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 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
  76 static struct hammer2_mntlist hammer2_mntlist;
  77 static struct hammer2_pfslist hammer2_pfslist;
  78 static struct lock hammer2_mntlk;
  79
  80 int hammer2_debug;
  81 int hammer2_cluster_enable = 4;
  82 int hammer2_hardlink_enable = 1;
  83 int hammer2_flush_pipe = 100;
  84 int hammer2_synchronous_flush = 1;
  85 int hammer2_dio_count;
  86 long hammer2_chain_allocs;
  87 long hammer2_chain_frees;
  88 long hammer2_limit_dirty_chains;
  89 long hammer2_count_modified_chains;
  90 long hammer2_iod_file_read;
  91 long hammer2_iod_meta_read;
  92 long hammer2_iod_indr_read;
  93 long hammer2_iod_fmap_read;
  94 long hammer2_iod_volu_read;
  95 long hammer2_iod_file_write;
  96 long hammer2_iod_file_wembed;
  97 long hammer2_iod_file_wzero;
  98 long hammer2_iod_file_wdedup;
  99 long hammer2_iod_meta_write;
 100 long hammer2_iod_indr_write;
 101 long hammer2_iod_fmap_write;
 102 long hammer2_iod_volu_write;
 103 long hammer2_ioa_file_read;
 104 long hammer2_ioa_meta_read;
 105 long hammer2_ioa_indr_read;
 106 long hammer2_ioa_fmap_read;
 107 long hammer2_ioa_volu_read;
 108 long hammer2_ioa_fmap_write;
 109 long hammer2_ioa_file_write;
 110 long hammer2_ioa_meta_write;
 111 long hammer2_ioa_indr_write;
 112 long hammer2_ioa_volu_write;
 113
 114 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 115 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 116                 "Buffer used for compression.");
 117
 118 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 119 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 120                 "Buffer used for decompression.");
 121
 122 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 123
 124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 125            &hammer2_debug, 0, "");
 126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW,
 127            &hammer2_cluster_enable, 0, "");
 128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
 129            &hammer2_hardlink_enable, 0, "");
 130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 131            &hammer2_flush_pipe, 0, "");
 132 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 133            &hammer2_synchronous_flush, 0, "");
 134 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 135            &hammer2_chain_allocs, 0, "");
 136 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 137            &hammer2_chain_frees, 0, "");
 138 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 139            &hammer2_limit_dirty_chains, 0, "");
 140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 141            &hammer2_count_modified_chains, 0, "");
 142 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 143            &hammer2_dio_count, 0, "");
 144
 145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 146            &hammer2_iod_file_read, 0, "");
 147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 148            &hammer2_iod_meta_read, 0, "");
 149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 150            &hammer2_iod_indr_read, 0, "");
 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 152            &hammer2_iod_fmap_read, 0, "");
 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 154            &hammer2_iod_volu_read, 0, "");
 155
 156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 157            &hammer2_iod_file_write, 0, "");
 158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 159            &hammer2_iod_file_wembed, 0, "");
 160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 161            &hammer2_iod_file_wzero, 0, "");
 162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 163            &hammer2_iod_file_wdedup, 0, "");
 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 165            &hammer2_iod_meta_write, 0, "");
 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 167            &hammer2_iod_indr_write, 0, "");
 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 169            &hammer2_iod_fmap_write, 0, "");
 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 171            &hammer2_iod_volu_write, 0, "");
 172
 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
 174            &hammer2_ioa_file_read, 0, "");
 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
 176            &hammer2_ioa_meta_read, 0, "");
 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
 178            &hammer2_ioa_indr_read, 0, "");
 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
 180            &hammer2_ioa_fmap_read, 0, "");
 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
 182            &hammer2_ioa_volu_read, 0, "");
 183
 184 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
 185            &hammer2_ioa_file_write, 0, "");
 186 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
 187            &hammer2_ioa_meta_write, 0, "");
 188 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
 189            &hammer2_ioa_indr_write, 0, "");
 190 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
 191            &hammer2_ioa_fmap_write, 0, "");
 192 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
 193            &hammer2_ioa_volu_write, 0, "");
 194
 195 long hammer2_check_icrc32;
 196 long hammer2_check_xxhash64;
 197 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 198            &hammer2_check_icrc32, 0, "");
 199 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 200            &hammer2_check_xxhash64, 0, "");
 201
 202 static int hammer2_vfs_init(struct vfsconf *conf);
 203 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 204 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 205                                 struct ucred *cred);
 206 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 207                                 struct vnode *, struct ucred *);
 208 static int hammer2_recovery(hammer2_dev_t *hmp);
 209 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 210 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 211 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 212                                 struct ucred *cred);
 213 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 214                                 struct ucred *cred);
 215 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
 216                                 ino_t ino, struct vnode **vpp);
 217 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 218                                 struct fid *fhp, struct vnode **vpp);
 219 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 220 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 221                                 int *exflagsp, struct ucred **credanonp);
 222
 223 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 224 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 225
 226 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 227
 228 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 229 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 230                                 hammer2_dev_t *hmp);
 231
 232 /*
 233  * HAMMER2 vfs operations.
 234  */
 235 static struct vfsops hammer2_vfsops = {
 236         .vfs_init       = hammer2_vfs_init,
 237         .vfs_uninit     = hammer2_vfs_uninit,
 238         .vfs_sync       = hammer2_vfs_sync,
 239         .vfs_mount      = hammer2_vfs_mount,
 240         .vfs_unmount    = hammer2_vfs_unmount,
 241         .vfs_root       = hammer2_vfs_root,
 242         .vfs_statfs     = hammer2_vfs_statfs,
 243         .vfs_statvfs    = hammer2_vfs_statvfs,
 244         .vfs_vget       = hammer2_vfs_vget,
 245         .vfs_vptofh     = hammer2_vfs_vptofh,
 246         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 247         .vfs_checkexp   = hammer2_vfs_checkexp
 248 };
 249
 250 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 251
 252 VFS_SET(hammer2_vfsops, hammer2, 0);
 253 MODULE_VERSION(hammer2, 1);
 254
 255 static
 256 int
 257 hammer2_vfs_init(struct vfsconf *conf)
 258 {
 259         static struct objcache_malloc_args margs_read;
 260         static struct objcache_malloc_args margs_write;
 261         static struct objcache_malloc_args margs_vop;
 262
 263         int error;
 264
 265         error = 0;
 266
 267         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 268                 error = EINVAL;
 269         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 270                 error = EINVAL;
 271         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 272                 error = EINVAL;
 273
 274         if (error)
 275                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 276
 277         margs_read.objsize = 65536;
 278         margs_read.mtype = M_HAMMER2_DEBUFFER;
 279
 280         margs_write.objsize = 32768;
 281         margs_write.mtype = M_HAMMER2_CBUFFER;
 282
 283         margs_vop.objsize = sizeof(hammer2_xop_t);
 284         margs_vop.mtype = M_HAMMER2;
 285
 286         /*
 287          * Note thaht for the XOPS cache we want backing store allocations
 288          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 289          * confusion), so use the backing store function that does it.  This
 290          * means that initial XOPS objects are zerod but REUSED objects are
 291          * not.  So we are responsible for cleaning the object up sufficiently
 292          * for our needs before objcache_put()ing it back (typically just the
 293          * FIFO indices).
 294          */
 295         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 296                                 0, 1, NULL, NULL, NULL,
 297                                 objcache_malloc_alloc,
 298                                 objcache_malloc_free,
 299                                 &margs_read);
 300         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 301                                 0, 1, NULL, NULL, NULL,
 302                                 objcache_malloc_alloc,
 303                                 objcache_malloc_free,
 304                                 &margs_write);
 305         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 306                                 0, 1, NULL, NULL, NULL,
 307                                 objcache_malloc_alloc_zero,
 308                                 objcache_malloc_free,
 309                                 &margs_vop);
 310
 311
 312         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 313         TAILQ_INIT(&hammer2_mntlist);
 314         TAILQ_INIT(&hammer2_pfslist);
 315
 316         hammer2_limit_dirty_chains = desiredvnodes / 10;
 317
 318         return (error);
 319 }
 320
 321 static
 322 int
 323 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 324 {
 325         objcache_destroy(cache_buffer_read);
 326         objcache_destroy(cache_buffer_write);
 327         objcache_destroy(cache_xops);
 328         return 0;
 329 }
 330
 331 /*
 332  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
 333  * mounts and the spmp structure for media (hmp) structures.
 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, const hammer2_inode_data_t *ripdata,
 344                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 345 {
 346         hammer2_inode_t *iroot;
 347         hammer2_pfs_t *pmp;
 348         int count;
 349         int i;
 350         int j;
 351
 352         /*
 353          * Locate or create the PFS based on the cluster id.  If ripdata
 354          * is NULL this is a spmp which is unique and is always allocated.
 355          *
 356          * If the device is mounted in local mode all PFSs are considered
 357          * independent and not part of any cluster (for debugging only).
