sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72 };
  73
  74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  75 static struct hammer2_mntlist hammer2_mntlist;
  76
  77 struct hammer2_pfslist hammer2_pfslist;
  78 struct lock hammer2_mntlk;
  79
  80 int hammer2_debug;
  81 int hammer2_cluster_read = 4;           /* physical read-ahead */
  82 int hammer2_cluster_write = 0;          /* bdwrite() so later inval works */
  83 int hammer2_dedup_enable = 1;
  84 int hammer2_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 long hammer2_chain_allocs;
  91 long hammer2_chain_frees;
  92 long hammer2_limit_dirty_chains;
  93 long hammer2_count_modified_chains;
  94 long hammer2_iod_invals;
  95 long hammer2_iod_file_read;
  96 long hammer2_iod_meta_read;
  97 long hammer2_iod_indr_read;
  98 long hammer2_iod_fmap_read;
  99 long hammer2_iod_volu_read;
 100 long hammer2_iod_file_write;
 101 long hammer2_iod_file_wembed;
 102 long hammer2_iod_file_wzero;
 103 long hammer2_iod_file_wdedup;
 104 long hammer2_iod_meta_write;
 105 long hammer2_iod_indr_write;
 106 long hammer2_iod_fmap_write;
 107 long hammer2_iod_volu_write;
 108
 109 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 110 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 111                 "Buffer used for compression.");
 112
 113 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 114 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 115                 "Buffer used for decompression.");
 116
 117 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 118
 119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 120            &hammer2_debug, 0, "");
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW,
 122            &hammer2_cluster_read, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 124            &hammer2_cluster_write, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 126            &hammer2_dedup_enable, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 128            &hammer2_always_compress, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 130            &hammer2_inval_enable, 0, "");
 131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 132            &hammer2_flush_pipe, 0, "");
 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 134            &hammer2_synchronous_flush, 0, "");
 135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 136            &hammer2_chain_allocs, 0, "");
 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 138            &hammer2_chain_frees, 0, "");
 139 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 140            &hammer2_limit_dirty_chains, 0, "");
 141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 142            &hammer2_count_modified_chains, 0, "");
 143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 144            &hammer2_dio_count, 0, "");
 145 SYSCTL_INT(_vfs_hammer2, OID_AUTO, limit_dio, CTLFLAG_RW,
 146            &hammer2_limit_dio, 0, "");
 147
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 149            &hammer2_iod_invals, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 151            &hammer2_iod_file_read, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 153            &hammer2_iod_meta_read, 0, "");
 154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 155            &hammer2_iod_indr_read, 0, "");
 156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 157            &hammer2_iod_fmap_read, 0, "");
 158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 159            &hammer2_iod_volu_read, 0, "");
 160
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 162            &hammer2_iod_file_write, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 164            &hammer2_iod_file_wembed, 0, "");
 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 166            &hammer2_iod_file_wzero, 0, "");
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 168            &hammer2_iod_file_wdedup, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 170            &hammer2_iod_meta_write, 0, "");
 171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 172            &hammer2_iod_indr_write, 0, "");
 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 174            &hammer2_iod_fmap_write, 0, "");
 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 176            &hammer2_iod_volu_write, 0, "");
 177
 178 long hammer2_check_icrc32;
 179 long hammer2_check_xxhash64;
 180 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 181            &hammer2_check_icrc32, 0, "");
 182 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 183            &hammer2_check_xxhash64, 0, "");
 184
 185 static int hammer2_vfs_init(struct vfsconf *conf);
 186 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 187 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 188                                 struct ucred *cred);
 189 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 190                                 struct vnode *, struct ucred *);
 191 static int hammer2_recovery(hammer2_dev_t *hmp);
 192 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 193 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 194 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 195                                 struct ucred *cred);
 196 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 197                                 struct ucred *cred);
 198 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 199                                 struct fid *fhp, struct vnode **vpp);
 200 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 201 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 202                                 int *exflagsp, struct ucred **credanonp);
 203
 204 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 205 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 206
 207 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 208
 209 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 210 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 211                                 hammer2_dev_t *hmp);
 212
 213 /*
 214  * HAMMER2 vfs operations.
 215  */
 216 static struct vfsops hammer2_vfsops = {
 217         .vfs_init       = hammer2_vfs_init,
 218         .vfs_uninit     = hammer2_vfs_uninit,
 219         .vfs_sync       = hammer2_vfs_sync,
 220         .vfs_mount      = hammer2_vfs_mount,
 221         .vfs_unmount    = hammer2_vfs_unmount,
 222         .vfs_root       = hammer2_vfs_root,
 223         .vfs_statfs     = hammer2_vfs_statfs,
 224         .vfs_statvfs    = hammer2_vfs_statvfs,
 225         .vfs_vget       = hammer2_vfs_vget,
 226         .vfs_vptofh     = hammer2_vfs_vptofh,
 227         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 228         .vfs_checkexp   = hammer2_vfs_checkexp
 229 };
 230
 231 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 232
 233 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 234 MODULE_VERSION(hammer2, 1);
 235
 236 static
 237 int
 238 hammer2_vfs_init(struct vfsconf *conf)
 239 {
 240         static struct objcache_malloc_args margs_read;
 241         static struct objcache_malloc_args margs_write;
 242         static struct objcache_malloc_args margs_vop;
 243
 244         int error;
 245
 246         error = 0;
 247
 248         /*
 249          * A large DIO cache is needed to retain dedup enablement masks.
 250          * The bulkfree code clears related masks as part of the disk block
 251          * recycling algorithm, preventing it from being used for a later
 252          * dedup.
 253          *
 254          * NOTE: A large buffer cache can actually interfere with dedup
 255          *       operation because we dedup based on media physical buffers
 256          *       and not logical buffers.  Try to make the DIO chace large
 257          *       enough to avoid this problem, but also cap it.
 258          */
 259         hammer2_limit_dio = nbuf * 2;
 260         if (hammer2_limit_dio > 100000)
 261                 hammer2_limit_dio = 100000;
 262
 263         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 264                 error = EINVAL;
 265         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 266                 error = EINVAL;
 267         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 268                 error = EINVAL;
 269
 270         if (error)
 271                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 272
 273         margs_read.objsize = 65536;
 274         margs_read.mtype = M_HAMMER2_DEBUFFER;
 275
 276         margs_write.objsize = 32768;
 277         margs_write.mtype = M_HAMMER2_CBUFFER;
 278
 279         margs_vop.objsize = sizeof(hammer2_xop_t);
 280         margs_vop.mtype = M_HAMMER2;
 281
 282         /*
 283          * Note thaht for the XOPS cache we want backing store allocations
 284          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 285          * confusion), so use the backing store function that does it.  This
 286          * means that initial XOPS objects are zerod but REUSED objects are
 287          * not.  So we are responsible for cleaning the object up sufficiently
 288          * for our needs before objcache_put()ing it back (typically just the
 289          * FIFO indices).
 290          */
 291         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 292                                 0, 1, NULL, NULL, NULL,
 293                                 objcache_malloc_alloc,
 294                                 objcache_malloc_free,
 295                                 &margs_read);
 296         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 297                                 0, 1, NULL, NULL, NULL,
 298                                 objcache_malloc_alloc,
 299                                 objcache_malloc_free,
 300                                 &margs_write);
 301         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 302                                 0, 1, NULL, NULL, NULL,
 303                                 objcache_malloc_alloc_zero,
 304                                 objcache_malloc_free,
 305                                 &margs_vop);
 306
 307
 308         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 309         TAILQ_INIT(&hammer2_mntlist);
 310         TAILQ_INIT(&hammer2_pfslist);
 311
 312         hammer2_limit_dirty_chains = maxvnodes / 10;
 313         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 314                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 315
 316         return (error);
 317 }
 318
 319 static
 320 int
 321 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 322 {
 323         objcache_destroy(cache_buffer_read);
 324         objcache_destroy(cache_buffer_write);
 325         objcache_destroy(cache_xops);
 326         return 0;
 327 }
 328
 329 /*
 330  * Core PFS allocator.  Used to allocate or reference the pmp structure
 331  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 332  * The pmp can be passed in or loaded by this function using the chain and
 333  * inode data.
 334  *
 335  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 336  * transactions.  Note that synchronization does not use this field.
 337  * (typically frontend operations and synchronization cannot run on the
 338  * same PFS node at the same time).
 339  *
 340  * XXX check locking
 341  */
 342 hammer2_pfs_t *
 343 hammer2_pfsalloc(hammer2_chain_t *chain,
 344                  const hammer2_inode_data_t *ripdata,
 345                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 346 {
 347         hammer2_pfs_t *pmp;
 348         hammer2_inode_t *iroot;
 349         int count;
 350         int i;
 351         int j;
 352
 353         pmp = NULL;
 354
 355         /*
 356          * Locate or create the PFS based on the cluster id.  If ripdata
 357          * is NULL this is a spmp which is unique and is always allocated.
 358          *
 359          * If the device is mounted in local mode all PFSs are considered
 360          * independent and not part of any cluster (for debugging only).
 361          */
 362         if (ripdata) {
 363                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 364                         if (force_local != pmp->force_local)
 365                                 continue;
 366                         if (force_local == NULL &&
 367                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 368                                  sizeof(pmp->pfs_clid)) == 0) {
 369                                         break;
 370                         } else if (force_local && pmp->pfs_names[0] &&
 371                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 372                                         break;
 373                         }
 374                 }
 375         }
 376
 377         if (pmp == NULL) {
 378                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 379                 pmp->force_local = force_local;
 380                 hammer2_trans_manage_init(pmp);
 381                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 382                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 383                 lockinit(&pmp->lock, "pfslk", 0, 0);
 384                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 385                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 386                 spin_init(&pmp->xop_spin, "h2xop");
 387                 spin_init(&pmp->lru_spin, "h2lru");
 388                 RB_INIT(&pmp->inum_tree);
 389                 TAILQ_INIT(&pmp->sideq);
 390                 TAILQ_INIT(&pmp->lru_list);
 391                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 392
 393                 /*
 394                  * Distribute backend operations to threads
 395                  */
 396                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 397                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 398
 399                 /*
 400                  * Save the last media transaction id for the flusher.  Set
 401                  * initial
 402                  */
 403                 if (ripdata)
 404                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 405                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 406
 407                 /*
 408                  * The synchronization thread may start too early, make
 409                  * sure it stays frozen until we are ready to let it go.
