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_read = 4;           /* physical read-ahead */
  82 int hammer2_cluster_write = 0;          /* bdwrite() so later inval works */
  83 int hammer2_hardlink_enable = 1;
  84 int hammer2_flush_pipe = 100;
  85 int hammer2_synchronous_flush = 1;
  86 int hammer2_dio_count;
  87 long hammer2_chain_allocs;
  88 long hammer2_chain_frees;
  89 long hammer2_limit_dirty_chains;
  90 long hammer2_count_modified_chains;
  91 long hammer2_iod_file_read;
  92 long hammer2_iod_meta_read;
  93 long hammer2_iod_indr_read;
  94 long hammer2_iod_fmap_read;
  95 long hammer2_iod_volu_read;
  96 long hammer2_iod_file_write;
  97 long hammer2_iod_file_wembed;
  98 long hammer2_iod_file_wzero;
  99 long hammer2_iod_file_wdedup;
 100 long hammer2_iod_meta_write;
 101 long hammer2_iod_indr_write;
 102 long hammer2_iod_fmap_write;
 103 long hammer2_iod_volu_write;
 104 long hammer2_ioa_file_read;
 105 long hammer2_ioa_meta_read;
 106 long hammer2_ioa_indr_read;
 107 long hammer2_ioa_fmap_read;
 108 long hammer2_ioa_volu_read;
 109 long hammer2_ioa_fmap_write;
 110 long hammer2_ioa_file_write;
 111 long hammer2_ioa_meta_write;
 112 long hammer2_ioa_indr_write;
 113 long hammer2_ioa_volu_write;
 114
 115 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 116 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 117                 "Buffer used for compression.");
 118
 119 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 120 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 121                 "Buffer used for decompression.");
 122
 123 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 124
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 126            &hammer2_debug, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW,
 128            &hammer2_cluster_read, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 130            &hammer2_cluster_write, 0, "");
 131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
 132            &hammer2_hardlink_enable, 0, "");
 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 134            &hammer2_flush_pipe, 0, "");
 135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 136            &hammer2_synchronous_flush, 0, "");
 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 138            &hammer2_chain_allocs, 0, "");
 139 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 140            &hammer2_chain_frees, 0, "");
 141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 142            &hammer2_limit_dirty_chains, 0, "");
 143 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 144            &hammer2_count_modified_chains, 0, "");
 145 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 146            &hammer2_dio_count, 0, "");
 147
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 149            &hammer2_iod_file_read, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 151            &hammer2_iod_meta_read, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 153            &hammer2_iod_indr_read, 0, "");
 154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 155            &hammer2_iod_fmap_read, 0, "");
 156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 157            &hammer2_iod_volu_read, 0, "");
 158
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 160            &hammer2_iod_file_write, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 162            &hammer2_iod_file_wembed, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 164            &hammer2_iod_file_wzero, 0, "");
 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 166            &hammer2_iod_file_wdedup, 0, "");
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 168            &hammer2_iod_meta_write, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 170            &hammer2_iod_indr_write, 0, "");
 171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 172            &hammer2_iod_fmap_write, 0, "");
 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 174            &hammer2_iod_volu_write, 0, "");
 175
 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
 177            &hammer2_ioa_file_read, 0, "");
 178 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
 179            &hammer2_ioa_meta_read, 0, "");
 180 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
 181            &hammer2_ioa_indr_read, 0, "");
 182 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
 183            &hammer2_ioa_fmap_read, 0, "");
 184 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
 185            &hammer2_ioa_volu_read, 0, "");
 186
 187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
 188            &hammer2_ioa_file_write, 0, "");
 189 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
 190            &hammer2_ioa_meta_write, 0, "");
 191 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
 192            &hammer2_ioa_indr_write, 0, "");
 193 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
 194            &hammer2_ioa_fmap_write, 0, "");
 195 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
 196            &hammer2_ioa_volu_write, 0, "");
 197
 198 long hammer2_check_icrc32;
 199 long hammer2_check_xxhash64;
 200 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 201            &hammer2_check_icrc32, 0, "");
 202 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 203            &hammer2_check_xxhash64, 0, "");
 204
 205 static int hammer2_vfs_init(struct vfsconf *conf);
 206 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 207 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 208                                 struct ucred *cred);
 209 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 210                                 struct vnode *, struct ucred *);
 211 static int hammer2_recovery(hammer2_dev_t *hmp);
 212 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 213 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 214 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 215                                 struct ucred *cred);
 216 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 217                                 struct ucred *cred);
 218 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 219                                 struct fid *fhp, struct vnode **vpp);
 220 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 221 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 222                                 int *exflagsp, struct ucred **credanonp);
 223
 224 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 225 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 226
 227 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 228
 229 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 230 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 231                                 hammer2_dev_t *hmp);
 232
 233 /*
 234  * HAMMER2 vfs operations.