 358          */
 359         if (ripdata) {
 360                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 361                         if (force_local != pmp->force_local)
 362                                 continue;
 363                         if (force_local == NULL &&
 364                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 365                                  sizeof(pmp->pfs_clid)) == 0) {
 366                                         break;
 367                         } else if (force_local && pmp->pfs_names[0] &&
 368                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 369                                         break;
 370                         }
 371                 }
 372         } else {
 373                 pmp = NULL;
 374         }
 375
 376         if (pmp == NULL) {
 377                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 378                 pmp->force_local = force_local;
 379                 hammer2_trans_manage_init(pmp);
 380                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 381                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 382                 lockinit(&pmp->lock, "pfslk", 0, 0);
 383                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 384                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 385                 spin_init(&pmp->xop_spin, "h2xop");
 386                 spin_init(&pmp->lru_spin, "h2lru");
 387                 RB_INIT(&pmp->inum_tree);
 388                 TAILQ_INIT(&pmp->sideq);
 389                 TAILQ_INIT(&pmp->lru_list);
 390                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 391
 392                 /*
 393                  * Distribute backend operations to threads
 394                  */
 395                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 396                         for (j = 0; j < HAMMER2_XOPGROUPS +
 397                                         HAMMER2_SPECTHREADS; ++j) {
 398                                 TAILQ_INIT(&pmp->xopq[i][j]);
 399                         }
 400                 }
 401                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 402                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 403
 404                 /*
 405                  * Save the last media transaction id for the flusher.  Set
 406                  * initial
 407                  */
 408                 if (ripdata)
 409                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 410                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 411
 412                 /*
 413                  * The synchronization thread may start too early, make
 414                  * sure it stays frozen until we are ready to let it go.
 415                  * XXX
 416                  */
 417                 /*
 418                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 419                                          HAMMER2_THREAD_REMASTER;
 420                 */
 421         }
 422
 423         /*
 424          * Create the PFS's root inode.
 425          */
 426         if ((iroot = pmp->iroot) == NULL) {
 427                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 428                 pmp->iroot = iroot;
 429                 hammer2_inode_ref(iroot);
 430                 hammer2_inode_unlock(iroot);
 431         }
 432
 433         /*
 434          * Stop here if no chain is passed in.
 435          */
 436         if (chain == NULL)
 437                 goto done;
 438
 439         /*
 440          * When a chain is passed in we must add it to the PFS's root
 441          * inode, update pmp->pfs_types[], and update the syncronization
 442          * threads.
 443          *
 444          * When forcing local mode, mark the PFS as a MASTER regardless.
 445          *
 446          * At the moment empty spots can develop due to removals or failures.
 447          * Ultimately we want to re-fill these spots but doing so might
 448          * confused running code. XXX
 449          */
 450         hammer2_inode_ref(iroot);
 451         hammer2_mtx_ex(&iroot->lock);
 452         j = iroot->cluster.nchains;
 453
 454         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 455
 456         if (j == HAMMER2_MAXCLUSTER) {
 457                 kprintf("hammer2_mount: cluster full!\n");
 458                 /* XXX fatal error? */
 459         } else {
 460                 KKASSERT(chain->pmp == NULL);
 461                 chain->pmp = pmp;
 462                 hammer2_chain_ref(chain);
 463                 iroot->cluster.array[j].chain = chain;
 464                 if (force_local)
 465                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 466                 else
 467                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 468                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 469                 pmp->pfs_hmps[j] = chain->hmp;
 470
 471                 /*
 472                  * If the PFS is already mounted we must account
 473                  * for the mount_count here.
 474                  */
 475                 if (pmp->mp)
 476                         ++chain->hmp->mount_count;
 477
 478                 /*
 479                  * May have to fixup dirty chain tracking.  Previous
 480                  * pmp was NULL so nothing to undo.
 481                  */
 482                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 483                         hammer2_pfs_memory_inc(pmp);
 484                 ++j;
 485         }
 486         iroot->cluster.nchains = j;
 487
 488         /*
 489          * Update nmasters from any PFS inode which is part of the cluster.
 490          * It is possible that this will result in a value which is too
 491          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 492          * override this value later on.
 493          *
 494          * (This informs us of masters that might not currently be
 495          *  discoverable by this mount).
 496          */
 497         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 498                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 499         }
 500
 501         /*
 502          * Count visible masters.  Masters are usually added with
 503          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 504          * are more (XXX and must update the master inodes).
 505          */
 506         count = 0;
 507         for (i = 0; i < iroot->cluster.nchains; ++i) {
 508                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 509                         ++count;
 510         }
 511         if (pmp->pfs_nmasters < count)
 512                 pmp->pfs_nmasters = count;
 513
 514         /*
 515          * Create missing synchronization and support threads.
 516          *
 517          * Single-node masters (including snapshots) have nothing to
 518          * synchronize and do not require this thread.
 519          *
 520          * Multi-node masters or any number of soft masters, slaves, copy,
 521          * or other PFS types need the thread.
 522          *
 523          * Each thread is responsible for its particular cluster index.
 524          * We use independent threads so stalls or mismatches related to
 525          * any given target do not affect other targets.
 526          */
 527         for (i = 0; i < iroot->cluster.nchains; ++i) {
 528                 /*
 529                  * Single-node masters (including snapshots) have nothing
 530                  * to synchronize and will make direct xops support calls,
 531                  * thus they do not require this thread.
 532                  *
 533                  * Note that there can be thousands of snapshots.  We do not
 534                  * want to create thousands of threads.
 535                  */
 536                 if (pmp->pfs_nmasters <= 1 &&
 537                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 538                         continue;
 539                 }
 540
 541                 /*
 542                  * Sync support thread
 543                  */
 544                 if (pmp->sync_thrs[i].td == NULL) {
 545                         hammer2_thr_create(&pmp->sync_thrs[i], pmp,
 546                                            "h2nod", i, -1,
 547                                            hammer2_primary_sync_thread);
 548                 }
 549         }
 550
 551         /*
 552          * Create missing Xop threads
 553          */
 554         if (pmp->mp)
 555                 hammer2_xop_helper_create(pmp);
 556
 557         hammer2_mtx_unlock(&iroot->lock);
 558         hammer2_inode_drop(iroot);
 559 done:
 560         return pmp;
 561 }
 562
 563 /*
 564  * Destroy a PFS, typically only occurs after the last mount on a device
 565  * has gone away.
 566  */
 567 static void
 568 hammer2_pfsfree(hammer2_pfs_t *pmp)
 569 {
 570         hammer2_inode_t *iroot;
 571         hammer2_chain_t *chain;
 572         int i;
 573         int j;
 574
 575         /*
 576          * Cleanup our reference on iroot.  iroot is (should) not be needed
 577          * by the flush code.
 578          */
 579         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 580
 581         iroot = pmp->iroot;
 582         if (iroot) {
 583                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 584                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 585                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 586                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 587                 }
 588 #if REPORT_REFS_ERRORS
 589                 if (pmp->iroot->refs != 1)
 590                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 591                                 pmp->iroot, pmp->iroot->refs);
 592 #else
 593                 KKASSERT(pmp->iroot->refs == 1);
 594 #endif
 595                 /* ref for pmp->iroot */
 596                 hammer2_inode_drop(pmp->iroot);
 597                 pmp->iroot = NULL;
 598         }
 599
 600         /*
 601          * Cleanup chains remaining on LRU list.
 602          */
 603         kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
 604         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 605                 hammer2_chain_ref(chain);
 606                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 607                 hammer2_chain_drop(chain);
 608         }
 609
 610         /*
 611          * Free remaining pmp resources
 612          */
 613         kmalloc_destroy(&pmp->mmsg);
 614         kmalloc_destroy(&pmp->minode);
 615
 616         kfree(pmp, M_HAMMER2);
 617 }
 618
 619 /*
 620  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 621  * empty is terminated and freed.
 622  *
 623  * XXX inefficient.
 624  */
 625 static void
 626 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 627 {
 628         hammer2_pfs_t *pmp;
 629         hammer2_inode_t *iroot;
 630         hammer2_chain_t *rchain;
 631         int didfreeze;
 632         int i;
 633         int j;
 634
 635 again:
 636         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 637                 if ((iroot = pmp->iroot) == NULL)
 638                         continue;
 639                 if (hmp->spmp == pmp) {
 640                         kprintf("unmount hmp %p remove spmp %p\n",
 641                                 hmp, pmp);
 642                         hmp->spmp = NULL;
 643                 }
 644
 645                 /*
 646                  * Determine if this PFS is affected.  If it is we must
 647                  * freeze all management threads and lock its iroot.
 648                  *
 649                  * Freezing a management thread forces it idle, operations
 650                  * in-progress will be aborted and it will have to start
 651                  * over again when unfrozen, or exit if told to exit.
 652                  */
 653                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 654                         if (pmp->pfs_hmps[i] == hmp)
 655                                 break;
 656                 }
 657                 if (i != HAMMER2_MAXCLUSTER) {
 658                         /*
 659                          * Make sure all synchronization threads are locked
 660                          * down.
 661                          */
 662                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 663                                 if (pmp->pfs_hmps[i] == NULL)
 664                                         continue;
 665                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 666                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 667                                         hammer2_thr_freeze_async(
 668                                                 &pmp->xop_groups[j].thrs[i]);
 669                                 }
 670                         }
 671                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 672                                 if (pmp->pfs_hmps[i] == NULL)
 673                                         continue;
 674                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 675                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 676                                         hammer2_thr_freeze(
 677                                                 &pmp->xop_groups[j].thrs[i]);
 678                                 }
 679                         }
 680
 681                         /*
 682                          * Lock the inode and clean out matching chains.
 683                          * Note that we cannot use hammer2_inode_lock_*()
 684                          * here because that would attempt to validate the
 685                          * cluster that we are in the middle of ripping
 686                          * apart.
 687                          *
 688                          * WARNING! We are working directly on the inodes
 689                          *          embedded cluster.