 410                  * XXX
 411                  */
 412                 /*
 413                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 414                                          HAMMER2_THREAD_REMASTER;
 415                 */
 416         }
 417
 418         /*
 419          * Create the PFS's root inode and any missing XOP helper threads.
 420          */
 421         if ((iroot = pmp->iroot) == NULL) {
 422                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 423                 if (ripdata)
 424                         iroot->meta = ripdata->meta;
 425                 pmp->iroot = iroot;
 426                 hammer2_inode_ref(iroot);
 427                 hammer2_inode_unlock(iroot);
 428         }
 429
 430         /*
 431          * Stop here if no chain is passed in.
 432          */
 433         if (chain == NULL)
 434                 goto done;
 435
 436         /*
 437          * When a chain is passed in we must add it to the PFS's root
 438          * inode, update pmp->pfs_types[], and update the syncronization
 439          * threads.
 440          *
 441          * When forcing local mode, mark the PFS as a MASTER regardless.
 442          *
 443          * At the moment empty spots can develop due to removals or failures.
 444          * Ultimately we want to re-fill these spots but doing so might
 445          * confused running code. XXX
 446          */
 447         hammer2_inode_ref(iroot);
 448         hammer2_mtx_ex(&iroot->lock);
 449         j = iroot->cluster.nchains;
 450
 451         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 452
 453         if (j == HAMMER2_MAXCLUSTER) {
 454                 kprintf("hammer2_mount: cluster full!\n");
 455                 /* XXX fatal error? */
 456         } else {
 457                 KKASSERT(chain->pmp == NULL);
 458                 chain->pmp = pmp;
 459                 hammer2_chain_ref(chain);
 460                 iroot->cluster.array[j].chain = chain;
 461                 if (force_local)
 462                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 463                 else
 464                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 465                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 466                 pmp->pfs_hmps[j] = chain->hmp;
 467
 468                 /*
 469                  * If the PFS is already mounted we must account
 470                  * for the mount_count here.
 471                  */
 472                 if (pmp->mp)
 473                         ++chain->hmp->mount_count;
 474
 475                 /*
 476                  * May have to fixup dirty chain tracking.  Previous
 477                  * pmp was NULL so nothing to undo.
 478                  */
 479                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 480                         hammer2_pfs_memory_inc(pmp);
 481                 ++j;
 482         }
 483         iroot->cluster.nchains = j;
 484
 485         /*
 486          * Update nmasters from any PFS inode which is part of the cluster.
 487          * It is possible that this will result in a value which is too
 488          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 489          * override this value later on.
 490          *
 491          * (This informs us of masters that might not currently be
 492          *  discoverable by this mount).
 493          */
 494         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 495                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 496         }
 497
 498         /*
 499          * Count visible masters.  Masters are usually added with
 500          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 501          * are more (XXX and must update the master inodes).
 502          */
 503         count = 0;
 504         for (i = 0; i < iroot->cluster.nchains; ++i) {
 505                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 506                         ++count;
 507         }
 508         if (pmp->pfs_nmasters < count)
 509                 pmp->pfs_nmasters = count;
 510
 511         /*
 512          * Create missing synchronization and support threads.
 513          *
 514          * Single-node masters (including snapshots) have nothing to
 515          * synchronize and do not require this thread.
 516          *
 517          * Multi-node masters or any number of soft masters, slaves, copy,
 518          * or other PFS types need the thread.
 519          *
 520          * Each thread is responsible for its particular cluster index.
 521          * We use independent threads so stalls or mismatches related to
 522          * any given target do not affect other targets.
 523          */
 524         for (i = 0; i < iroot->cluster.nchains; ++i) {
 525                 /*
 526                  * Single-node masters (including snapshots) have nothing
 527                  * to synchronize and will make direct xops support calls,
 528                  * thus they do not require this thread.
 529                  *
 530                  * Note that there can be thousands of snapshots.  We do not
 531                  * want to create thousands of threads.
 532                  */
 533                 if (pmp->pfs_nmasters <= 1 &&
 534                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 535                         continue;
 536                 }
 537
 538                 /*
 539                  * Sync support thread
 540                  */
 541                 if (pmp->sync_thrs[i].td == NULL) {
 542                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 543                                            "h2nod", i, -1,
 544                                            hammer2_primary_sync_thread);
 545                 }
 546         }
 547
 548         /*
 549          * Create missing Xop threads
 550          *
 551          * NOTE: We create helper threads for all mounted PFSs or any
 552          *       PFSs with 2+ nodes (so the sync thread can update them,
 553          *       even if not mounted).
 554          */
 555         if (pmp->mp || iroot->cluster.nchains >= 2)
 556                 hammer2_xop_helper_create(pmp);
 557
 558         hammer2_mtx_unlock(&iroot->lock);
 559         hammer2_inode_drop(iroot);
 560 done:
 561         return pmp;
 562 }
 563
 564 /*
 565  * Deallocate an element of a probed PFS.  If destroying and this is a
 566  * MASTER, adjust nmasters.
 567  *
 568  * This function does not physically destroy the PFS element in its device
 569  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 570  */
 571 void
 572 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 573 {
 574         hammer2_inode_t *iroot;
 575         hammer2_chain_t *chain;
 576         int j;
 577
 578         /*
 579          * Cleanup our reference on iroot.  iroot is (should) not be needed
 580          * by the flush code.
 581          */
 582         iroot = pmp->iroot;
 583         if (iroot) {
 584                 /*
 585                  * Stop synchronizing
 586                  *
 587                  * XXX flush after acquiring the iroot lock.
 588                  * XXX clean out the cluster index from all inode structures.
 589                  */
 590                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 591
 592                 /*
 593                  * Remove the cluster index from the group.  If destroying
 594                  * the PFS and this is a master, adjust pfs_nmasters.
 595                  */
 596                 hammer2_mtx_ex(&iroot->lock);
 597                 chain = iroot->cluster.array[clindex].chain;
 598                 iroot->cluster.array[clindex].chain = NULL;
 599
 600                 switch(pmp->pfs_types[clindex]) {
 601                 case HAMMER2_PFSTYPE_MASTER:
 602                         if (destroying && pmp->pfs_nmasters > 0)
 603                                 --pmp->pfs_nmasters;
 604                         /* XXX adjust ripdata->meta.pfs_nmasters */
 605                         break;
 606                 default:
 607                         break;
 608                 }
 609                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 610
 611                 hammer2_mtx_unlock(&iroot->lock);
 612
 613                 /*
 614                  * Release the chain.
 615                  */
 616                 if (chain) {
 617                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 618                         hammer2_chain_drop(chain);
 619                 }
 620
 621                 /*
 622                  * Terminate all XOP threads for the cluster index.
 623                  */
 624                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 625                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 626         }
 627 }
 628
 629 /*
 630  * Destroy a PFS, typically only occurs after the last mount on a device
 631  * has gone away.
 632  */
 633 static void
 634 hammer2_pfsfree(hammer2_pfs_t *pmp)
 635 {
 636         hammer2_inode_t *iroot;
 637         hammer2_chain_t *chain;
 638         int i;
 639         int j;
 640
 641         /*
 642          * Cleanup our reference on iroot.  iroot is (should) not be needed
 643          * by the flush code.
 644          */
 645         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 646
 647         iroot = pmp->iroot;
 648         if (iroot) {
 649                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 650                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 651                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 652                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 653                 }
 654 #if REPORT_REFS_ERRORS
 655                 if (pmp->iroot->refs != 1)
 656                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 657                                 pmp->iroot, pmp->iroot->refs);
 658 #else
 659                 KKASSERT(pmp->iroot->refs == 1);
 660 #endif
 661                 /* ref for pmp->iroot */
 662                 hammer2_inode_drop(pmp->iroot);
 663                 pmp->iroot = NULL;
 664         }
 665
 666         /*
 667          * Cleanup chains remaining on LRU list.
 668          */
 669         kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
 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                         kprintf("unmount hmp %p remove spmp %p\n",
 707                                 hmp, pmp);
 708                         hmp->spmp = NULL;
 709                 }
 710
 711                 /*
 712                  * Determine if this PFS is affected.  If it is we must
 713                  * freeze all management threads and lock its iroot.
 714                  *
 715                  * Freezing a management thread forces it idle, operations
 716                  * in-progress will be aborted and it will have to start
 717                  * over again when unfrozen, or exit if told to exit.
 718                  */
 719                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 720                         if (pmp->pfs_hmps[i] == hmp)
 721                                 break;
 722                 }
 723                 if (i != HAMMER2_MAXCLUSTER) {
 724                         /*
 725                          * Make sure all synchronization threads are locked
 726                          * down.
 727                          */
 728                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 729                                 if (pmp->pfs_hmps[i] == NULL)
 730                                         continue;
 731                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 732                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 733                                         hammer2_thr_freeze_async(
 734                                                 &pmp->xop_groups[j].thrs[i]);
 735                                 }
 736                         }
 737                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 738                                 if (pmp->pfs_hmps[i] == NULL)
 739                                         continue;
 740                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 741                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 742                                         hammer2_thr_freeze(
 743                                                 &pmp->xop_groups[j].thrs[i]);
 744                                 }
 745                         }
 746
 747                         /*
 748                          * Lock the inode and clean out matching chains.
 749                          * Note that we cannot use hammer2_inode_lock_*()
 750                          * here because that would attempt to validate the
 751                          * cluster that we are in the middle of ripping
 752                          * apart.
 753                          *
 754                          * WARNING! We are working directly on the inodes
 755                          *          embedded cluster.
 756                          */
 757                         hammer2_mtx_ex(&iroot->lock);
 758
 759                         /*
 760                          * Remove the chain from matching elements of the PFS.