 235  */
 236 static struct vfsops hammer2_vfsops = {
 237         .vfs_init       = hammer2_vfs_init,
 238         .vfs_uninit     = hammer2_vfs_uninit,
 239         .vfs_sync       = hammer2_vfs_sync,
 240         .vfs_mount      = hammer2_vfs_mount,
 241         .vfs_unmount    = hammer2_vfs_unmount,
 242         .vfs_root       = hammer2_vfs_root,
 243         .vfs_statfs     = hammer2_vfs_statfs,
 244         .vfs_statvfs    = hammer2_vfs_statvfs,
 245         .vfs_vget       = hammer2_vfs_vget,
 246         .vfs_vptofh     = hammer2_vfs_vptofh,
 247         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 248         .vfs_checkexp   = hammer2_vfs_checkexp
 249 };
 250
 251 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 252
 253 VFS_SET(hammer2_vfsops, hammer2, 0);
 254 MODULE_VERSION(hammer2, 1);
 255
 256 static
 257 int
 258 hammer2_vfs_init(struct vfsconf *conf)
 259 {
 260         static struct objcache_malloc_args margs_read;
 261         static struct objcache_malloc_args margs_write;
 262         static struct objcache_malloc_args margs_vop;
 263
 264         int error;
 265
 266         error = 0;
 267
 268         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 269                 error = EINVAL;
 270         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 271                 error = EINVAL;
 272         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 273                 error = EINVAL;
 274
 275         if (error)
 276                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 277
 278         margs_read.objsize = 65536;
 279         margs_read.mtype = M_HAMMER2_DEBUFFER;
 280
 281         margs_write.objsize = 32768;
 282         margs_write.mtype = M_HAMMER2_CBUFFER;
 283
 284         margs_vop.objsize = sizeof(hammer2_xop_t);
 285         margs_vop.mtype = M_HAMMER2;
 286
 287         /*
 288          * Note thaht for the XOPS cache we want backing store allocations
 289          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 290          * confusion), so use the backing store function that does it.  This
 291          * means that initial XOPS objects are zerod but REUSED objects are
 292          * not.  So we are responsible for cleaning the object up sufficiently
 293          * for our needs before objcache_put()ing it back (typically just the
 294          * FIFO indices).
 295          */
 296         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 297                                 0, 1, NULL, NULL, NULL,
 298                                 objcache_malloc_alloc,
 299                                 objcache_malloc_free,
 300                                 &margs_read);
 301         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 302                                 0, 1, NULL, NULL, NULL,
 303                                 objcache_malloc_alloc,
 304                                 objcache_malloc_free,
 305                                 &margs_write);
 306         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 307                                 0, 1, NULL, NULL, NULL,
 308                                 objcache_malloc_alloc_zero,
 309                                 objcache_malloc_free,
 310                                 &margs_vop);
 311
 312
 313         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 314         TAILQ_INIT(&hammer2_mntlist);
 315         TAILQ_INIT(&hammer2_pfslist);
 316
 317         hammer2_limit_dirty_chains = desiredvnodes / 10;
 318
 319         return (error);
 320 }
 321
 322 static
 323 int
 324 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 325 {
 326         objcache_destroy(cache_buffer_read);
 327         objcache_destroy(cache_buffer_write);
 328         objcache_destroy(cache_xops);
 329         return 0;
 330 }
 331
 332 /*
 333  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
 334  * mounts and the spmp structure for media (hmp) structures.
 335  *
 336  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 337  * transactions.  Note that synchronization does not use this field.
 338  * (typically frontend operations and synchronization cannot run on the
 339  * same PFS node at the same time).
 340  *
 341  * XXX check locking
 342  */
 343 hammer2_pfs_t *
 344 hammer2_pfsalloc(hammer2_chain_t *chain, const hammer2_inode_data_t *ripdata,
 345                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 346 {
 347         hammer2_inode_t *iroot;
 348         hammer2_pfs_t *pmp;
 349         int count;
 350         int i;
 351         int j;
 352
 353         /*
 354          * Locate or create the PFS based on the cluster id.  If ripdata
 355          * is NULL this is a spmp which is unique and is always allocated.