 690                          */
 691                         hammer2_mtx_ex(&iroot->lock);
 692
 693                         /*
 694                          * Remove the chain from matching elements of the PFS.
 695                          */
 696                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 697                                 if (pmp->pfs_hmps[i] != hmp)
 698                                         continue;
 699                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 700                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 701                                         hammer2_thr_delete(
 702                                                 &pmp->xop_groups[j].thrs[i]);
 703                                 }
 704                                 rchain = iroot->cluster.array[i].chain;
 705                                 iroot->cluster.array[i].chain = NULL;
 706                                 pmp->pfs_types[i] = 0;
 707                                 if (pmp->pfs_names[i]) {
 708                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 709                                         pmp->pfs_names[i] = NULL;
 710                                 }
 711                                 if (rchain) {
 712                                         hammer2_chain_drop(rchain);
 713                                         /* focus hint */
 714                                         if (iroot->cluster.focus == rchain)
 715                                                 iroot->cluster.focus = NULL;
 716                                 }
 717                                 pmp->pfs_hmps[i] = NULL;
 718                         }
 719                         hammer2_mtx_unlock(&iroot->lock);
 720                         didfreeze = 1;  /* remaster, unfreeze down below */
 721                 } else {
 722                         didfreeze = 0;
 723                 }
 724
 725                 /*
 726                  * Cleanup trailing chains.  Gaps may remain.
 727                  */
 728                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 729                         if (pmp->pfs_hmps[i])
 730                                 break;
 731                 }
 732                 iroot->cluster.nchains = i + 1;
 733
 734                 /*
 735                  * If the PMP has no elements remaining we can destroy it.
 736                  * (this will transition management threads from frozen->exit).
 737                  */
 738                 if (iroot->cluster.nchains == 0) {
 739                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 740                                 hmp, pmp);
 741                         hammer2_pfsfree(pmp);
 742                         goto again;
 743                 }
 744
 745                 /*
 746                  * If elements still remain we need to set the REMASTER
 747                  * flag and unfreeze it.
 748                  */
 749                 if (didfreeze) {
 750                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 751                                 if (pmp->pfs_hmps[i] == NULL)
 752                                         continue;
 753                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 754                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 755                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 756                                         hammer2_thr_remaster(
 757                                                 &pmp->xop_groups[j].thrs[i]);
 758                                         hammer2_thr_unfreeze(
 759                                                 &pmp->xop_groups[j].thrs[i]);
 760                                 }
 761                         }
 762                 }
 763         }
 764 }
 765
 766 /*
 767  * Mount or remount HAMMER2 fileystem from physical media
 768  *
 769  *      mountroot
 770  *              mp              mount point structure
 771  *              path            NULL
 772  *              data            <unused>
 773  *              cred            <unused>
 774  *
 775  *      mount
 776  *              mp              mount point structure
 777  *              path            path to mount point
 778  *              data            pointer to argument structure in user space
 779  *                      volume  volume path (device@LABEL form)
 780  *                      hflags  user mount flags
 781  *              cred            user credentials
 782  *
 783  * RETURNS:     0       Success
 784  *              !0      error number
 785  */
 786 static
 787 int
 788 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 789                   struct ucred *cred)
 790 {
 791         struct hammer2_mount_info info;
 792         hammer2_pfs_t *pmp;
 793         hammer2_pfs_t *spmp;
 794         hammer2_dev_t *hmp;
 795         hammer2_dev_t *force_local;
 796         hammer2_key_t key_next;
 797         hammer2_key_t key_dummy;
 798         hammer2_key_t lhc;
 799         struct vnode *devvp;
 800         struct nlookupdata nd;
 801         hammer2_chain_t *parent;
 802         hammer2_chain_t *chain;
 803         hammer2_cluster_t *cluster;
 804         const hammer2_inode_data_t *ripdata;
 805         hammer2_blockref_t bref;
 806         struct file *fp;
 807         char devstr[MNAMELEN];
 808         size_t size;
 809         size_t done;
 810         char *dev;
 811         char *label;
 812         int ronly = 1;
 813         int error;
 814         int cache_index;
 815         int i;
 816
 817         hmp = NULL;
 818         pmp = NULL;
 819         dev = NULL;
 820         label = NULL;
 821         devvp = NULL;
 822         cache_index = -1;
 823
 824         kprintf("hammer2_mount\n");
 825
 826         if (path == NULL) {
 827                 /*
 828                  * Root mount
 829                  */
 830                 bzero(&info, sizeof(info));
 831                 info.cluster_fd = -1;
 832                 ksnprintf(devstr, sizeof(devstr), "%s",
 833                           mp->mnt_stat.f_mntfromname);
 834                 kprintf("hammer2_mount: root '%s'\n", devstr);
 835         } else {
 836                 /*
 837                  * Non-root mount or updating a mount
 838                  */
 839                 error = copyin(data, &info, sizeof(info));
 840                 if (error)
 841                         return (error);
 842
 843                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 844                 if (error)
 845                         return (error);
 846         }
 847
 848         /* Extract device and label */
 849         dev = devstr;
 850         label = strchr(devstr, '@');
 851         if (label == NULL ||
 852             ((label + 1) - dev) > done) {
 853                 return (EINVAL);
 854         }
 855         *label = '\0';
 856         label++;
 857         if (*label == '\0')
 858                 return (EINVAL);
 859
 860         if (mp->mnt_flag & MNT_UPDATE) {
 861                 /*
 862                  * Update mount.  Note that pmp->iroot->cluster is
 863                  * an inode-embedded cluster and thus cannot be
 864                  * directly locked.
 865                  *
 866                  * XXX HAMMER2 needs to implement NFS export via
 867                  *     mountctl.
 868                  */
 869                 pmp = MPTOPMP(mp);
 870                 pmp->hflags = info.hflags;
 871                 cluster = &pmp->iroot->cluster;
 872                 for (i = 0; i < cluster->nchains; ++i) {
 873                         if (cluster->array[i].chain == NULL)
 874                                 continue;
 875                         hmp = cluster->array[i].chain->hmp;
 876                         devvp = hmp->devvp;
 877                         error = hammer2_remount(hmp, mp, path,
 878                                                 devvp, cred);
 879                         if (error)
 880                                 break;
 881                 }
 882
 883                 return error;
 884         }
 885
 886         /*
 887          * HMP device mount
 888          *
 889          * Lookup name and verify it refers to a block device.
 890          */
 891         if (path) {
 892                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 893                 if (error == 0)
 894                         error = nlookup(&nd);
 895                 if (error == 0)
 896                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 897                 nlookup_done(&nd);
 898         } else {
 899                 /* root mount */
 900                 cdev_t cdev = kgetdiskbyname(dev);
 901                 error = bdevvp(cdev, &devvp);
 902                 if (error)
 903                         kprintf("hammer2: cannot find '%s'\n", dev);
 904         }
 905
 906         if (error == 0) {
 907                 if (vn_isdisk(devvp, &error))
 908                         error = vfs_mountedon(devvp);
 909         }
 910
 911         /*
 912          * Determine if the device has already been mounted.  After this
 913          * check hmp will be non-NULL if we are doing the second or more
 914          * hammer2 mounts from the same device.
 915          */
 916         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 917         TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 918                 if (hmp->devvp == devvp)
 919                         break;
 920         }
 921
 922         /*
 923          * Open the device if this isn't a secondary mount and construct
 924          * the H2 device mount (hmp).
 925          */
 926         if (hmp == NULL) {
 927                 hammer2_chain_t *schain;
 928                 hammer2_xid_t xid;
 929
 930                 if (error == 0 && vcount(devvp) > 0)
 931                         error = EBUSY;
 932
 933                 /*
 934                  * Now open the device
 935                  */
 936                 if (error == 0) {
 937                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
 938                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
 939                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
 940                         if (error == 0) {
 941                                 error = VOP_OPEN(devvp,
 942                                                  ronly ? FREAD : FREAD | FWRITE,
 943                                                  FSCRED, NULL);
 944                         }
 945                         vn_unlock(devvp);
 946                 }
 947                 if (error && devvp) {
 948                         vrele(devvp);
 949                         devvp = NULL;
 950                 }
 951                 if (error) {
 952                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 953                         return error;
 954                 }
 955                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
 956                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
 957                 hmp->ronly = ronly;
 958                 hmp->devvp = devvp;
 959                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
 960                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
 961                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
 962                 RB_INIT(&hmp->iotree);
 963                 spin_init(&hmp->io_spin, "hm2mount_io");
 964                 spin_init(&hmp->list_spin, "hm2mount_list");
 965                 TAILQ_INIT(&hmp->flushq);
 966
 967                 lockinit(&hmp->vollk, "h2vol", 0, 0);
 968                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
 969
 970                 /*
 971                  * vchain setup. vchain.data is embedded.
 972                  * vchain.refs is initialized and will never drop to 0.
 973                  *
 974                  * NOTE! voldata is not yet loaded.
 975                  */
 976                 hmp->vchain.hmp = hmp;
 977                 hmp->vchain.refs = 1;
 978                 hmp->vchain.data = (void *)&hmp->voldata;
 979                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
 980                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 981                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 982
 983                 hammer2_chain_core_init(&hmp->vchain);
 984                 /* hmp->vchain.u.xxx is left NULL */
 985
 986                 /*
 987                  * fchain setup.  fchain.data is embedded.
 988                  * fchain.refs is initialized and will never drop to 0.
