sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72 };
  73
  74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  75 static struct hammer2_mntlist hammer2_mntlist;
  76
  77 struct hammer2_pfslist hammer2_pfslist;
  78 struct lock hammer2_mntlk;
  79
  80 int hammer2_debug;
  81 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
  82 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
  83 int hammer2_dedup_enable = 1;
  84 int hammer2_always_compress = 0;        /* always try to compress */
  85 int hammer2_inval_enable = 0;
  86 int hammer2_flush_pipe = 100;
  87 int hammer2_synchronous_flush = 1;
  88 int hammer2_dio_count;
  89 int hammer2_limit_dio = 256;
  90 int hammer2_bulkfree_tps = 5000;
  91 long hammer2_chain_allocs;
  92 long hammer2_chain_frees;
  93 long hammer2_limit_dirty_chains;
  94 long hammer2_count_modified_chains;
  95 long hammer2_iod_invals;
  96 long hammer2_iod_file_read;
  97 long hammer2_iod_meta_read;
  98 long hammer2_iod_indr_read;
  99 long hammer2_iod_fmap_read;
 100 long hammer2_iod_volu_read;
 101 long hammer2_iod_file_write;
 102 long hammer2_iod_file_wembed;
 103 long hammer2_iod_file_wzero;
 104 long hammer2_iod_file_wdedup;
 105 long hammer2_iod_meta_write;
 106 long hammer2_iod_indr_write;
 107 long hammer2_iod_fmap_write;
 108 long hammer2_iod_volu_write;
 109
 110 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 111 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 112                 "Buffer used for compression.");
 113
 114 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 115 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 116                 "Buffer used for decompression.");
 117
 118 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 119
 120 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 121            &hammer2_debug, 0, "");
 122 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 123            &hammer2_cluster_meta_read, 0, "");
 124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 125            &hammer2_cluster_data_read, 0, "");
 126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 127            &hammer2_dedup_enable, 0, "");
 128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 129            &hammer2_always_compress, 0, "");
 130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 131            &hammer2_inval_enable, 0, "");
 132 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 133            &hammer2_flush_pipe, 0, "");
 134 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 135            &hammer2_synchronous_flush, 0, "");
 136 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
 137            &hammer2_bulkfree_tps, 0, "");
 138 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 139            &hammer2_chain_allocs, 0, "");
 140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 141            &hammer2_chain_frees, 0, "");
 142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 143            &hammer2_limit_dirty_chains, 0, "");
 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 145            &hammer2_count_modified_chains, 0, "");
 146 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 147            &hammer2_dio_count, 0, "");
 148 SYSCTL_INT(_vfs_hammer2, OID_AUTO, limit_dio, CTLFLAG_RW,
 149            &hammer2_limit_dio, 0, "");
 150
 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 152            &hammer2_iod_invals, 0, "");
 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 154            &hammer2_iod_file_read, 0, "");
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 156            &hammer2_iod_meta_read, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 158            &hammer2_iod_indr_read, 0, "");
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 160            &hammer2_iod_fmap_read, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 162            &hammer2_iod_volu_read, 0, "");
 163
 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 165            &hammer2_iod_file_write, 0, "");
 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 167            &hammer2_iod_file_wembed, 0, "");
 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 169            &hammer2_iod_file_wzero, 0, "");
 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 171            &hammer2_iod_file_wdedup, 0, "");
 172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 173            &hammer2_iod_meta_write, 0, "");
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 175            &hammer2_iod_indr_write, 0, "");
 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 177            &hammer2_iod_fmap_write, 0, "");
 178 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 179            &hammer2_iod_volu_write, 0, "");
 180
 181 long hammer2_check_icrc32;
 182 long hammer2_check_xxhash64;
 183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 184            &hammer2_check_icrc32, 0, "");
 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 186            &hammer2_check_xxhash64, 0, "");
 187
 188 static int hammer2_vfs_init(struct vfsconf *conf);
 189 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 190 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 191                                 struct ucred *cred);
 192 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 193                                 struct vnode *, struct ucred *);
 194 static int hammer2_recovery(hammer2_dev_t *hmp);
 195 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 196 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 197 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 198                                 struct ucred *cred);
 199 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 200                                 struct ucred *cred);
 201 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 202                                 struct fid *fhp, struct vnode **vpp);
 203 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 204 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 205                                 int *exflagsp, struct ucred **credanonp);
 206
 207 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 208 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 209
 210 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 211
 212 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 213 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 214                                 hammer2_dev_t *hmp);
 215
 216 /*
 217  * HAMMER2 vfs operations.
 218  */
 219 static struct vfsops hammer2_vfsops = {
 220         .vfs_init       = hammer2_vfs_init,
 221         .vfs_uninit     = hammer2_vfs_uninit,
 222         .vfs_sync       = hammer2_vfs_sync,
 223         .vfs_mount      = hammer2_vfs_mount,
 224         .vfs_unmount    = hammer2_vfs_unmount,
 225         .vfs_root       = hammer2_vfs_root,
 226         .vfs_statfs     = hammer2_vfs_statfs,
 227         .vfs_statvfs    = hammer2_vfs_statvfs,
 228         .vfs_vget       = hammer2_vfs_vget,
 229         .vfs_vptofh     = hammer2_vfs_vptofh,
 230         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 231         .vfs_checkexp   = hammer2_vfs_checkexp
 232 };
 233
 234 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 235
 236 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 237 MODULE_VERSION(hammer2, 1);
 238
 239 static
 240 int
 241 hammer2_vfs_init(struct vfsconf *conf)
 242 {
 243         static struct objcache_malloc_args margs_read;
 244         static struct objcache_malloc_args margs_write;
 245         static struct objcache_malloc_args margs_vop;
 246
 247         int error;
 248
 249         error = 0;
 250
 251         /*
 252          * A large DIO cache is needed to retain dedup enablement masks.
 253          * The bulkfree code clears related masks as part of the disk block
 254          * recycling algorithm, preventing it from being used for a later
 255          * dedup.
 256          *
 257          * NOTE: A large buffer cache can actually interfere with dedup
 258          *       operation because we dedup based on media physical buffers
 259          *       and not logical buffers.  Try to make the DIO chace large
 260          *       enough to avoid this problem, but also cap it.
 261          */
 262         hammer2_limit_dio = nbuf * 2;
 263         if (hammer2_limit_dio > 100000)
 264                 hammer2_limit_dio = 100000;
 265
 266         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 267                 error = EINVAL;
 268         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 269                 error = EINVAL;
 270         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 271                 error = EINVAL;
 272
 273         if (error)
 274                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 275
 276         margs_read.objsize = 65536;
 277         margs_read.mtype = M_HAMMER2_DEBUFFER;
 278
 279         margs_write.objsize = 32768;
 280         margs_write.mtype = M_HAMMER2_CBUFFER;
 281
 282         margs_vop.objsize = sizeof(hammer2_xop_t);
 283         margs_vop.mtype = M_HAMMER2;
 284
 285         /*
 286          * Note thaht for the XOPS cache we want backing store allocations
 287          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 288          * confusion), so use the backing store function that does it.  This
 289          * means that initial XOPS objects are zerod but REUSED objects are
 290          * not.  So we are responsible for cleaning the object up sufficiently
 291          * for our needs before objcache_put()ing it back (typically just the
 292          * FIFO indices).
 293          */
 294         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 295                                 0, 1, NULL, NULL, NULL,
 296                                 objcache_malloc_alloc,
 297                                 objcache_malloc_free,
 298                                 &margs_read);
 299         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 300                                 0, 1, NULL, NULL, NULL,
 301                                 objcache_malloc_alloc,
 302                                 objcache_malloc_free,
 303                                 &margs_write);
 304         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 305                                 0, 1, NULL, NULL, NULL,
 306                                 objcache_malloc_alloc_zero,
 307                                 objcache_malloc_free,
 308                                 &margs_vop);
 309
 310
 311         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 312         TAILQ_INIT(&hammer2_mntlist);
 313         TAILQ_INIT(&hammer2_pfslist);
 314
 315         hammer2_limit_dirty_chains = maxvnodes / 10;
 316         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 317                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 318
 319         return (error);
 320 }
 321
 322 static
 323 int
 324 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 325 {
 326         objcache_destroy(cache_buffer_read);
 327         objcache_destroy(cache_buffer_write);
 328         objcache_destroy(cache_xops);
 329         return 0;
 330 }
 331
 332 /*
 333  * Core PFS allocator.  Used to allocate or reference the pmp structure
 334  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 335  * The pmp can be passed in or loaded by this function using the chain and
 336  * inode data.
 337  *
 338  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 339  * transactions.  Note that synchronization does not use this field.
 340  * (typically frontend operations and synchronization cannot run on the
 341  * same PFS node at the same time).
 342  *
 343  * XXX check locking
 344  */
 345 hammer2_pfs_t *
 346 hammer2_pfsalloc(hammer2_chain_t *chain,
 347                  const hammer2_inode_data_t *ripdata,
 348                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 349 {
 350         hammer2_pfs_t *pmp;
 351         hammer2_inode_t *iroot;
 352         int count;
 353         int i;
 354         int j;
 355
 356         pmp = NULL;
 357
 358         /*
 359          * Locate or create the PFS based on the cluster id.  If ripdata
 360          * is NULL this is a spmp which is unique and is always allocated.
 361          *
 362          * If the device is mounted in local mode all PFSs are considered
 363          * independent and not part of any cluster (for debugging only).
 364          */
 365         if (ripdata) {
 366                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 367                         if (force_local != pmp->force_local)
 368                                 continue;
 369                         if (force_local == NULL &&
 370                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 371                                  sizeof(pmp->pfs_clid)) == 0) {
 372                                         break;
 373                         } else if (force_local && pmp->pfs_names[0] &&
 374                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 375                                         break;
 376                         }
 377                 }
 378         }
 379
 380         if (pmp == NULL) {
 381                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 382                 pmp->force_local = force_local;
 383                 hammer2_trans_manage_init(pmp);
 384                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 385                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 386                 lockinit(&pmp->lock, "pfslk", 0, 0);
 387                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 388                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 389                 spin_init(&pmp->xop_spin, "h2xop");
 390                 spin_init(&pmp->lru_spin, "h2lru");
 391                 RB_INIT(&pmp->inum_tree);
 392                 TAILQ_INIT(&pmp->sideq);
 393                 TAILQ_INIT(&pmp->lru_list);
 394                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 395
 396                 /*
 397                  * Distribute backend operations to threads
 398                  */
 399                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 400                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 401
 402                 /*
 403                  * Save the last media transaction id for the flusher.  Set
 404                  * initial
 405                  */
 406                 if (ripdata)
 407                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 408                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 409
 410                 /*
 411                  * The synchronization thread may start too early, make
 412                  * sure it stays frozen until we are ready to let it go.