 761                          */
 762                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 763                                 if (pmp->pfs_hmps[i] != hmp)
 764                                         continue;
 765                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 766                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 767                                         hammer2_thr_delete(
 768                                                 &pmp->xop_groups[j].thrs[i]);
 769                                 }
 770                                 rchain = iroot->cluster.array[i].chain;
 771                                 iroot->cluster.array[i].chain = NULL;
 772                                 pmp->pfs_types[i] = 0;
 773                                 if (pmp->pfs_names[i]) {
 774                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 775                                         pmp->pfs_names[i] = NULL;
 776                                 }
 777                                 if (rchain) {
 778                                         hammer2_chain_drop(rchain);
 779                                         /* focus hint */
 780                                         if (iroot->cluster.focus == rchain)
 781                                                 iroot->cluster.focus = NULL;
 782                                 }
 783                                 pmp->pfs_hmps[i] = NULL;
 784                         }
 785                         hammer2_mtx_unlock(&iroot->lock);
 786                         didfreeze = 1;  /* remaster, unfreeze down below */
 787                 } else {
 788                         didfreeze = 0;
 789                 }
 790
 791                 /*
 792                  * Cleanup trailing chains.  Gaps may remain.
 793                  */
 794                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 795                         if (pmp->pfs_hmps[i])
 796                                 break;
 797                 }
 798                 iroot->cluster.nchains = i + 1;
 799
 800                 /*
 801                  * If the PMP has no elements remaining we can destroy it.
 802                  * (this will transition management threads from frozen->exit).
 803                  */
 804                 if (iroot->cluster.nchains == 0) {
 805                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 806                                 hmp, pmp);
 807                         hammer2_pfsfree(pmp);
 808                         goto again;
 809                 }
 810
 811                 /*
 812                  * If elements still remain we need to set the REMASTER
 813                  * flag and unfreeze it.
 814                  */
 815                 if (didfreeze) {
 816                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 817                                 if (pmp->pfs_hmps[i] == NULL)
 818                                         continue;
 819                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 820                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 821                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 822                                         hammer2_thr_remaster(
 823                                                 &pmp->xop_groups[j].thrs[i]);
 824                                         hammer2_thr_unfreeze(
 825                                                 &pmp->xop_groups[j].thrs[i]);
 826                                 }
 827                         }
 828                 }
 829         }
 830 }
 831
 832 /*
 833  * Mount or remount HAMMER2 fileystem from physical media
 834  *
 835  *      mountroot
 836  *              mp              mount point structure
 837  *              path            NULL
 838  *              data            <unused>
 839  *              cred            <unused>
 840  *
 841  *      mount
 842  *              mp              mount point structure
 843  *              path            path to mount point
 844  *              data            pointer to argument structure in user space
 845  *                      volume  volume path (device@LABEL form)
 846  *                      hflags  user mount flags
 847  *              cred            user credentials
 848  *
 849  * RETURNS:     0       Success
 850  *              !0      error number
 851  */
 852 static
 853 int
 854 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 855                   struct ucred *cred)
 856 {
 857         struct hammer2_mount_info info;
 858         hammer2_pfs_t *pmp;
 859         hammer2_pfs_t *spmp;
 860         hammer2_dev_t *hmp;
 861         hammer2_dev_t *force_local;
 862         hammer2_key_t key_next;
 863         hammer2_key_t key_dummy;
 864         hammer2_key_t lhc;
 865         struct vnode *devvp;
 866         struct nlookupdata nd;
 867         hammer2_chain_t *parent;
 868         hammer2_chain_t *chain;
 869         hammer2_cluster_t *cluster;
 870         const hammer2_inode_data_t *ripdata;
 871         hammer2_blockref_t bref;
 872         struct file *fp;
 873         char devstr[MNAMELEN];
 874         size_t size;
 875         size_t done;
 876         char *dev;
 877         char *label;
 878         int ronly = 1;
 879         int error;
 880         int i;
 881
 882         hmp = NULL;
 883         pmp = NULL;
 884         dev = NULL;
 885         label = NULL;
 886         devvp = NULL;
 887
 888         kprintf("hammer2_mount\n");
 889
 890         if (path == NULL) {
 891                 /*
 892                  * Root mount
 893                  */
 894                 bzero(&info, sizeof(info));
 895                 info.cluster_fd = -1;
 896                 ksnprintf(devstr, sizeof(devstr), "%s",
 897                           mp->mnt_stat.f_mntfromname);
 898                 kprintf("hammer2_mount: root '%s'\n", devstr);
 899         } else {
 900                 /*
 901                  * Non-root mount or updating a mount
 902                  */
 903                 error = copyin(data, &info, sizeof(info));
 904                 if (error)
 905                         return (error);
 906
 907                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 908                 if (error)
 909                         return (error);
 910         }
 911
 912         /*
 913          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 914          * if no label specified, based on the partition id.  Error out if no
 915          * partition id.  This is strictly a convenience to match the
 916          * default label created by newfs_hammer2, our preference is
 917          * that a label always be specified.
 918          */
 919         dev = devstr;
 920         label = strchr(devstr, '@');
 921         if (label && ((label + 1) - dev) > done)
 922                 return (EINVAL);
 923         if (label && label == devstr)
 924                 return (EINVAL);
 925         if (label == NULL || label[1] == 0) {
 926                 char slice;
 927
 928                 if (label == NULL)
 929                         label = devstr + strlen(devstr);
 930                 slice = label[-1];
 931                 switch(slice) {
 932                 case 'a':
 933                         label = "BOOT";
 934                         break;
 935                 case 'd':
 936                         label = "ROOT";
 937                         break;
 938                 default:
 939                         label = "DATA";
 940                         break;
 941                 }
 942         } else {
 943                 *label = '\0';
 944                 label++;
 945         }
 946
 947         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
 948                 dev, label, (mp->mnt_flag & MNT_RDONLY));
 949
 950         if (mp->mnt_flag & MNT_UPDATE) {
 951                 /*
 952                  * Update mount.  Note that pmp->iroot->cluster is
 953                  * an inode-embedded cluster and thus cannot be
 954                  * directly locked.
 955                  *
 956                  * XXX HAMMER2 needs to implement NFS export via
 957                  *     mountctl.
 958                  */
 959                 pmp = MPTOPMP(mp);
 960                 pmp->hflags = info.hflags;
 961                 cluster = &pmp->iroot->cluster;
 962                 for (i = 0; i < cluster->nchains; ++i) {
 963                         if (cluster->array[i].chain == NULL)
 964                                 continue;
 965                         hmp = cluster->array[i].chain->hmp;
 966                         devvp = hmp->devvp;
 967                         error = hammer2_remount(hmp, mp, path,
 968                                                 devvp, cred);
 969                         if (error)
 970                                 break;
 971                 }
 972
 973                 return error;
 974         }
 975
 976         /*
 977          * HMP device mount
 978          *
 979          * If a path is specified and dev is not an empty string, lookup the
 980          * name and verify that it referes to a block device.
 981          *
 982          * If a path is specified and dev is an empty string we fall through
 983          * and locate the label in the hmp search.
 984          */
 985         if (path && *dev != 0) {
 986                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 987                 if (error == 0)
 988                         error = nlookup(&nd);
 989                 if (error == 0)
 990                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 991                 nlookup_done(&nd);
 992         } else if (path == NULL) {
 993                 /* root mount */
 994                 cdev_t cdev = kgetdiskbyname(dev);
 995                 error = bdevvp(cdev, &devvp);
 996                 if (error)
 997                         kprintf("hammer2: cannot find '%s'\n", dev);
 998         } else {
 999                 /*
1000                  * We will locate the hmp using the label in the hmp loop.
1001                  */
1002                 error = 0;
1003         }
1004
1005         /*
1006          * Make sure its a block device.  Do not check to see if it is
1007          * already mounted until we determine that its a fresh H2 device.
1008          */
1009         if (error == 0 && devvp) {
1010                 vn_isdisk(devvp, &error);
1011         }
1012
1013         /*
1014          * Determine if the device has already been mounted.  After this
1015          * check hmp will be non-NULL if we are doing the second or more
1016          * hammer2 mounts from the same device.
1017          */
1018         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1019         if (devvp) {
1020                 /*
1021                  * Match the device.  Due to the way devfs works,
1022                  * we may not be able to directly match the vnode pointer,
1023                  * so also check to see if the underlying device matches.
1024                  */
1025                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1026                         if (hmp->devvp == devvp)
1027                                 break;
1028                         if (devvp->v_rdev &&
1029                             hmp->devvp->v_rdev == devvp->v_rdev) {
1030                                 break;
1031                         }
1032                 }
1033
1034                 /*
1035                  * If no match this may be a fresh H2 mount, make sure
1036                  * the device is not mounted on anything else.
1037                  */
1038                 if (hmp == NULL)
1039                         error = vfs_mountedon(devvp);
1040         } else if (error == 0) {
1041                 /*
1042                  * Match the label to a pmp already probed.
1043                  */
1044                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1045                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1046                                 if (pmp->pfs_names[i] &&
1047                                     strcmp(pmp->pfs_names[i], label) == 0) {
1048                                         hmp = pmp->pfs_hmps[i];
1049                                         break;
1050                                 }
1051                         }
1052                         if (hmp)
1053                                 break;
1054                 }
1055                 if (hmp == NULL)
1056                         error = ENOENT;
1057         }
1058
1059         /*
1060          * Open the device if this isn't a secondary mount and construct
1061          * the H2 device mount (hmp).
1062          */
1063         if (hmp == NULL) {
1064                 hammer2_chain_t *schain;
1065                 hammer2_xid_t xid;
1066
1067                 if (error == 0 && vcount(devvp) > 0) {
1068                         kprintf("Primary device already has references\n");
1069                         error = EBUSY;
1070                 }
1071
1072                 /*
1073                  * Now open the device
1074                  */
1075                 if (error == 0) {
1076                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1077                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1078                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1079                         if (error == 0) {
1080                                 error = VOP_OPEN(devvp,
1081                                              (ronly ? FREAD : FREAD | FWRITE),
1082                                              FSCRED, NULL);
1083                         }
1084                         vn_unlock(devvp);
1085                 }
1086                 if (error && devvp) {
1087                         vrele(devvp);
1088                         devvp = NULL;
1089                 }
1090                 if (error) {
1091                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1092                         return error;
1093                 }
1094                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1095                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1096                 hmp->ronly = ronly;
1097                 hmp->devvp = devvp;
1098                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1099                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1100                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1101                 RB_INIT(&hmp->iotree);
1102                 spin_init(&hmp->io_spin, "hm2mount_io");
1103                 spin_init(&hmp->list_spin, "hm2mount_list");
1104                 TAILQ_INIT(&hmp->flushq);
1105
1106                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1107                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1108                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1109
1110                 /*
1111                  * vchain setup. vchain.data is embedded.