 989                  *
 990                  * The data is not used but needs to be initialized to
 991                  * pass assertion muster.  We use this chain primarily
 992                  * as a placeholder for the freemap's top-level RBTREE
 993                  * so it does not interfere with the volume's topology
 994                  * RBTREE.
 995                  */
 996                 hmp->fchain.hmp = hmp;
 997                 hmp->fchain.refs = 1;
 998                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
 999                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1000                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1001                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1002                 hmp->fchain.bref.methods =
1003                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1004                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1005
1006                 hammer2_chain_core_init(&hmp->fchain);
1007                 /* hmp->fchain.u.xxx is left NULL */
1008
1009                 /*
1010                  * Install the volume header and initialize fields from
1011                  * voldata.
1012                  */
1013                 error = hammer2_install_volume_header(hmp);
1014                 if (error) {
1015                         hammer2_unmount_helper(mp, NULL, hmp);
1016                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1017                         hammer2_vfs_unmount(mp, MNT_FORCE);
1018                         return error;
1019                 }
1020
1021                 /*
1022                  * Really important to get these right or flush will get
1023                  * confused.
1024                  */
1025                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1026                 kprintf("alloc spmp %p tid %016jx\n",
1027                         hmp->spmp, hmp->voldata.mirror_tid);
1028                 spmp = hmp->spmp;
1029
1030                 /*
1031                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1032                  * is inherited from the volume header.
1033                  */
1034                 xid = 0;
1035                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1036                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1037                 hmp->vchain.pmp = spmp;
1038                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1039                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1040                 hmp->fchain.pmp = spmp;
1041
1042                 /*
1043                  * First locate the super-root inode, which is key 0
1044                  * relative to the volume header's blockset.
1045                  *
1046                  * Then locate the root inode by scanning the directory keyspace
1047                  * represented by the label.
1048                  */
1049                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1050                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1051                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1052                                       &cache_index, 0);
1053                 hammer2_chain_lookup_done(parent);
1054                 if (schain == NULL) {
1055                         kprintf("hammer2_mount: invalid super-root\n");
1056                         hammer2_unmount_helper(mp, NULL, hmp);
1057                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1058                         hammer2_vfs_unmount(mp, MNT_FORCE);
1059                         return EINVAL;
1060                 }
1061                 if (schain->error) {
1062                         kprintf("hammer2_mount: error %s reading super-root\n",
1063                                 hammer2_error_str(schain->error));
1064                         hammer2_chain_unlock(schain);
1065                         hammer2_chain_drop(schain);
1066                         schain = NULL;
1067                         hammer2_unmount_helper(mp, NULL, hmp);
1068                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1069                         hammer2_vfs_unmount(mp, MNT_FORCE);
1070                         return EINVAL;
1071                 }
1072
1073                 /*
1074                  * The super-root always uses an inode_tid of 1 when
1075                  * creating PFSs.
1076                  */
1077                 spmp->inode_tid = 1;
1078                 spmp->modify_tid = schain->bref.modify_tid + 1;
1079
1080                 /*
1081                  * Sanity-check schain's pmp and finish initialization.
1082                  * Any chain belonging to the super-root topology should
1083                  * have a NULL pmp (not even set to spmp).
1084                  */
1085                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1086                 KKASSERT(schain->pmp == NULL);
1087                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1088
1089                 /*
1090                  * Replace the dummy spmp->iroot with a real one.  It's
1091                  * easier to just do a wholesale replacement than to try
1092                  * to update the chain and fixup the iroot fields.
1093                  *
1094                  * The returned inode is locked with the supplied cluster.
1095                  */
1096                 cluster = hammer2_cluster_from_chain(schain);
1097                 hammer2_inode_drop(spmp->iroot);
1098                 spmp->iroot = NULL;
1099                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1100                 spmp->spmp_hmp = hmp;
1101                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1102                 spmp->pfs_hmps[0] = hmp;
1103                 hammer2_inode_ref(spmp->iroot);
1104                 hammer2_inode_unlock(spmp->iroot);
1105                 hammer2_cluster_unlock(cluster);
1106                 hammer2_cluster_drop(cluster);
1107                 schain = NULL;
1108                 /* leave spmp->iroot with one ref */
1109
1110                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1111                         error = hammer2_recovery(hmp);
1112                         /* XXX do something with error */
1113                 }
1114                 hammer2_update_pmps(hmp);
1115                 hammer2_iocom_init(hmp);
1116
1117                 /*
1118                  * Ref the cluster management messaging descriptor.  The mount
1119                  * program deals with the other end of the communications pipe.
1120                  *
1121                  * Root mounts typically do not supply one.
1122                  */
1123                 if (info.cluster_fd >= 0) {
1124                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1125                         if (fp) {
1126                                 hammer2_cluster_reconnect(hmp, fp);
1127                         } else {
1128                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1129                         }
1130                 }
1131         } else {
1132                 spmp = hmp->spmp;
1133                 if (info.hflags & HMNT2_DEVFLAGS) {
1134                         kprintf("hammer2: Warning: mount flags pertaining "
1135                                 "to the whole device may only be specified "
1136                                 "on the first mount of the device: %08x\n",
1137                                 info.hflags & HMNT2_DEVFLAGS);
1138                 }
1139         }
1140
1141         /*
1142          * Force local mount (disassociate all PFSs from their clusters).
1143          * Used primarily for debugging.
1144          */
1145         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1146
1147         /*
1148          * Lookup the mount point under the media-localized super-root.
1149          * Scanning hammer2_pfslist doesn't help us because it represents
1150          * PFS cluster ids which can aggregate several named PFSs together.
1151          *
1152          * cluster->pmp will incorrectly point to spmp and must be fixed
1153          * up later on.
1154          */
1155         hammer2_inode_lock(spmp->iroot, 0);
1156         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1157         lhc = hammer2_dirhash(label, strlen(label));
1158         chain = hammer2_chain_lookup(&parent, &key_next,
1159                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1160                                      &cache_index, 0);
1161         while (chain) {
1162                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1163                     strcmp(label, chain->data->ipdata.filename) == 0) {
1164                         break;
1165                 }
1166                 chain = hammer2_chain_next(&parent, chain, &key_next,
1167                                             key_next,
1168                                             lhc + HAMMER2_DIRHASH_LOMASK,
1169                                             &cache_index, 0);
1170         }
1171         if (parent) {
1172                 hammer2_chain_unlock(parent);
1173                 hammer2_chain_drop(parent);
1174         }
1175         hammer2_inode_unlock(spmp->iroot);
1176
1177         /*
1178          * PFS could not be found?
1179          */
1180         if (chain == NULL) {
1181                 kprintf("hammer2_mount: PFS label not found\n");
1182                 hammer2_unmount_helper(mp, NULL, hmp);
1183                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1184                 hammer2_vfs_unmount(mp, MNT_FORCE);
1185
1186                 return EINVAL;
1187         }
1188
1189         /*
1190          * Acquire the pmp structure (it should have already been allocated
1191          * via hammer2_update_pmps() so do not pass cluster in to add to
1192          * available chains).
1193          *
1194          * Check if the cluster has already been mounted.  A cluster can
1195          * only be mounted once, use null mounts to mount additional copies.
1196          */
1197         ripdata = &chain->data->ipdata;
1198         bref = chain->bref;
1199         pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid, force_local);
1200         hammer2_chain_unlock(chain);
1201         hammer2_chain_drop(chain);
1202
1203         if (pmp->mp) {
1204                 kprintf("hammer2_mount: PFS already mounted!\n");
1205                 hammer2_unmount_helper(mp, NULL, hmp);
1206                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1207                 hammer2_vfs_unmount(mp, MNT_FORCE);
1208
1209                 return EBUSY;
1210         }
1211
1212         /*
1213          * Finish the mount
1214          */
1215         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1216
1217         pmp->hflags = info.hflags;
1218         mp->mnt_flag = MNT_LOCAL;
1219         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1220         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1221
1222         /*
1223          * required mount structure initializations
1224          */
1225         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1226         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1227
1228         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1229         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1230
1231         /*
1232          * Optional fields
1233          */
1234         mp->mnt_iosize_max = MAXPHYS;
1235
1236         /*
1237          * Connect up mount pointers.
1238          */
1239         hammer2_mount_helper(mp, pmp);
1240
1241         lockmgr(&hammer2_mntlk, LK_RELEASE);
1242
1243         /*
1244          * Finish setup
1245          */
1246         vfs_getnewfsid(mp);
1247         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1248         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1249         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1250
1251         if (path) {
1252                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1253                           MNAMELEN - 1, &size);
1254                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1255         } /* else root mount, already in there */
1256
1257         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1258         if (path) {
1259                 copyinstr(path, mp->mnt_stat.f_mntonname,
1260                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1261                           &size);
1262         } else {
1263                 /* root mount */
1264                 mp->mnt_stat.f_mntonname[0] = '/';
1265         }
1266
1267         /*
1268          * Initial statfs to prime mnt_stat.
1269          */
1270         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1271
1272         return 0;
1273 }
1274
1275 /*
1276  * Scan PFSs under the super-root and create hammer2_pfs structures.
1277  */
1278 static
1279 void
1280 hammer2_update_pmps(hammer2_dev_t *hmp)
1281 {
1282         const hammer2_inode_data_t *ripdata;
1283         hammer2_chain_t *parent;
1284         hammer2_chain_t *chain;
1285         hammer2_blockref_t bref;
1286         hammer2_dev_t *force_local;
1287         hammer2_pfs_t *spmp;
1288         hammer2_pfs_t *pmp;
1289         hammer2_key_t key_next;
1290         int cache_index = -1;
1291
1292         /*
1293          * Force local mount (disassociate all PFSs from their clusters).