 413                  * XXX
 414                  */
 415                 /*
 416                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 417                                          HAMMER2_THREAD_REMASTER;
 418                 */
 419         }
 420
 421         /*
 422          * Create the PFS's root inode and any missing XOP helper threads.
 423          */
 424         if ((iroot = pmp->iroot) == NULL) {
 425                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 426                 if (ripdata)
 427                         iroot->meta = ripdata->meta;
 428                 pmp->iroot = iroot;
 429                 hammer2_inode_ref(iroot);
 430                 hammer2_inode_unlock(iroot);
 431         }
 432
 433         /*
 434          * Stop here if no chain is passed in.
 435          */
 436         if (chain == NULL)
 437                 goto done;
 438
 439         /*
 440          * When a chain is passed in we must add it to the PFS's root
 441          * inode, update pmp->pfs_types[], and update the syncronization
 442          * threads.
 443          *
 444          * When forcing local mode, mark the PFS as a MASTER regardless.
 445          *
 446          * At the moment empty spots can develop due to removals or failures.
 447          * Ultimately we want to re-fill these spots but doing so might
 448          * confused running code. XXX
 449          */
 450         hammer2_inode_ref(iroot);
 451         hammer2_mtx_ex(&iroot->lock);
 452         j = iroot->cluster.nchains;
 453
 454         if (j == HAMMER2_MAXCLUSTER) {
 455                 kprintf("hammer2_mount: cluster full!\n");
 456                 /* XXX fatal error? */
 457         } else {
 458                 KKASSERT(chain->pmp == NULL);
 459                 chain->pmp = pmp;
 460                 hammer2_chain_ref(chain);
 461                 iroot->cluster.array[j].chain = chain;
 462                 if (force_local)
 463                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 464                 else
 465                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 466                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 467                 pmp->pfs_hmps[j] = chain->hmp;
 468
 469                 /*
 470                  * If the PFS is already mounted we must account
 471                  * for the mount_count here.
 472                  */
 473                 if (pmp->mp)
 474                         ++chain->hmp->mount_count;
 475
 476                 /*
 477                  * May have to fixup dirty chain tracking.  Previous
 478                  * pmp was NULL so nothing to undo.
 479                  */
 480                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 481                         hammer2_pfs_memory_inc(pmp);
 482                 ++j;
 483         }
 484         iroot->cluster.nchains = j;
 485
 486         /*
 487          * Update nmasters from any PFS inode which is part of the cluster.
 488          * It is possible that this will result in a value which is too
 489          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 490          * override this value later on.
 491          *
 492          * (This informs us of masters that might not currently be
 493          *  discoverable by this mount).
 494          */
 495         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 496                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 497         }
 498
 499         /*
 500          * Count visible masters.  Masters are usually added with
 501          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 502          * are more (XXX and must update the master inodes).
 503          */
 504         count = 0;
 505         for (i = 0; i < iroot->cluster.nchains; ++i) {
 506                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 507                         ++count;
 508         }
 509         if (pmp->pfs_nmasters < count)
 510                 pmp->pfs_nmasters = count;
 511
 512         /*
 513          * Create missing synchronization and support threads.
 514          *
 515          * Single-node masters (including snapshots) have nothing to
 516          * synchronize and do not require this thread.
 517          *
 518          * Multi-node masters or any number of soft masters, slaves, copy,
 519          * or other PFS types need the thread.
 520          *
 521          * Each thread is responsible for its particular cluster index.
 522          * We use independent threads so stalls or mismatches related to
 523          * any given target do not affect other targets.
 524          */
 525         for (i = 0; i < iroot->cluster.nchains; ++i) {
 526                 /*
 527                  * Single-node masters (including snapshots) have nothing
 528                  * to synchronize and will make direct xops support calls,
 529                  * thus they do not require this thread.
 530                  *
 531                  * Note that there can be thousands of snapshots.  We do not
 532                  * want to create thousands of threads.
 533                  */
 534                 if (pmp->pfs_nmasters <= 1 &&
 535                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 536                         continue;
 537                 }
 538
 539                 /*
 540                  * Sync support thread
 541                  */
 542                 if (pmp->sync_thrs[i].td == NULL) {
 543                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 544                                            "h2nod", i, -1,
 545                                            hammer2_primary_sync_thread);
 546                 }
 547         }
 548
 549         /*
 550          * Create missing Xop threads
 551          *
 552          * NOTE: We create helper threads for all mounted PFSs or any
 553          *       PFSs with 2+ nodes (so the sync thread can update them,
 554          *       even if not mounted).
 555          */
 556         if (pmp->mp || iroot->cluster.nchains >= 2)
 557                 hammer2_xop_helper_create(pmp);
 558
 559         hammer2_mtx_unlock(&iroot->lock);
 560         hammer2_inode_drop(iroot);
 561 done:
 562         return pmp;
 563 }
 564
 565 /*
 566  * Deallocate an element of a probed PFS.  If destroying and this is a
 567  * MASTER, adjust nmasters.
 568  *
 569  * This function does not physically destroy the PFS element in its device
 570  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 571  */
 572 void
 573 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 574 {
 575         hammer2_inode_t *iroot;
 576         hammer2_chain_t *chain;
 577         int j;
 578
 579         /*
 580          * Cleanup our reference on iroot.  iroot is (should) not be needed
 581          * by the flush code.
 582          */
 583         iroot = pmp->iroot;
 584         if (iroot) {
 585                 /*
 586                  * Stop synchronizing
 587                  *
 588                  * XXX flush after acquiring the iroot lock.
 589                  * XXX clean out the cluster index from all inode structures.
 590                  */
 591                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 592
 593                 /*
 594                  * Remove the cluster index from the group.  If destroying
 595                  * the PFS and this is a master, adjust pfs_nmasters.
 596                  */
 597                 hammer2_mtx_ex(&iroot->lock);
 598                 chain = iroot->cluster.array[clindex].chain;
 599                 iroot->cluster.array[clindex].chain = NULL;
 600
 601                 switch(pmp->pfs_types[clindex]) {
 602                 case HAMMER2_PFSTYPE_MASTER:
 603                         if (destroying && pmp->pfs_nmasters > 0)
 604                                 --pmp->pfs_nmasters;
 605                         /* XXX adjust ripdata->meta.pfs_nmasters */
 606                         break;
 607                 default:
 608                         break;
 609                 }
 610                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 611
 612                 hammer2_mtx_unlock(&iroot->lock);
 613
 614                 /*
 615                  * Release the chain.
 616                  */
 617                 if (chain) {
 618                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 619                         hammer2_chain_drop(chain);
 620                 }
 621
 622                 /*
 623                  * Terminate all XOP threads for the cluster index.
 624                  */
 625                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 626                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 627         }
 628 }
 629
 630 /*
 631  * Destroy a PFS, typically only occurs after the last mount on a device
 632  * has gone away.
 633  */
 634 static void
 635 hammer2_pfsfree(hammer2_pfs_t *pmp)
 636 {
 637         hammer2_inode_t *iroot;
 638         hammer2_chain_t *chain;
 639         int i;
 640         int j;
 641
 642         /*
 643          * Cleanup our reference on iroot.  iroot is (should) not be needed
 644          * by the flush code.
 645          */
 646         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 647
 648         iroot = pmp->iroot;
 649         if (iroot) {
 650                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 651                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 652                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 653                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 654                 }
 655 #if REPORT_REFS_ERRORS
 656                 if (pmp->iroot->refs != 1)
 657                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 658                                 pmp->iroot, pmp->iroot->refs);
 659 #else
 660                 KKASSERT(pmp->iroot->refs == 1);
 661 #endif
 662                 /* ref for pmp->iroot */
 663                 hammer2_inode_drop(pmp->iroot);
 664                 pmp->iroot = NULL;
 665         }
 666
 667         /*
 668          * Cleanup chains remaining on LRU list.
 669          */
 670         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 671                 hammer2_chain_ref(chain);
 672                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 673                 hammer2_chain_drop(chain);
 674         }
 675
 676         /*
 677          * Free remaining pmp resources
 678          */
 679         kmalloc_destroy(&pmp->mmsg);
 680         kmalloc_destroy(&pmp->minode);
 681
 682         kfree(pmp, M_HAMMER2);
 683 }
 684
 685 /*
 686  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 687  * empty is terminated and freed.
 688  *
 689  * XXX inefficient.
 690  */
 691 static void
 692 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 693 {
 694         hammer2_pfs_t *pmp;
 695         hammer2_inode_t *iroot;
 696         hammer2_chain_t *rchain;
 697         int didfreeze;
 698         int i;
 699         int j;
 700
 701 again:
 702         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 703                 if ((iroot = pmp->iroot) == NULL)
 704                         continue;
 705                 if (hmp->spmp == pmp)
 706                         hmp->spmp = NULL;
 707
 708                 /*
 709                  * Determine if this PFS is affected.  If it is we must
 710                  * freeze all management threads and lock its iroot.
 711                  *
 712                  * Freezing a management thread forces it idle, operations
 713                  * in-progress will be aborted and it will have to start
 714                  * over again when unfrozen, or exit if told to exit.
 715                  */
 716                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 717                         if (pmp->pfs_hmps[i] == hmp)
 718                                 break;
 719                 }
 720                 if (i != HAMMER2_MAXCLUSTER) {
 721                         /*
 722                          * Make sure all synchronization threads are locked
 723                          * down.
 724                          */
 725                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 726                                 if (pmp->pfs_hmps[i] == NULL)
 727                                         continue;
 728                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 729                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 730                                         hammer2_thr_freeze_async(
 731                                                 &pmp->xop_groups[j].thrs[i]);
 732                                 }
 733                         }
 734                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 735                                 if (pmp->pfs_hmps[i] == NULL)
 736                                         continue;
 737                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 738                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 739                                         hammer2_thr_freeze(
 740                                                 &pmp->xop_groups[j].thrs[i]);
 741                                 }
 742                         }
 743
 744                         /*
 745                          * Lock the inode and clean out matching chains.
