAILQ_FOREACH(pmp, wlist, mntentry) {
 767                 if ((iroot = pmp->iroot) == NULL)
 768                         continue;
 769
 770                 /*
 771                  * Determine if this PFS is affected.  If it is we must
 772                  * freeze all management threads and lock its iroot.
 773                  *
 774                  * Freezing a management thread forces it idle, operations
 775                  * in-progress will be aborted and it will have to start
 776                  * over again when unfrozen, or exit if told to exit.
 777                  */
 778                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 779                         if (pmp->pfs_hmps[i] == hmp)
 780                                 break;
 781                 }
 782                 if (i == HAMMER2_MAXCLUSTER)
 783                         continue;
 784
 785                 hammer2_vfs_sync_pmp(pmp, MNT_WAIT);
 786
 787                 /*
 788                  * Make sure all synchronization threads are locked
 789                  * down.
 790                  */
 791                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 792                         if (pmp->pfs_hmps[i] == NULL)
 793                                 continue;
 794                         hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 795                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 796                                 hammer2_thr_freeze_async(
 797                                         &pmp->xop_groups[j].thrs[i]);
 798                         }
 799                 }
 800                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 801                         if (pmp->pfs_hmps[i] == NULL)
 802                                 continue;
 803                         hammer2_thr_freeze(&pmp->sync_thrs[i]);
 804                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 805                                 hammer2_thr_freeze(
 806                                         &pmp->xop_groups[j].thrs[i]);
 807                         }
 808                 }
 809
 810                 /*
 811                  * Lock the inode and clean out matching chains.
 812                  * Note that we cannot use hammer2_inode_lock_*()
 813                  * here because that would attempt to validate the
 814                  * cluster that we are in the middle of ripping
 815                  * apart.
 816                  *
 817                  * WARNING! We are working directly on the inodes
 818                  *          embedded cluster.
 819                  */
 820                 hammer2_mtx_ex(&iroot->lock);
 821
 822                 /*
 823                  * Remove the chain from matching elements of the PFS.
 824                  */
 825                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 826                         if (pmp->pfs_hmps[i] != hmp)
 827                                 continue;
 828                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 829                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 830                                 hammer2_thr_delete(
 831                                         &pmp->xop_groups[j].thrs[i]);
 832                         }
 833                         rchain = iroot->cluster.array[i].chain;
 834                         iroot->cluster.array[i].chain = NULL;
 835                         pmp->pfs_types[i] = 0;
 836                         if (pmp->pfs_names[i]) {
 837                                 kfree(pmp->pfs_names[i], M_HAMMER2);
 838                                 pmp->pfs_names[i] = NULL;
 839                         }
 840                         if (rchain) {
 841                                 hammer2_chain_drop(rchain);
 842                                 /* focus hint */
 843                                 if (iroot->cluster.focus == rchain)
 844                                         iroot->cluster.focus = NULL;
 845                         }
 846                         pmp->pfs_hmps[i] = NULL;
 847                 }
 848                 hammer2_mtx_unlock(&iroot->lock);
 849
 850                 /*
 851                  * Cleanup trailing chains.  Gaps may remain.
 852                  */
 853                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 854                         if (pmp->pfs_hmps[i])
 855                                 break;
 856                 }
 857                 iroot->cluster.nchains = i + 1;
 858
 859                 /*
 860                  * If the PMP has no elements remaining we can destroy it.
 861                  * (this will transition management threads from frozen->exit).
 862                  */
 863                 if (iroot->cluster.nchains == 0) {
 864                         /*
 865                          * If this was the hmp's spmp, we need to clean
 866                          * a little more stuff out.
 867                          */
 868                         if (hmp->spmp == pmp) {
 869                                 hmp->spmp = NULL;
 870                                 hmp->vchain.pmp = NULL;
 871                                 hmp->fchain.pmp = NULL;
 872                         }
 873
 874                         /*
 875                          * Free the pmp and restart the loop
 876                          */
 877                         KKASSERT(TAILQ_EMPTY(&pmp->syncq));
 878                         KKASSERT(TAILQ_EMPTY(&pmp->depq));
 879                         hammer2_pfsfree(pmp);
 880                         goto again;
 881                 }
 882
 883                 /*
 884                  * If elements still remain we need to set the REMASTER
 885                  * flag and unfreeze it.
 886                  */
 887                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 888                         if (pmp->pfs_hmps[i] == NULL)
 889                                 continue;
 890                         hammer2_thr_remaster(&pmp->sync_thrs[i]);
 891                         hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 892                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 893                                 hammer2_thr_remaster(
 894                                         &pmp->xop_groups[j].thrs[i]);
 895                                 hammer2_thr_unfreeze(
 896                                         &pmp->xop_groups[j].thrs[i]);
 897                         }
 898                 }
 899         }
 900 }
 901
 902 /*
 903  * Mount or remount HAMMER2 fileystem from physical media
 904  *
 905  *      mountroot
 906  *              mp              mount point structure
 907  *              path            NULL
 908  *              data            <unused>
 909  *              cred            <unused>
 910  *
 911  *      mount
 912  *              mp              mount point structure
 913  *              path            path to mount point
 914  *              data            pointer to argument structure in user space
 915  *                      volume  volume path (device@LABEL form)
 916  *                      hflags  user mount flags
 917  *              cred            user credentials
 918  *
 919  * RETURNS:     0       Success
 920  *              !0      error number
 921  */
 922 static
 923 int
 924 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 925                   struct ucred *cred)
 926 {
 927         struct hammer2_mount_info info;
 928         hammer2_pfs_t *pmp;
 929         hammer2_pfs_t *spmp;
 930         hammer2_dev_t *hmp;
 931         hammer2_dev_t *force_local;
 932         hammer2_key_t key_next;
 933         hammer2_key_t key_dummy;
 934         hammer2_key_t lhc;
 935         struct vnode *devvp;
 936         struct nlookupdata nd;
 937         hammer2_chain_t *parent;
 938         hammer2_chain_t *chain;
 939         const hammer2_inode_data_t *ripdata;
 940         hammer2_blockref_t bref;
 941         struct file *fp;
 942         char devstr[MNAMELEN];
 943         size_t size;
 944         size_t done;
 945         char *dev;
 946         char *label;
 947         int ronly = 1;
 948         int error;
 949         int i;
 950
 951         hmp = NULL;
 952         pmp = NULL;
 953         dev = NULL;
 954         label = NULL;
 955         devvp = NULL;
 956
 957         if (path == NULL) {
 958                 /*
 959                  * Root mount
 960                  */
 961                 bzero(&info, sizeof(info));
 962                 info.cluster_fd = -1;
 963                 ksnprintf(devstr, sizeof(devstr), "%s",
 964                           mp->mnt_stat.f_mntfromname);
 965                 kprintf("hammer2_mount: root '%s'\n", devstr);
 966                 done = strlen(devstr) + 1;
 967         } else {
 968                 /*
 969                  * Non-root mount or updating a mount
 970                  */
 971                 error = copyin(data, &info, sizeof(info));
 972                 if (error)
 973                         return (error);
 974
 975                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 976                 if (error)
 977                         return (error);
 978                 kprintf("hammer2_mount: '%s'\n", devstr);
 979         }
 980
 981         /*
 982          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 983          * if no label specified, based on the partition id.  Error out if no
 984          * label or device (with partition id) is specified.  This is strictly
 985          * a convenience to match the default label created by newfs_hammer2,
 986          * our preference is that a label always be specified.
 987          *
 988          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
 989          *       that does not specify a device, as long as some H2 label
 990          *       has already been mounted from that device.  This makes
 991          *       mounting snapshots a lot easier.
 992          */
 993         dev = devstr;
 994         label = strchr(devstr, '@');
 995         if (label && ((label + 1) - dev) > done) {
 996                 kprintf("hammer2: mount: bad label %s/%zd\n",
 997                         devstr, done);
 998                 return (EINVAL);
 999         }
1000         if (label == NULL || label[1] == 0) {
1001                 char slice;
1002
1003                 if (label == NULL)
1004                         label = devstr + strlen(devstr);
1005                 else
1006                         *label = '\0';          /* clean up trailing @ */
1007
1008                 slice = label[-1];
1009                 switch(slice) {
1010                 case 'a':
1011                         label = "BOOT";
1012                         break;
1013                 case 'd':
1014                         label = "ROOT";
1015                         break;
1016                 default:
1017                         label = "DATA";
1018                         break;
1019                 }
1020         } else {
1021                 *label = '\0';
1022                 label++;
1023         }
1024
1025         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
1026                 dev, label, (mp->mnt_flag & MNT_RDONLY));
1027
1028         if (mp->mnt_flag & MNT_UPDATE) {
1029                 /*
1030                  * Update mount.  Note that pmp->iroot->cluster is
1031                  * an inode-embedded cluster and thus cannot be
1032                  * directly locked.
1033                  *
1034                  * XXX HAMMER2 needs to implement NFS export via
1035                  *     mountctl.
1036                  */
1037                 hammer2_cluster_t *cluster;
1038
1039                 pmp = MPTOPMP(mp);
1040                 pmp->hflags = info.hflags;
1041                 cluster = &pmp->iroot->cluster;
1042                 for (i = 0; i < cluster->nchains; ++i) {
1043                         if (cluster->array[i].chain == NULL)
1044                                 continue;
1045                         hmp = cluster->array[i].chain->hmp;
1046                         devvp = hmp->devvp;
1047                         error = hammer2_remount(hmp, mp, path,
1048                                                 devvp, cred);
1049                         if (error)
1050                                 break;
1051                 }
1052
1053                 return error;
1054         }
1055
1056         /*
1057          * HMP device mount
1058          *
1059          * If a path is specified and dev is not an empty string, lookup the
1060          * name and verify that it referes to a block device.
1061          *
1062          * If a path is specified and dev is an empty string we fall through
1063          * and locate the label in the hmp search.