1294          * Used primarily for debugging.
1295          */
1296         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1297
1298         /*
1299          * Lookup mount point under the media-localized super-root.
1300          *
1301          * cluster->pmp will incorrectly point to spmp and must be fixed
1302          * up later on.
1303          */
1304         spmp = hmp->spmp;
1305         hammer2_inode_lock(spmp->iroot, 0);
1306         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1307         chain = hammer2_chain_lookup(&parent, &key_next,
1308                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1309                                          &cache_index, 0);
1310         while (chain) {
1311                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1312                         continue;
1313                 ripdata = &chain->data->ipdata;
1314                 bref = chain->bref;
1315                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1316
1317                 pmp = hammer2_pfsalloc(chain, ripdata,
1318                                        bref.modify_tid, force_local);
1319                 chain = hammer2_chain_next(&parent, chain, &key_next,
1320                                            key_next, HAMMER2_KEY_MAX,
1321                                            &cache_index, 0);
1322         }
1323         if (parent) {
1324                 hammer2_chain_unlock(parent);
1325                 hammer2_chain_drop(parent);
1326         }
1327         hammer2_inode_unlock(spmp->iroot);
1328 }
1329
1330 static
1331 int
1332 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1333                 struct vnode *devvp, struct ucred *cred)
1334 {
1335         int error;
1336
1337         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1338                 error = hammer2_recovery(hmp);
1339         } else {
1340                 error = 0;
1341         }
1342         return error;
1343 }
1344
1345 static
1346 int
1347 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1348 {
1349         hammer2_pfs_t *pmp;
1350         int flags;
1351         int error = 0;
1352
1353         pmp = MPTOPMP(mp);
1354
1355         if (pmp == NULL)
1356                 return(0);
1357
1358         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1359
1360         /*
1361          * If mount initialization proceeded far enough we must flush
1362          * its vnodes and sync the underlying mount points.  Three syncs
1363          * are required to fully flush the filesystem (freemap updates lag
1364          * by one flush, and one extra for safety).
1365          */
1366         if (mntflags & MNT_FORCE)
1367                 flags = FORCECLOSE;
1368         else
1369                 flags = 0;
1370         if (pmp->iroot) {
1371                 error = vflush(mp, 0, flags);
1372                 if (error)
1373                         goto failed;
1374                 hammer2_vfs_sync(mp, MNT_WAIT);
1375                 hammer2_vfs_sync(mp, MNT_WAIT);
1376                 hammer2_vfs_sync(mp, MNT_WAIT);
1377         }
1378
1379         /*
1380          * Cleanup the frontend support XOPS threads
1381          */
1382         hammer2_xop_helper_cleanup(pmp);
1383
1384         /*
1385          * Cleanup our reference on ihidden.
1386          */
1387         if (pmp->ihidden) {
1388                 hammer2_inode_drop(pmp->ihidden);
1389                 pmp->ihidden = NULL;
1390         }
1391         if (pmp->mp)
1392                 hammer2_unmount_helper(mp, pmp, NULL);
1393
1394         error = 0;
1395 failed:
1396         lockmgr(&hammer2_mntlk, LK_RELEASE);
1397
1398         return (error);
1399 }
1400
1401 /*
1402  * Mount helper, hook the system mount into our PFS.
1403  * The mount lock is held.
1404  *
1405  * We must bump the mount_count on related devices for any
1406  * mounted PFSs.
1407  */
1408 static
1409 void
1410 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1411 {
1412         hammer2_cluster_t *cluster;
1413         hammer2_chain_t *rchain;
1414         int i;
1415
1416         mp->mnt_data = (qaddr_t)pmp;
1417         pmp->mp = mp;
1418
1419         /*
1420          * After pmp->mp is set we have to adjust hmp->mount_count.
1421          */
1422         cluster = &pmp->iroot->cluster;
1423         for (i = 0; i < cluster->nchains; ++i) {
1424                 rchain = cluster->array[i].chain;
1425                 if (rchain == NULL)
1426                         continue;
1427                 ++rchain->hmp->mount_count;
1428                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1429                         rchain->hmp, rchain->hmp->mount_count);
1430         }
1431
1432         /*
1433          * Create missing Xop threads
1434          */
1435         hammer2_xop_helper_create(pmp);
1436 }
1437
1438 /*
1439  * Mount helper, unhook the system mount from our PFS.
1440  * The mount lock is held.
1441  *
1442  * If hmp is supplied a mount responsible for being the first to open
1443  * the block device failed and the block device and all PFSs using the
1444  * block device must be cleaned up.
1445  *
1446  * If pmp is supplied multiple devices might be backing the PFS and each
1447  * must be disconnected.  This might not be the last PFS using some of the
1448  * underlying devices.  Also, we have to adjust our hmp->mount_count
1449  * accounting for the devices backing the pmp which is now undergoing an
1450  * unmount.
1451  */
1452 static
1453 void
1454 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1455 {
1456         hammer2_cluster_t *cluster;
1457         hammer2_chain_t *rchain;
1458         struct vnode *devvp;
1459         int dumpcnt;
1460         int ronly = 0;
1461         int i;
1462
1463         /*
1464          * If no device supplied this is a high-level unmount and we have to
1465          * to disconnect the mount, adjust mount_count, and locate devices
1466          * that might now have no mounts.
1467          */
1468         if (pmp) {
1469                 KKASSERT(hmp == NULL);
1470                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1471                 pmp->mp = NULL;
1472                 mp->mnt_data = NULL;
1473
1474                 /*
1475                  * After pmp->mp is cleared we have to account for
1476                  * mount_count.
1477                  */
1478                 cluster = &pmp->iroot->cluster;
1479                 for (i = 0; i < cluster->nchains; ++i) {
1480                         rchain = cluster->array[i].chain;
1481                         if (rchain == NULL)
1482                                 continue;
1483                         --rchain->hmp->mount_count;
1484                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1485                                 rchain->hmp, rchain->hmp->mount_count);
1486                         /* scrapping hmp now may invalidate the pmp */
1487                 }
1488 again:
1489                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1490                         if (hmp->mount_count == 0) {
1491                                 hammer2_unmount_helper(NULL, NULL, hmp);
1492                                 goto again;
1493                         }
1494                 }
1495                 return;
1496         }
1497
1498         /*
1499          * Try to terminate the block device.  We can't terminate it if
1500          * there are still PFSs referencing it.
1501          */
1502         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1503                 hmp, hmp->mount_count);
1504         if (hmp->mount_count)
1505                 return;
1506
1507         hammer2_pfsfree_scan(hmp);
1508         hammer2_dev_exlock(hmp);        /* XXX order */
1509
1510         /*
1511          * Cycle the volume data lock as a safety (probably not needed any
1512          * more).  To ensure everything is out we need to flush at least
1513          * three times.  (1) The running of the sideq can dirty the
1514          * filesystem, (2) A normal flush can dirty the freemap, and
1515          * (3) ensure that the freemap is fully synchronized.
1516          *
1517          * The next mount's recovery scan can clean everything up but we want
1518          * to leave the filesystem in a 100% clean state on a normal unmount.
1519          */
1520 #if 0
1521         hammer2_voldata_lock(hmp);
1522         hammer2_voldata_unlock(hmp);
1523 #endif
1524         hammer2_iocom_uninit(hmp);
1525
1526         if ((hmp->vchain.flags | hmp->fchain.flags) &
1527             HAMMER2_CHAIN_FLUSH_MASK) {
1528                 kprintf("hammer2_unmount: chains left over "
1529                         "after final sync\n");
1530                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1531                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1532
1533                 if (hammer2_debug & 0x0010)
1534                         Debugger("entered debugger");
1535         }
1536
1537         KKASSERT(hmp->spmp == NULL);
1538
1539         /*
1540          * Finish up with the device vnode
1541          */
1542         if ((devvp = hmp->devvp) != NULL) {
1543                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1544                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1545                 hmp->devvp = NULL;
1546                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1547                 vn_unlock(devvp);
1548                 vrele(devvp);
1549                 devvp = NULL;
1550         }
1551
1552         /*
1553          * Clear vchain/fchain flags that might prevent final cleanup
1554          * of these chains.
1555          */
1556         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1557                 atomic_add_long(&hammer2_count_modified_chains, -1);
1558                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1559                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1560         }
1561         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1562                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1563         }
1564
1565         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1566                 atomic_add_long(&hammer2_count_modified_chains, -1);
1567                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1568                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1569         }
1570         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1571                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1572         }
1573
1574         /*
1575          * Final drop of embedded freemap root chain to
1576          * clean up fchain.core (fchain structure is not
1577          * flagged ALLOCATED so it is cleaned out and then
1578          * left to rot).
1579          */
1580         hammer2_chain_drop(&hmp->fchain);
1581
1582         /*
1583          * Final drop of embedded volume root chain to clean
1584          * up vchain.core (vchain structure is not flagged
1585          * ALLOCATED so it is cleaned out and then left to
1586          * rot).