 746                          * Note that we cannot use hammer2_inode_lock_*()
 747                          * here because that would attempt to validate the
 748                          * cluster that we are in the middle of ripping
 749                          * apart.
 750                          *
 751                          * WARNING! We are working directly on the inodes
 752                          *          embedded cluster.
 753                          */
 754                         hammer2_mtx_ex(&iroot->lock);
 755
 756                         /*
 757                          * Remove the chain from matching elements of the PFS.
 758                          */
 759                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 760                                 if (pmp->pfs_hmps[i] != hmp)
 761                                         continue;
 762                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 763                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 764                                         hammer2_thr_delete(
 765                                                 &pmp->xop_groups[j].thrs[i]);
 766                                 }
 767                                 rchain = iroot->cluster.array[i].chain;
 768                                 iroot->cluster.array[i].chain = NULL;
 769                                 pmp->pfs_types[i] = 0;
 770                                 if (pmp->pfs_names[i]) {
 771                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 772                                         pmp->pfs_names[i] = NULL;
 773                                 }
 774                                 if (rchain) {
 775                                         hammer2_chain_drop(rchain);
 776                                         /* focus hint */
 777                                         if (iroot->cluster.focus == rchain)
 778                                                 iroot->cluster.focus = NULL;
 779                                 }
 780                                 pmp->pfs_hmps[i] = NULL;
 781                         }
 782                         hammer2_mtx_unlock(&iroot->lock);
 783                         didfreeze = 1;  /* remaster, unfreeze down below */
 784                 } else {
 785                         didfreeze = 0;
 786                 }
 787
 788                 /*
 789                  * Cleanup trailing chains.  Gaps may remain.
 790                  */
 791                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 792                         if (pmp->pfs_hmps[i])
 793                                 break;
 794                 }
 795                 iroot->cluster.nchains = i + 1;
 796
 797                 /*
 798                  * If the PMP has no elements remaining we can destroy it.
 799                  * (this will transition management threads from frozen->exit).
 800                  */
 801                 if (iroot->cluster.nchains == 0) {
 802                         hammer2_pfsfree(pmp);
 803                         goto again;
 804                 }
 805
 806                 /*
 807                  * If elements still remain we need to set the REMASTER
 808                  * flag and unfreeze it.
 809                  */
 810                 if (didfreeze) {
 811                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 812                                 if (pmp->pfs_hmps[i] == NULL)
 813                                         continue;
 814                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 815                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 816                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 817                                         hammer2_thr_remaster(
 818                                                 &pmp->xop_groups[j].thrs[i]);
 819                                         hammer2_thr_unfreeze(
 820                                                 &pmp->xop_groups[j].thrs[i]);
 821                                 }
 822                         }
 823                 }
 824         }
 825 }
 826
 827 /*
 828  * Mount or remount HAMMER2 fileystem from physical media
 829  *
 830  *      mountroot
 831  *              mp              mount point structure
 832  *              path            NULL
 833  *              data            <unused>
 834  *              cred            <unused>
 835  *
 836  *      mount
 837  *              mp              mount point structure
 838  *              path            path to mount point
 839  *              data            pointer to argument structure in user space
 840  *                      volume  volume path (device@LABEL form)
 841  *                      hflags  user mount flags
 842  *              cred            user credentials
 843  *
 844  * RETURNS:     0       Success
 845  *              !0      error number
 846  */
 847 static
 848 int
 849 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 850                   struct ucred *cred)
 851 {
 852         struct hammer2_mount_info info;
 853         hammer2_pfs_t *pmp;
 854         hammer2_pfs_t *spmp;
 855         hammer2_dev_t *hmp;
 856         hammer2_dev_t *force_local;
 857         hammer2_key_t key_next;
 858         hammer2_key_t key_dummy;
 859         hammer2_key_t lhc;
 860         struct vnode *devvp;
 861         struct nlookupdata nd;
 862         hammer2_chain_t *parent;
 863         hammer2_chain_t *chain;
 864         hammer2_cluster_t *cluster;
 865         const hammer2_inode_data_t *ripdata;
 866         hammer2_blockref_t bref;
 867         struct file *fp;
 868         char devstr[MNAMELEN];
 869         size_t size;
 870         size_t done;
 871         char *dev;
 872         char *label;
 873         int ronly = 1;
 874         int error;
 875         int i;
 876
 877         hmp = NULL;
 878         pmp = NULL;
 879         dev = NULL;
 880         label = NULL;
 881         devvp = NULL;
 882
 883         kprintf("hammer2_mount\n");
 884
 885         if (path == NULL) {
 886                 /*
 887                  * Root mount
 888                  */
 889                 bzero(&info, sizeof(info));
 890                 info.cluster_fd = -1;
 891                 ksnprintf(devstr, sizeof(devstr), "%s",
 892                           mp->mnt_stat.f_mntfromname);
 893                 kprintf("hammer2_mount: root '%s'\n", devstr);
 894         } else {
 895                 /*
 896                  * Non-root mount or updating a mount
 897                  */
 898                 error = copyin(data, &info, sizeof(info));
 899                 if (error)
 900                         return (error);
 901
 902                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 903                 if (error)
 904                         return (error);
 905         }
 906
 907         /*
 908          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 909          * if no label specified, based on the partition id.  Error out if no
 910          * label or device (with partition id) is specified.  This is strictly
 911          * a convenience to match the default label created by newfs_hammer2,
 912          * our preference is that a label always be specified.
 913          *
 914          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
 915          *       that does not specify a device, as long as some H2 label
 916          *       has already been mounted from that device.  This makes
 917          *       mounting snapshots a lot easier.
 918          */
 919         dev = devstr;
 920         label = strchr(devstr, '@');
 921         if (label && ((label + 1) - dev) > done)
 922                 return (EINVAL);
 923         if (label == NULL || label[1] == 0) {
 924                 char slice;
 925
 926                 if (label == NULL)
 927                         label = devstr + strlen(devstr);
 928                 slice = label[-1];
 929                 switch(slice) {
 930                 case 'a':
 931                         label = "BOOT";
 932                         break;
 933                 case 'd':
 934                         label = "ROOT";
 935                         break;
 936                 default:
 937                         label = "DATA";
 938                         break;
 939                 }
 940         } else {
 941                 *label = '\0';
 942                 label++;
 943         }
 944
 945         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
 946                 dev, label, (mp->mnt_flag & MNT_RDONLY));
 947
 948         if (mp->mnt_flag & MNT_UPDATE) {
 949                 /*
 950                  * Update mount.  Note that pmp->iroot->cluster is
 951                  * an inode-embedded cluster and thus cannot be
 952                  * directly locked.
 953                  *
 954                  * XXX HAMMER2 needs to implement NFS export via
 955                  *     mountctl.
 956                  */
 957                 pmp = MPTOPMP(mp);
 958                 pmp->hflags = info.hflags;
 959                 cluster = &pmp->iroot->cluster;
 960                 for (i = 0; i < cluster->nchains; ++i) {
 961                         if (cluster->array[i].chain == NULL)
 962                                 continue;
 963                         hmp = cluster->array[i].chain->hmp;
 964                         devvp = hmp->devvp;
 965                         error = hammer2_remount(hmp, mp, path,
 966                                                 devvp, cred);
 967                         if (error)
 968                                 break;
 969                 }
 970
 971                 return error;
 972         }
 973
 974         /*
 975          * HMP device mount
 976          *
 977          * If a path is specified and dev is not an empty string, lookup the
 978          * name and verify that it referes to a block device.
 979          *
 980          * If a path is specified and dev is an empty string we fall through
 981          * and locate the label in the hmp search.
 982          */
 983         if (path && *dev != 0) {
 984                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 985                 if (error == 0)
 986                         error = nlookup(&nd);
 987                 if (error == 0)
 988                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 989                 nlookup_done(&nd);
 990         } else if (path == NULL) {
 991                 /* root mount */
 992                 cdev_t cdev = kgetdiskbyname(dev);
 993                 error = bdevvp(cdev, &devvp);
 994                 if (error)
 995                         kprintf("hammer2: cannot find '%s'\n", dev);
 996         } else {
 997                 /*
 998                  * We will locate the hmp using the label in the hmp loop.
 999                  */
1000                 error = 0;
1001         }
1002
1003         /*
1004          * Make sure its a block device.  Do not check to see if it is
1005          * already mounted until we determine that its a fresh H2 device.
1006          */
1007         if (error == 0 && devvp) {
1008                 vn_isdisk(devvp, &error);
1009         }
1010
1011         /*
1012          * Determine if the device has already been mounted.  After this
1013          * check hmp will be non-NULL if we are doing the second or more
1014          * hammer2 mounts from the same device.
1015          */
1016         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1017         if (devvp) {
1018                 /*
1019                  * Match the device.  Due to the way devfs works,
1020                  * we may not be able to directly match the vnode pointer,
1021                  * so also check to see if the underlying device matches.
1022                  */
1023                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1024                         if (hmp->devvp == devvp)
1025                                 break;
1026                         if (devvp->v_rdev &&
1027                             hmp->devvp->v_rdev == devvp->v_rdev) {
1028                                 break;
1029                         }
1030                 }
1031
1032                 /*
1033                  * If no match this may be a fresh H2 mount, make sure
1034                  * the device is not mounted on anything else.
1035                  */
1036                 if (hmp == NULL)
1037                         error = vfs_mountedon(devvp);
1038         } else if (error == 0) {
1039                 /*
1040                  * Match the label to a pmp already probed.
1041                  */
1042                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1043                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1044                                 if (pmp->pfs_names[i] &&
1045                                     strcmp(pmp->pfs_names[i], label) == 0) {
1046                                         hmp = pmp->pfs_hmps[i];
1047                                         break;
1048                                 }
1049                         }
1050                         if (hmp)
1051                                 break;
1052                 }
1053                 if (hmp == NULL)
1054                         error = ENOENT;
1055         }
1056
1057         /*
1058          * Open the device if this isn't a secondary mount and construct
1059          * the H2 device mount (hmp).