1112                  * vchain.refs is initialized and will never drop to 0.
1113                  *
1114                  * NOTE! voldata is not yet loaded.
1115                  */
1116                 hmp->vchain.hmp = hmp;
1117                 hmp->vchain.refs = 1;
1118                 hmp->vchain.data = (void *)&hmp->voldata;
1119                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1120                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1121                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1122
1123                 hammer2_chain_core_init(&hmp->vchain);
1124                 /* hmp->vchain.u.xxx is left NULL */
1125
1126                 /*
1127                  * fchain setup.  fchain.data is embedded.
1128                  * fchain.refs is initialized and will never drop to 0.
1129                  *
1130                  * The data is not used but needs to be initialized to
1131                  * pass assertion muster.  We use this chain primarily
1132                  * as a placeholder for the freemap's top-level RBTREE
1133                  * so it does not interfere with the volume's topology
1134                  * RBTREE.
1135                  */
1136                 hmp->fchain.hmp = hmp;
1137                 hmp->fchain.refs = 1;
1138                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1139                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1140                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1141                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1142                 hmp->fchain.bref.methods =
1143                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1144                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1145
1146                 hammer2_chain_core_init(&hmp->fchain);
1147                 /* hmp->fchain.u.xxx is left NULL */
1148
1149                 /*
1150                  * Install the volume header and initialize fields from
1151                  * voldata.
1152                  */
1153                 error = hammer2_install_volume_header(hmp);
1154                 if (error) {
1155                         hammer2_unmount_helper(mp, NULL, hmp);
1156                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1157                         hammer2_vfs_unmount(mp, MNT_FORCE);
1158                         return error;
1159                 }
1160
1161                 /*
1162                  * Really important to get these right or flush will get
1163                  * confused.
1164                  */
1165                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1166                 kprintf("alloc spmp %p tid %016jx\n",
1167                         hmp->spmp, hmp->voldata.mirror_tid);
1168                 spmp = hmp->spmp;
1169
1170                 /*
1171                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1172                  * is inherited from the volume header.
1173                  */
1174                 xid = 0;
1175                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1176                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1177                 hmp->vchain.pmp = spmp;
1178                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1179                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1180                 hmp->fchain.pmp = spmp;
1181
1182                 /*
1183                  * First locate the super-root inode, which is key 0
1184                  * relative to the volume header's blockset.
1185                  *
1186                  * Then locate the root inode by scanning the directory keyspace
1187                  * represented by the label.
1188                  */
1189                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1190                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1191                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1192                                       &error, 0);
1193                 hammer2_chain_lookup_done(parent);
1194                 if (schain == NULL) {
1195                         kprintf("hammer2_mount: invalid super-root\n");
1196                         hammer2_unmount_helper(mp, NULL, hmp);
1197                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1198                         hammer2_vfs_unmount(mp, MNT_FORCE);
1199                         return EINVAL;
1200                 }
1201                 if (schain->error) {
1202                         kprintf("hammer2_mount: error %s reading super-root\n",
1203                                 hammer2_error_str(schain->error));
1204                         hammer2_chain_unlock(schain);
1205                         hammer2_chain_drop(schain);
1206                         schain = NULL;
1207                         hammer2_unmount_helper(mp, NULL, hmp);
1208                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1209                         hammer2_vfs_unmount(mp, MNT_FORCE);
1210                         return EINVAL;
1211                 }
1212
1213                 /*
1214                  * The super-root always uses an inode_tid of 1 when
1215                  * creating PFSs.
1216                  */
1217                 spmp->inode_tid = 1;
1218                 spmp->modify_tid = schain->bref.modify_tid + 1;
1219
1220                 /*
1221                  * Sanity-check schain's pmp and finish initialization.
1222                  * Any chain belonging to the super-root topology should
1223                  * have a NULL pmp (not even set to spmp).
1224                  */
1225                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1226                 KKASSERT(schain->pmp == NULL);
1227                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1228
1229                 /*
1230                  * Replace the dummy spmp->iroot with a real one.  It's
1231                  * easier to just do a wholesale replacement than to try
1232                  * to update the chain and fixup the iroot fields.
1233                  *
1234                  * The returned inode is locked with the supplied cluster.
1235                  */
1236                 cluster = hammer2_cluster_from_chain(schain);
1237                 hammer2_inode_drop(spmp->iroot);
1238                 spmp->iroot = NULL;
1239                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1240                 spmp->spmp_hmp = hmp;
1241                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1242                 spmp->pfs_hmps[0] = hmp;
1243                 hammer2_inode_ref(spmp->iroot);
1244                 hammer2_inode_unlock(spmp->iroot);
1245                 hammer2_cluster_unlock(cluster);
1246                 hammer2_cluster_drop(cluster);
1247                 schain = NULL;
1248                 /* leave spmp->iroot with one ref */
1249
1250                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1251                         error = hammer2_recovery(hmp);
1252                         /* XXX do something with error */
1253                 }
1254                 hammer2_update_pmps(hmp);
1255                 hammer2_iocom_init(hmp);
1256                 hammer2_bulkfree_init(hmp);
1257
1258                 /*
1259                  * Ref the cluster management messaging descriptor.  The mount
1260                  * program deals with the other end of the communications pipe.
1261                  *
1262                  * Root mounts typically do not supply one.
1263                  */
1264                 if (info.cluster_fd >= 0) {
1265                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1266                         if (fp) {
1267                                 hammer2_cluster_reconnect(hmp, fp);
1268                         } else {
1269                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1270                         }
1271                 }
1272         } else {
1273                 spmp = hmp->spmp;
1274                 if (info.hflags & HMNT2_DEVFLAGS) {
1275                         kprintf("hammer2: Warning: mount flags pertaining "
1276                                 "to the whole device may only be specified "
1277                                 "on the first mount of the device: %08x\n",
1278                                 info.hflags & HMNT2_DEVFLAGS);
1279                 }
1280         }
1281
1282         /*
1283          * Force local mount (disassociate all PFSs from their clusters).
1284          * Used primarily for debugging.
1285          */
1286         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1287
1288         /*
1289          * Lookup the mount point under the media-localized super-root.
1290          * Scanning hammer2_pfslist doesn't help us because it represents
1291          * PFS cluster ids which can aggregate several named PFSs together.
1292          *
1293          * cluster->pmp will incorrectly point to spmp and must be fixed
1294          * up later on.
1295          */
1296         hammer2_inode_lock(spmp->iroot, 0);
1297         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1298         lhc = hammer2_dirhash(label, strlen(label));
1299         chain = hammer2_chain_lookup(&parent, &key_next,
1300                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1301                                      &error, 0);
1302         while (chain) {
1303                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1304                     strcmp(label, chain->data->ipdata.filename) == 0) {
1305                         break;
1306                 }
1307                 chain = hammer2_chain_next(&parent, chain, &key_next,
1308                                             key_next,
1309                                             lhc + HAMMER2_DIRHASH_LOMASK,
1310                                             &error, 0);
1311         }
1312         if (parent) {
1313                 hammer2_chain_unlock(parent);
1314                 hammer2_chain_drop(parent);
1315         }
1316         hammer2_inode_unlock(spmp->iroot);
1317
1318         /*
1319          * PFS could not be found?
1320          */
1321         if (chain == NULL) {
1322                 if (error)
1323                         kprintf("hammer2_mount: PFS label I/O error\n");
1324                 else
1325                         kprintf("hammer2_mount: PFS label not found\n");
1326                 hammer2_unmount_helper(mp, NULL, hmp);
1327                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1328                 hammer2_vfs_unmount(mp, MNT_FORCE);
1329
1330                 return EINVAL;
1331         }
1332
1333         /*
1334          * Acquire the pmp structure (it should have already been allocated
1335          * via hammer2_update_pmps() so do not pass cluster in to add to
1336          * available chains).
1337          *
1338          * Check if the cluster has already been mounted.  A cluster can
1339          * only be mounted once, use null mounts to mount additional copies.
1340          */
1341         if (chain->error) {
1342                 kprintf("hammer2_mount: PFS label I/O error\n");
1343         } else {
1344                 ripdata = &chain->data->ipdata;
1345                 bref = chain->bref;
1346                 pmp = hammer2_pfsalloc(NULL, ripdata,
1347                                        bref.modify_tid, force_local);
1348         }
1349         hammer2_chain_unlock(chain);
1350         hammer2_chain_drop(chain);
1351
1352         /*
1353          * Finish the mount
1354          */
1355         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1356
1357         if (pmp->mp) {
1358                 kprintf("hammer2_mount: PFS already mounted!\n");
1359                 hammer2_unmount_helper(mp, NULL, hmp);
1360                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1361                 hammer2_vfs_unmount(mp, MNT_FORCE);
1362
1363                 return EBUSY;
1364         }
1365
1366         pmp->hflags = info.hflags;
1367         mp->mnt_flag |= MNT_LOCAL;
1368         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1369         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1370
1371         /*
1372          * required mount structure initializations
1373          */
1374         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1375         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1376
1377         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1378         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1379
1380         /*
1381          * Optional fields
1382          */
1383         mp->mnt_iosize_max = MAXPHYS;
1384
1385         /*
1386          * Connect up mount pointers.
1387          */
1388         hammer2_mount_helper(mp, pmp);
1389
1390         lockmgr(&hammer2_mntlk, LK_RELEASE);
1391
1392         /*
1393          * Finish setup
1394          */
1395         vfs_getnewfsid(mp);
1396         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1397         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1398         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1399
1400         if (path) {
1401                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1402                           MNAMELEN - 1, &size);
1403                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1404         } /* else root mount, already in there */
1405
1406         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1407         if (path) {
1408                 copyinstr(path, mp->mnt_stat.f_mntonname,
1409                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1410                           &size);
1411         } else {
1412                 /* root mount */
1413                 mp->mnt_stat.f_mntonname[0] = '/';
1414         }
1415
1416         /*
1417          * Initial statfs to prime mnt_stat.
1418          */
1419         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1420
1421         return 0;
1422 }
1423
1424 /*
1425  * Scan PFSs under the super-root and create hammer2_pfs structures.