1587          */
1588         dumpcnt = 50;
1589         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1590         dumpcnt = 50;
1591         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1592         hammer2_dev_unlock(hmp);
1593         hammer2_chain_drop(&hmp->vchain);
1594
1595         hammer2_io_cleanup(hmp, &hmp->iotree);
1596         if (hmp->iofree_count) {
1597                 kprintf("io_cleanup: %d I/O's left hanging\n",
1598                         hmp->iofree_count);
1599         }
1600
1601         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1602         kmalloc_destroy(&hmp->mchain);
1603         kfree(hmp, M_HAMMER2);
1604 }
1605
1606 static
1607 int
1608 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1609                  ino_t ino, struct vnode **vpp)
1610 {
1611         hammer2_xop_lookup_t *xop;
1612         hammer2_pfs_t *pmp;
1613         hammer2_inode_t *ip;
1614         hammer2_tid_t inum;
1615         int error;
1616
1617         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1618
1619         error = 0;
1620         pmp = MPTOPMP(mp);
1621
1622         /*
1623          * Easy if we already have it cached
1624          */
1625         ip = hammer2_inode_lookup(pmp, inum);
1626         if (ip) {
1627                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1628                 *vpp = hammer2_igetv(ip, &error);
1629                 hammer2_inode_unlock(ip);
1630                 hammer2_inode_drop(ip);         /* from lookup */
1631
1632                 return error;
1633         }
1634
1635         /*
1636          * Otherwise we have to find the inode
1637          */
1638         xop = hammer2_xop_alloc(pmp->iroot, 0);
1639         xop->lhc = inum;
1640         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1641         error = hammer2_xop_collect(&xop->head, 0);
1642
1643         if (error == 0)
1644                 ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1645         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1646
1647         if (ip) {
1648                 hammer2_inode_resolve_pip(ip);
1649                 *vpp = hammer2_igetv(ip, &error);
1650                 hammer2_inode_unlock(ip);
1651         } else {
1652                 *vpp = NULL;
1653                 error = ENOENT;
1654         }
1655         return (error);
1656 }
1657
1658 static
1659 int
1660 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1661 {
1662         hammer2_pfs_t *pmp;
1663         struct vnode *vp;
1664         int error;
1665
1666         pmp = MPTOPMP(mp);
1667         if (pmp->iroot == NULL) {
1668                 *vpp = NULL;
1669                 return EINVAL;
1670         }
1671
1672         error = 0;
1673         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1674
1675         while (pmp->inode_tid == 0) {
1676                 hammer2_xop_ipcluster_t *xop;
1677                 hammer2_inode_meta_t *meta;
1678
1679                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1680                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1681                 error = hammer2_xop_collect(&xop->head, 0);
1682
1683                 if (error == 0) {
1684                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1685                         pmp->iroot->meta = *meta;
1686                         pmp->inode_tid = meta->pfs_inum + 1;
1687                         if (pmp->inode_tid < HAMMER2_INODE_START)
1688                                 pmp->inode_tid = HAMMER2_INODE_START;
1689                         pmp->modify_tid =
1690                                 xop->head.cluster.focus->bref.modify_tid + 1;
1691                         kprintf("PFS: Starting inode %jd\n",
1692                                 (intmax_t)pmp->inode_tid);
1693                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1694                                 pmp->inode_tid, pmp->modify_tid);
1695                         wakeup(&pmp->iroot);
1696
1697                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1698
1699                         /*
1700                          * Prime the mount info.
1701                          */
1702                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1703
1704                         /*
1705                          * With the cluster operational, check for and
1706                          * install ihidden if needed.  The install_hidden
1707                          * code needs to get a transaction so we must unlock
1708                          * iroot around it.
1709                          *
1710                          * This is only applicable PFS mounts, there is no
1711                          * hidden directory in the spmp.
1712                          */
1713                         hammer2_inode_unlock(pmp->iroot);
1714                         hammer2_inode_install_hidden(pmp);
1715                         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1716
1717                         break;
1718                 }
1719
1720                 /*
1721                  * Loop, try again
1722                  */
1723                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1724                 hammer2_inode_unlock(pmp->iroot);
1725                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1726                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1727                 if (error == EINTR)
1728                         break;
1729         }
1730
1731         if (error) {
1732                 hammer2_inode_unlock(pmp->iroot);
1733                 *vpp = NULL;
1734         } else {
1735                 vp = hammer2_igetv(pmp->iroot, &error);
1736                 hammer2_inode_unlock(pmp->iroot);
1737                 *vpp = vp;
1738         }
1739
1740         return (error);
1741 }
1742
1743 /*
1744  * Filesystem status
1745  *
1746  * XXX incorporate ipdata->meta.inode_quota and data_quota
1747  */
1748 static
1749 int
1750 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1751 {
1752         hammer2_pfs_t *pmp;
1753         hammer2_dev_t *hmp;
1754         hammer2_blockref_t bref;
1755         int i;
1756
1757         /*
1758          * NOTE: iroot might not have validated the cluster yet.
1759          */
1760         pmp = MPTOPMP(mp);
1761
1762         mp->mnt_stat.f_files = 0;
1763         mp->mnt_stat.f_ffree = 0;
1764         mp->mnt_stat.f_blocks = 0;
1765         mp->mnt_stat.f_bfree = 0;
1766         mp->mnt_stat.f_bavail = 0;
1767
1768         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1769                 hmp = pmp->pfs_hmps[i];
1770                 if (hmp == NULL)
1771                         continue;
1772                 if (pmp->iroot->cluster.array[i].chain)
1773                         bref = pmp->iroot->cluster.array[i].chain->bref;
1774                 else
1775                         bzero(&bref, sizeof(bref));
1776
1777                 mp->mnt_stat.f_files = bref.inode_count;
1778                 mp->mnt_stat.f_ffree = 0;
1779                 mp->mnt_stat.f_blocks = (bref.data_count +
1780                                          hmp->voldata.allocator_free) /
1781                                         mp->mnt_vstat.f_bsize;
1782                 mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free /
1783                                         mp->mnt_vstat.f_bsize;
1784                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1785
1786                 *sbp = mp->mnt_stat;
1787         }
1788         return (0);
1789 }
1790
1791 static
1792 int
1793 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1794 {
1795         hammer2_pfs_t *pmp;
1796         hammer2_dev_t *hmp;
1797         hammer2_blockref_t bref;
1798         int i;
1799
1800         /*
1801          * NOTE: iroot might not have validated the cluster yet.
1802          */
1803         pmp = MPTOPMP(mp);
1804
1805         mp->mnt_vstat.f_bsize = 0;
1806         mp->mnt_vstat.f_files = 0;
1807         mp->mnt_vstat.f_ffree = 0;
1808         mp->mnt_vstat.f_blocks = 0;
1809         mp->mnt_vstat.f_bfree = 0;
1810         mp->mnt_vstat.f_bavail = 0;
1811
1812         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1813                 hmp = pmp->pfs_hmps[i];
1814                 if (hmp == NULL)
1815                         continue;
1816                 if (pmp->iroot->cluster.array[i].chain)
1817                         bref = pmp->iroot->cluster.array[i].chain->bref;
1818                 else
1819                         bzero(&bref, sizeof(bref));
1820
1821                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1822                 mp->mnt_vstat.f_files = bref.inode_count;
1823                 mp->mnt_vstat.f_ffree = 0;
1824                 mp->mnt_vstat.f_blocks = (bref.data_count +
1825                                          hmp->voldata.allocator_free) /
1826                                         mp->mnt_vstat.f_bsize;
1827                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1828                                         mp->mnt_vstat.f_bsize;
1829                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1830
1831                 *sbp = mp->mnt_vstat;
1832         }
1833         return (0);
1834 }
1835
1836 /*
1837  * Mount-time recovery (RW mounts)
1838  *
1839  * Updates to the free block table are allowed to lag flushes by one
1840  * transaction.  In case of a crash, then on a fresh mount we must do an
1841  * incremental scan of the last committed transaction id and make sure that
1842  * all related blocks have been marked allocated.
1843  *
1844  * The super-root topology and each PFS has its own transaction id domain,
1845  * so we must track PFS boundary transitions.
1846  */
1847 struct hammer2_recovery_elm {
1848         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1849         hammer2_chain_t *chain;
1850         hammer2_tid_t sync_tid;
1851 };
1852
1853 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1854
1855 struct hammer2_recovery_info {
1856         struct hammer2_recovery_list list;
1857         hammer2_tid_t   mtid;
1858         int     depth;
1859 };
1860
1861 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1862                         hammer2_chain_t *parent,
1863                         struct hammer2_recovery_info *info,
1864                         hammer2_tid_t sync_tid);
1865
1866 #define HAMMER2_RECOVERY_MAXDEPTH       10
1867
1868 static
1869 int
1870 hammer2_recovery(hammer2_dev_t *hmp)
1871 {
1872         struct hammer2_recovery_info info;
1873         struct hammer2_recovery_elm *elm;
1874         hammer2_chain_t *parent;
1875         hammer2_tid_t sync_tid;
1876         hammer2_tid_t mirror_tid;
1877         int error;
1878         int cumulative_error = 0;
1879
1880         hammer2_trans_init(hmp->spmp, 0);
1881
1882         sync_tid = hmp->voldata.freemap_tid;
1883         mirror_tid = hmp->voldata.mirror_tid;
1884
1885         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1886         if (sync_tid >= mirror_tid) {
1887                 kprintf(" no recovery needed\n");
1888         } else {
1889                 kprintf(" freemap recovery %016jx-%016jx\n",
1890                         sync_tid + 1, mirror_tid);
1891         }
1892
1893         TAILQ_INIT(&info.list);
1894         info.depth = 0;
1895         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1896         cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
1897         hammer2_chain_lookup_done(parent);
1898
1899         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1900                 TAILQ_REMOVE(&info.list, elm, entry);
1901                 parent = elm->chain;
1902                 sync_tid = elm->sync_tid;
1903                 kfree(elm, M_HAMMER2);
1904
1905                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1906                 error = hammer2_recovery_scan(hmp, parent, &info,
1907                                               hmp->voldata.freemap_tid);
1908                 hammer2_chain_unlock(parent);
1909                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
1910                 if (error)
1911                         cumulative_error = error;
1912         }
1913         hammer2_trans_done(hmp->spmp);
1914
1915         return cumulative_error;
1916 }
1917
1918 static
1919 int
1920 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1921                       struct hammer2_recovery_info *info,
1922                       hammer2_tid_t sync_tid)
1923 {
1924         const hammer2_inode_data_t *ripdata;
1925         hammer2_chain_t *chain;
1926         hammer2_blockref_t bref;
1927         int cache_index;
1928         int cumulative_error = 0;
1929         int error;
1930         int first;
1931
1932         /*
1933          * Adjust freemap to ensure that the block(s) are marked allocated.