1060          */
1061         if (hmp == NULL) {
1062                 hammer2_chain_t *schain;
1063                 hammer2_xid_t xid;
1064
1065                 if (error == 0 && vcount(devvp) > 0) {
1066                         kprintf("Primary device already has references\n");
1067                         error = EBUSY;
1068                 }
1069
1070                 /*
1071                  * Now open the device
1072                  */
1073                 if (error == 0) {
1074                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1075                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1076                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1077                         if (error == 0) {
1078                                 error = VOP_OPEN(devvp,
1079                                              (ronly ? FREAD : FREAD | FWRITE),
1080                                              FSCRED, NULL);
1081                         }
1082                         vn_unlock(devvp);
1083                 }
1084                 if (error && devvp) {
1085                         vrele(devvp);
1086                         devvp = NULL;
1087                 }
1088                 if (error) {
1089                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1090                         return error;
1091                 }
1092                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1093                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1094                 hmp->ronly = ronly;
1095                 hmp->devvp = devvp;
1096                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1097                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1098                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1099                 RB_INIT(&hmp->iotree);
1100                 spin_init(&hmp->io_spin, "hm2mount_io");
1101                 spin_init(&hmp->list_spin, "hm2mount_list");
1102                 TAILQ_INIT(&hmp->flushq);
1103
1104                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1105                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1106                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1107
1108                 /*
1109                  * vchain setup. vchain.data is embedded.
1110                  * vchain.refs is initialized and will never drop to 0.
1111                  *
1112                  * NOTE! voldata is not yet loaded.
1113                  */
1114                 hmp->vchain.hmp = hmp;
1115                 hmp->vchain.refs = 1;
1116                 hmp->vchain.data = (void *)&hmp->voldata;
1117                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1118                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1119                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1120
1121                 hammer2_chain_core_init(&hmp->vchain);
1122                 /* hmp->vchain.u.xxx is left NULL */
1123
1124                 /*
1125                  * fchain setup.  fchain.data is embedded.
1126                  * fchain.refs is initialized and will never drop to 0.
1127                  *
1128                  * The data is not used but needs to be initialized to
1129                  * pass assertion muster.  We use this chain primarily
1130                  * as a placeholder for the freemap's top-level RBTREE
1131                  * so it does not interfere with the volume's topology
1132                  * RBTREE.
1133                  */
1134                 hmp->fchain.hmp = hmp;
1135                 hmp->fchain.refs = 1;
1136                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1137                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1138                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1139                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1140                 hmp->fchain.bref.methods =
1141                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1142                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1143
1144                 hammer2_chain_core_init(&hmp->fchain);
1145                 /* hmp->fchain.u.xxx is left NULL */
1146
1147                 /*
1148                  * Install the volume header and initialize fields from
1149                  * voldata.
1150                  */
1151                 error = hammer2_install_volume_header(hmp);
1152                 if (error) {
1153                         hammer2_unmount_helper(mp, NULL, hmp);
1154                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1155                         hammer2_vfs_unmount(mp, MNT_FORCE);
1156                         return error;
1157                 }
1158
1159                 /*
1160                  * Really important to get these right or flush will get
1161                  * confused.
1162                  */
1163                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1164                 kprintf("alloc spmp %p tid %016jx\n",
1165                         hmp->spmp, hmp->voldata.mirror_tid);
1166                 spmp = hmp->spmp;
1167
1168                 /*
1169                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1170                  * is inherited from the volume header.
1171                  */
1172                 xid = 0;
1173                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1174                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1175                 hmp->vchain.pmp = spmp;
1176                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1177                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1178                 hmp->fchain.pmp = spmp;
1179
1180                 /*
1181                  * First locate the super-root inode, which is key 0
1182                  * relative to the volume header's blockset.
1183                  *
1184                  * Then locate the root inode by scanning the directory keyspace
1185                  * represented by the label.
1186                  */
1187                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1188                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1189                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1190                                       &error, 0);
1191                 hammer2_chain_lookup_done(parent);
1192                 if (schain == NULL) {
1193                         kprintf("hammer2_mount: invalid super-root\n");
1194                         hammer2_unmount_helper(mp, NULL, hmp);
1195                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1196                         hammer2_vfs_unmount(mp, MNT_FORCE);
1197                         return EINVAL;
1198                 }
1199                 if (schain->error) {
1200                         kprintf("hammer2_mount: error %s reading super-root\n",
1201                                 hammer2_error_str(schain->error));
1202                         hammer2_chain_unlock(schain);
1203                         hammer2_chain_drop(schain);
1204                         schain = NULL;
1205                         hammer2_unmount_helper(mp, NULL, hmp);
1206                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1207                         hammer2_vfs_unmount(mp, MNT_FORCE);
1208                         return EINVAL;
1209                 }
1210
1211                 /*
1212                  * The super-root always uses an inode_tid of 1 when
1213                  * creating PFSs.
1214                  */
1215                 spmp->inode_tid = 1;
1216                 spmp->modify_tid = schain->bref.modify_tid + 1;
1217
1218                 /*
1219                  * Sanity-check schain's pmp and finish initialization.
1220                  * Any chain belonging to the super-root topology should
1221                  * have a NULL pmp (not even set to spmp).
1222                  */
1223                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1224                 KKASSERT(schain->pmp == NULL);
1225                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1226
1227                 /*
1228                  * Replace the dummy spmp->iroot with a real one.  It's
1229                  * easier to just do a wholesale replacement than to try
1230                  * to update the chain and fixup the iroot fields.
1231                  *
1232                  * The returned inode is locked with the supplied cluster.
1233                  */
1234                 cluster = hammer2_cluster_from_chain(schain);
1235                 hammer2_inode_drop(spmp->iroot);
1236                 spmp->iroot = NULL;
1237                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1238                 spmp->spmp_hmp = hmp;
1239                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1240                 spmp->pfs_hmps[0] = hmp;
1241                 hammer2_inode_ref(spmp->iroot);
1242                 hammer2_inode_unlock(spmp->iroot);
1243                 hammer2_cluster_unlock(cluster);
1244                 hammer2_cluster_drop(cluster);
1245                 schain = NULL;
1246                 /* leave spmp->iroot with one ref */
1247
1248                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1249                         error = hammer2_recovery(hmp);
1250                         /* XXX do something with error */
1251                 }
1252                 hammer2_update_pmps(hmp);
1253                 hammer2_iocom_init(hmp);
1254                 hammer2_bulkfree_init(hmp);
1255
1256                 /*
1257                  * Ref the cluster management messaging descriptor.  The mount
1258                  * program deals with the other end of the communications pipe.
1259                  *
1260                  * Root mounts typically do not supply one.
1261                  */
1262                 if (info.cluster_fd >= 0) {
1263                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1264                         if (fp) {
1265                                 hammer2_cluster_reconnect(hmp, fp);
1266                         } else {
1267                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1268                         }
1269                 }
1270         } else {
1271                 spmp = hmp->spmp;
1272                 if (info.hflags & HMNT2_DEVFLAGS) {
1273                         kprintf("hammer2: Warning: mount flags pertaining "
1274                                 "to the whole device may only be specified "
1275                                 "on the first mount of the device: %08x\n",
1276                                 info.hflags & HMNT2_DEVFLAGS);
1277                 }
1278         }
1279
1280         /*
1281          * Force local mount (disassociate all PFSs from their clusters).
1282          * Used primarily for debugging.
1283          */
1284         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1285
1286         /*
1287          * Lookup the mount point under the media-localized super-root.
1288          * Scanning hammer2_pfslist doesn't help us because it represents
1289          * PFS cluster ids which can aggregate several named PFSs together.
1290          *
1291          * cluster->pmp will incorrectly point to spmp and must be fixed
1292          * up later on.
1293          */
1294         hammer2_inode_lock(spmp->iroot, 0);
1295         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1296         lhc = hammer2_dirhash(label, strlen(label));
1297         chain = hammer2_chain_lookup(&parent, &key_next,
1298                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1299                                      &error, 0);
1300         while (chain) {
1301                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1302                     strcmp(label, chain->data->ipdata.filename) == 0) {
1303                         break;
1304                 }
1305                 chain = hammer2_chain_next(&parent, chain, &key_next,
1306                                             key_next,
1307                                             lhc + HAMMER2_DIRHASH_LOMASK,
1308                                             &error, 0);
1309         }
1310         if (parent) {
1311                 hammer2_chain_unlock(parent);
1312                 hammer2_chain_drop(parent);
1313         }
1314         hammer2_inode_unlock(spmp->iroot);
1315
1316         /*
1317          * PFS could not be found?
1318          */
1319         if (chain == NULL) {
1320                 if (error)
1321                         kprintf("hammer2_mount: PFS label I/O error\n");
1322                 else
1323                         kprintf("hammer2_mount: PFS label not found\n");
1324                 hammer2_unmount_helper(mp, NULL, hmp);
1325                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1326                 hammer2_vfs_unmount(mp, MNT_FORCE);
1327
1328                 return EINVAL;
1329         }
1330
1331         /*
1332          * Acquire the pmp structure (it should have already been allocated
1333          * via hammer2_update_pmps() so do not pass cluster in to add to
1334          * available chains).
1335          *
1336          * Check if the cluster has already been mounted.  A cluster can
1337          * only be mounted once, use null mounts to mount additional copies.
1338          */
1339         if (chain->error) {
1340                 kprintf("hammer2_mount: PFS label I/O error\n");
1341         } else {
1342                 ripdata = &chain->data->ipdata;
1343                 bref = chain->bref;
1344                 pmp = hammer2_pfsalloc(NULL, ripdata,
1345                                        bref.modify_tid, force_local);
1346         }
1347         hammer2_chain_unlock(chain);
1348         hammer2_chain_drop(chain);
1349
1350         /*
1351          * Finish the mount
1352          */
1353         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1354
1355         if (pmp->mp) {
1356                 kprintf("hammer2_mount: PFS already mounted!\n");
1357                 hammer2_unmount_helper(mp, NULL, hmp);
1358                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1359                 hammer2_vfs_unmount(mp, MNT_FORCE);
1360
1361                 return EBUSY;
1362         }
1363
1364         pmp->hflags = info.hflags;
1365         mp->mnt_flag |= MNT_LOCAL;
1366         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1367         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1368
1369         /*
1370          * required mount structure initializations
1371          */
1372         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1373         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1374
1375         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1376         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1377
1378         /*
1379          * Optional fields
1380          */
1381         mp->mnt_iosize_max = MAXPHYS;
1382
1383         /*
1384          * Connect up mount pointers.
1385          */
1386         hammer2_mount_helper(mp, pmp);
1387
1388         lockmgr(&hammer2_mntlk, LK_RELEASE);
1389
1390         /*
1391          * Finish setup
1392          */
1393         vfs_getnewfsid(mp);
1394         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1395         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1396         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1397
1398         if (path) {
1399                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1400                           MNAMELEN - 1, &size);
1401                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1402         } /* else root mount, already in there */
1403
1404         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1405         if (path) {
1406                 copyinstr(path, mp->mnt_stat.f_mntonname,
1407                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1408                           &size);
1409         } else {
1410                 /* root mount */
1411                 mp->mnt_stat.f_mntonname[0] = '/';
1412         }
1413
1414         /*
1415          * Initial statfs to prime mnt_stat.