1064          */
1065         if (path && *dev != 0) {
1066                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1067                 if (error == 0)
1068                         error = nlookup(&nd);
1069                 if (error == 0)
1070                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1071                 nlookup_done(&nd);
1072         } else if (path == NULL) {
1073                 /* root mount */
1074                 cdev_t cdev = kgetdiskbyname(dev);
1075                 error = bdevvp(cdev, &devvp);
1076                 if (error)
1077                         kprintf("hammer2: cannot find '%s'\n", dev);
1078         } else {
1079                 /*
1080                  * We will locate the hmp using the label in the hmp loop.
1081                  */
1082                 error = 0;
1083         }
1084
1085         /*
1086          * Make sure its a block device.  Do not check to see if it is
1087          * already mounted until we determine that its a fresh H2 device.
1088          */
1089         if (error == 0 && devvp) {
1090                 vn_isdisk(devvp, &error);
1091         }
1092
1093         /*
1094          * Determine if the device has already been mounted.  After this
1095          * check hmp will be non-NULL if we are doing the second or more
1096          * hammer2 mounts from the same device.
1097          */
1098         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1099         if (devvp) {
1100                 /*
1101                  * Match the device.  Due to the way devfs works,
1102                  * we may not be able to directly match the vnode pointer,
1103                  * so also check to see if the underlying device matches.
1104                  */
1105                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1106                         if (hmp->devvp == devvp)
1107                                 break;
1108                         if (devvp->v_rdev &&
1109                             hmp->devvp->v_rdev == devvp->v_rdev) {
1110                                 break;
1111                         }
1112                 }
1113
1114                 /*
1115                  * If no match this may be a fresh H2 mount, make sure
1116                  * the device is not mounted on anything else.
1117                  */
1118                 if (hmp == NULL)
1119                         error = vfs_mountedon(devvp);
1120         } else if (error == 0) {
1121                 /*
1122                  * Match the label to a pmp already probed.
1123                  */
1124                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1125                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1126                                 if (pmp->pfs_names[i] &&
1127                                     strcmp(pmp->pfs_names[i], label) == 0) {
1128                                         hmp = pmp->pfs_hmps[i];
1129                                         break;
1130                                 }
1131                         }
1132                         if (hmp)
1133                                 break;
1134                 }
1135                 if (hmp == NULL)
1136                         error = ENOENT;
1137         }
1138
1139         /*
1140          * Open the device if this isn't a secondary mount and construct
1141          * the H2 device mount (hmp).
1142          */
1143         if (hmp == NULL) {
1144                 hammer2_chain_t *schain;
1145                 hammer2_xid_t xid;
1146                 hammer2_xop_head_t xop;
1147
1148                 if (error == 0 && vcount(devvp) > 0) {
1149                         kprintf("Primary device already has references\n");
1150                         error = EBUSY;
1151                 }
1152
1153                 /*
1154                  * Now open the device
1155                  */
1156                 if (error == 0) {
1157                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1158                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1159                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1160                         if (error == 0) {
1161                                 error = VOP_OPEN(devvp,
1162                                              (ronly ? FREAD : FREAD | FWRITE),
1163                                              FSCRED, NULL);
1164                         }
1165                         vn_unlock(devvp);
1166                 }
1167                 if (error && devvp) {
1168                         vrele(devvp);
1169                         devvp = NULL;
1170                 }
1171                 if (error) {
1172                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1173                         return error;
1174                 }
1175                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1176                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1177                 hmp->ronly = ronly;
1178                 hmp->devvp = devvp;
1179                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1180                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1181                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1182                 RB_INIT(&hmp->iotree);
1183                 spin_init(&hmp->io_spin, "h2mount_io");
1184                 spin_init(&hmp->list_spin, "h2mount_list");
1185
1186                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1187                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1188                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1189
1190                 /*
1191                  * vchain setup. vchain.data is embedded.
1192                  * vchain.refs is initialized and will never drop to 0.
1193                  *
1194                  * NOTE! voldata is not yet loaded.
1195                  */
1196                 hmp->vchain.hmp = hmp;
1197                 hmp->vchain.refs = 1;
1198                 hmp->vchain.data = (void *)&hmp->voldata;
1199                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1200                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1201                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1202
1203                 hammer2_chain_core_init(&hmp->vchain);
1204                 /* hmp->vchain.u.xxx is left NULL */
1205
1206                 /*
1207                  * fchain setup.  fchain.data is embedded.
1208                  * fchain.refs is initialized and will never drop to 0.
1209                  *
1210                  * The data is not used but needs to be initialized to
1211                  * pass assertion muster.  We use this chain primarily
1212                  * as a placeholder for the freemap's top-level RBTREE
1213                  * so it does not interfere with the volume's topology
1214                  * RBTREE.
1215                  */
1216                 hmp->fchain.hmp = hmp;
1217                 hmp->fchain.refs = 1;
1218                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1219                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1220                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1221                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1222                 hmp->fchain.bref.methods =
1223                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1224                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1225
1226                 hammer2_chain_core_init(&hmp->fchain);
1227                 /* hmp->fchain.u.xxx is left NULL */
1228
1229                 /*
1230                  * Install the volume header and initialize fields from
1231                  * voldata.
1232                  */
1233                 error = hammer2_install_volume_header(hmp);
1234                 if (error) {
1235                         hammer2_unmount_helper(mp, NULL, hmp);
1236                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1237                         hammer2_vfs_unmount(mp, MNT_FORCE);
1238                         return error;
1239                 }
1240
1241                 /*
1242                  * Really important to get these right or flush will get
1243                  * confused.
1244                  */
1245                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1246                 spmp = hmp->spmp;
1247
1248                 /*
1249                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1250                  * is inherited from the volume header.
1251                  */
1252                 xid = 0;
1253                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1254                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1255                 hmp->vchain.pmp = spmp;
1256                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1257                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1258                 hmp->fchain.pmp = spmp;
1259
1260                 /*
1261                  * First locate the super-root inode, which is key 0
1262                  * relative to the volume header's blockset.
1263                  *
1264                  * Then locate the root inode by scanning the directory keyspace
1265                  * represented by the label.
1266                  */
1267                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1268                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1269                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1270                                       &error, 0);
1271                 hammer2_chain_lookup_done(parent);
1272                 if (schain == NULL) {
1273                         kprintf("hammer2_mount: invalid super-root\n");
1274                         hammer2_unmount_helper(mp, NULL, hmp);
1275                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1276                         hammer2_vfs_unmount(mp, MNT_FORCE);
1277                         return EINVAL;
1278                 }
1279                 if (schain->error) {
1280                         kprintf("hammer2_mount: error %s reading super-root\n",
1281                                 hammer2_error_str(schain->error));
1282                         hammer2_chain_unlock(schain);
1283                         hammer2_chain_drop(schain);
1284                         schain = NULL;
1285                         hammer2_unmount_helper(mp, NULL, hmp);
1286                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1287                         hammer2_vfs_unmount(mp, MNT_FORCE);
1288                         return EINVAL;
1289                 }
1290
1291                 /*
1292                  * The super-root always uses an inode_tid of 1 when
1293                  * creating PFSs.
1294                  */
1295                 spmp->inode_tid = 1;
1296                 spmp->modify_tid = schain->bref.modify_tid + 1;
1297
1298                 /*
1299                  * Sanity-check schain's pmp and finish initialization.
1300                  * Any chain belonging to the super-root topology should
1301                  * have a NULL pmp (not even set to spmp).
1302                  */
1303                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1304                 KKASSERT(schain->pmp == NULL);
1305                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1306
1307                 /*
1308                  * Replace the dummy spmp->iroot with a real one.  It's
1309                  * easier to just do a wholesale replacement than to try
1310                  * to update the chain and fixup the iroot fields.
1311                  *
1312                  * The returned inode is locked with the supplied cluster.
1313                  */
1314                 hammer2_dummy_xop_from_chain(&xop, schain);
1315                 hammer2_inode_drop(spmp->iroot);
1316                 spmp->iroot = NULL;
1317                 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1);
1318                 spmp->spmp_hmp = hmp;
1319                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1320                 spmp->pfs_hmps[0] = hmp;
1321                 hammer2_inode_ref(spmp->iroot);
1322                 hammer2_inode_unlock(spmp->iroot);
1323                 hammer2_cluster_unlock(&xop.cluster);
1324                 hammer2_chain_drop(schain);
1325                 /* do not call hammer2_cluster_drop() on an embedded cluster */
1326                 schain = NULL;  /* now invalid */
1327                 /* leave spmp->iroot with one ref */
1328
1329                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1330                         error = hammer2_recovery(hmp);
1331                         if (error == 0)
1332                                 error |= hammer2_fixup_pfses(hmp);
1333                         /* XXX do something with error */
1334                 }
1335                 hammer2_update_pmps(hmp);
1336                 hammer2_iocom_init(hmp);
1337                 hammer2_bulkfree_init(hmp);
1338
1339                 /*
1340                  * Ref the cluster management messaging descriptor.  The mount
1341                  * program deals with the other end of the communications pipe.
1342                  *
1343                  * Root mounts typically do not supply one.
1344                  */
1345                 if (info.cluster_fd >= 0) {
1346                         fp = holdfp(curthread, info.cluster_fd, -1);
1347                         if (fp) {
1348                                 hammer2_cluster_reconnect(hmp, fp);
1349                         } else {
1350                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1351                         }
1352                 }
1353         } else {
1354                 spmp = hmp->spmp;
1355                 if (info.hflags & HMNT2_DEVFLAGS) {
1356                         kprintf("hammer2: Warning: mount flags pertaining "
1357                                 "to the whole device may only be specified "
1358                                 "on the first mount of the device: %08x\n",
1359                                 info.hflags & HMNT2_DEVFLAGS);
1360                 }
1361         }
1362
1363         /*
1364          * Force local mount (disassociate all PFSs from their clusters).
1365          * Used primarily for debugging.
1366          */
1367         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1368
1369         /*
1370          * Lookup the mount point under the media-localized super-root.
1371          * Scanning hammer2_pfslist doesn't help us because it represents
1372          * PFS cluster ids which can aggregate several named PFSs together.