 356          *
 357          * If the device is mounted in local mode all PFSs are considered
 358          * independent and not part of any cluster (for debugging only).
 359          */
 360         if (ripdata) {
 361                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 362                         if (force_local != pmp->force_local)
 363                                 continue;
 364                         if (force_local == NULL &&
 365                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 366                                  sizeof(pmp->pfs_clid)) == 0) {
 367                                         break;
 368                         } else if (force_local && pmp->pfs_names[0] &&
 369                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 370                                         break;
 371                         }
 372                 }
 373         } else {
 374                 pmp = NULL;
 375         }
 376
 377         if (pmp == NULL) {
 378                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 379                 pmp->force_local = force_local;
 380                 hammer2_trans_manage_init(pmp);
 381                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 382                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 383                 lockinit(&pmp->lock, "pfslk", 0, 0);
 384                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 385                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 386                 spin_init(&pmp->xop_spin, "h2xop");
 387                 spin_init(&pmp->lru_spin, "h2lru");
 388                 RB_INIT(&pmp->inum_tree);
 389                 TAILQ_INIT(&pmp->sideq);
 390                 TAILQ_INIT(&pmp->lru_list);
 391                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 392
 393                 /*
 394                  * Distribute backend operations to threads
 395                  */
 396                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 397                         for (j = 0; j < HAMMER2_XOPGROUPS +
 398                                         HAMMER2_SPECTHREADS; ++j) {
 399                                 TAILQ_INIT(&pmp->xopq[i][j]);
 400                         }
 401                 }
 402                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 403                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 404
 405                 /*
 406                  * Save the last media transaction id for the flusher.  Set
 407                  * initial
 408                  */
 409                 if (ripdata)
 410                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 411                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 412
 413                 /*
 414                  * The synchronization thread may start too early, make
 415                  * sure it stays frozen until we are ready to let it go.
 416                  * XXX
 417                  */
 418                 /*
 419                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 420                                          HAMMER2_THREAD_REMASTER;
 421                 */
 422         }
 423
 424         /*
 425          * Create the PFS's root inode.
 426          */
 427         if ((iroot = pmp->iroot) == NULL) {
 428                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 429                 pmp->iroot = iroot;
 430                 hammer2_inode_ref(iroot);
 431                 hammer2_inode_unlock(iroot);
 432         }
 433
 434         /*
 435          * Stop here if no chain is passed in.
 436          */
 437         if (chain == NULL)
 438                 goto done;
 439
 440         /*
 441          * When a chain is passed in we must add it to the PFS's root
 442          * inode, update pmp->pfs_types[], and update the syncronization
 443          * threads.
 444          *
 445          * When forcing local mode, mark the PFS as a MASTER regardless.
 446          *
 447          * At the moment empty spots can develop due to removals or failures.
 448          * Ultimately we want to re-fill these spots but doing so might
 449          * confused running code. XXX
 450          */
 451         hammer2_inode_ref(iroot);
 452         hammer2_mtx_ex(&iroot->lock);
 453         j = iroot->cluster.nchains;
 454
 455         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 456
 457         if (j == HAMMER2_MAXCLUSTER) {
 458                 kprintf("hammer2_mount: cluster full!\n");
 459                 /* XXX fatal error? */
 460         } else {
 461                 KKASSERT(chain->pmp == NULL);
 462                 chain->pmp = pmp;
 463                 hammer2_chain_ref(chain);
 464                 iroot->cluster.array[j].chain = chain;
 465                 if (force_local)
 466                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 467                 else
 468                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 469                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 470                 pmp->pfs_hmps[j] = chain->hmp;
 471
 472                 /*
 473                  * If the PFS is already mounted we must account
 474                  * for the mount_count here.
 475                  */
 476                 if (pmp->mp)
 477                         ++chain->hmp->mount_count;
 478
 479                 /*
 480                  * May have to fixup dirty chain tracking.  Previous
 481                  * pmp was NULL so nothing to undo.
 482                  */
 483                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 484                         hammer2_pfs_memory_inc(pmp);
 485                 ++j;
 486         }
 487         iroot->cluster.nchains = j;
 488
 489         /*
 490          * Update nmasters from any PFS inode which is part of the cluster.
 491          * It is possible that this will result in a value which is too
 492          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 493          * override this value later on.
 494          *
 495          * (This informs us of masters that might not currently be
 496          *  discoverable by this mount).
 497          */
 498         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 499                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 500         }
 501
 502         /*
 503          * Count visible masters.  Masters are usually added with
 504          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 505          * are more (XXX and must update the master inodes).