1416          */
1417         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1418
1419         return 0;
1420 }
1421
1422 /*
1423  * Scan PFSs under the super-root and create hammer2_pfs structures.
1424  */
1425 static
1426 void
1427 hammer2_update_pmps(hammer2_dev_t *hmp)
1428 {
1429         const hammer2_inode_data_t *ripdata;
1430         hammer2_chain_t *parent;
1431         hammer2_chain_t *chain;
1432         hammer2_blockref_t bref;
1433         hammer2_dev_t *force_local;
1434         hammer2_pfs_t *spmp;
1435         hammer2_pfs_t *pmp;
1436         hammer2_key_t key_next;
1437         int error;
1438
1439         /*
1440          * Force local mount (disassociate all PFSs from their clusters).
1441          * Used primarily for debugging.
1442          */
1443         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1444
1445         /*
1446          * Lookup mount point under the media-localized super-root.
1447          *
1448          * cluster->pmp will incorrectly point to spmp and must be fixed
1449          * up later on.
1450          */
1451         spmp = hmp->spmp;
1452         hammer2_inode_lock(spmp->iroot, 0);
1453         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1454         chain = hammer2_chain_lookup(&parent, &key_next,
1455                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1456                                          &error, 0);
1457         while (chain) {
1458                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1459                         continue;
1460                 if (chain->error) {
1461                         kprintf("I/O error scanning PFS labels\n");
1462                 } else {
1463                         ripdata = &chain->data->ipdata;
1464                         bref = chain->bref;
1465
1466                         pmp = hammer2_pfsalloc(chain, ripdata,
1467                                                bref.modify_tid, force_local);
1468                 }
1469                 chain = hammer2_chain_next(&parent, chain, &key_next,
1470                                            key_next, HAMMER2_KEY_MAX,
1471                                            &error, 0);
1472         }
1473         if (parent) {
1474                 hammer2_chain_unlock(parent);
1475                 hammer2_chain_drop(parent);
1476         }
1477         hammer2_inode_unlock(spmp->iroot);
1478 }
1479
1480 static
1481 int
1482 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1483                 struct vnode *devvp, struct ucred *cred)
1484 {
1485         int error;
1486
1487         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1488                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1489                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1490                 vn_unlock(devvp);
1491                 error = hammer2_recovery(hmp);
1492                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1493                 if (error == 0) {
1494                         VOP_CLOSE(devvp, FREAD, NULL);
1495                         hmp->ronly = 0;
1496                 } else {
1497                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1498                 }
1499                 vn_unlock(devvp);
1500         } else {
1501                 error = 0;
1502         }
1503         return error;
1504 }
1505
1506 static
1507 int
1508 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1509 {
1510         hammer2_pfs_t *pmp;
1511         int flags;
1512         int error = 0;
1513
1514         pmp = MPTOPMP(mp);
1515
1516         if (pmp == NULL)
1517                 return(0);
1518
1519         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1520
1521         /*
1522          * If mount initialization proceeded far enough we must flush
1523          * its vnodes and sync the underlying mount points.  Three syncs
1524          * are required to fully flush the filesystem (freemap updates lag
1525          * by one flush, and one extra for safety).
1526          */
1527         if (mntflags & MNT_FORCE)
1528                 flags = FORCECLOSE;
1529         else
1530                 flags = 0;
1531         if (pmp->iroot) {
1532                 error = vflush(mp, 0, flags);
1533                 if (error)
1534                         goto failed;
1535                 hammer2_vfs_sync(mp, MNT_WAIT);
1536                 hammer2_vfs_sync(mp, MNT_WAIT);
1537                 hammer2_vfs_sync(mp, MNT_WAIT);
1538         }
1539
1540         /*
1541          * Cleanup the frontend support XOPS threads
1542          */
1543         hammer2_xop_helper_cleanup(pmp);
1544
1545         if (pmp->mp)
1546                 hammer2_unmount_helper(mp, pmp, NULL);
1547
1548         error = 0;
1549 failed:
1550         lockmgr(&hammer2_mntlk, LK_RELEASE);
1551
1552         return (error);
1553 }
1554
1555 /*
1556  * Mount helper, hook the system mount into our PFS.
1557  * The mount lock is held.
1558  *
1559  * We must bump the mount_count on related devices for any
1560  * mounted PFSs.
1561  */
1562 static
1563 void
1564 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1565 {
1566         hammer2_cluster_t *cluster;
1567         hammer2_chain_t *rchain;
1568         int i;
1569
1570         mp->mnt_data = (qaddr_t)pmp;
1571         pmp->mp = mp;
1572
1573         /*
1574          * After pmp->mp is set we have to adjust hmp->mount_count.
1575          */
1576         cluster = &pmp->iroot->cluster;
1577         for (i = 0; i < cluster->nchains; ++i) {
1578                 rchain = cluster->array[i].chain;
1579                 if (rchain == NULL)
1580                         continue;
1581                 ++rchain->hmp->mount_count;
1582                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1583                         rchain->hmp, rchain->hmp->mount_count);
1584         }
1585
1586         /*
1587          * Create missing Xop threads
1588          */
1589         hammer2_xop_helper_create(pmp);
1590 }
1591
1592 /*
1593  * Mount helper, unhook the system mount from our PFS.
1594  * The mount lock is held.
1595  *
1596  * If hmp is supplied a mount responsible for being the first to open
1597  * the block device failed and the block device and all PFSs using the
1598  * block device must be cleaned up.
1599  *
1600  * If pmp is supplied multiple devices might be backing the PFS and each
1601  * must be disconnected.  This might not be the last PFS using some of the
1602  * underlying devices.  Also, we have to adjust our hmp->mount_count
1603  * accounting for the devices backing the pmp which is now undergoing an
1604  * unmount.
1605  */
1606 static
1607 void
1608 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1609 {
1610         hammer2_cluster_t *cluster;
1611         hammer2_chain_t *rchain;
1612         struct vnode *devvp;
1613         int dumpcnt;
1614         int ronly;
1615         int i;
1616
1617         /*
1618          * If no device supplied this is a high-level unmount and we have to
1619          * to disconnect the mount, adjust mount_count, and locate devices
1620          * that might now have no mounts.
1621          */
1622         if (pmp) {
1623                 KKASSERT(hmp == NULL);
1624                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1625                 pmp->mp = NULL;
1626                 mp->mnt_data = NULL;
1627
1628                 /*
1629                  * After pmp->mp is cleared we have to account for
1630                  * mount_count.
1631                  */
1632                 cluster = &pmp->iroot->cluster;
1633                 for (i = 0; i < cluster->nchains; ++i) {
1634                         rchain = cluster->array[i].chain;
1635                         if (rchain == NULL)
1636                                 continue;
1637                         --rchain->hmp->mount_count;
1638                         /* scrapping hmp now may invalidate the pmp */
1639                 }
1640 again:
1641                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1642                         if (hmp->mount_count == 0) {
1643                                 hammer2_unmount_helper(NULL, NULL, hmp);
1644                                 goto again;
1645                         }
1646                 }
1647                 return;
1648         }
1649
1650         /*
1651          * Try to terminate the block device.  We can't terminate it if
1652          * there are still PFSs referencing it.
1653          */
1654         if (hmp->mount_count)
1655                 return;
1656
1657         /*
1658          * Decomission the network before we start messing with the
1659          * device and PFS.
1660          */
1661         hammer2_iocom_uninit(hmp);
1662
1663         hammer2_bulkfree_uninit(hmp);
1664         hammer2_pfsfree_scan(hmp);
1665         hammer2_dev_exlock(hmp);        /* XXX order */
1666
1667         /*
1668          * Cycle the volume data lock as a safety (probably not needed any
1669          * more).  To ensure everything is out we need to flush at least
1670          * three times.  (1) The running of the sideq can dirty the
1671          * filesystem, (2) A normal flush can dirty the freemap, and
1672          * (3) ensure that the freemap is fully synchronized.
1673          *
1674          * The next mount's recovery scan can clean everything up but we want
1675          * to leave the filesystem in a 100% clean state on a normal unmount.
1676          */
1677 #if 0
1678         hammer2_voldata_lock(hmp);
1679         hammer2_voldata_unlock(hmp);
1680 #endif
1681
1682         if ((hmp->vchain.flags | hmp->fchain.flags) &
1683             HAMMER2_CHAIN_FLUSH_MASK) {
1684                 kprintf("hammer2_unmount: chains left over "
1685                         "after final sync\n");
1686                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1687                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1688
1689                 if (hammer2_debug & 0x0010)
1690                         Debugger("entered debugger");
1691         }
1692
1693         KKASSERT(hmp->spmp == NULL);
1694
1695         /*
1696          * Finish up with the device vnode
1697          */
1698         if ((devvp = hmp->devvp) != NULL) {
1699                 ronly = hmp->ronly;
1700                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1701                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1702                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1703                         ronly);
1704                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1705                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1706                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1707                 hmp->devvp = NULL;
1708                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1709                 vn_unlock(devvp);
1710                 vrele(devvp);
1711                 devvp = NULL;
1712         }
1713
1714         /*
1715          * Clear vchain/fchain flags that might prevent final cleanup
1716          * of these chains.
1717          */
1718         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1719                 atomic_add_long(&hammer2_count_modified_chains, -1);
1720                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1721                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1722         }
1723         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1724                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1725         }
1726
1727         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1728                 atomic_add_long(&hammer2_count_modified_chains, -1);
1729                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1730                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1731         }
1732         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1733                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1734         }
1735
1736         /*
1737          * Final drop of embedded freemap root chain to
1738          * clean up fchain.core (fchain structure is not
1739          * flagged ALLOCATED so it is cleaned out and then
1740          * left to rot).
1741          */
1742         hammer2_chain_drop(&hmp->fchain);
1743
1744         /*
1745          * Final drop of embedded volume root chain to clean
1746          * up vchain.core (vchain structure is not flagged
1747          * ALLOCATED so it is cleaned out and then left to
1748          * rot).