1373          *
1374          * cluster->pmp will incorrectly point to spmp and must be fixed
1375          * up later on.
1376          */
1377         hammer2_inode_lock(spmp->iroot, 0);
1378         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1379         lhc = hammer2_dirhash(label, strlen(label));
1380         chain = hammer2_chain_lookup(&parent, &key_next,
1381                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1382                                      &error, 0);
1383         while (chain) {
1384                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1385                     strcmp(label, chain->data->ipdata.filename) == 0) {
1386                         break;
1387                 }
1388                 chain = hammer2_chain_next(&parent, chain, &key_next,
1389                                             key_next,
1390                                             lhc + HAMMER2_DIRHASH_LOMASK,
1391                                             &error, 0);
1392         }
1393         if (parent) {
1394                 hammer2_chain_unlock(parent);
1395                 hammer2_chain_drop(parent);
1396         }
1397         hammer2_inode_unlock(spmp->iroot);
1398
1399         /*
1400          * PFS could not be found?
1401          */
1402         if (chain == NULL) {
1403                 if (error)
1404                         kprintf("hammer2_mount: PFS label I/O error\n");
1405                 else
1406                         kprintf("hammer2_mount: PFS label not found\n");
1407                 hammer2_unmount_helper(mp, NULL, hmp);
1408                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1409                 hammer2_vfs_unmount(mp, MNT_FORCE);
1410
1411                 return EINVAL;
1412         }
1413
1414         /*
1415          * Acquire the pmp structure (it should have already been allocated
1416          * via hammer2_update_pmps() so do not pass cluster in to add to
1417          * available chains).
1418          *
1419          * Check if the cluster has already been mounted.  A cluster can
1420          * only be mounted once, use null mounts to mount additional copies.
1421          */
1422         if (chain->error) {
1423                 kprintf("hammer2_mount: PFS label I/O error\n");
1424         } else {
1425                 ripdata = &chain->data->ipdata;
1426                 bref = chain->bref;
1427                 pmp = hammer2_pfsalloc(NULL, ripdata,
1428                                        bref.modify_tid, force_local);
1429         }
1430         hammer2_chain_unlock(chain);
1431         hammer2_chain_drop(chain);
1432
1433         /*
1434          * Finish the mount
1435          */
1436         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1437
1438         if (pmp->mp) {
1439                 kprintf("hammer2_mount: PFS already mounted!\n");
1440                 hammer2_unmount_helper(mp, NULL, hmp);
1441                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1442                 hammer2_vfs_unmount(mp, MNT_FORCE);
1443
1444                 return EBUSY;
1445         }
1446
1447         pmp->hflags = info.hflags;
1448         mp->mnt_flag |= MNT_LOCAL;
1449         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1450         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1451
1452         /*
1453          * required mount structure initializations
1454          */
1455         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1456         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1457
1458         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1459         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1460
1461         /*
1462          * Optional fields
1463          */
1464         mp->mnt_iosize_max = MAXPHYS;
1465
1466         /*
1467          * Connect up mount pointers.
1468          */
1469         hammer2_mount_helper(mp, pmp);
1470
1471         lockmgr(&hammer2_mntlk, LK_RELEASE);
1472
1473         /*
1474          * Finish setup
1475          */
1476         vfs_getnewfsid(mp);
1477         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1478         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1479         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1480
1481         if (path) {
1482                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1483                           MNAMELEN - 1, &size);
1484                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1485         } /* else root mount, already in there */
1486
1487         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1488         if (path) {
1489                 copyinstr(path, mp->mnt_stat.f_mntonname,
1490                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1491                           &size);
1492         } else {
1493                 /* root mount */
1494                 mp->mnt_stat.f_mntonname[0] = '/';
1495         }
1496
1497         /*
1498          * Initial statfs to prime mnt_stat.
1499          */
1500         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1501
1502         return 0;
1503 }
1504
1505 /*
1506  * Scan PFSs under the super-root and create hammer2_pfs structures.
1507  */
1508 static
1509 void
1510 hammer2_update_pmps(hammer2_dev_t *hmp)
1511 {
1512         const hammer2_inode_data_t *ripdata;
1513         hammer2_chain_t *parent;
1514         hammer2_chain_t *chain;
1515         hammer2_blockref_t bref;
1516         hammer2_dev_t *force_local;
1517         hammer2_pfs_t *spmp;
1518         hammer2_pfs_t *pmp;
1519         hammer2_key_t key_next;
1520         int error;
1521
1522         /*
1523          * Force local mount (disassociate all PFSs from their clusters).
1524          * Used primarily for debugging.
1525          */
1526         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1527
1528         /*
1529          * Lookup mount point under the media-localized super-root.
1530          *
1531          * cluster->pmp will incorrectly point to spmp and must be fixed
1532          * up later on.
1533          */
1534         spmp = hmp->spmp;
1535         hammer2_inode_lock(spmp->iroot, 0);
1536         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1537         chain = hammer2_chain_lookup(&parent, &key_next,
1538                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1539                                          &error, 0);
1540         while (chain) {
1541                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1542                         continue;
1543                 if (chain->error) {
1544                         kprintf("I/O error scanning PFS labels\n");
1545                 } else {
1546                         ripdata = &chain->data->ipdata;
1547                         bref = chain->bref;
1548
1549                         pmp = hammer2_pfsalloc(chain, ripdata,
1550                                                bref.modify_tid, force_local);
1551                 }
1552                 chain = hammer2_chain_next(&parent, chain, &key_next,
1553                                            key_next, HAMMER2_KEY_MAX,
1554                                            &error, 0);
1555         }
1556         if (parent) {
1557                 hammer2_chain_unlock(parent);
1558                 hammer2_chain_drop(parent);
1559         }
1560         hammer2_inode_unlock(spmp->iroot);
1561 }
1562
1563 static
1564 int
1565 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1566                 struct vnode *devvp, struct ucred *cred)
1567 {
1568         int error;
1569
1570         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1571                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1572                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1573                 vn_unlock(devvp);
1574                 error = hammer2_recovery(hmp);
1575                 if (error == 0)
1576                         error |= hammer2_fixup_pfses(hmp);
1577                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1578                 if (error == 0) {
1579                         VOP_CLOSE(devvp, FREAD, NULL);
1580                         hmp->ronly = 0;
1581                 } else {
1582                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1583                 }
1584                 vn_unlock(devvp);
1585         } else {
1586                 error = 0;
1587         }
1588         return error;
1589 }
1590
1591 static
1592 int
1593 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1594 {
1595         hammer2_pfs_t *pmp;
1596         int flags;
1597         int error = 0;
1598
1599         pmp = MPTOPMP(mp);
1600
1601         if (pmp == NULL)
1602                 return(0);
1603
1604         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1605
1606         /*
1607          * If mount initialization proceeded far enough we must flush
1608          * its vnodes and sync the underlying mount points.  Three syncs
1609          * are required to fully flush the filesystem (freemap updates lag
1610          * by one flush, and one extra for safety).
1611          */
1612         if (mntflags & MNT_FORCE)
1613                 flags = FORCECLOSE;
1614         else
1615                 flags = 0;
1616         if (pmp->iroot) {
1617                 error = vflush(mp, 0, flags);
1618                 if (error)
1619                         goto failed;
1620                 hammer2_vfs_sync(mp, MNT_WAIT);
1621                 hammer2_vfs_sync(mp, MNT_WAIT);
1622                 hammer2_vfs_sync(mp, MNT_WAIT);
1623         }
1624
1625         /*
1626          * Cleanup the frontend support XOPS threads
1627          */
1628         hammer2_xop_helper_cleanup(pmp);
1629
1630         if (pmp->mp)
1631                 hammer2_unmount_helper(mp, pmp, NULL);
1632
1633         error = 0;
1634 failed:
1635         lockmgr(&hammer2_mntlk, LK_RELEASE);
1636
1637         return (error);
1638 }
1639
1640 /*
1641  * Mount helper, hook the system mount into our PFS.
1642  * The mount lock is held.
1643  *
1644  * We must bump the mount_count on related devices for any
1645  * mounted PFSs.
1646  */
1647 static
1648 void
1649 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1650 {
1651         hammer2_cluster_t *cluster;
1652         hammer2_chain_t *rchain;
1653         int i;
1654
1655         mp->mnt_data = (qaddr_t)pmp;
1656         pmp->mp = mp;
1657
1658         /*
1659          * After pmp->mp is set we have to adjust hmp->mount_count.
1660          */
1661         cluster = &pmp->iroot->cluster;
1662         for (i = 0; i < cluster->nchains; ++i) {
1663                 rchain = cluster->array[i].chain;
1664                 if (rchain == NULL)
1665                         continue;
1666                 ++rchain->hmp->mount_count;
1667         }
1668
1669         /*
1670          * Create missing Xop threads
1671          */
1672         hammer2_xop_helper_create(pmp);
1673 }
1674
1675 /*
1676  * Mount helper, unhook the system mount from our PFS.
1677  * The mount lock is held.
1678  *
1679  * If hmp is supplied a mount responsible for being the first to open
1680  * the block device failed and the block device and all PFSs using the
1681  * block device must be cleaned up.
1682  *
1683  * If pmp is supplied multiple devices might be backing the PFS and each
1684  * must be disconnected.  This might not be the last PFS using some of the
1685  * underlying devices.  Also, we have to adjust our hmp->mount_count
1686  * accounting for the devices backing the pmp which is now undergoing an
1687  * unmount.
1688  */
1689 static
1690 void
1691 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1692 {
1693         hammer2_cluster_t *cluster;
1694         hammer2_chain_t *rchain;
1695         struct vnode *devvp;
1696         int dumpcnt;
1697         int ronly;
1698         int i;
1699
1700         /*
1701          * If no device supplied this is a high-level unmount and we have to
1702          * to disconnect the mount, adjust mount_count, and locate devices
1703          * that might now have no mounts.