 506          */
 507         count = 0;
 508         for (i = 0; i < iroot->cluster.nchains; ++i) {
 509                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 510                         ++count;
 511         }
 512         if (pmp->pfs_nmasters < count)
 513                 pmp->pfs_nmasters = count;
 514
 515         /*
 516          * Create missing synchronization and support threads.
 517          *
 518          * Single-node masters (including snapshots) have nothing to
 519          * synchronize and do not require this thread.
 520          *
 521          * Multi-node masters or any number of soft masters, slaves, copy,
 522          * or other PFS types need the thread.
 523          *
 524          * Each thread is responsible for its particular cluster index.
 525          * We use independent threads so stalls or mismatches related to
 526          * any given target do not affect other targets.
 527          */
 528         for (i = 0; i < iroot->cluster.nchains; ++i) {
 529                 /*
 530                  * Single-node masters (including snapshots) have nothing
 531                  * to synchronize and will make direct xops support calls,
 532                  * thus they do not require this thread.
 533                  *
 534                  * Note that there can be thousands of snapshots.  We do not
 535                  * want to create thousands of threads.
 536                  */
 537                 if (pmp->pfs_nmasters <= 1 &&
 538                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 539                         continue;
 540                 }
 541
 542                 /*
 543                  * Sync support thread
 544                  */
 545                 if (pmp->sync_thrs[i].td == NULL) {
 546                         hammer2_thr_create(&pmp->sync_thrs[i], pmp,
 547                                            "h2nod", i, -1,
 548                                            hammer2_primary_sync_thread);
 549                 }
 550         }
 551
 552         /*
 553          * Create missing Xop threads
 554          */
 555         if (pmp->mp)
 556                 hammer2_xop_helper_create(pmp);
 557
 558         hammer2_mtx_unlock(&iroot->lock);
 559         hammer2_inode_drop(iroot);
 560 done:
 561         return pmp;
 562 }
 563
 564 /*
 565  * Destroy a PFS, typically only occurs after the last mount on a device
 566  * has gone away.
 567  */
 568 static void
 569 hammer2_pfsfree(hammer2_pfs_t *pmp)
 570 {
 571         hammer2_inode_t *iroot;
 572         hammer2_chain_t *chain;
 573         int i;
 574         int j;
 575
 576         /*
 577          * Cleanup our reference on iroot.  iroot is (should) not be needed
 578          * by the flush code.
 579          */
 580         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 581
 582         iroot = pmp->iroot;
 583         if (iroot) {
 584                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 585                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 586                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 587                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 588                 }
 589 #if REPORT_REFS_ERRORS
 590                 if (pmp->iroot->refs != 1)
 591                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 592                                 pmp->iroot, pmp->iroot->refs);
 593 #else
 594                 KKASSERT(pmp->iroot->refs == 1);
 595 #endif
 596                 /* ref for pmp->iroot */
 597                 hammer2_inode_drop(pmp->iroot);
 598                 pmp->iroot = NULL;
 599         }
 600
 601         /*
 602          * Cleanup chains remaining on LRU list.
 603          */
 604         kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
 605         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 606                 hammer2_chain_ref(chain);
 607                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 608                 hammer2_chain_drop(chain);
 609         }
 610
 611         /*
 612          * Free remaining pmp resources
 613          */
 614         kmalloc_destroy(&pmp->mmsg);
 615         kmalloc_destroy(&pmp->minode);
 616
 617         kfree(pmp, M_HAMMER2);
 618 }
 619
 620 /*
 621  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 622  * empty is terminated and freed.
 623  *
 624  * XXX inefficient.
 625  */
 626 static void
 627 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 628 {
 629         hammer2_pfs_t *pmp;
 630         hammer2_inode_t *iroot;
 631         hammer2_chain_t *rchain;
 632         int didfreeze;
 633         int i;
 634         int j;
 635
 636 again:
 637         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 638                 if ((iroot = pmp->iroot) == NULL)
 639                         continue;
 640                 if (hmp->spmp == pmp) {
 641                         kprintf("unmount hmp %p remove spmp %p\n",
 642                                 hmp, pmp);
 643                         hmp->spmp = NULL;
 644                 }
 645
 646                 /*
 647                  * Determine if this PFS is affected.  If it is we must
 648                  * freeze all management threads and lock its iroot.
 649                  *
 650                  * Freezing a management thread forces it idle, operations
 651                  * in-progress will be aborted and it will have to start
 652                  * over again when unfrozen, or exit if told to exit.