1749          */
1750         dumpcnt = 50;
1751         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1752         dumpcnt = 50;
1753         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1754         hammer2_dev_unlock(hmp);
1755         hammer2_chain_drop(&hmp->vchain);
1756
1757         hammer2_io_cleanup(hmp, &hmp->iotree);
1758         if (hmp->iofree_count) {
1759                 kprintf("io_cleanup: %d I/O's left hanging\n",
1760                         hmp->iofree_count);
1761         }
1762
1763         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1764         kmalloc_destroy(&hmp->mchain);
1765         kfree(hmp, M_HAMMER2);
1766 }
1767
1768 int
1769 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1770                  ino_t ino, struct vnode **vpp)
1771 {
1772         hammer2_xop_lookup_t *xop;
1773         hammer2_pfs_t *pmp;
1774         hammer2_inode_t *ip;
1775         hammer2_tid_t inum;
1776         int error;
1777
1778         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1779
1780         error = 0;
1781         pmp = MPTOPMP(mp);
1782
1783         /*
1784          * Easy if we already have it cached
1785          */
1786         ip = hammer2_inode_lookup(pmp, inum);
1787         if (ip) {
1788                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1789                 *vpp = hammer2_igetv(ip, &error);
1790                 hammer2_inode_unlock(ip);
1791                 hammer2_inode_drop(ip);         /* from lookup */
1792
1793                 return error;
1794         }
1795
1796         /*
1797          * Otherwise we have to find the inode
1798          */
1799         xop = hammer2_xop_alloc(pmp->iroot, 0);
1800         xop->lhc = inum;
1801         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1802         error = hammer2_xop_collect(&xop->head, 0);
1803
1804         if (error == 0) {
1805                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1806                         kprintf("vget: no collect error but also no rdata\n");
1807                         kprintf("xop %p\n", xop);
1808                         while ((hammer2_debug & 0x80000) == 0) {
1809                                 tsleep(xop, PCATCH, "wait", hz * 10);
1810                         }
1811                         ip = NULL;
1812                 } else {
1813                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1814                 }
1815         }
1816         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1817
1818         if (ip) {
1819                 *vpp = hammer2_igetv(ip, &error);
1820                 hammer2_inode_unlock(ip);
1821         } else {
1822                 *vpp = NULL;
1823                 error = ENOENT;
1824         }
1825         return (error);
1826 }
1827
1828 static
1829 int
1830 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1831 {
1832         hammer2_pfs_t *pmp;
1833         struct vnode *vp;
1834         int error;
1835
1836         pmp = MPTOPMP(mp);
1837         if (pmp->iroot == NULL) {
1838                 *vpp = NULL;
1839                 return EINVAL;
1840         }
1841
1842         error = 0;
1843         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1844
1845         while (pmp->inode_tid == 0) {
1846                 hammer2_xop_ipcluster_t *xop;
1847                 hammer2_inode_meta_t *meta;
1848
1849                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1850                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1851                 error = hammer2_xop_collect(&xop->head, 0);
1852
1853                 if (error == 0) {
1854                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1855                         pmp->iroot->meta = *meta;
1856                         pmp->inode_tid = meta->pfs_inum + 1;
1857                         if (pmp->inode_tid < HAMMER2_INODE_START)
1858                                 pmp->inode_tid = HAMMER2_INODE_START;
1859                         pmp->modify_tid =
1860                                 xop->head.cluster.focus->bref.modify_tid + 1;
1861                         kprintf("PFS: Starting inode %jd\n",
1862                                 (intmax_t)pmp->inode_tid);
1863                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1864                                 pmp->inode_tid, pmp->modify_tid);
1865                         wakeup(&pmp->iroot);
1866
1867                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1868
1869                         /*
1870                          * Prime the mount info.
1871                          */
1872                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1873                         break;
1874                 }
1875
1876                 /*
1877                  * Loop, try again
1878                  */
1879                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1880                 hammer2_inode_unlock(pmp->iroot);
1881                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1882                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1883                 if (error == EINTR)
1884                         break;
1885         }
1886
1887         if (error) {
1888                 hammer2_inode_unlock(pmp->iroot);
1889                 *vpp = NULL;
1890         } else {
1891                 vp = hammer2_igetv(pmp->iroot, &error);
1892                 hammer2_inode_unlock(pmp->iroot);
1893                 *vpp = vp;
1894         }
1895
1896         return (error);
1897 }
1898
1899 /*
1900  * Filesystem status
1901  *
1902  * XXX incorporate ipdata->meta.inode_quota and data_quota
1903  */
1904 static
1905 int
1906 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1907 {
1908         hammer2_pfs_t *pmp;
1909         hammer2_dev_t *hmp;
1910         hammer2_blockref_t bref;
1911         int i;
1912
1913         /*
1914          * NOTE: iroot might not have validated the cluster yet.
1915          */
1916         pmp = MPTOPMP(mp);
1917
1918         mp->mnt_stat.f_files = 0;
1919         mp->mnt_stat.f_ffree = 0;
1920         mp->mnt_stat.f_blocks = 0;
1921         mp->mnt_stat.f_bfree = 0;
1922         mp->mnt_stat.f_bavail = 0;
1923
1924         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1925                 hmp = pmp->pfs_hmps[i];
1926                 if (hmp == NULL)
1927                         continue;
1928                 if (pmp->iroot->cluster.array[i].chain)
1929                         bref = pmp->iroot->cluster.array[i].chain->bref;
1930                 else
1931                         bzero(&bref, sizeof(bref));
1932
1933                 mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1934                 mp->mnt_stat.f_ffree = 0;
1935                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1936                                         mp->mnt_vstat.f_bsize;
1937                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1938                                         mp->mnt_vstat.f_bsize;
1939                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1940
1941                 if (cred && cred->cr_uid != 0) {
1942                         uint64_t adj;
1943
1944                         /* 5% */
1945                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
1946                         mp->mnt_stat.f_blocks -= adj;
1947                         mp->mnt_stat.f_bfree -= adj;
1948                         mp->mnt_stat.f_bavail -= adj;
1949                 }
1950
1951                 *sbp = mp->mnt_stat;
1952         }
1953         return (0);
1954 }
1955
1956 static
1957 int
1958 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1959 {
1960         hammer2_pfs_t *pmp;
1961         hammer2_dev_t *hmp;
1962         hammer2_blockref_t bref;
1963         int i;
1964
1965         /*
1966          * NOTE: iroot might not have validated the cluster yet.
1967          */
1968         pmp = MPTOPMP(mp);
1969
1970         mp->mnt_vstat.f_bsize = 0;
1971         mp->mnt_vstat.f_files = 0;
1972         mp->mnt_vstat.f_ffree = 0;
1973         mp->mnt_vstat.f_blocks = 0;
1974         mp->mnt_vstat.f_bfree = 0;
1975         mp->mnt_vstat.f_bavail = 0;
1976
1977         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1978                 hmp = pmp->pfs_hmps[i];
1979                 if (hmp == NULL)
1980                         continue;
1981                 if (pmp->iroot->cluster.array[i].chain)
1982                         bref = pmp->iroot->cluster.array[i].chain->bref;
1983                 else
1984                         bzero(&bref, sizeof(bref));
1985
1986                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1987                 mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
1988                 mp->mnt_vstat.f_ffree = 0;
1989                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1990                                         mp->mnt_vstat.f_bsize;
1991                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1992                                         mp->mnt_vstat.f_bsize;
1993                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1994
1995                 if (cred && cred->cr_uid != 0) {
1996                         uint64_t adj;
1997
1998                         /* 5% */
1999                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2000                         mp->mnt_vstat.f_blocks -= adj;
2001                         mp->mnt_vstat.f_bfree -= adj;
2002                         mp->mnt_vstat.f_bavail -= adj;
2003                 }
2004
2005                 *sbp = mp->mnt_vstat;
2006         }
2007         return (0);
2008 }
2009
2010 /*
2011  * Mount-time recovery (RW mounts)
2012  *
2013  * Updates to the free block table are allowed to lag flushes by one
2014  * transaction.  In case of a crash, then on a fresh mount we must do an
2015  * incremental scan of the last committed transaction id and make sure that
2016  * all related blocks have been marked allocated.
2017  *
2018  * The super-root topology and each PFS has its own transaction id domain,
2019  * so we must track PFS boundary transitions.
2020  */
2021 struct hammer2_recovery_elm {
2022         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2023         hammer2_chain_t *chain;
2024         hammer2_tid_t sync_tid;
2025 };
2026
2027 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2028
2029 struct hammer2_recovery_info {
2030         struct hammer2_recovery_list list;
2031         hammer2_tid_t   mtid;
2032         int     depth;
2033 };
2034
2035 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2036                         hammer2_chain_t *parent,
2037                         struct hammer2_recovery_info *info,
2038                         hammer2_tid_t sync_tid);
2039
2040 #define HAMMER2_RECOVERY_MAXDEPTH       10
2041
2042 static
2043 int
2044 hammer2_recovery(hammer2_dev_t *hmp)
2045 {
2046         struct hammer2_recovery_info info;
2047         struct hammer2_recovery_elm *elm;
2048         hammer2_chain_t *parent;
2049         hammer2_tid_t sync_tid;
2050         hammer2_tid_t mirror_tid;
2051         int error;
2052
2053         hammer2_trans_init(hmp->spmp, 0);
2054
2055         sync_tid = hmp->voldata.freemap_tid;
2056         mirror_tid = hmp->voldata.mirror_tid;
2057
2058         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2059         if (sync_tid >= mirror_tid) {
2060                 kprintf(" no recovery needed\n");
2061         } else {
2062                 kprintf(" freemap recovery %016jx-%016jx\n",
2063                         sync_tid + 1, mirror_tid);
2064         }
2065
2066         TAILQ_INIT(&info.list);
2067         info.depth = 0;
2068         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2069         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2070         hammer2_chain_lookup_done(parent);
2071
2072         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2073                 TAILQ_REMOVE(&info.list, elm, entry);
2074                 parent = elm->chain;
2075                 sync_tid = elm->sync_tid;
2076                 kfree(elm, M_HAMMER2);
2077
2078                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2079                 error |= hammer2_recovery_scan(hmp, parent, &info,
2080                                               hmp->voldata.freemap_tid);
2081                 hammer2_chain_unlock(parent);
2082                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2083         }
2084         hammer2_trans_done(hmp->spmp);
2085
2086         return error;
2087 }
2088
2089 static
2090 int
2091 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2092                       struct hammer2_recovery_info *info,
2093                       hammer2_tid_t sync_tid)
2094 {
2095         const hammer2_inode_data_t *ripdata;
2096         hammer2_chain_t *chain;
2097         hammer2_blockref_t bref;
2098         int tmp_error;
2099         int rup_error;
2100         int error;
2101         int first;
2102
2103         /*
2104          * Adjust freemap to ensure that the block(s) are marked allocated.