1426  */
1427 static
1428 void
1429 hammer2_update_pmps(hammer2_dev_t *hmp)
1430 {
1431         const hammer2_inode_data_t *ripdata;
1432         hammer2_chain_t *parent;
1433         hammer2_chain_t *chain;
1434         hammer2_blockref_t bref;
1435         hammer2_dev_t *force_local;
1436         hammer2_pfs_t *spmp;
1437         hammer2_pfs_t *pmp;
1438         hammer2_key_t key_next;
1439         int error;
1440
1441         /*
1442          * Force local mount (disassociate all PFSs from their clusters).
1443          * Used primarily for debugging.
1444          */
1445         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1446
1447         /*
1448          * Lookup mount point under the media-localized super-root.
1449          *
1450          * cluster->pmp will incorrectly point to spmp and must be fixed
1451          * up later on.
1452          */
1453         spmp = hmp->spmp;
1454         hammer2_inode_lock(spmp->iroot, 0);
1455         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1456         chain = hammer2_chain_lookup(&parent, &key_next,
1457                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1458                                          &error, 0);
1459         while (chain) {
1460                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1461                         continue;
1462                 if (chain->error) {
1463                         kprintf("I/O error scanning PFS labels\n");
1464                 } else {
1465                         ripdata = &chain->data->ipdata;
1466                         bref = chain->bref;
1467                         kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1468
1469                         pmp = hammer2_pfsalloc(chain, ripdata,
1470                                                bref.modify_tid, force_local);
1471                 }
1472                 chain = hammer2_chain_next(&parent, chain, &key_next,
1473                                            key_next, HAMMER2_KEY_MAX,
1474                                            &error, 0);
1475         }
1476         if (parent) {
1477                 hammer2_chain_unlock(parent);
1478                 hammer2_chain_drop(parent);
1479         }
1480         hammer2_inode_unlock(spmp->iroot);
1481 }
1482
1483 static
1484 int
1485 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1486                 struct vnode *devvp, struct ucred *cred)
1487 {
1488         int error;
1489
1490         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1491                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1492                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1493                 vn_unlock(devvp);
1494                 error = hammer2_recovery(hmp);
1495                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1496                 if (error == 0) {
1497                         VOP_CLOSE(devvp, FREAD, NULL);
1498                         hmp->ronly = 0;
1499                 } else {
1500                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1501                 }
1502                 vn_unlock(devvp);
1503         } else {
1504                 error = 0;
1505         }
1506         return error;
1507 }
1508
1509 static
1510 int
1511 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1512 {
1513         hammer2_pfs_t *pmp;
1514         int flags;
1515         int error = 0;
1516
1517         pmp = MPTOPMP(mp);
1518
1519         if (pmp == NULL)
1520                 return(0);
1521
1522         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1523
1524         /*
1525          * If mount initialization proceeded far enough we must flush
1526          * its vnodes and sync the underlying mount points.  Three syncs
1527          * are required to fully flush the filesystem (freemap updates lag
1528          * by one flush, and one extra for safety).
1529          */
1530         if (mntflags & MNT_FORCE)
1531                 flags = FORCECLOSE;
1532         else
1533                 flags = 0;
1534         if (pmp->iroot) {
1535                 error = vflush(mp, 0, flags);
1536                 if (error)
1537                         goto failed;
1538                 hammer2_vfs_sync(mp, MNT_WAIT);
1539                 hammer2_vfs_sync(mp, MNT_WAIT);
1540                 hammer2_vfs_sync(mp, MNT_WAIT);
1541         }
1542
1543         /*
1544          * Cleanup the frontend support XOPS threads
1545          */
1546         hammer2_xop_helper_cleanup(pmp);
1547
1548         if (pmp->mp)
1549                 hammer2_unmount_helper(mp, pmp, NULL);
1550
1551         error = 0;
1552 failed:
1553         lockmgr(&hammer2_mntlk, LK_RELEASE);
1554
1555         return (error);
1556 }
1557
1558 /*
1559  * Mount helper, hook the system mount into our PFS.
1560  * The mount lock is held.
1561  *
1562  * We must bump the mount_count on related devices for any
1563  * mounted PFSs.
1564  */
1565 static
1566 void
1567 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1568 {
1569         hammer2_cluster_t *cluster;
1570         hammer2_chain_t *rchain;
1571         int i;
1572
1573         mp->mnt_data = (qaddr_t)pmp;
1574         pmp->mp = mp;
1575
1576         /*
1577          * After pmp->mp is set we have to adjust hmp->mount_count.
1578          */
1579         cluster = &pmp->iroot->cluster;
1580         for (i = 0; i < cluster->nchains; ++i) {
1581                 rchain = cluster->array[i].chain;
1582                 if (rchain == NULL)
1583                         continue;
1584                 ++rchain->hmp->mount_count;
1585                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1586                         rchain->hmp, rchain->hmp->mount_count);
1587         }
1588
1589         /*
1590          * Create missing Xop threads
1591          */
1592         hammer2_xop_helper_create(pmp);
1593 }
1594
1595 /*
1596  * Mount helper, unhook the system mount from our PFS.
1597  * The mount lock is held.
1598  *
1599  * If hmp is supplied a mount responsible for being the first to open
1600  * the block device failed and the block device and all PFSs using the
1601  * block device must be cleaned up.
1602  *
1603  * If pmp is supplied multiple devices might be backing the PFS and each
1604  * must be disconnected.  This might not be the last PFS using some of the
1605  * underlying devices.  Also, we have to adjust our hmp->mount_count
1606  * accounting for the devices backing the pmp which is now undergoing an
1607  * unmount.
1608  */
1609 static
1610 void
1611 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1612 {
1613         hammer2_cluster_t *cluster;
1614         hammer2_chain_t *rchain;
1615         struct vnode *devvp;
1616         int dumpcnt;
1617         int ronly;
1618         int i;
1619
1620         /*
1621          * If no device supplied this is a high-level unmount and we have to
1622          * to disconnect the mount, adjust mount_count, and locate devices
1623          * that might now have no mounts.
1624          */
1625         if (pmp) {
1626                 KKASSERT(hmp == NULL);
1627                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1628                 pmp->mp = NULL;
1629                 mp->mnt_data = NULL;
1630
1631                 /*
1632                  * After pmp->mp is cleared we have to account for
1633                  * mount_count.
1634                  */
1635                 cluster = &pmp->iroot->cluster;
1636                 for (i = 0; i < cluster->nchains; ++i) {
1637                         rchain = cluster->array[i].chain;
1638                         if (rchain == NULL)
1639                                 continue;
1640                         --rchain->hmp->mount_count;
1641                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1642                                 rchain->hmp, rchain->hmp->mount_count);
1643                         /* scrapping hmp now may invalidate the pmp */
1644                 }
1645 again:
1646                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1647                         if (hmp->mount_count == 0) {
1648                                 hammer2_unmount_helper(NULL, NULL, hmp);
1649                                 goto again;
1650                         }
1651                 }
1652                 return;
1653         }
1654
1655         /*
1656          * Try to terminate the block device.  We can't terminate it if
1657          * there are still PFSs referencing it.
1658          */
1659         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1660                 hmp, hmp->mount_count);
1661         if (hmp->mount_count)
1662                 return;
1663
1664         hammer2_bulkfree_uninit(hmp);
1665         hammer2_pfsfree_scan(hmp);
1666         hammer2_dev_exlock(hmp);        /* XXX order */
1667
1668         /*
1669          * Cycle the volume data lock as a safety (probably not needed any
1670          * more).  To ensure everything is out we need to flush at least
1671          * three times.  (1) The running of the sideq can dirty the
1672          * filesystem, (2) A normal flush can dirty the freemap, and
1673          * (3) ensure that the freemap is fully synchronized.
1674          *
1675          * The next mount's recovery scan can clean everything up but we want
1676          * to leave the filesystem in a 100% clean state on a normal unmount.
1677          */
1678 #if 0
1679         hammer2_voldata_lock(hmp);
1680         hammer2_voldata_unlock(hmp);
1681 #endif
1682         hammer2_iocom_uninit(hmp);
1683
1684         if ((hmp->vchain.flags | hmp->fchain.flags) &
1685             HAMMER2_CHAIN_FLUSH_MASK) {
1686                 kprintf("hammer2_unmount: chains left over "
1687                         "after final sync\n");
1688                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1689                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1690
1691                 if (hammer2_debug & 0x0010)
1692                         Debugger("entered debugger");
1693         }
1694
1695         KKASSERT(hmp->spmp == NULL);
1696
1697         /*
1698          * Finish up with the device vnode
1699          */
1700         if ((devvp = hmp->devvp) != NULL) {
1701                 ronly = hmp->ronly;
1702                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1703                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1704                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1705                         ronly);
1706                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1707                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1708                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1709                 hmp->devvp = NULL;
1710                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1711                 vn_unlock(devvp);
1712                 vrele(devvp);
1713                 devvp = NULL;
1714         }
1715
1716         /*
1717          * Clear vchain/fchain flags that might prevent final cleanup
1718          * of these chains.
1719          */
1720         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1721                 atomic_add_long(&hammer2_count_modified_chains, -1);
1722                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1723                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1724         }
1725         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1726                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1727         }
1728
1729         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1730                 atomic_add_long(&hammer2_count_modified_chains, -1);
1731                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1732                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1733         }
1734         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1735                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1736         }
1737
1738         /*
1739          * Final drop of embedded freemap root chain to
1740          * clean up fchain.core (fchain structure is not
1741          * flagged ALLOCATED so it is cleaned out and then
1742          * left to rot).
1743          */
1744         hammer2_chain_drop(&hmp->fchain);
1745
1746         /*
1747          * Final drop of embedded volume root chain to clean
1748          * up vchain.core (vchain structure is not flagged
1749          * ALLOCATED so it is cleaned out and then left to
1750          * rot).