1704          */
1705         if (pmp) {
1706                 KKASSERT(hmp == NULL);
1707                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1708                 pmp->mp = NULL;
1709                 mp->mnt_data = NULL;
1710
1711                 /*
1712                  * After pmp->mp is cleared we have to account for
1713                  * mount_count.
1714                  */
1715                 cluster = &pmp->iroot->cluster;
1716                 for (i = 0; i < cluster->nchains; ++i) {
1717                         rchain = cluster->array[i].chain;
1718                         if (rchain == NULL)
1719                                 continue;
1720                         --rchain->hmp->mount_count;
1721                         /* scrapping hmp now may invalidate the pmp */
1722                 }
1723 again:
1724                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1725                         if (hmp->mount_count == 0) {
1726                                 hammer2_unmount_helper(NULL, NULL, hmp);
1727                                 goto again;
1728                         }
1729                 }
1730                 return;
1731         }
1732
1733         /*
1734          * Try to terminate the block device.  We can't terminate it if
1735          * there are still PFSs referencing it.
1736          */
1737         if (hmp->mount_count)
1738                 return;
1739
1740         /*
1741          * Decomission the network before we start messing with the
1742          * device and PFS.
1743          */
1744         hammer2_iocom_uninit(hmp);
1745
1746         hammer2_bulkfree_uninit(hmp);
1747         hammer2_pfsfree_scan(hmp, 0);
1748 #if 0
1749         hammer2_dev_exlock(hmp);        /* XXX order */
1750 #endif
1751
1752         /*
1753          * Cycle the volume data lock as a safety (probably not needed any
1754          * more).  To ensure everything is out we need to flush at least
1755          * three times.  (1) The running of the sideq can dirty the
1756          * filesystem, (2) A normal flush can dirty the freemap, and
1757          * (3) ensure that the freemap is fully synchronized.
1758          *
1759          * The next mount's recovery scan can clean everything up but we want
1760          * to leave the filesystem in a 100% clean state on a normal unmount.
1761          */
1762 #if 0
1763         hammer2_voldata_lock(hmp);
1764         hammer2_voldata_unlock(hmp);
1765 #endif
1766
1767         /*
1768          * Flush whatever is left.  Unmounted but modified PFS's might still
1769          * have some dirty chains on them.
1770          */
1771         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1772         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1773
1774         if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1775                 hammer2_voldata_modify(hmp);
1776                 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1777                                             HAMMER2_FLUSH_ALL);
1778         }
1779         hammer2_chain_unlock(&hmp->fchain);
1780
1781         if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1782                 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1783                                             HAMMER2_FLUSH_ALL);
1784         }
1785         hammer2_chain_unlock(&hmp->vchain);
1786
1787         if ((hmp->vchain.flags | hmp->fchain.flags) &
1788             HAMMER2_CHAIN_FLUSH_MASK) {
1789                 kprintf("hammer2_unmount: chains left over "
1790                         "after final sync\n");
1791                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1792                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1793
1794                 if (hammer2_debug & 0x0010)
1795                         Debugger("entered debugger");
1796         }
1797
1798         hammer2_pfsfree_scan(hmp, 1);
1799
1800         KKASSERT(hmp->spmp == NULL);
1801
1802         /*
1803          * Finish up with the device vnode
1804          */
1805         if ((devvp = hmp->devvp) != NULL) {
1806                 ronly = hmp->ronly;
1807                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1808                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1809                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1810                         ronly);
1811                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1812                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1813                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1814                 hmp->devvp = NULL;
1815                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1816                 vn_unlock(devvp);
1817                 vrele(devvp);
1818                 devvp = NULL;
1819         }
1820
1821         /*
1822          * Clear vchain/fchain flags that might prevent final cleanup
1823          * of these chains.
1824          */
1825         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1826                 atomic_add_long(&hammer2_count_modified_chains, -1);
1827                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1828                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1829         }
1830         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1831                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1832         }
1833
1834         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1835                 atomic_add_long(&hammer2_count_modified_chains, -1);
1836                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1837                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1838         }
1839         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1840                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1841         }
1842
1843         /*
1844          * Final drop of embedded freemap root chain to
1845          * clean up fchain.core (fchain structure is not
1846          * flagged ALLOCATED so it is cleaned out and then
1847          * left to rot).
1848          */
1849         hammer2_chain_drop(&hmp->fchain);
1850
1851         /*
1852          * Final drop of embedded volume root chain to clean
1853          * up vchain.core (vchain structure is not flagged
1854          * ALLOCATED so it is cleaned out and then left to
1855          * rot).
1856          */
1857         dumpcnt = 50;
1858         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1859         dumpcnt = 50;
1860         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1861 #if 0
1862         hammer2_dev_unlock(hmp);
1863 #endif
1864         hammer2_chain_drop(&hmp->vchain);
1865
1866         hammer2_io_cleanup(hmp, &hmp->iotree);
1867         if (hmp->iofree_count) {
1868                 kprintf("io_cleanup: %d I/O's left hanging\n",
1869                         hmp->iofree_count);
1870         }
1871
1872         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1873         kmalloc_destroy(&hmp->mchain);
1874         kfree(hmp, M_HAMMER2);
1875 }
1876
1877 int
1878 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1879                  ino_t ino, struct vnode **vpp)
1880 {
1881         hammer2_xop_lookup_t *xop;
1882         hammer2_pfs_t *pmp;
1883         hammer2_inode_t *ip;
1884         hammer2_tid_t inum;
1885         int error;
1886
1887         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1888
1889         error = 0;
1890         pmp = MPTOPMP(mp);
1891
1892         /*
1893          * Easy if we already have it cached
1894          */
1895         ip = hammer2_inode_lookup(pmp, inum);
1896         if (ip) {
1897                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1898                 *vpp = hammer2_igetv(ip, &error);
1899                 hammer2_inode_unlock(ip);
1900                 hammer2_inode_drop(ip);         /* from lookup */
1901
1902                 return error;
1903         }
1904
1905         /*
1906          * Otherwise we have to find the inode
1907          */
1908         xop = hammer2_xop_alloc(pmp->iroot, 0);
1909         xop->lhc = inum;
1910         hammer2_xop_start(&xop->head, &hammer2_lookup_desc);
1911         error = hammer2_xop_collect(&xop->head, 0);
1912
1913         if (error == 0)
1914                 ip = hammer2_inode_get(pmp, &xop->head, -1, -1);
1915         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1916
1917         if (ip) {
1918                 *vpp = hammer2_igetv(ip, &error);
1919                 hammer2_inode_unlock(ip);
1920         } else {
1921                 *vpp = NULL;
1922                 error = ENOENT;
1923         }
1924         return (error);
1925 }
1926
1927 static
1928 int
1929 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1930 {
1931         hammer2_pfs_t *pmp;
1932         struct vnode *vp;
1933         int error;
1934
1935         pmp = MPTOPMP(mp);
1936         if (pmp->iroot == NULL) {
1937                 kprintf("hammer2 (%s): no root inode\n",
1938                         mp->mnt_stat.f_mntfromname);
1939                 *vpp = NULL;
1940                 return EINVAL;
1941         }
1942
1943         error = 0;
1944         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1945
1946         while (pmp->inode_tid == 0) {
1947                 hammer2_xop_ipcluster_t *xop;
1948                 const hammer2_inode_meta_t *meta;
1949
1950                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1951                 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc);
1952                 error = hammer2_xop_collect(&xop->head, 0);
1953
1954                 if (error == 0) {
1955                         meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1956                         pmp->iroot->meta = *meta;
1957                         pmp->inode_tid = meta->pfs_inum + 1;
1958                         hammer2_xop_pdata(&xop->head);
1959                         /* meta invalid */
1960
1961                         if (pmp->inode_tid < HAMMER2_INODE_START)
1962                                 pmp->inode_tid = HAMMER2_INODE_START;
1963                         pmp->modify_tid =
1964                                 xop->head.cluster.focus->bref.modify_tid + 1;
1965 #if 0
1966                         kprintf("PFS: Starting inode %jd\n",
1967                                 (intmax_t)pmp->inode_tid);
1968                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1969                                 pmp->inode_tid, pmp->modify_tid);
1970 #endif
1971                         wakeup(&pmp->iroot);
1972
1973                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1974
1975                         /*
1976                          * Prime the mount info.
1977                          */
1978                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1979                         break;
1980                 }
1981
1982                 /*
1983                  * Loop, try again
1984                  */
1985                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1986                 hammer2_inode_unlock(pmp->iroot);
1987                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1988                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1989                 if (error == EINTR)
1990                         break;
1991         }
1992
1993         if (error) {
1994                 hammer2_inode_unlock(pmp->iroot);
1995                 *vpp = NULL;
1996         } else {
1997                 vp = hammer2_igetv(pmp->iroot, &error);
1998                 hammer2_inode_unlock(pmp->iroot);
1999                 *vpp = vp;
2000         }
2001
2002         return (error);
2003 }
2004
2005 /*
2006  * Filesystem status
2007  *
2008  * XXX incorporate ipdata->meta.inode_quota and data_quota
2009  */
2010 static
2011 int
2012 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
2013 {
2014         hammer2_pfs_t *pmp;
2015         hammer2_dev_t *hmp;
2016         hammer2_blockref_t bref;
2017         struct statfs tmp;
2018         int i;
2019
2020         /*
2021          * NOTE: iroot might not have validated the cluster yet.