2105          */
2106         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2107                 hammer2_freemap_adjust(hmp, &parent->bref,
2108                                        HAMMER2_FREEMAP_DORECOVER);
2109         }
2110
2111         /*
2112          * Check type for recursive scan
2113          */
2114         switch(parent->bref.type) {
2115         case HAMMER2_BREF_TYPE_VOLUME:
2116                 /* data already instantiated */
2117                 break;
2118         case HAMMER2_BREF_TYPE_INODE:
2119                 /*
2120                  * Must instantiate data for DIRECTDATA test and also
2121                  * for recursion.
2122                  */
2123                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2124                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2125                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2126                         /* not applicable to recovery scan */
2127                         hammer2_chain_unlock(parent);
2128                         return 0;
2129                 }
2130                 hammer2_chain_unlock(parent);
2131                 break;
2132         case HAMMER2_BREF_TYPE_INDIRECT:
2133                 /*
2134                  * Must instantiate data for recursion
2135                  */
2136                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2137                 hammer2_chain_unlock(parent);
2138                 break;
2139         case HAMMER2_BREF_TYPE_DIRENT:
2140         case HAMMER2_BREF_TYPE_DATA:
2141         case HAMMER2_BREF_TYPE_FREEMAP:
2142         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2143         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2144                 /* not applicable to recovery scan */
2145                 return 0;
2146                 break;
2147         default:
2148                 return HAMMER2_ERROR_BADBREF;
2149         }
2150
2151         /*
2152          * Defer operation if depth limit reached or if we are crossing a
2153          * PFS boundary.
2154          */
2155         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2156                 struct hammer2_recovery_elm *elm;
2157
2158                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2159                 elm->chain = parent;
2160                 elm->sync_tid = sync_tid;
2161                 hammer2_chain_ref(parent);
2162                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2163                 /* unlocked by caller */
2164
2165                 return(0);
2166         }
2167
2168
2169         /*
2170          * Recursive scan of the last flushed transaction only.  We are
2171          * doing this without pmp assignments so don't leave the chains
2172          * hanging around after we are done with them.
2173          *
2174          * error        Cumulative error this level only
2175          * rup_error    Cumulative error for recursion
2176          * tmp_error    Specific non-cumulative recursion error
2177          */
2178         chain = NULL;
2179         first = 1;
2180         rup_error = 0;
2181         error = 0;
2182
2183         for (;;) {
2184                 error |= hammer2_chain_scan(parent, &chain, &bref,
2185                                             &first,
2186                                             HAMMER2_LOOKUP_NODATA);
2187
2188                 /*
2189                  * Problem during scan or EOF
2190                  */
2191                 if (error)
2192                         break;
2193
2194                 /*
2195                  * If this is a leaf
2196                  */
2197                 if (chain == NULL) {
2198                         if (bref.mirror_tid > sync_tid) {
2199                                 hammer2_freemap_adjust(hmp, &bref,
2200                                                      HAMMER2_FREEMAP_DORECOVER);
2201                         }
2202                         continue;
2203                 }
2204
2205                 /*
2206                  * This may or may not be a recursive node.
2207                  */
2208                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2209                 if (bref.mirror_tid > sync_tid) {
2210                         ++info->depth;
2211                         tmp_error = hammer2_recovery_scan(hmp, chain,
2212                                                            info, sync_tid);
2213                         --info->depth;
2214                 } else {
2215                         tmp_error = 0;
2216                 }
2217
2218                 /*
2219                  * Flush the recovery at the PFS boundary to stage it for
2220                  * the final flush of the super-root topology.
2221                  */
2222                 if (tmp_error == 0 &&
2223                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2224                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2225                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2226                 }
2227                 rup_error |= tmp_error;
2228         }
2229         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2230 }
2231
2232 /*
2233  * Sync a mount point; this is called on a per-mount basis from the
2234  * filesystem syncer process periodically and whenever a user issues
2235  * a sync.
2236  */
2237 int
2238 hammer2_vfs_sync(struct mount *mp, int waitfor)
2239 {
2240         hammer2_xop_flush_t *xop;
2241         struct hammer2_sync_info info;
2242         hammer2_inode_t *iroot;
2243         hammer2_pfs_t *pmp;
2244         int flags;
2245         int error;
2246
2247         pmp = MPTOPMP(mp);
2248         iroot = pmp->iroot;
2249         KKASSERT(iroot);
2250         KKASSERT(iroot->pmp == pmp);
2251
2252         /*
2253          * We can't acquire locks on existing vnodes while in a transaction
2254          * without risking a deadlock.  This assumes that vfsync() can be
2255          * called without the vnode locked (which it can in DragonFly).
2256          * Otherwise we'd have to implement a multi-pass or flag the lock
2257          * failures and retry.
2258          *
2259          * The reclamation code interlocks with the sync list's token
2260          * (by removing the vnode from the scan list) before unlocking
2261          * the inode, giving us time to ref the inode.
2262          */
2263         /*flags = VMSC_GETVP;*/
2264         flags = 0;
2265         if (waitfor & MNT_LAZY)
2266                 flags |= VMSC_ONEPASS;
2267
2268         /*
2269          * Preflush the vnodes using a normal transaction before interlocking
2270          * with a flush transaction.  We do this to try to run as much of
2271          * the compression as possible outside the flush transaction.
2272          *
2273          * For efficiency do an async pass before making sure with a
2274          * synchronous pass on all related buffer cache buffers.
2275          */
2276         hammer2_trans_init(pmp, 0);
2277         info.error = 0;
2278         info.waitfor = MNT_NOWAIT;
2279         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2280         info.waitfor = MNT_WAIT;
2281         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2282         hammer2_trans_done(pmp);
2283
2284         /*
2285          * Start our flush transaction.  This does not return until all
2286          * concurrent transactions have completed and will prevent any
2287          * new transactions from running concurrently, except for the
2288          * buffer cache transactions.
2289          *
2290          * (1) vfsync() all dirty vnodes via vfsyncscan().
2291          *
2292          * (2) Flush any remaining dirty inodes (the sideq), including any
2293          *     which may have been created during or raced against the
2294          *     vfsync().  To catch all cases this must be done after the
2295          *     vfsync().
2296          *
2297          * (3) Wait for any pending BIO I/O to complete (hammer2_bioq_sync()).
2298          *
2299          * NOTE! It is still possible for the paging code to push pages
2300          *       out via a UIO_NOCOPY hammer2_vop_write() during the main
2301          *       flush.
2302          */
2303         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2304
2305         info.error = 0;
2306         info.waitfor = MNT_NOWAIT;
2307         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2308         info.waitfor = MNT_WAIT;
2309         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2310         hammer2_inode_run_sideq(pmp, 1);
2311         hammer2_bioq_sync(pmp);
2312
2313         /*
2314          * Use the XOP interface to concurrently flush all nodes to
2315          * synchronize the PFSROOT subtopology to the media.  A standard
2316          * end-of-scan ENOENT error indicates cluster sufficiency.
2317          *
2318          * Note that this flush will not be visible on crash recovery until
2319          * we flush the super-root topology in the next loop.
2320          *
2321          * XXX For now wait for all flushes to complete.
2322          */
2323         if (iroot) {
2324                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2325                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2326                 error = hammer2_xop_collect(&xop->head,
2327                                             HAMMER2_XOP_COLLECT_WAITALL);
2328                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2329                 if (error == HAMMER2_ERROR_ENOENT)
2330                         error = 0;
2331                 else
2332                         error = hammer2_error_to_errno(error);
2333         } else {
2334                 error = 0;
2335         }
2336         hammer2_trans_done(pmp);
2337
2338         return (error);
2339 }
2340
2341 /*
2342  * Sync passes.
2343  *
2344  * Note that we ignore the tranasction mtid we got above.  Instead,
2345  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2346  * transactions.
2347  */
2348 static int
2349 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2350 {
2351         struct hammer2_sync_info *info = data;
2352         hammer2_inode_t *ip;
2353         int error;
2354
2355         /*
2356          * Degenerate cases.  Note that ip == NULL typically means the
2357          * syncer vnode itself and we don't want to vclrisdirty() in that
2358          * situation.
2359          */
2360         ip = VTOI(vp);
2361         if (ip == NULL) {
2362                 return(0);
2363         }
2364         if (vp->v_type == VNON || vp->v_type == VBAD) {
2365                 vclrisdirty(vp);
2366                 return(0);
2367         }
2368
2369         /*
2370          * VOP_FSYNC will start a new transaction so replicate some code
2371          * here to do it inline (see hammer2_vop_fsync()).
2372          *
2373          * WARNING: The vfsync interacts with the buffer cache and might
2374          *          block, we can't hold the inode lock at that time.
2375          *          However, we MUST ref ip before blocking to ensure that
2376          *          it isn't ripped out from under us (since we do not
2377          *          hold a lock on the vnode).
2378          */
2379         hammer2_inode_ref(ip);
2380         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2381             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2382                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2383                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2384                                  HAMMER2_INODE_MODIFIED)) {
2385                         hammer2_inode_lock(ip, 0);
2386                         if (ip->flags & (HAMMER2_INODE_RESIZED |
2387                                          HAMMER2_INODE_MODIFIED)) {
2388                                 hammer2_inode_chain_sync(ip);
2389                         }
2390                         hammer2_inode_unlock(ip);
2391                 }
2392         }
2393         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2394             RB_EMPTY(&vp->v_rbdirty_tree)) {
2395                 vclrisdirty(vp);
2396         }
2397
2398         hammer2_inode_drop(ip);
2399 #if 1
2400         error = 0;
2401         if (error)
2402                 info->error = error;
2403 #endif
2404         return(0);
2405 }
2406
2407 static
2408 int
2409 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2410 {
2411         hammer2_inode_t *ip;
2412
2413         KKASSERT(MAXFIDSZ >= 16);
2414         ip = VTOI(vp);
2415         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2416         fhp->fid_ext = 0;
2417         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2418         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2419
2420         return 0;
2421 }
2422
2423 static
2424 int
2425 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2426                struct fid *fhp, struct vnode **vpp)
2427 {
2428         hammer2_pfs_t *pmp;
2429         hammer2_tid_t inum;
2430         int error;
2431
2432         pmp = MPTOPMP(mp);
2433         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2434         if (vpp) {
2435                 if (inum == 1)
2436                         error = hammer2_vfs_root(mp, vpp);
2437                 else
2438                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2439         } else {
2440                 error = 0;
2441         }
2442         if (error)
2443                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2444         return error;
2445 }
2446
2447 static
2448 int
2449 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2450                  int *exflagsp, struct ucred **credanonp)
2451 {
2452         hammer2_pfs_t *pmp;
2453         struct netcred *np;
2454         int error;
2455
2456         pmp = MPTOPMP(mp);
2457         np = vfs_export_lookup(mp, &pmp->export, nam);
2458         if (np) {
2459                 *exflagsp = np->netc_exflags;
2460                 *credanonp = &np->netc_anon;
2461                 error = 0;
2462         } else {
2463                 error = EACCES;
2464         }
2465         return error;
2466 }
2467
2468 /*
2469  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2470  * header into the HMP
2471  *
2472  * XXX read four volhdrs and use the one with the highest TID whos CRC
2473  *     matches.