1751          */
1752         dumpcnt = 50;
1753         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1754         dumpcnt = 50;
1755         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1756         hammer2_dev_unlock(hmp);
1757         hammer2_chain_drop(&hmp->vchain);
1758
1759         hammer2_io_cleanup(hmp, &hmp->iotree);
1760         if (hmp->iofree_count) {
1761                 kprintf("io_cleanup: %d I/O's left hanging\n",
1762                         hmp->iofree_count);
1763         }
1764
1765         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1766         kmalloc_destroy(&hmp->mchain);
1767         kfree(hmp, M_HAMMER2);
1768 }
1769
1770 int
1771 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1772                  ino_t ino, struct vnode **vpp)
1773 {
1774         hammer2_xop_lookup_t *xop;
1775         hammer2_pfs_t *pmp;
1776         hammer2_inode_t *ip;
1777         hammer2_tid_t inum;
1778         int error;
1779
1780         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1781
1782         error = 0;
1783         pmp = MPTOPMP(mp);
1784
1785         /*
1786          * Easy if we already have it cached
1787          */
1788         ip = hammer2_inode_lookup(pmp, inum);
1789         if (ip) {
1790                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1791                 *vpp = hammer2_igetv(ip, &error);
1792                 hammer2_inode_unlock(ip);
1793                 hammer2_inode_drop(ip);         /* from lookup */
1794
1795                 return error;
1796         }
1797
1798         /*
1799          * Otherwise we have to find the inode
1800          */
1801         xop = hammer2_xop_alloc(pmp->iroot, 0);
1802         xop->lhc = inum;
1803         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1804         error = hammer2_xop_collect(&xop->head, 0);
1805
1806         if (error == 0) {
1807                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1808                         kprintf("vget: no collect error but also no rdata\n");
1809                         kprintf("xop %p\n", xop);
1810                         while ((hammer2_debug & 0x80000) == 0) {
1811                                 tsleep(xop, PCATCH, "wait", hz * 10);
1812                         }
1813                         ip = NULL;
1814                 } else {
1815                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1816                 }
1817         }
1818         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1819
1820         if (ip) {
1821                 *vpp = hammer2_igetv(ip, &error);
1822                 hammer2_inode_unlock(ip);
1823         } else {
1824                 *vpp = NULL;
1825                 error = ENOENT;
1826         }
1827         return (error);
1828 }
1829
1830 static
1831 int
1832 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1833 {
1834         hammer2_pfs_t *pmp;
1835         struct vnode *vp;
1836         int error;
1837
1838         pmp = MPTOPMP(mp);
1839         if (pmp->iroot == NULL) {
1840                 *vpp = NULL;
1841                 return EINVAL;
1842         }
1843
1844         error = 0;
1845         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1846
1847         while (pmp->inode_tid == 0) {
1848                 hammer2_xop_ipcluster_t *xop;
1849                 hammer2_inode_meta_t *meta;
1850
1851                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1852                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1853                 error = hammer2_xop_collect(&xop->head, 0);
1854
1855                 if (error == 0) {
1856                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1857                         pmp->iroot->meta = *meta;
1858                         pmp->inode_tid = meta->pfs_inum + 1;
1859                         if (pmp->inode_tid < HAMMER2_INODE_START)
1860                                 pmp->inode_tid = HAMMER2_INODE_START;
1861                         pmp->modify_tid =
1862                                 xop->head.cluster.focus->bref.modify_tid + 1;
1863                         kprintf("PFS: Starting inode %jd\n",
1864                                 (intmax_t)pmp->inode_tid);
1865                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1866                                 pmp->inode_tid, pmp->modify_tid);
1867                         wakeup(&pmp->iroot);
1868
1869                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1870
1871                         /*
1872                          * Prime the mount info.
1873                          */
1874                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1875                         break;
1876                 }
1877
1878                 /*
1879                  * Loop, try again
1880                  */
1881                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1882                 hammer2_inode_unlock(pmp->iroot);
1883                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1884                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1885                 if (error == EINTR)
1886                         break;
1887         }
1888
1889         if (error) {
1890                 hammer2_inode_unlock(pmp->iroot);
1891                 *vpp = NULL;
1892         } else {
1893                 vp = hammer2_igetv(pmp->iroot, &error);
1894                 hammer2_inode_unlock(pmp->iroot);
1895                 *vpp = vp;
1896         }
1897
1898         return (error);
1899 }
1900
1901 /*
1902  * Filesystem status
1903  *
1904  * XXX incorporate ipdata->meta.inode_quota and data_quota
1905  */
1906 static
1907 int
1908 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1909 {
1910         hammer2_pfs_t *pmp;
1911         hammer2_dev_t *hmp;
1912         hammer2_blockref_t bref;
1913         int i;
1914
1915         /*
1916          * NOTE: iroot might not have validated the cluster yet.
1917          */
1918         pmp = MPTOPMP(mp);
1919
1920         mp->mnt_stat.f_files = 0;
1921         mp->mnt_stat.f_ffree = 0;
1922         mp->mnt_stat.f_blocks = 0;
1923         mp->mnt_stat.f_bfree = 0;
1924         mp->mnt_stat.f_bavail = 0;
1925
1926         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1927                 hmp = pmp->pfs_hmps[i];
1928                 if (hmp == NULL)
1929                         continue;
1930                 if (pmp->iroot->cluster.array[i].chain)
1931                         bref = pmp->iroot->cluster.array[i].chain->bref;
1932                 else
1933                         bzero(&bref, sizeof(bref));
1934
1935                 mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1936                 mp->mnt_stat.f_ffree = 0;
1937                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1938                                         mp->mnt_vstat.f_bsize;
1939                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1940                                         mp->mnt_vstat.f_bsize;
1941                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1942
1943                 if (cred && cred->cr_uid != 0) {
1944                         uint64_t adj;
1945
1946                         /* 5% */
1947                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
1948                         mp->mnt_stat.f_blocks -= adj;
1949                         mp->mnt_stat.f_bfree -= adj;
1950                         mp->mnt_stat.f_bavail -= adj;
1951                 }
1952
1953                 *sbp = mp->mnt_stat;
1954         }
1955         return (0);
1956 }
1957
1958 static
1959 int
1960 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1961 {
1962         hammer2_pfs_t *pmp;
1963         hammer2_dev_t *hmp;
1964         hammer2_blockref_t bref;
1965         int i;
1966
1967         /*
1968          * NOTE: iroot might not have validated the cluster yet.
1969          */
1970         pmp = MPTOPMP(mp);
1971
1972         mp->mnt_vstat.f_bsize = 0;
1973         mp->mnt_vstat.f_files = 0;
1974         mp->mnt_vstat.f_ffree = 0;
1975         mp->mnt_vstat.f_blocks = 0;
1976         mp->mnt_vstat.f_bfree = 0;
1977         mp->mnt_vstat.f_bavail = 0;
1978
1979         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1980                 hmp = pmp->pfs_hmps[i];
1981                 if (hmp == NULL)
1982                         continue;
1983                 if (pmp->iroot->cluster.array[i].chain)
1984                         bref = pmp->iroot->cluster.array[i].chain->bref;
1985                 else
1986                         bzero(&bref, sizeof(bref));
1987
1988                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1989                 mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
1990                 mp->mnt_vstat.f_ffree = 0;
1991                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1992                                         mp->mnt_vstat.f_bsize;
1993                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1994                                         mp->mnt_vstat.f_bsize;
1995                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1996
1997                 if (cred && cred->cr_uid != 0) {
1998                         uint64_t adj;
1999
2000                         /* 5% */
2001                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2002                         mp->mnt_vstat.f_blocks -= adj;
2003                         mp->mnt_vstat.f_bfree -= adj;
2004                         mp->mnt_vstat.f_bavail -= adj;
2005                 }
2006
2007                 *sbp = mp->mnt_vstat;
2008         }
2009         return (0);
2010 }
2011
2012 /*
2013  * Mount-time recovery (RW mounts)
2014  *
2015  * Updates to the free block table are allowed to lag flushes by one
2016  * transaction.  In case of a crash, then on a fresh mount we must do an
2017  * incremental scan of the last committed transaction id and make sure that
2018  * all related blocks have been marked allocated.
2019  *
2020  * The super-root topology and each PFS has its own transaction id domain,
2021  * so we must track PFS boundary transitions.
2022  */
2023 struct hammer2_recovery_elm {
2024         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2025         hammer2_chain_t *chain;
2026         hammer2_tid_t sync_tid;
2027 };
2028
2029 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2030
2031 struct hammer2_recovery_info {
2032         struct hammer2_recovery_list list;
2033         hammer2_tid_t   mtid;
2034         int     depth;
2035 };
2036
2037 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2038                         hammer2_chain_t *parent,
2039                         struct hammer2_recovery_info *info,
2040                         hammer2_tid_t sync_tid);
2041
2042 #define HAMMER2_RECOVERY_MAXDEPTH       10
2043
2044 static
2045 int
2046 hammer2_recovery(hammer2_dev_t *hmp)
2047 {
2048         struct hammer2_recovery_info info;
2049         struct hammer2_recovery_elm *elm;
2050         hammer2_chain_t *parent;
2051         hammer2_tid_t sync_tid;
2052         hammer2_tid_t mirror_tid;
2053         int error;
2054
2055         hammer2_trans_init(hmp->spmp, 0);
2056
2057         sync_tid = hmp->voldata.freemap_tid;
2058         mirror_tid = hmp->voldata.mirror_tid;
2059
2060         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2061         if (sync_tid >= mirror_tid) {
2062                 kprintf(" no recovery needed\n");
2063         } else {
2064                 kprintf(" freemap recovery %016jx-%016jx\n",
2065                         sync_tid + 1, mirror_tid);
2066         }
2067
2068         TAILQ_INIT(&info.list);
2069         info.depth = 0;
2070         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2071         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2072         hammer2_chain_lookup_done(parent);
2073
2074         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2075                 TAILQ_REMOVE(&info.list, elm, entry);
2076                 parent = elm->chain;
2077                 sync_tid = elm->sync_tid;
2078                 kfree(elm, M_HAMMER2);
2079
2080                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2081                 error |= hammer2_recovery_scan(hmp, parent, &info,
2082                                               hmp->voldata.freemap_tid);
2083                 hammer2_chain_unlock(parent);
2084                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2085         }
2086         hammer2_trans_done(hmp->spmp);
2087
2088         return error;
2089 }
2090
2091 static
2092 int
2093 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2094                       struct hammer2_recovery_info *info,
2095                       hammer2_tid_t sync_tid)
2096 {
2097         const hammer2_inode_data_t *ripdata;
2098         hammer2_chain_t *chain;
2099         hammer2_blockref_t bref;
2100         int tmp_error;
2101         int rup_error;
2102         int error;
2103         int first;
2104
2105         /*
2106          * Adjust freemap to ensure that the block(s) are marked allocated.