2022          */
2023         pmp = MPTOPMP(mp);
2024
2025         bzero(&tmp, sizeof(tmp));
2026
2027         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2028                 hmp = pmp->pfs_hmps[i];
2029                 if (hmp == NULL)
2030                         continue;
2031                 if (pmp->iroot->cluster.array[i].chain)
2032                         bref = pmp->iroot->cluster.array[i].chain->bref;
2033                 else
2034                         bzero(&bref, sizeof(bref));
2035
2036                 tmp.f_files = bref.embed.stats.inode_count;
2037                 tmp.f_ffree = 0;
2038                 tmp.f_blocks = hmp->voldata.allocator_size /
2039                                mp->mnt_vstat.f_bsize;
2040                 tmp.f_bfree = hmp->voldata.allocator_free /
2041                               mp->mnt_vstat.f_bsize;
2042                 tmp.f_bavail = tmp.f_bfree;
2043
2044                 if (cred && cred->cr_uid != 0) {
2045                         uint64_t adj;
2046
2047                         /* 5% */
2048                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2049                         tmp.f_blocks -= adj;
2050                         tmp.f_bfree -= adj;
2051                         tmp.f_bavail -= adj;
2052                 }
2053
2054                 mp->mnt_stat.f_blocks = tmp.f_blocks;
2055                 mp->mnt_stat.f_bfree = tmp.f_bfree;
2056                 mp->mnt_stat.f_bavail = tmp.f_bavail;
2057                 mp->mnt_stat.f_files = tmp.f_files;
2058                 mp->mnt_stat.f_ffree = tmp.f_ffree;
2059
2060                 *sbp = mp->mnt_stat;
2061         }
2062         return (0);
2063 }
2064
2065 static
2066 int
2067 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2068 {
2069         hammer2_pfs_t *pmp;
2070         hammer2_dev_t *hmp;
2071         hammer2_blockref_t bref;
2072         struct statvfs tmp;
2073         int i;
2074
2075         /*
2076          * NOTE: iroot might not have validated the cluster yet.
2077          */
2078         pmp = MPTOPMP(mp);
2079         bzero(&tmp, sizeof(tmp));
2080
2081         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2082                 hmp = pmp->pfs_hmps[i];
2083                 if (hmp == NULL)
2084                         continue;
2085                 if (pmp->iroot->cluster.array[i].chain)
2086                         bref = pmp->iroot->cluster.array[i].chain->bref;
2087                 else
2088                         bzero(&bref, sizeof(bref));
2089
2090                 tmp.f_files = bref.embed.stats.inode_count;
2091                 tmp.f_ffree = 0;
2092                 tmp.f_blocks = hmp->voldata.allocator_size /
2093                                mp->mnt_vstat.f_bsize;
2094                 tmp.f_bfree = hmp->voldata.allocator_free /
2095                               mp->mnt_vstat.f_bsize;
2096                 tmp.f_bavail = tmp.f_bfree;
2097
2098                 if (cred && cred->cr_uid != 0) {
2099                         uint64_t adj;
2100
2101                         /* 5% */
2102                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2103                         tmp.f_blocks -= adj;
2104                         tmp.f_bfree -= adj;
2105                         tmp.f_bavail -= adj;
2106                 }
2107
2108                 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2109                 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2110                 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2111                 mp->mnt_vstat.f_files = tmp.f_files;
2112                 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2113
2114                 *sbp = mp->mnt_vstat;
2115         }
2116         return (0);
2117 }
2118
2119 /*
2120  * Mount-time recovery (RW mounts)
2121  *
2122  * Updates to the free block table are allowed to lag flushes by one
2123  * transaction.  In case of a crash, then on a fresh mount we must do an
2124  * incremental scan of the last committed transaction id and make sure that
2125  * all related blocks have been marked allocated.
2126  *
2127  * The super-root topology and each PFS has its own transaction id domain,
2128  * so we must track PFS boundary transitions.
2129  */
2130 struct hammer2_recovery_elm {
2131         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2132         hammer2_chain_t *chain;
2133         hammer2_tid_t sync_tid;
2134 };
2135
2136 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2137
2138 struct hammer2_recovery_info {
2139         struct hammer2_recovery_list list;
2140         hammer2_tid_t   mtid;
2141         int     depth;
2142 };
2143
2144 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2145                         hammer2_chain_t *parent,
2146                         struct hammer2_recovery_info *info,
2147                         hammer2_tid_t sync_tid);
2148
2149 #define HAMMER2_RECOVERY_MAXDEPTH       10
2150
2151 static
2152 int
2153 hammer2_recovery(hammer2_dev_t *hmp)
2154 {
2155         struct hammer2_recovery_info info;
2156         struct hammer2_recovery_elm *elm;
2157         hammer2_chain_t *parent;
2158         hammer2_tid_t sync_tid;
2159         hammer2_tid_t mirror_tid;
2160         int error;
2161
2162         hammer2_trans_init(hmp->spmp, 0);
2163
2164         sync_tid = hmp->voldata.freemap_tid;
2165         mirror_tid = hmp->voldata.mirror_tid;
2166
2167         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2168         if (sync_tid >= mirror_tid) {
2169                 kprintf(" no recovery needed\n");
2170         } else {
2171                 kprintf(" freemap recovery %016jx-%016jx\n",
2172                         sync_tid + 1, mirror_tid);
2173         }
2174
2175         TAILQ_INIT(&info.list);
2176         info.depth = 0;
2177         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2178         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2179         hammer2_chain_lookup_done(parent);
2180
2181         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2182                 TAILQ_REMOVE(&info.list, elm, entry);
2183                 parent = elm->chain;
2184                 sync_tid = elm->sync_tid;
2185                 kfree(elm, M_HAMMER2);
2186
2187                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2188                 error |= hammer2_recovery_scan(hmp, parent, &info,
2189                                               hmp->voldata.freemap_tid);
2190                 hammer2_chain_unlock(parent);
2191                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2192         }
2193
2194         hammer2_trans_done(hmp->spmp, 0);
2195
2196         return error;
2197 }
2198
2199 static
2200 int
2201 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2202                       struct hammer2_recovery_info *info,
2203                       hammer2_tid_t sync_tid)
2204 {
2205         const hammer2_inode_data_t *ripdata;
2206         hammer2_chain_t *chain;
2207         hammer2_blockref_t bref;
2208         int tmp_error;
2209         int rup_error;
2210         int error;
2211         int first;
2212
2213         /*
2214          * Adjust freemap to ensure that the block(s) are marked allocated.
2215          */
2216         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2217                 hammer2_freemap_adjust(hmp, &parent->bref,
2218                                        HAMMER2_FREEMAP_DORECOVER);
2219         }
2220
2221         /*
2222          * Check type for recursive scan
2223          */
2224         switch(parent->bref.type) {
2225         case HAMMER2_BREF_TYPE_VOLUME:
2226                 /* data already instantiated */
2227                 break;
2228         case HAMMER2_BREF_TYPE_INODE:
2229                 /*
2230                  * Must instantiate data for DIRECTDATA test and also
2231                  * for recursion.
2232                  */
2233                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2234                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2235                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2236                         /* not applicable to recovery scan */
2237                         hammer2_chain_unlock(parent);
2238                         return 0;
2239                 }
2240                 hammer2_chain_unlock(parent);
2241                 break;
2242         case HAMMER2_BREF_TYPE_INDIRECT:
2243                 /*
2244                  * Must instantiate data for recursion
2245                  */
2246                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2247                 hammer2_chain_unlock(parent);
2248                 break;
2249         case HAMMER2_BREF_TYPE_DIRENT:
2250         case HAMMER2_BREF_TYPE_DATA:
2251         case HAMMER2_BREF_TYPE_FREEMAP:
2252         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2253         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2254                 /* not applicable to recovery scan */
2255                 return 0;
2256                 break;
2257         default:
2258                 return HAMMER2_ERROR_BADBREF;
2259         }
2260
2261         /*
2262          * Defer operation if depth limit reached or if we are crossing a
2263          * PFS boundary.
2264          */
2265         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2266                 struct hammer2_recovery_elm *elm;
2267
2268                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2269                 elm->chain = parent;
2270                 elm->sync_tid = sync_tid;
2271                 hammer2_chain_ref(parent);
2272                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2273                 /* unlocked by caller */
2274
2275                 return(0);
2276         }
2277
2278
2279         /*
2280          * Recursive scan of the last flushed transaction only.  We are
2281          * doing this without pmp assignments so don't leave the chains
2282          * hanging around after we are done with them.
2283          *
2284          * error        Cumulative error this level only
2285          * rup_error    Cumulative error for recursion
2286          * tmp_error    Specific non-cumulative recursion error
2287          */
2288         chain = NULL;
2289         first = 1;
2290         rup_error = 0;
2291         error = 0;
2292
2293         for (;;) {
2294                 error |= hammer2_chain_scan(parent, &chain, &bref,
2295                                             &first,
2296                                             HAMMER2_LOOKUP_NODATA);
2297
2298                 /*
2299                  * Problem during scan or EOF
2300                  */
2301                 if (error)
2302                         break;
2303
2304                 /*
2305                  * If this is a leaf
2306                  */
2307                 if (chain == NULL) {
2308                         if (bref.mirror_tid > sync_tid) {
2309                                 hammer2_freemap_adjust(hmp, &bref,
2310                                                      HAMMER2_FREEMAP_DORECOVER);
2311                         }
2312                         continue;
2313                 }
2314
2315                 /*
2316                  * This may or may not be a recursive node.
2317                  */
2318                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2319                 if (bref.mirror_tid > sync_tid) {
2320                         ++info->depth;
2321                         tmp_error = hammer2_recovery_scan(hmp, chain,
2322                                                            info, sync_tid);
2323                         --info->depth;
2324                 } else {
2325                         tmp_error = 0;
2326                 }
2327
2328                 /*
2329                  * Flush the recovery at the PFS boundary to stage it for
2330                  * the final flush of the super-root topology.
2331                  */
2332                 if (tmp_error == 0 &&
2333                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2334                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2335                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2336                                              HAMMER2_FLUSH_ALL);
2337                 }
2338                 rup_error |= tmp_error;
2339         }
2340         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2341 }
2342
2343 /*
2344  * This fixes up an error introduced in earlier H2 implementations where
2345  * moving a PFS inode into an indirect block wound up causing the
2346  * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2347  */
2348 static
2349 int
2350 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2351 {
2352         const hammer2_inode_data_t *ripdata;
2353         hammer2_chain_t *parent;
2354         hammer2_chain_t *chain;
2355         hammer2_key_t key_next;
2356         hammer2_pfs_t *spmp;
2357         int error;
2358
2359         error = 0;
2360
2361         /*
2362          * Lookup mount point under the media-localized super-root.