2474  *
2475  * XXX check iCRCs.
2476  *
2477  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2478  *     nonexistant locations.
2479  *
2480  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2481  */
2482 static
2483 int
2484 hammer2_install_volume_header(hammer2_dev_t *hmp)
2485 {
2486         hammer2_volume_data_t *vd;
2487         struct buf *bp;
2488         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2489         int error_reported;
2490         int error;
2491         int valid;
2492         int i;
2493
2494         error_reported = 0;
2495         error = 0;
2496         valid = 0;
2497         bp = NULL;
2498
2499         /*
2500          * There are up to 4 copies of the volume header (syncs iterate
2501          * between them so there is no single master).  We don't trust the
2502          * volu_size field so we don't know precisely how large the filesystem
2503          * is, so depend on the OS to return an error if we go beyond the
2504          * block device's EOF.
2505          */
2506         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2507                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2508                               HAMMER2_VOLUME_BYTES, &bp);
2509                 if (error) {
2510                         brelse(bp);
2511                         bp = NULL;
2512                         continue;
2513                 }
2514
2515                 vd = (struct hammer2_volume_data *) bp->b_data;
2516                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2517                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2518                         brelse(bp);
2519                         bp = NULL;
2520                         continue;
2521                 }
2522
2523                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2524                         /* XXX: Reversed-endianness filesystem */
2525                         kprintf("hammer2: reverse-endian filesystem detected");
2526                         brelse(bp);
2527                         bp = NULL;
2528                         continue;
2529                 }
2530
2531                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2532                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2533                                       HAMMER2_VOLUME_ICRC0_SIZE);
2534                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2535                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2536                                        HAMMER2_VOLUME_ICRC1_SIZE);
2537                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2538                         kprintf("hammer2 volume header crc "
2539                                 "mismatch copy #%d %08x/%08x\n",
2540                                 i, crc0, crc);
2541                         error_reported = 1;
2542                         brelse(bp);
2543                         bp = NULL;
2544                         continue;
2545                 }
2546                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2547                         valid = 1;
2548                         hmp->voldata = *vd;
2549                         hmp->volhdrno = i;
2550                 }
2551                 brelse(bp);
2552                 bp = NULL;
2553         }
2554         if (valid) {
2555                 hmp->volsync = hmp->voldata;
2556                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2557                 error = 0;
2558                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2559                         kprintf("hammer2: using volume header #%d\n",
2560                                 hmp->volhdrno);
2561                 }
2562         } else {
2563                 error = EINVAL;
2564                 kprintf("hammer2: no valid volume headers found!\n");
2565         }
2566         return (error);
2567 }
2568
2569 /*
2570  * This handles hysteresis on regular file flushes.  Because the BIOs are
2571  * routed to a thread it is possible for an excessive number to build up
2572  * and cause long front-end stalls long before the runningbuffspace limit
2573  * is hit, so we implement hammer2_flush_pipe to control the
2574  * hysteresis.
2575  *
2576  * This is a particular problem when compression is used.
2577  */
2578 void
2579 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2580 {
2581         atomic_add_int(&pmp->count_lwinprog, 1);
2582 }
2583
2584 void
2585 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2586 {
2587         int lwinprog;
2588
2589         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2590         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2591             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2592                 atomic_clear_int(&pmp->count_lwinprog,
2593                                  HAMMER2_LWINPROG_WAITING);
2594                 wakeup(&pmp->count_lwinprog);
2595         }
2596         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2597             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2598                 atomic_clear_int(&pmp->count_lwinprog,
2599                                  HAMMER2_LWINPROG_WAITING0);
2600                 wakeup(&pmp->count_lwinprog);
2601         }
2602 }
2603
2604 void
2605 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2606 {
2607         int lwinprog;
2608         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2609                                     HAMMER2_LWINPROG_WAITING0;
2610
2611         for (;;) {
2612                 lwinprog = pmp->count_lwinprog;
2613                 cpu_ccfence();
2614                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2615                         break;
2616                 tsleep_interlock(&pmp->count_lwinprog, 0);
2617                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2618                 lwinprog = pmp->count_lwinprog;
2619                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2620                         break;
2621                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2622         }
2623 }
2624
2625 /*
2626  * Manage excessive memory resource use for chain and related
2627  * structures.
2628  */
2629 void
2630 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2631 {
2632         uint32_t waiting;
2633         uint32_t count;
2634         uint32_t limit;
2635 #if 0
2636         static int zzticks;
2637 #endif
2638
2639         /*
2640          * Atomic check condition and wait.  Also do an early speedup of
2641          * the syncer to try to avoid hitting the wait.
2642          */
2643         for (;;) {
2644                 waiting = pmp->inmem_dirty_chains;
2645                 cpu_ccfence();
2646                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2647
2648                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2649                 if (limit < hammer2_limit_dirty_chains)
2650                         limit = hammer2_limit_dirty_chains;
2651                 if (limit < 1000)
2652                         limit = 1000;
2653
2654 #if 0
2655                 if ((int)(ticks - zzticks) > hz) {
2656                         zzticks = ticks;
2657                         kprintf("count %ld %ld\n", count, limit);
2658                 }
2659 #endif
2660
2661                 /*
2662                  * Block if there are too many dirty chains present, wait
2663                  * for the flush to clean some out.
2664                  */
2665                 if (count > limit) {
2666                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2667                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2668                                                waiting,
2669                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2670                                 speedup_syncer(pmp->mp);
2671                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2672                                        "chnmem", hz);
2673                         }
2674                         continue;       /* loop on success or fail */
2675                 }
2676
2677                 /*
2678                  * Try to start an early flush before we are forced to block.
2679                  */
2680                 if (count > limit * 7 / 10)
2681                         speedup_syncer(pmp->mp);
2682                 break;
2683         }
2684 }
2685
2686 void
2687 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2688 {
2689         if (pmp) {
2690                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2691         }
2692 }
2693
2694 void
2695 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2696 {
2697         uint32_t waiting;
2698
2699         if (pmp == NULL)
2700                 return;
2701
2702         for (;;) {
2703                 waiting = pmp->inmem_dirty_chains;
2704                 cpu_ccfence();
2705                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2706                                        waiting,
2707                                        (waiting - 1) &
2708                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2709                         break;
2710                 }
2711         }
2712
2713         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2714                 wakeup(&pmp->inmem_dirty_chains);
2715 }
2716
2717 /*
2718  * Returns 0 if the filesystem has tons of free space
2719  * Returns 1 if the filesystem has less than 10% remaining
2720  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2721  */
2722 int
2723 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2724 {
2725         hammer2_pfs_t *pmp;
2726         hammer2_dev_t *hmp;
2727         hammer2_off_t free_reserved;
2728         hammer2_off_t free_nominal;
2729         int i;
2730
2731         pmp = ip->pmp;
2732
2733         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2734                 free_reserved = HAMMER2_SEGSIZE;
2735                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2736                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2737                         hmp = pmp->pfs_hmps[i];
2738                         if (hmp == NULL)
2739                                 continue;
2740                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2741                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2742                                 continue;
2743
2744                         if (free_nominal > hmp->voldata.allocator_free)
2745                                 free_nominal = hmp->voldata.allocator_free;
2746                         if (free_reserved < hmp->free_reserved)
2747                                 free_reserved = hmp->free_reserved;
2748                 }
2749
2750                 /*
2751                  * SMP races ok
2752                  */
2753                 pmp->free_reserved = free_reserved;
2754                 pmp->free_nominal = free_nominal;
2755                 pmp->free_ticks = ticks;
2756         } else {
2757                 free_reserved = pmp->free_reserved;
2758                 free_nominal = pmp->free_nominal;
2759         }
2760         if (cred && cred->cr_uid != 0) {
2761                 if ((int64_t)(free_nominal - bytes) <
2762                     (int64_t)free_reserved) {
2763                         return 2;
2764                 }
2765         } else {
2766                 if ((int64_t)(free_nominal - bytes) <
2767                     (int64_t)free_reserved / 2) {
2768                         return 2;
2769                 }
2770         }
2771         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2772                 return 1;
2773         return 0;
2774 }
2775
2776 /*
2777  * Debugging
2778  */
2779 void
2780 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2781 {
2782         hammer2_chain_t *scan;
2783         hammer2_chain_t *parent;
2784
2785         --*countp;
2786         if (*countp == 0) {
2787                 kprintf("%*.*s...\n", tab, tab, "");
2788                 return;
2789         }
2790         if (*countp < 0)
2791                 return;
2792         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2793                 tab, tab, "", pfx,
2794                 chain, chain->bref.type,
2795                 chain->bref.key, chain->bref.keybits,
2796                 chain->bref.mirror_tid);
2797
2798         kprintf("%*.*s      [%08x] (%s) refs=%d",
2799                 tab, tab, "",
2800                 chain->flags,
2801                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2802                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2803                 chain->refs);
2804
2805         parent = chain->parent;
2806         if (parent)
2807                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2808                         tab, tab, "",
2809                         parent, parent->flags, parent->refs);
2810         if (RB_EMPTY(&chain->core.rbtree)) {
2811                 kprintf("\n");
2812         } else {
2813                 kprintf(" {\n");
2814                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2815                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2816                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2817                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2818                                 chain->data->ipdata.filename);
2819                 else
2820                         kprintf("%*.*s}\n", tab, tab, "");
2821         }
2822 }