2107          */
2108         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2109                 hammer2_freemap_adjust(hmp, &parent->bref,
2110                                        HAMMER2_FREEMAP_DORECOVER);
2111         }
2112
2113         /*
2114          * Check type for recursive scan
2115          */
2116         switch(parent->bref.type) {
2117         case HAMMER2_BREF_TYPE_VOLUME:
2118                 /* data already instantiated */
2119                 break;
2120         case HAMMER2_BREF_TYPE_INODE:
2121                 /*
2122                  * Must instantiate data for DIRECTDATA test and also
2123                  * for recursion.
2124                  */
2125                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2126                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2127                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2128                         /* not applicable to recovery scan */
2129                         hammer2_chain_unlock(parent);
2130                         return 0;
2131                 }
2132                 hammer2_chain_unlock(parent);
2133                 break;
2134         case HAMMER2_BREF_TYPE_INDIRECT:
2135                 /*
2136                  * Must instantiate data for recursion
2137                  */
2138                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2139                 hammer2_chain_unlock(parent);
2140                 break;
2141         case HAMMER2_BREF_TYPE_DIRENT:
2142         case HAMMER2_BREF_TYPE_DATA:
2143         case HAMMER2_BREF_TYPE_FREEMAP:
2144         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2145         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2146                 /* not applicable to recovery scan */
2147                 return 0;
2148                 break;
2149         default:
2150                 return HAMMER2_ERROR_BADBREF;
2151         }
2152
2153         /*
2154          * Defer operation if depth limit reached or if we are crossing a
2155          * PFS boundary.
2156          */
2157         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2158                 struct hammer2_recovery_elm *elm;
2159
2160                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2161                 elm->chain = parent;
2162                 elm->sync_tid = sync_tid;
2163                 hammer2_chain_ref(parent);
2164                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2165                 /* unlocked by caller */
2166
2167                 return(0);
2168         }
2169
2170
2171         /*
2172          * Recursive scan of the last flushed transaction only.  We are
2173          * doing this without pmp assignments so don't leave the chains
2174          * hanging around after we are done with them.
2175          *
2176          * error        Cumulative error this level only
2177          * rup_error    Cumulative error for recursion
2178          * tmp_error    Specific non-cumulative recursion error
2179          */
2180         chain = NULL;
2181         first = 1;
2182         rup_error = 0;
2183         error = 0;
2184
2185         for (;;) {
2186                 error |= hammer2_chain_scan(parent, &chain, &bref,
2187                                             &first,
2188                                             HAMMER2_LOOKUP_NODATA);
2189
2190                 /*
2191                  * Problem during scan or EOF
2192                  */
2193                 if (error)
2194                         break;
2195
2196                 /*
2197                  * If this is a leaf
2198                  */
2199                 if (chain == NULL) {
2200                         if (bref.mirror_tid > sync_tid) {
2201                                 hammer2_freemap_adjust(hmp, &bref,
2202                                                      HAMMER2_FREEMAP_DORECOVER);
2203                         }
2204                         continue;
2205                 }
2206
2207                 /*
2208                  * This may or may not be a recursive node.
2209                  */
2210                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2211                 if (bref.mirror_tid > sync_tid) {
2212                         ++info->depth;
2213                         tmp_error = hammer2_recovery_scan(hmp, chain,
2214                                                            info, sync_tid);
2215                         --info->depth;
2216                 } else {
2217                         tmp_error = 0;
2218                 }
2219
2220                 /*
2221                  * Flush the recovery at the PFS boundary to stage it for
2222                  * the final flush of the super-root topology.
2223                  */
2224                 if (tmp_error == 0 &&
2225                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2226                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2227                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2228                 }
2229                 rup_error |= tmp_error;
2230         }
2231         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2232 }
2233
2234 /*
2235  * Sync a mount point; this is called on a per-mount basis from the
2236  * filesystem syncer process periodically and whenever a user issues
2237  * a sync.
2238  */
2239 int
2240 hammer2_vfs_sync(struct mount *mp, int waitfor)
2241 {
2242         hammer2_xop_flush_t *xop;
2243         struct hammer2_sync_info info;
2244         hammer2_inode_t *iroot;
2245         hammer2_pfs_t *pmp;
2246         int flags;
2247         int error;
2248
2249         pmp = MPTOPMP(mp);
2250         iroot = pmp->iroot;
2251         KKASSERT(iroot);
2252         KKASSERT(iroot->pmp == pmp);
2253
2254         /*
2255          * We can't acquire locks on existing vnodes while in a transaction
2256          * without risking a deadlock.  This assumes that vfsync() can be
2257          * called without the vnode locked (which it can in DragonFly).
2258          * Otherwise we'd have to implement a multi-pass or flag the lock
2259          * failures and retry.
2260          *
2261          * The reclamation code interlocks with the sync list's token
2262          * (by removing the vnode from the scan list) before unlocking
2263          * the inode, giving us time to ref the inode.
2264          */
2265         /*flags = VMSC_GETVP;*/
2266         flags = 0;
2267         if (waitfor & MNT_LAZY)
2268                 flags |= VMSC_ONEPASS;
2269
2270         /*
2271          * Preflush the vnodes using a normal transaction before interlocking
2272          * with a flush transaction.  We do this to try to run as much of
2273          * the compression as possible outside the flush transaction.
2274          *
2275          * For efficiency do an async pass before making sure with a
2276          * synchronous pass on all related buffer cache buffers.
2277          */
2278         hammer2_trans_init(pmp, 0);
2279         info.error = 0;
2280         info.waitfor = MNT_NOWAIT;
2281         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2282         info.waitfor = MNT_WAIT;
2283         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2284         hammer2_trans_done(pmp);
2285
2286         /*
2287          * Start our flush transaction.  This does not return until all
2288          * concurrent transactions have completed and will prevent any
2289          * new transactions from running concurrently, except for the
2290          * buffer cache transactions.
2291          *
2292          * NOTE!  It is still possible for the paging code to push pages
2293          *        out via a UIO_NOCOPY hammer2_vop_write() during the main
2294          *        flush.
2295          */
2296         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2297         hammer2_inode_run_sideq(pmp);
2298
2299         info.error = 0;
2300         info.waitfor = MNT_NOWAIT;
2301         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2302         info.waitfor = MNT_WAIT;
2303         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2304         hammer2_bioq_sync(pmp);
2305
2306         /*
2307          * Use the XOP interface to concurrently flush all nodes to
2308          * synchronize the PFSROOT subtopology to the media.  A standard
2309          * end-of-scan ENOENT error indicates cluster sufficiency.
2310          *
2311          * Note that this flush will not be visible on crash recovery until
2312          * we flush the super-root topology in the next loop.
2313          *
2314          * XXX For now wait for all flushes to complete.
2315          */
2316         if (iroot) {
2317                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2318                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2319                 error = hammer2_xop_collect(&xop->head,
2320                                             HAMMER2_XOP_COLLECT_WAITALL);
2321                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2322                 if (error == HAMMER2_ERROR_ENOENT)
2323                         error = 0;
2324                 else
2325                         error = hammer2_error_to_errno(error);
2326         } else {
2327                 error = 0;
2328         }
2329         hammer2_trans_done(pmp);
2330
2331         return (error);
2332 }
2333
2334 /*
2335  * Sync passes.
2336  *
2337  * Note that we ignore the tranasction mtid we got above.  Instead,
2338  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2339  * transactions.
2340  */
2341 static int
2342 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2343 {
2344         struct hammer2_sync_info *info = data;
2345         hammer2_inode_t *ip;
2346         int error;
2347
2348         /*
2349          * Degenerate cases.  Note that ip == NULL typically means the
2350          * syncer vnode itself and we don't want to vclrisdirty() in that
2351          * situation.
2352          */
2353         ip = VTOI(vp);
2354         if (ip == NULL) {
2355                 return(0);
2356         }
2357         if (vp->v_type == VNON || vp->v_type == VBAD) {
2358                 vclrisdirty(vp);
2359                 return(0);
2360         }
2361
2362         /*
2363          * VOP_FSYNC will start a new transaction so replicate some code
2364          * here to do it inline (see hammer2_vop_fsync()).
2365          *
2366          * WARNING: The vfsync interacts with the buffer cache and might
2367          *          block, we can't hold the inode lock at that time.
2368          *          However, we MUST ref ip before blocking to ensure that
2369          *          it isn't ripped out from under us (since we do not
2370          *          hold a lock on the vnode).
2371          */
2372         hammer2_inode_ref(ip);
2373         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2374             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2375                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2376                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2377                                  HAMMER2_INODE_MODIFIED)) {
2378                         hammer2_inode_lock(ip, 0);
2379                         if (ip->flags & (HAMMER2_INODE_RESIZED |
2380                                          HAMMER2_INODE_MODIFIED)) {
2381                                 hammer2_inode_chain_sync(ip);
2382                         }
2383                         hammer2_inode_unlock(ip);
2384                 }
2385         }
2386         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2387             RB_EMPTY(&vp->v_rbdirty_tree)) {
2388                 vclrisdirty(vp);
2389         }
2390
2391         hammer2_inode_drop(ip);
2392 #if 1
2393         error = 0;
2394         if (error)
2395                 info->error = error;
2396 #endif
2397         return(0);
2398 }
2399
2400 static
2401 int
2402 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2403 {
2404         hammer2_inode_t *ip;
2405
2406         KKASSERT(MAXFIDSZ >= 16);
2407         ip = VTOI(vp);
2408         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2409         fhp->fid_ext = 0;
2410         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2411         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2412
2413         return 0;
2414 }
2415
2416 static
2417 int
2418 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2419                struct fid *fhp, struct vnode **vpp)
2420 {
2421         hammer2_pfs_t *pmp;
2422         hammer2_tid_t inum;
2423         int error;
2424
2425         pmp = MPTOPMP(mp);
2426         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2427         if (vpp) {
2428                 if (inum == 1)
2429                         error = hammer2_vfs_root(mp, vpp);
2430                 else
2431                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2432         } else {
2433                 error = 0;
2434         }
2435         if (error)
2436                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2437         return error;
2438 }
2439
2440 static
2441 int
2442 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2443                  int *exflagsp, struct ucred **credanonp)
2444 {
2445         hammer2_pfs_t *pmp;
2446         struct netcred *np;
2447         int error;
2448
2449         pmp = MPTOPMP(mp);
2450         np = vfs_export_lookup(mp, &pmp->export, nam);
2451         if (np) {
2452                 *exflagsp = np->netc_exflags;
2453                 *credanonp = &np->netc_anon;
2454                 error = 0;
2455         } else {
2456                 error = EACCES;
2457         }
2458         return error;
2459 }
2460
2461 /*
2462  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2463  * header into the HMP
2464  *
2465  * XXX read four volhdrs and use the one with the highest TID whos CRC
2466  *     matches.