2363          *
2364          * cluster->pmp will incorrectly point to spmp and must be fixed
2365          * up later on.
2366          */
2367         spmp = hmp->spmp;
2368         hammer2_inode_lock(spmp->iroot, 0);
2369         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2370         chain = hammer2_chain_lookup(&parent, &key_next,
2371                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2372                                          &error, 0);
2373         while (chain) {
2374                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2375                         continue;
2376                 if (chain->error) {
2377                         kprintf("I/O error scanning PFS labels\n");
2378                         error |= chain->error;
2379                 } else if ((chain->bref.flags &
2380                             HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2381                         int error2;
2382
2383                         ripdata = &chain->data->ipdata;
2384                         hammer2_trans_init(hmp->spmp, 0);
2385                         error2 = hammer2_chain_modify(chain,
2386                                                       chain->bref.modify_tid,
2387                                                       0, 0);
2388                         if (error2 == 0) {
2389                                 kprintf("hammer2: Correct mis-flagged PFS %s\n",
2390                                         ripdata->filename);
2391                                 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2392                         } else {
2393                                 error |= error2;
2394                         }
2395                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2396                                              HAMMER2_FLUSH_ALL);
2397                         hammer2_trans_done(hmp->spmp, 0);
2398                 }
2399                 chain = hammer2_chain_next(&parent, chain, &key_next,
2400                                            key_next, HAMMER2_KEY_MAX,
2401                                            &error, 0);
2402         }
2403         if (parent) {
2404                 hammer2_chain_unlock(parent);
2405                 hammer2_chain_drop(parent);
2406         }
2407         hammer2_inode_unlock(spmp->iroot);
2408
2409         return error;
2410 }
2411
2412 /*
2413  * Sync a mount point; this is called periodically on a per-mount basis from
2414  * the filesystem syncer, and whenever a user issues a sync.
2415  */
2416 int
2417 hammer2_vfs_sync(struct mount *mp, int waitfor)
2418 {
2419         int error;
2420
2421         error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor);
2422
2423         return error;
2424 }
2425
2426 /*
2427  * Because frontend operations lock vnodes before we get a chance to
2428  * lock the related inode, we can't just acquire a vnode lock without
2429  * risking a deadlock.  The frontend may be holding a vnode lock while
2430  * also blocked on our SYNCQ flag while trying to get the inode lock.
2431  *
2432  * To deal with this situation we can check the vnode lock situation
2433  * after locking the inode and perform a work-around.
2434  */
2435 int
2436 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor)
2437 {
2438         struct mount *mp;
2439         /*hammer2_xop_flush_t *xop;*/
2440         /*struct hammer2_sync_info info;*/
2441         hammer2_inode_t *ip;
2442         hammer2_depend_t *depend;
2443         hammer2_depend_t *depend_next;
2444         struct vnode *vp;
2445         uint32_t pass2;
2446         int error;
2447         int dorestart;
2448
2449         mp = pmp->mp;
2450
2451         /*
2452          * Move all inodes on sideq to syncq.  This will clear sideq.
2453          * This should represent all flushable inodes.  These inodes
2454          * will already have refs due to being on syncq or sideq.  We
2455          * must do this all at once with the spinlock held to ensure that
2456          * all inode dependencies are part of the same flush.
2457          *
2458          * We should be able to do this asynchronously from frontend
2459          * operations because we will be locking the inodes later on
2460          * to actually flush them, and that will partition any frontend
2461          * op using the same inode.  Either it has already locked the
2462          * inode and we will block, or it has not yet locked the inode
2463          * and it will block until we are finished flushing that inode.
2464          *
2465          * When restarting, only move the inodes flagged as PASS2 from
2466          * SIDEQ to SYNCQ.  PASS2 propagation by inode_lock4() and
2467          * inode_depend() are atomic with the spin-lock.
2468          */
2469         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2470 #ifdef HAMMER2_DEBUG_SYNC
2471         kprintf("FILESYSTEM SYNC BOUNDARY\n");
2472 #endif
2473         dorestart = 0;
2474
2475         /*
2476          * Move inodes from depq to syncq, releasing the related
2477          * depend structures.
2478          */
2479 restart:
2480 #ifdef HAMMER2_DEBUG_SYNC
2481         kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart);
2482 #endif
2483         hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/);
2484         hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN);
2485
2486         /*
2487          * Move inodes from depq to syncq.  When restarting, only depq's
2488          * marked pass2 are moved.
2489          */
2490         hammer2_spin_ex(&pmp->list_spin);
2491         depend_next = TAILQ_FIRST(&pmp->depq);
2492
2493         while ((depend = depend_next) != NULL) {
2494                 depend_next = TAILQ_NEXT(depend, entry);
2495                 if (dorestart && depend->pass2 == 0)
2496                         continue;
2497                 TAILQ_FOREACH(ip, &depend->sideq, entry) {
2498                         KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ);
2499                         atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ);
2500                         atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ);
2501                         ip->depend = NULL;
2502                 }
2503                 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry);
2504                 pmp->sideq_count -= depend->count;
2505                 depend->count = 0;
2506                 depend->pass2 = 0;
2507                 TAILQ_REMOVE(&pmp->depq, depend, entry);
2508         }
2509
2510         hammer2_spin_unex(&pmp->list_spin);
2511         hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/
2512                                       HAMMER2_TRANS_WAITING);
2513         dorestart = 0;
2514
2515         /*
2516          * sideq_count may have dropped enough to allow us to unstall
2517          * the frontend.
2518          */
2519         hammer2_pfs_memory_inc(pmp);
2520         hammer2_pfs_memory_wakeup(pmp);
2521
2522         /*
2523          * Now run through all inodes on syncq.
2524          *
2525          * Flush transactions only interlock with other flush transactions.
2526          * Any conflicting frontend operations will block on the inode, but
2527          * may hold a vnode lock while doing so.
2528          */
2529         hammer2_spin_ex(&pmp->list_spin);
2530         while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) {
2531                 /*
2532                  * Remove the inode from the SYNCQ, transfer the syncq ref
2533                  * to us.  We must clear SYNCQ to allow any potential
2534                  * front-end deadlock to proceed.  We must set PASS2 so
2535                  * the dependency code knows what to do.
2536                  */
2537                 pass2 = ip->flags;
2538                 cpu_ccfence();
2539                 if (atomic_cmpset_int(&ip->flags,
2540                               pass2,
2541                               (pass2 & ~(HAMMER2_INODE_SYNCQ |
2542                                          HAMMER2_INODE_SYNCQ_WAKEUP)) |
2543                               HAMMER2_INODE_SYNCQ_PASS2) == 0) {
2544                         continue;
2545                 }
2546                 TAILQ_REMOVE(&pmp->syncq, ip, entry);
2547                 hammer2_spin_unex(&pmp->list_spin);
2548                 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP)
2549                         wakeup(&ip->flags);
2550
2551                 /*
2552                  * Relock the inode, and we inherit a ref from the above.
2553                  * We will check for a race after we acquire the vnode.
2554                  */
2555                 hammer2_mtx_ex(&ip->lock);
2556
2557                 /*
2558                  * We need the vp in order to vfsync() dirty buffers, so if
2559                  * one isn't attached we can skip it.
2560                  *
2561                  * Ordering the inode lock and then the vnode lock has the
2562                  * potential to deadlock.  If we had left SYNCQ set that could
2563                  * also deadlock us against the frontend even if we don't hold
2564                  * any locks, but the latter is not a problem now since we
2565                  * cleared it.  igetv will temporarily release the inode lock
2566                  * in a safe manner to work-around the deadlock.
2567                  *
2568                  * Unfortunately it is still possible to deadlock when the
2569                  * frontend obtains multiple inode locks, because all the
2570                  * related vnodes are already locked (nor can the vnode locks
2571                  * be released and reacquired without messing up RECLAIM and
2572                  * INACTIVE sequencing).
2573                  *
2574                  * The solution for now is to move the vp back onto SIDEQ
2575                  * and set dorestart, which will restart the flush after we
2576                  * exhaust the current SYNCQ.  Note that additional
2577                  * dependencies may build up, so we definitely need to move
2578                  * the whole SIDEQ back to SYNCQ when we restart.
2579                  */
2580                 vp = ip->vp;
2581                 if (vp) {
2582                         if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) {
2583                                 /*
2584                                  * Failed to get the vnode, requeue the inode
2585                                  * (PASS2 is already set so it will be found
2586                                  * again on the restart).
2587                                  *
2588                                  * Then unlock, possibly sleep, and retry
2589                                  * later.  We sleep if PASS2 was *previously*
2590                                  * set, before we set it again above.
2591                                  */
2592                                 vp = NULL;
2593                                 dorestart = 1;
2594 #ifdef HAMMER2_DEBUG_SYNC
2595                                 kprintf("inum %ld (sync delayed by vnode)\n",
2596                                         (long)ip->meta.inum);
2597 #endif
2598                                 hammer2_inode_delayed_sideq(ip);
2599
2600                                 hammer2_mtx_unlock(&ip->lock);
2601                                 hammer2_inode_drop(ip);
2602
2603                                 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) {
2604                                         tsleep(&dorestart, 0, "h2syndel", 2);
2605                                 }
2606                                 hammer2_spin_ex(&pmp->list_spin);
2607                                 continue;
2608                         }
2609                 } else {
2610                         vp = NULL;
2611                 }
2612
2613                 /*
2614                  * If the inode wound up on a SIDEQ again it will already be
2615                  * prepped for another PASS2.  In this situation if we flush
2616                  * it now we will just wind up flushing it again in the same
2617                  * syncer run, so we might as well not flush it now.