1934          */
1935         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1936                 hammer2_freemap_adjust(hmp, &parent->bref,
1937                                        HAMMER2_FREEMAP_DORECOVER);
1938         }
1939
1940         /*
1941          * Check type for recursive scan
1942          */
1943         switch(parent->bref.type) {
1944         case HAMMER2_BREF_TYPE_VOLUME:
1945                 /* data already instantiated */
1946                 break;
1947         case HAMMER2_BREF_TYPE_INODE:
1948                 /*
1949                  * Must instantiate data for DIRECTDATA test and also
1950                  * for recursion.
1951                  */
1952                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1953                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
1954                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1955                         /* not applicable to recovery scan */
1956                         hammer2_chain_unlock(parent);
1957                         return 0;
1958                 }
1959                 hammer2_chain_unlock(parent);
1960                 break;
1961         case HAMMER2_BREF_TYPE_INDIRECT:
1962                 /*
1963                  * Must instantiate data for recursion
1964                  */
1965                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1966                 hammer2_chain_unlock(parent);
1967                 break;
1968         case HAMMER2_BREF_TYPE_DATA:
1969         case HAMMER2_BREF_TYPE_FREEMAP:
1970         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1971         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1972                 /* not applicable to recovery scan */
1973                 return 0;
1974                 break;
1975         default:
1976                 return EDOM;
1977         }
1978
1979         /*
1980          * Defer operation if depth limit reached or if we are crossing a
1981          * PFS boundary.
1982          */
1983         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1984                 struct hammer2_recovery_elm *elm;
1985
1986                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1987                 elm->chain = parent;
1988                 elm->sync_tid = sync_tid;
1989                 hammer2_chain_ref(parent);
1990                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
1991                 /* unlocked by caller */
1992
1993                 return(0);
1994         }
1995
1996
1997         /*
1998          * Recursive scan of the last flushed transaction only.  We are
1999          * doing this without pmp assignments so don't leave the chains
2000          * hanging around after we are done with them.
2001          */
2002         cache_index = 0;
2003         chain = NULL;
2004         first = 1;
2005
2006         while (hammer2_chain_scan(parent, &chain, &bref,
2007                                   &first, &cache_index,
2008                                   HAMMER2_LOOKUP_NODATA) != NULL) {
2009                 /*
2010                  * If this is a leaf
2011                  */
2012                 if (chain == NULL) {
2013                         if (bref.mirror_tid > sync_tid) {
2014                                 hammer2_freemap_adjust(hmp, &bref,
2015                                                      HAMMER2_FREEMAP_DORECOVER);
2016                         }
2017                         continue;
2018                 }
2019
2020                 /*
2021                  * This may or may not be a recursive node.
2022                  */
2023                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2024                 if (bref.mirror_tid > sync_tid) {
2025                         ++info->depth;
2026                         error = hammer2_recovery_scan(hmp, chain,
2027                                                       info, sync_tid);
2028                         --info->depth;
2029                         if (error)
2030                                 cumulative_error = error;
2031                 }
2032
2033                 /*
2034                  * Flush the recovery at the PFS boundary to stage it for
2035                  * the final flush of the super-root topology.
2036                  */
2037                 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2038                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2039                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2040                 }
2041         }
2042
2043         return cumulative_error;
2044 }
2045
2046 /*
2047  * Sync a mount point; this is called on a per-mount basis from the
2048  * filesystem syncer process periodically and whenever a user issues
2049  * a sync.
2050  */
2051 int
2052 hammer2_vfs_sync(struct mount *mp, int waitfor)
2053 {
2054         hammer2_xop_flush_t *xop;
2055         struct hammer2_sync_info info;
2056         hammer2_inode_t *iroot;
2057         hammer2_pfs_t *pmp;
2058         int flags;
2059         int error;
2060
2061         pmp = MPTOPMP(mp);
2062         iroot = pmp->iroot;
2063         KKASSERT(iroot);
2064         KKASSERT(iroot->pmp == pmp);
2065
2066         /*
2067          * We can't acquire locks on existing vnodes while in a transaction
2068          * without risking a deadlock.  This assumes that vfsync() can be
2069          * called without the vnode locked (which it can in DragonFly).
2070          * Otherwise we'd have to implement a multi-pass or flag the lock
2071          * failures and retry.
2072          *
2073          * The reclamation code interlocks with the sync list's token
2074          * (by removing the vnode from the scan list) before unlocking
2075          * the inode, giving us time to ref the inode.
2076          */
2077         /*flags = VMSC_GETVP;*/
2078         flags = 0;
2079         if (waitfor & MNT_LAZY)
2080                 flags |= VMSC_ONEPASS;
2081
2082 #if 0
2083         /*
2084          * Preflush the vnodes using a normal transaction before interlocking
2085          * with a flush transaction.
2086          */
2087         hammer2_trans_init(pmp, 0);
2088         info.error = 0;
2089         info.waitfor = MNT_NOWAIT;
2090         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2091         hammer2_trans_done(pmp);
2092 #endif
2093
2094         /*
2095          * Start our flush transaction.  This does not return until all
2096          * concurrent transactions have completed and will prevent any
2097          * new transactions from running concurrently, except for the
2098          * buffer cache transactions.
2099          *
2100          * For efficiency do an async pass before making sure with a
2101          * synchronous pass on all related buffer cache buffers.  It
2102          * should theoretically not be possible for any new file buffers
2103          * to be instantiated during this sequence.
2104          */
2105         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH |
2106                                 HAMMER2_TRANS_PREFLUSH);
2107         hammer2_inode_run_sideq(pmp);
2108
2109         info.error = 0;
2110         info.waitfor = MNT_NOWAIT;
2111         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2112         info.waitfor = MNT_WAIT;
2113         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2114
2115         /*
2116          * Clear PREFLUSH.  This prevents (or asserts on) any new logical
2117          * buffer cache flushes which occur during the flush.  Device buffers
2118          * are not affected.
2119          */
2120         hammer2_bioq_sync(pmp);
2121         hammer2_trans_clear_preflush(pmp);
2122
2123         /*
2124          * Use the XOP interface to concurrently flush all nodes to
2125          * synchronize the PFSROOT subtopology to the media.  A standard
2126          * end-of-scan ENOENT error indicates cluster sufficiency.
2127          *
2128          * Note that this flush will not be visible on crash recovery until
2129          * we flush the super-root topology in the next loop.
2130          *
2131          * XXX For now wait for all flushes to complete.
2132          */
2133         if (iroot) {
2134                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2135                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2136                 error = hammer2_xop_collect(&xop->head,
2137                                             HAMMER2_XOP_COLLECT_WAITALL);
2138                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2139                 if (error == ENOENT)
2140                         error = 0;
2141         } else {
2142                 error = 0;
2143         }
2144         hammer2_trans_done(pmp);
2145
2146         return (error);
2147 }
2148
2149 /*
2150  * Sync passes.
2151  *
2152  * Note that we ignore the tranasction mtid we got above.  Instead,
2153  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2154  * transactions.
2155  */
2156 static int
2157 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2158 {
2159         struct hammer2_sync_info *info = data;
2160         hammer2_inode_t *ip;
2161         int error;
2162
2163         /*
2164          * Degenerate cases.  Note that ip == NULL typically means the
2165          * syncer vnode itself and we don't want to vclrisdirty() in that
2166          * situation.
2167          */
2168         ip = VTOI(vp);
2169         if (ip == NULL) {
2170                 return(0);
2171         }
2172         if (vp->v_type == VNON || vp->v_type == VBAD) {
2173                 vclrisdirty(vp);
2174                 return(0);
2175         }
2176
2177         /*
2178          * VOP_FSYNC will start a new transaction so replicate some code
2179          * here to do it inline (see hammer2_vop_fsync()).
2180          *
2181          * WARNING: The vfsync interacts with the buffer cache and might
2182          *          block, we can't hold the inode lock at that time.
2183          *          However, we MUST ref ip before blocking to ensure that
2184          *          it isn't ripped out from under us (since we do not
2185          *          hold a lock on the vnode).