2467  *
2468  * XXX check iCRCs.
2469  *
2470  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2471  *     nonexistant locations.
2472  *
2473  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2474  */
2475 static
2476 int
2477 hammer2_install_volume_header(hammer2_dev_t *hmp)
2478 {
2479         hammer2_volume_data_t *vd;
2480         struct buf *bp;
2481         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2482         int error_reported;
2483         int error;
2484         int valid;
2485         int i;
2486
2487         error_reported = 0;
2488         error = 0;
2489         valid = 0;
2490         bp = NULL;
2491
2492         /*
2493          * There are up to 4 copies of the volume header (syncs iterate
2494          * between them so there is no single master).  We don't trust the
2495          * volu_size field so we don't know precisely how large the filesystem
2496          * is, so depend on the OS to return an error if we go beyond the
2497          * block device's EOF.
2498          */
2499         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2500                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2501                               HAMMER2_VOLUME_BYTES, &bp);
2502                 if (error) {
2503                         brelse(bp);
2504                         bp = NULL;
2505                         continue;
2506                 }
2507
2508                 vd = (struct hammer2_volume_data *) bp->b_data;
2509                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2510                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2511                         brelse(bp);
2512                         bp = NULL;
2513                         continue;
2514                 }
2515
2516                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2517                         /* XXX: Reversed-endianness filesystem */
2518                         kprintf("hammer2: reverse-endian filesystem detected");
2519                         brelse(bp);
2520                         bp = NULL;
2521                         continue;
2522                 }
2523
2524                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2525                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2526                                       HAMMER2_VOLUME_ICRC0_SIZE);
2527                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2528                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2529                                        HAMMER2_VOLUME_ICRC1_SIZE);
2530                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2531                         kprintf("hammer2 volume header crc "
2532                                 "mismatch copy #%d %08x/%08x\n",
2533                                 i, crc0, crc);
2534                         error_reported = 1;
2535                         brelse(bp);
2536                         bp = NULL;
2537                         continue;
2538                 }
2539                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2540                         valid = 1;
2541                         hmp->voldata = *vd;
2542                         hmp->volhdrno = i;
2543                 }
2544                 brelse(bp);
2545                 bp = NULL;
2546         }
2547         if (valid) {
2548                 hmp->volsync = hmp->voldata;
2549                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2550                 error = 0;
2551                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2552                         kprintf("hammer2: using volume header #%d\n",
2553                                 hmp->volhdrno);
2554                 }
2555         } else {
2556                 error = EINVAL;
2557                 kprintf("hammer2: no valid volume headers found!\n");
2558         }
2559         return (error);
2560 }
2561
2562 /*
2563  * This handles hysteresis on regular file flushes.  Because the BIOs are
2564  * routed to a thread it is possible for an excessive number to build up
2565  * and cause long front-end stalls long before the runningbuffspace limit
2566  * is hit, so we implement hammer2_flush_pipe to control the
2567  * hysteresis.
2568  *
2569  * This is a particular problem when compression is used.
2570  */
2571 void
2572 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2573 {
2574         atomic_add_int(&pmp->count_lwinprog, 1);
2575 }
2576
2577 void
2578 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2579 {
2580         int lwinprog;
2581
2582         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2583         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2584             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2585                 atomic_clear_int(&pmp->count_lwinprog,
2586                                  HAMMER2_LWINPROG_WAITING);
2587                 wakeup(&pmp->count_lwinprog);
2588         }
2589         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2590             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2591                 atomic_clear_int(&pmp->count_lwinprog,
2592                                  HAMMER2_LWINPROG_WAITING0);
2593                 wakeup(&pmp->count_lwinprog);
2594         }
2595 }
2596
2597 void
2598 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2599 {
2600         int lwinprog;
2601         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2602                                     HAMMER2_LWINPROG_WAITING0;
2603
2604         for (;;) {
2605                 lwinprog = pmp->count_lwinprog;
2606                 cpu_ccfence();
2607                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2608                         break;
2609                 tsleep_interlock(&pmp->count_lwinprog, 0);
2610                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2611                 lwinprog = pmp->count_lwinprog;
2612                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2613                         break;
2614                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2615         }
2616 }
2617
2618 /*
2619  * Manage excessive memory resource use for chain and related
2620  * structures.
2621  */
2622 void
2623 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2624 {
2625         uint32_t waiting;
2626         uint32_t count;
2627         uint32_t limit;
2628 #if 0
2629         static int zzticks;
2630 #endif
2631
2632         /*
2633          * Atomic check condition and wait.  Also do an early speedup of
2634          * the syncer to try to avoid hitting the wait.
2635          */
2636         for (;;) {
2637                 waiting = pmp->inmem_dirty_chains;
2638                 cpu_ccfence();
2639                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2640
2641                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2642                 if (limit < hammer2_limit_dirty_chains)
2643                         limit = hammer2_limit_dirty_chains;
2644                 if (limit < 1000)
2645                         limit = 1000;
2646
2647 #if 0
2648                 if ((int)(ticks - zzticks) > hz) {
2649                         zzticks = ticks;
2650                         kprintf("count %ld %ld\n", count, limit);
2651                 }
2652 #endif
2653
2654                 /*
2655                  * Block if there are too many dirty chains present, wait
2656                  * for the flush to clean some out.
2657                  */
2658                 if (count > limit) {
2659                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2660                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2661                                                waiting,
2662                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2663                                 speedup_syncer(pmp->mp);
2664                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2665                                        "chnmem", hz);
2666                         }
2667                         continue;       /* loop on success or fail */
2668                 }
2669
2670                 /*
2671                  * Try to start an early flush before we are forced to block.
2672                  */
2673                 if (count > limit * 7 / 10)
2674                         speedup_syncer(pmp->mp);
2675                 break;
2676         }
2677 }
2678
2679 void
2680 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2681 {
2682         if (pmp) {
2683                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2684         }
2685 }
2686
2687 void
2688 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2689 {
2690         uint32_t waiting;
2691
2692         if (pmp == NULL)
2693                 return;
2694
2695         for (;;) {
2696                 waiting = pmp->inmem_dirty_chains;
2697                 cpu_ccfence();
2698                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2699                                        waiting,
2700                                        (waiting - 1) &
2701                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2702                         break;
2703                 }
2704         }
2705
2706         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2707                 wakeup(&pmp->inmem_dirty_chains);
2708 }
2709
2710 /*
2711  * Returns 0 if the filesystem has tons of free space
2712  * Returns 1 if the filesystem has less than 10% remaining
2713  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2714  */
2715 int
2716 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2717 {
2718         hammer2_pfs_t *pmp;
2719         hammer2_dev_t *hmp;
2720         hammer2_off_t free_reserved;
2721         hammer2_off_t free_nominal;
2722         int i;
2723
2724         pmp = ip->pmp;
2725
2726         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2727                 free_reserved = HAMMER2_SEGSIZE;
2728                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2729                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2730                         hmp = pmp->pfs_hmps[i];
2731                         if (hmp == NULL)
2732                                 continue;
2733                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2734                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2735                                 continue;
2736
2737                         if (free_nominal > hmp->voldata.allocator_free)
2738                                 free_nominal = hmp->voldata.allocator_free;
2739                         if (free_reserved < hmp->free_reserved)
2740                                 free_reserved = hmp->free_reserved;
2741                 }
2742
2743                 /*
2744                  * SMP races ok
2745                  */
2746                 pmp->free_reserved = free_reserved;
2747                 pmp->free_nominal = free_nominal;
2748                 pmp->free_ticks = ticks;
2749         } else {
2750                 free_reserved = pmp->free_reserved;
2751                 free_nominal = pmp->free_nominal;
2752         }
2753         if (cred && cred->cr_uid != 0) {
2754                 if ((int64_t)(free_nominal - bytes) <
2755                     (int64_t)free_reserved) {
2756                         return 2;
2757                 }
2758         } else {
2759                 if ((int64_t)(free_nominal - bytes) <
2760                     (int64_t)free_reserved / 2) {
2761                         return 2;
2762                 }
2763         }
2764         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2765                 return 1;
2766         return 0;
2767 }
2768
2769 /*
2770  * Debugging
2771  */
2772 void
2773 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2774 {
2775         hammer2_chain_t *scan;
2776         hammer2_chain_t *parent;
2777
2778         --*countp;
2779         if (*countp == 0) {
2780                 kprintf("%*.*s...\n", tab, tab, "");
2781                 return;
2782         }
2783         if (*countp < 0)
2784                 return;
2785         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2786                 tab, tab, "", pfx,
2787                 chain, chain->bref.type,
2788                 chain->bref.key, chain->bref.keybits,
2789                 chain->bref.mirror_tid);
2790
2791         kprintf("%*.*s      [%08x] (%s) refs=%d",
2792                 tab, tab, "",
2793                 chain->flags,
2794                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2795                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2796                 chain->refs);
2797
2798         parent = chain->parent;
2799         if (parent)
2800                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2801                         tab, tab, "",
2802                         parent, parent->flags, parent->refs);
2803         if (RB_EMPTY(&chain->core.rbtree)) {
2804                 kprintf("\n");
2805         } else {
2806                 kprintf(" {\n");
2807                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2808                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2809                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2810                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2811                                 chain->data->ipdata.filename);
2812                 else
2813                         kprintf("%*.*s}\n", tab, tab, "");
2814         }
2815 }