2618                  */
2619                 if (ip->flags & HAMMER2_INODE_SIDEQ) {
2620                         hammer2_mtx_unlock(&ip->lock);
2621                         hammer2_inode_drop(ip);
2622                         if (vp)
2623                                 vput(vp);
2624                         dorestart = 1;
2625                         hammer2_spin_ex(&pmp->list_spin);
2626                         continue;
2627                 }
2628
2629                 /*
2630                  * Ok we have the inode exclusively locked and if vp is
2631                  * not NULL that will also be exclusively locked.  Do the
2632                  * meat of the flush.
2633                  *
2634                  * vp token needed for v_rbdirty_tree check / vclrisdirty
2635                  * sequencing.  Though we hold the vnode exclusively so
2636                  * we shouldn't need to hold the token also in this case.
2637                  */
2638                 if (vp) {
2639                         vfsync(vp, MNT_WAIT, 1, NULL, NULL);
2640                         bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */
2641                 }
2642
2643                 /*
2644                  * If the inode has not yet been inserted into the tree
2645                  * we must do so.  Then sync and flush it.  The flush should
2646                  * update the parent.
2647                  */
2648                 if (ip->flags & HAMMER2_INODE_DELETING) {
2649 #ifdef HAMMER2_DEBUG_SYNC
2650                         kprintf("inum %ld destroy\n", (long)ip->meta.inum);
2651 #endif
2652                         hammer2_inode_chain_des(ip);
2653                         atomic_add_long(&hammer2_iod_inode_deletes, 1);
2654                 } else if (ip->flags & HAMMER2_INODE_CREATING) {
2655 #ifdef HAMMER2_DEBUG_SYNC
2656                         kprintf("inum %ld insert\n", (long)ip->meta.inum);
2657 #endif
2658                         hammer2_inode_chain_ins(ip);
2659                         atomic_add_long(&hammer2_iod_inode_creates, 1);
2660                 }
2661 #ifdef HAMMER2_DEBUG_SYNC
2662                 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum);
2663 #endif
2664
2665                 /*
2666                  * Because I kinda messed up the design and index the inodes
2667                  * under the root inode, along side the directory entries,
2668                  * we can't flush the inode index under the iroot until the
2669                  * end.  If we do it now we might miss effects created by
2670                  * other inodes on the SYNCQ.
2671                  *
2672                  * Do a normal (non-FSSYNC) flush instead, which allows the
2673                  * vnode code to work the same.  We don't want to force iroot
2674                  * back onto the SIDEQ, and we also don't want the flush code
2675                  * to update pfs_iroot_blocksets until the final flush later.
2676                  *
2677                  * XXX at the moment this will likely result in a double-flush
2678                  * of the iroot chain.
2679                  */
2680                 hammer2_inode_chain_sync(ip);
2681                 if (ip == pmp->iroot) {
2682                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP);
2683                 } else {
2684                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2685                                                       HAMMER2_XOP_FSSYNC);
2686                 }
2687                 if (vp) {
2688                         lwkt_gettoken(&vp->v_token);
2689                         if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2690                                           HAMMER2_INODE_RESIZED |
2691                                           HAMMER2_INODE_DIRTYDATA)) == 0 &&
2692                             RB_EMPTY(&vp->v_rbdirty_tree) &&
2693                             !bio_track_active(&vp->v_track_write)) {
2694                                 vclrisdirty(vp);
2695                         } else {
2696                                 hammer2_inode_delayed_sideq(ip);
2697                         }
2698                         lwkt_reltoken(&vp->v_token);
2699                         vput(vp);
2700                         vp = NULL;      /* safety */
2701                 }
2702                 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2);
2703                 hammer2_inode_unlock(ip);       /* unlock+drop */
2704                 /* ip pointer invalid */
2705
2706                 /*
2707                  * If the inode got dirted after we dropped our locks,
2708                  * it will have already been moved back to the SIDEQ.
2709                  */
2710                 hammer2_spin_ex(&pmp->list_spin);
2711         }
2712         hammer2_spin_unex(&pmp->list_spin);
2713         if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) {
2714 #ifdef HAMMER2_DEBUG_SYNC
2715                 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n");
2716                 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/
2717 #endif
2718                 dorestart = 1;
2719                 goto restart;
2720         }
2721 #ifdef HAMMER2_DEBUG_SYNC
2722         kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n");
2723         /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/
2724 #endif
2725
2726         /*
2727          * We have to flush the PFS root last, even if it does not appear to
2728          * be dirty, because all the inodes in the PFS are indexed under it.
2729          * The normal flushing of iroot above would only occur if directory
2730          * entries under the root were changed.
2731          *
2732          * Specifying VOLHDR will cause an additionl flush of hmp->spmp
2733          * for the media making up the cluster.
2734          */
2735         if ((ip = pmp->iroot) != NULL) {
2736                 hammer2_inode_ref(ip);
2737                 hammer2_mtx_ex(&ip->lock);
2738                 hammer2_inode_chain_sync(ip);
2739                 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2740                                               HAMMER2_XOP_FSSYNC |
2741                                               HAMMER2_XOP_VOLHDR);
2742                 hammer2_inode_unlock(ip);       /* unlock+drop */
2743         }
2744 #ifdef HAMMER2_DEBUG_SYNC
2745         kprintf("FILESYSTEM SYNC STAGE 2 DONE\n");
2746 #endif
2747
2748         /*
2749          * device bioq sync
2750          */
2751         hammer2_bioq_sync(pmp);
2752
2753 #if 0
2754         info.pass = 1;
2755         info.waitfor = MNT_WAIT;
2756         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2757
2758         info.pass = 2;
2759         info.waitfor = MNT_WAIT;
2760         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2761 #endif
2762 #if 0
2763         /*
2764          * Generally speaking we now want to flush the media topology from
2765          * the iroot through to the inodes.  The flush stops at any inode
2766          * boundary, which allows the frontend to continue running concurrent
2767          * modifying operations on inodes (including kernel flushes of
2768          * buffers) without interfering with the main sync.
2769          *
2770          * Use the XOP interface to concurrently flush all nodes to
2771          * synchronize the PFSROOT subtopology to the media.  A standard
2772          * end-of-scan ENOENT error indicates cluster sufficiency.
2773          *
2774          * Note that this flush will not be visible on crash recovery until
2775          * we flush the super-root topology in the next loop.
2776          *
2777          * XXX For now wait for all flushes to complete.
2778          */
2779         if (mp && (ip = pmp->iroot) != NULL) {
2780                 /*
2781                  * If unmounting try to flush everything including any
2782                  * sub-trees under inodes, just in case there is dangling
2783                  * modified data, as a safety.  Otherwise just flush up to
2784                  * the inodes in this stage.
2785                  */
2786                 kprintf("MP & IROOT\n");
2787 #ifdef HAMMER2_DEBUG_SYNC
2788                 kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n");
2789 #endif
2790                 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2791                         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2792                                                     HAMMER2_XOP_VOLHDR |
2793                                                     HAMMER2_XOP_FSSYNC |
2794                                                     HAMMER2_XOP_INODE_STOP);
2795                 } else {
2796                         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2797                                                     HAMMER2_XOP_INODE_STOP |
2798                                                     HAMMER2_XOP_VOLHDR |
2799                                                     HAMMER2_XOP_FSSYNC |
2800                                                     HAMMER2_XOP_INODE_STOP);
2801                 }
2802                 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc);
2803                 error = hammer2_xop_collect(&xop->head,
2804                                             HAMMER2_XOP_COLLECT_WAITALL);
2805                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2806 #ifdef HAMMER2_DEBUG_SYNC
2807                 kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n");
2808 #endif
2809                 if (error == HAMMER2_ERROR_ENOENT)
2810                         error = 0;
2811                 else
2812                         error = hammer2_error_to_errno(error);
2813         } else {
2814                 error = 0;
2815         }
2816 #endif
2817         error = 0;      /* XXX */
2818         hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH);
2819
2820         return (error);
2821 }
2822
2823 static
2824 int
2825 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2826 {
2827         hammer2_inode_t *ip;
2828
2829         KKASSERT(MAXFIDSZ >= 16);
2830         ip = VTOI(vp);
2831         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2832         fhp->fid_ext = 0;
2833         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2834         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2835
2836         return 0;
2837 }
2838
2839 static
2840 int
2841 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2842                struct fid *fhp, struct vnode **vpp)
2843 {
2844         hammer2_pfs_t *pmp;
2845         hammer2_tid_t inum;
2846         int error;
2847
2848         pmp = MPTOPMP(mp);
2849         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2850         if (vpp) {
2851                 if (inum == 1)
2852                         error = hammer2_vfs_root(mp, vpp);
2853                 else
2854                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2855         } else {
2856                 error = 0;
2857         }
2858         if (error)
2859                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2860         return error;
2861 }
2862
2863 static
2864 int
2865 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2866                  int *exflagsp, struct ucred **credanonp)
2867 {
2868         hammer2_pfs_t *pmp;
2869         struct netcred *np;
2870         int error;
2871
2872         pmp = MPTOPMP(mp);
2873         np = vfs_export_lookup(mp, &pmp->export, nam);
2874         if (np) {
2875                 *exflagsp = np->netc_exflags;
2876                 *credanonp = &np->netc_anon;
2877                 error = 0;
2878         } else {
2879                 error = EACCES;
2880         }
2881         return error;
2882 }
2883
2884 /*
2885  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2886  * header into the HMP
2887  *
2888  * XXX read four volhdrs and use the one with the highest TID whos CRC
2889  *     matches.
2890  *
2891  * XXX check iCRCs.
2892  *
2893  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2894  *     nonexistant locations.
2895  *
2896  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2897  */
2898 static
2899 int
2900 hammer2_install_volume_header(hammer2_dev_t *hmp)
2901 {
2902         hammer2_volume_data_t *vd;
2903         struct buf *bp;
2904         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2905         int error_reported;
2906         int error;
2907         int valid;
2908         int i;
2909
2910         error_reported = 0;
2911         error = 0;
2912         valid = 0;
2913         bp = NULL;
2914
2915         /*
2916          * There are up to 4 copies of the volume header (syncs iterate
2917          * between them so there is no single master).  We don't trust the
2918          * volu_size field so we don't know precisely how large the filesystem
2919          * is, so depend on the OS to return an error if we go beyond the
2920          * block device's EOF.