2186          */
2187         hammer2_inode_ref(ip);
2188         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2189             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2190                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2191                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2192                                  HAMMER2_INODE_MODIFIED)) {
2193                         hammer2_inode_lock(ip, 0);
2194                         hammer2_inode_chain_sync(ip);
2195                         hammer2_inode_unlock(ip);
2196                 }
2197         }
2198         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2199             RB_EMPTY(&vp->v_rbdirty_tree)) {
2200                 vclrisdirty(vp);
2201         }
2202
2203         hammer2_inode_drop(ip);
2204 #if 1
2205         error = 0;
2206         if (error)
2207                 info->error = error;
2208 #endif
2209         return(0);
2210 }
2211
2212 static
2213 int
2214 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2215 {
2216         hammer2_inode_t *ip;
2217
2218         KKASSERT(MAXFIDSZ >= 16);
2219         ip = VTOI(vp);
2220         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2221         fhp->fid_ext = 0;
2222         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2223         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2224
2225         return 0;
2226 }
2227
2228 static
2229 int
2230 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2231                struct fid *fhp, struct vnode **vpp)
2232 {
2233         hammer2_pfs_t *pmp;
2234         hammer2_tid_t inum;
2235         int error;
2236
2237         pmp = MPTOPMP(mp);
2238         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2239         if (vpp) {
2240                 if (inum == 1)
2241                         error = hammer2_vfs_root(mp, vpp);
2242                 else
2243                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2244         } else {
2245                 error = 0;
2246         }
2247         if (error)
2248                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2249         return error;
2250 }
2251
2252 static
2253 int
2254 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2255                  int *exflagsp, struct ucred **credanonp)
2256 {
2257         hammer2_pfs_t *pmp;
2258         struct netcred *np;
2259         int error;
2260
2261         pmp = MPTOPMP(mp);
2262         np = vfs_export_lookup(mp, &pmp->export, nam);
2263         if (np) {
2264                 *exflagsp = np->netc_exflags;
2265                 *credanonp = &np->netc_anon;
2266                 error = 0;
2267         } else {
2268                 error = EACCES;
2269         }
2270         return error;
2271 }
2272
2273 /*
2274  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2275  * header into the HMP
2276  *
2277  * XXX read four volhdrs and use the one with the highest TID whos CRC
2278  *     matches.
2279  *
2280  * XXX check iCRCs.
2281  *
2282  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2283  *     nonexistant locations.
2284  *
2285  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2286  */
2287 static
2288 int
2289 hammer2_install_volume_header(hammer2_dev_t *hmp)
2290 {
2291         hammer2_volume_data_t *vd;
2292         struct buf *bp;
2293         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2294         int error_reported;
2295         int error;
2296         int valid;
2297         int i;
2298
2299         error_reported = 0;
2300         error = 0;
2301         valid = 0;
2302         bp = NULL;
2303
2304         /*
2305          * There are up to 4 copies of the volume header (syncs iterate
2306          * between them so there is no single master).  We don't trust the
2307          * volu_size field so we don't know precisely how large the filesystem
2308          * is, so depend on the OS to return an error if we go beyond the
2309          * block device's EOF.
2310          */
2311         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2312                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2313                               HAMMER2_VOLUME_BYTES, &bp);
2314                 if (error) {
2315                         brelse(bp);
2316                         bp = NULL;
2317                         continue;
2318                 }
2319
2320                 vd = (struct hammer2_volume_data *) bp->b_data;
2321                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2322                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2323                         brelse(bp);
2324                         bp = NULL;
2325                         continue;
2326                 }
2327
2328                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2329                         /* XXX: Reversed-endianness filesystem */
2330                         kprintf("hammer2: reverse-endian filesystem detected");
2331                         brelse(bp);
2332                         bp = NULL;
2333                         continue;
2334                 }
2335
2336                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2337                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2338                                       HAMMER2_VOLUME_ICRC0_SIZE);
2339                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2340                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2341                                        HAMMER2_VOLUME_ICRC1_SIZE);
2342                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2343                         kprintf("hammer2 volume header crc "
2344                                 "mismatch copy #%d %08x/%08x\n",
2345                                 i, crc0, crc);
2346                         error_reported = 1;
2347                         brelse(bp);
2348                         bp = NULL;
2349                         continue;
2350                 }
2351                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2352                         valid = 1;
2353                         hmp->voldata = *vd;
2354                         hmp->volhdrno = i;
2355                 }
2356                 brelse(bp);
2357                 bp = NULL;
2358         }
2359         if (valid) {
2360                 hmp->volsync = hmp->voldata;
2361                 error = 0;
2362                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2363                         kprintf("hammer2: using volume header #%d\n",
2364                                 hmp->volhdrno);
2365                 }
2366         } else {
2367                 error = EINVAL;
2368                 kprintf("hammer2: no valid volume headers found!\n");
2369         }
2370         return (error);
2371 }
2372
2373 /*
2374  * This handles hysteresis on regular file flushes.  Because the BIOs are
2375  * routed to a thread it is possible for an excessive number to build up
2376  * and cause long front-end stalls long before the runningbuffspace limit
2377  * is hit, so we implement hammer2_flush_pipe to control the
2378  * hysteresis.
2379  *
2380  * This is a particular problem when compression is used.
2381  */
2382 void
2383 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2384 {
2385         atomic_add_int(&pmp->count_lwinprog, 1);
2386 }
2387
2388 void
2389 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2390 {
2391         int lwinprog;
2392
2393         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2394         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2395             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2396                 atomic_clear_int(&pmp->count_lwinprog,
2397                                  HAMMER2_LWINPROG_WAITING);
2398                 wakeup(&pmp->count_lwinprog);
2399         }
2400         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2401             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2402                 atomic_clear_int(&pmp->count_lwinprog,
2403                                  HAMMER2_LWINPROG_WAITING0);
2404                 wakeup(&pmp->count_lwinprog);
2405         }
2406 }
2407
2408 void
2409 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2410 {
2411         int lwinprog;
2412         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2413                                     HAMMER2_LWINPROG_WAITING0;
2414
2415         for (;;) {
2416                 lwinprog = pmp->count_lwinprog;
2417                 cpu_ccfence();
2418                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2419                         break;
2420                 tsleep_interlock(&pmp->count_lwinprog, 0);
2421                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2422                 lwinprog = pmp->count_lwinprog;
2423                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2424                         break;
2425                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2426         }
2427 }
2428
2429 /*
2430  * Manage excessive memory resource use for chain and related
2431  * structures.
2432  */
2433 void
2434 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2435 {
2436         uint32_t waiting;
2437         uint32_t count;
2438         uint32_t limit;
2439 #if 0
2440         static int zzticks;
2441 #endif
2442
2443         /*
2444          * Atomic check condition and wait.  Also do an early speedup of
2445          * the syncer to try to avoid hitting the wait.
2446          */
2447         for (;;) {
2448                 waiting = pmp->inmem_dirty_chains;
2449                 cpu_ccfence();
2450                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2451
2452                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2453                 if (limit < hammer2_limit_dirty_chains)
2454                         limit = hammer2_limit_dirty_chains;
2455                 if (limit < 1000)
2456                         limit = 1000;
2457
2458 #if 0
2459                 if ((int)(ticks - zzticks) > hz) {
2460                         zzticks = ticks;
2461                         kprintf("count %ld %ld\n", count, limit);
2462                 }
2463 #endif
2464
2465                 /*
2466                  * Block if there are too many dirty chains present, wait
2467                  * for the flush to clean some out.
2468                  */
2469                 if (count > limit) {
2470                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2471                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2472                                                waiting,
2473                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2474                                 speedup_syncer(pmp->mp);
2475                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2476                                        "chnmem", hz);
2477                         }
2478                         continue;       /* loop on success or fail */
2479                 }
2480
2481                 /*
2482                  * Try to start an early flush before we are forced to block.
2483                  */
2484                 if (count > limit * 7 / 10)
2485                         speedup_syncer(pmp->mp);
2486                 break;
2487         }
2488 }
2489
2490 void
2491 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2492 {
2493         if (pmp) {
2494                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2495         }
2496 }
2497
2498 void
2499 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2500 {
2501         uint32_t waiting;
2502
2503         if (pmp == NULL)
2504                 return;
2505
2506         for (;;) {
2507                 waiting = pmp->inmem_dirty_chains;
2508                 cpu_ccfence();
2509                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2510                                        waiting,
2511                                        (waiting - 1) &
2512                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2513                         break;
2514                 }
2515         }
2516
2517         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2518                 wakeup(&pmp->inmem_dirty_chains);
2519 }
2520
2521 /*
2522  * Debugging
2523  */
2524 void
2525 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2526 {
2527         hammer2_chain_t *scan;
2528         hammer2_chain_t *parent;
2529
2530         --*countp;
2531         if (*countp == 0) {
2532                 kprintf("%*.*s...\n", tab, tab, "");
2533                 return;
2534         }
2535         if (*countp < 0)
2536                 return;
2537         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2538                 tab, tab, "", pfx,
2539                 chain, chain->bref.type,
2540                 chain->bref.key, chain->bref.keybits,
2541                 chain->bref.mirror_tid);
2542
2543         kprintf("%*.*s      [%08x] (%s) refs=%d",
2544                 tab, tab, "",
2545                 chain->flags,
2546                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2547                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2548                 chain->refs);
2549
2550         parent = chain->parent;
2551         if (parent)
2552                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2553                         tab, tab, "",
2554                         parent, parent->flags, parent->refs);
2555         if (RB_EMPTY(&chain->core.rbtree)) {
2556                 kprintf("\n");
2557         } else {
2558                 kprintf(" {\n");
2559                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2560                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2561                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2562                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2563                                 chain->data->ipdata.filename);
2564                 else
2565                         kprintf("%*.*s}\n", tab, tab, "");
2566         }
2567 }