2921          */
2922         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2923                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2924                               HAMMER2_VOLUME_BYTES, &bp);
2925                 if (error) {
2926                         brelse(bp);
2927                         bp = NULL;
2928                         continue;
2929                 }
2930
2931                 vd = (struct hammer2_volume_data *) bp->b_data;
2932                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2933                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2934                         brelse(bp);
2935                         bp = NULL;
2936                         continue;
2937                 }
2938
2939                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2940                         /* XXX: Reversed-endianness filesystem */
2941                         kprintf("hammer2: reverse-endian filesystem detected");
2942                         brelse(bp);
2943                         bp = NULL;
2944                         continue;
2945                 }
2946
2947                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2948                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2949                                       HAMMER2_VOLUME_ICRC0_SIZE);
2950                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2951                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2952                                        HAMMER2_VOLUME_ICRC1_SIZE);
2953                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2954                         kprintf("hammer2 volume header crc "
2955                                 "mismatch copy #%d %08x/%08x\n",
2956                                 i, crc0, crc);
2957                         error_reported = 1;
2958                         brelse(bp);
2959                         bp = NULL;
2960                         continue;
2961                 }
2962                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2963                         valid = 1;
2964                         hmp->voldata = *vd;
2965                         hmp->volhdrno = i;
2966                 }
2967                 brelse(bp);
2968                 bp = NULL;
2969         }
2970         if (valid) {
2971                 hmp->volsync = hmp->voldata;
2972                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2973                 error = 0;
2974                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2975                         kprintf("hammer2: using volume header #%d\n",
2976                                 hmp->volhdrno);
2977                 }
2978         } else {
2979                 error = EINVAL;
2980                 kprintf("hammer2: no valid volume headers found!\n");
2981         }
2982         return (error);
2983 }
2984
2985 /*
2986  * This handles hysteresis on regular file flushes.  Because the BIOs are
2987  * routed to a thread it is possible for an excessive number to build up
2988  * and cause long front-end stalls long before the runningbuffspace limit
2989  * is hit, so we implement hammer2_flush_pipe to control the
2990  * hysteresis.
2991  *
2992  * This is a particular problem when compression is used.
2993  */
2994 void
2995 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2996 {
2997         atomic_add_int(&pmp->count_lwinprog, 1);
2998 }
2999
3000 void
3001 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
3002 {
3003         int lwinprog;
3004
3005         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
3006         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
3007             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
3008                 atomic_clear_int(&pmp->count_lwinprog,
3009                                  HAMMER2_LWINPROG_WAITING);
3010                 wakeup(&pmp->count_lwinprog);
3011         }
3012         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
3013             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
3014                 atomic_clear_int(&pmp->count_lwinprog,
3015                                  HAMMER2_LWINPROG_WAITING0);
3016                 wakeup(&pmp->count_lwinprog);
3017         }
3018 }
3019
3020 void
3021 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
3022 {
3023         int lwinprog;
3024         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
3025                                     HAMMER2_LWINPROG_WAITING0;
3026
3027         for (;;) {
3028                 lwinprog = pmp->count_lwinprog;
3029                 cpu_ccfence();
3030                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3031                         break;
3032                 tsleep_interlock(&pmp->count_lwinprog, 0);
3033                 atomic_set_int(&pmp->count_lwinprog, lwflag);
3034                 lwinprog = pmp->count_lwinprog;
3035                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3036                         break;
3037                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
3038         }
3039 }
3040
3041 /*
3042  * It is possible for an excessive number of dirty chains or dirty inodes
3043  * to build up.  When this occurs we start an asynchronous filesystem sync.
3044  * If the level continues to build up, we stall, waiting for it to drop,
3045  * with some hysteresis.
3046  *
3047  * We limit the stall to two seconds per call.
3048  *
3049  * This relies on the kernel calling hammer2_vfs_modifying() prior to
3050  * obtaining any vnode locks before making a modifying VOP call.
3051  */
3052 static void
3053 hammer2_vfs_modifying(struct mount *mp)
3054 {
3055         hammer2_pfs_memory_wait(MPTOPMP(mp));
3056 }
3057
3058 /*
3059  * Initiate an asynchronous filesystem sync and, with hysteresis,
3060  * stall if the internal data structure count becomes too bloated.
3061  */
3062 void
3063 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
3064 {
3065         uint32_t waiting;
3066         int loops;
3067
3068         if (pmp == NULL || pmp->mp == NULL)
3069                 return;
3070
3071         for (loops = 0; loops < 2; ++loops) {
3072                 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK;
3073                 cpu_ccfence();
3074
3075                 /*
3076                  * Start the syncer running at 1/2 the limit
3077                  */
3078                 if (waiting > hammer2_limit_dirty_chains / 2 ||
3079                     pmp->sideq_count > hammer2_limit_dirty_inodes / 2) {
3080                         trigger_syncer(pmp->mp);
3081                 }
3082
3083                 /*
3084                  * Stall at the limit waiting for the counts to drop.
3085                  * This code will typically be woken up once the count
3086                  * drops below 3/4 the limit, or in one second.
3087                  */
3088                 if (waiting < hammer2_limit_dirty_chains &&
3089                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
3090                         break;
3091                 }
3092                 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
3093                 atomic_set_int(&pmp->inmem_dirty_chains,
3094                                HAMMER2_DIRTYCHAIN_WAITING);
3095                 if (waiting < hammer2_limit_dirty_chains &&
3096                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
3097                         break;
3098                 }
3099                 trigger_syncer(pmp->mp);
3100                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED, "h2memw", hz);
3101 #if 0
3102                 limit = pmp->mp->mnt_nvnodelistsize / 10;
3103                 if (limit < hammer2_limit_dirty_chains)
3104                         limit = hammer2_limit_dirty_chains;
3105                 if (limit < 1000)
3106                         limit = 1000;
3107 #endif
3108         }
3109 }
3110
3111 void
3112 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
3113 {
3114         if (pmp) {
3115                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
3116         }
3117 }
3118
3119 void
3120 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
3121 {
3122         uint32_t waiting;
3123
3124         if (pmp) {
3125                 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
3126                 /* don't need --waiting to test flag */
3127
3128                 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) &&
3129                     (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <=
3130                     hammer2_limit_dirty_chains * 2 / 3 &&
3131                     pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) {
3132                         atomic_clear_int(&pmp->inmem_dirty_chains,
3133                                          HAMMER2_DIRTYCHAIN_WAITING);
3134                         wakeup(&pmp->inmem_dirty_chains);
3135                 }
3136         }
3137 }
3138
3139 /*
3140  * Returns 0 if the filesystem has tons of free space
3141  * Returns 1 if the filesystem has less than 10% remaining
3142  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
3143  */
3144 int
3145 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
3146 {
3147         hammer2_pfs_t *pmp;
3148         hammer2_dev_t *hmp;
3149         hammer2_off_t free_reserved;
3150         hammer2_off_t free_nominal;
3151         int i;
3152
3153         pmp = ip->pmp;
3154
3155         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
3156                 free_reserved = HAMMER2_SEGSIZE;
3157                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
3158                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
3159                         hmp = pmp->pfs_hmps[i];
3160                         if (hmp == NULL)
3161                                 continue;
3162                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
3163                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
3164                                 continue;
3165
3166                         if (free_nominal > hmp->voldata.allocator_free)
3167                                 free_nominal = hmp->voldata.allocator_free;
3168                         if (free_reserved < hmp->free_reserved)
3169                                 free_reserved = hmp->free_reserved;
3170                 }
3171
3172                 /*
3173                  * SMP races ok
3174                  */
3175                 pmp->free_reserved = free_reserved;
3176                 pmp->free_nominal = free_nominal;
3177                 pmp->free_ticks = ticks;
3178         } else {
3179                 free_reserved = pmp->free_reserved;
3180                 free_nominal = pmp->free_nominal;
3181         }
3182         if (cred && cred->cr_uid != 0) {
3183                 if ((int64_t)(free_nominal - bytes) <
3184                     (int64_t)free_reserved) {
3185                         return 2;
3186                 }
3187         } else {
3188                 if ((int64_t)(free_nominal - bytes) <
3189                     (int64_t)free_reserved / 2) {
3190                         return 2;
3191                 }
3192         }
3193         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
3194                 return 1;
3195         return 0;
3196 }
3197
3198 /*
3199  * Debugging
3200  */
3201 void
3202 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
3203                    u_int flags)
3204 {
3205         hammer2_chain_t *scan;
3206         hammer2_chain_t *parent;
3207
3208         --*countp;
3209         if (*countp == 0) {
3210                 kprintf("%*.*s...\n", tab, tab, "");
3211                 return;
3212         }
3213         if (*countp < 0)
3214                 return;
3215         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
3216                 tab, tab, "", pfx,
3217                 chain, chain->bref.type,
3218                 chain->bref.key, chain->bref.keybits,
3219                 chain->bref.mirror_tid);
3220
3221         kprintf("%*.*s      [%08x] (%s) refs=%d",
3222                 tab, tab, "",
3223                 chain->flags,
3224                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3225                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
3226                 chain->refs);
3227
3228         parent = chain->parent;
3229         if (parent)
3230                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
3231                         tab, tab, "",
3232                         parent, parent->flags, parent->refs);
3233         if (RB_EMPTY(&chain->core.rbtree)) {
3234                 kprintf("\n");
3235         } else {
3236                 kprintf(" {\n");
3237                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3238                         if ((scan->flags & flags) || flags == (u_int)-1) {
3239                                 hammer2_dump_chain(scan, tab + 4, countp, 'a',
3240                                                    flags);
3241                         }
3242                 }
3243                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3244                         kprintf("%*.*s}(%s)\n", tab, tab, "",
3245                                 chain->data->ipdata.filename);
3246                 else
3247                         kprintf("%*.*s}\n", tab, tab, "");
3248         }
3249 }