 653                  */
 654                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 655                         if (pmp->pfs_hmps[i] == hmp)
 656                                 break;
 657                 }
 658                 if (i != HAMMER2_MAXCLUSTER) {
 659                         /*
 660                          * Make sure all synchronization threads are locked
 661                          * down.
 662                          */
 663                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 664                                 if (pmp->pfs_hmps[i] == NULL)
 665                                         continue;
 666                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 667                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 668                                         hammer2_thr_freeze_async(
 669                                                 &pmp->xop_groups[j].thrs[i]);
 670                                 }
 671                         }
 672                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 673                                 if (pmp->pfs_hmps[i] == NULL)
 674                                         continue;
 675                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 676                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 677                                         hammer2_thr_freeze(
 678                                                 &pmp->xop_groups[j].thrs[i]);
 679                                 }
 680                         }
 681
 682                         /*
 683                          * Lock the inode and clean out matching chains.
 684                          * Note that we cannot use hammer2_inode_lock_*()
 685                          * here because that would attempt to validate the
 686                          * cluster that we are in the middle of ripping
 687                          * apart.
 688                          *
 689                          * WARNING! We are working directly on the inodes
 690                          *          embedded cluster.
 691                          */
 692                         hammer2_mtx_ex(&iroot->lock);
 693
 694                         /*
 695                          * Remove the chain from matching elements of the PFS.
 696                          */
 697                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 698                                 if (pmp->pfs_hmps[i] != hmp)
 699                                         continue;
 700                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 701                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 702                                         hammer2_thr_delete(
 703                                                 &pmp->xop_groups[j].thrs[i]);
 704                                 }
 705                                 rchain = iroot->cluster.array[i].chain;
 706                                 iroot->cluster.array[i].chain = NULL;
 707                                 pmp->pfs_types[i] = 0;
 708                                 if (pmp->pfs_names[i]) {
 709                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 710                                         pmp->pfs_names[i] = NULL;
 711                                 }
 712                                 if (rchain) {
 713                                         hammer2_chain_drop(rchain);
 714                                         /* focus hint */
 715                                         if (iroot->cluster.focus == rchain)
 716                                                 iroot->cluster.focus = NULL;
 717                                 }
 718                                 pmp->pfs_hmps[i] = NULL;
 719                         }
 720                         hammer2_mtx_unlock(&iroot->lock);
 721                         didfreeze = 1;  /* remaster, unfreeze down below */
 722                 } else {
 723                         didfreeze = 0;
 724                 }
 725
 726                 /*
 727                  * Cleanup trailing chains.  Gaps may remain.
 728                  */
 729                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 730                         if (pmp->pfs_hmps[i])
 731                                 break;
 732                 }
 733                 iroot->cluster.nchains = i + 1;
 734
 735                 /*
 736                  * If the PMP has no elements remaining we can destroy it.
 737                  * (this will transition management threads from frozen->exit).
 738                  */
 739                 if (iroot->cluster.nchains == 0) {
 740                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 741                                 hmp, pmp);
 742                         hammer2_pfsfree(pmp);
 743                         goto again;
 744                 }
 745
 746                 /*
 747                  * If elements still remain we need to set the REMASTER
 748                  * flag and unfreeze it.
 749                  */
 750                 if (didfreeze) {
 751                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 752                                 if (pmp->pfs_hmps[i] == NULL)
 753                                         continue;
 754                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 755                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 756                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 757                                         hammer2_thr_remaster(
 758                                                 &pmp->xop_groups[j].thrs[i]);
 759                                         hammer2_thr_unfreeze(
 760                                                 &pmp->xop_groups[j].thrs[i]);
 761                                 }
 762                         }
 763                 }
 764         }
 765 }
 766
 767 /*
 768  * Mount or remount HAMMER2 fileystem from physical media
 769  *
 770  *      mountroot
 771  *              mp              mount point structure
 772  *              path            NULL
 773  *              data            <unused>
 774  *              cred            <unused>
 775  *
 776  *      mount
 777  *              mp              mount point structure
 778  *              path            path to mount point
 779  *              data            pointer to argument structure in user space
 780  *                      volume  volume path (device@LABEL form)
 781  *                      hflags  user mount flags
 782  *              cred            user credentials
 783  *
 784  * RETURNS:     0       Success
 785  *              !0      error number
 786  */
 787 static
 788 int
 789 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 790                   struct ucred *cred)
 791 {
 792         struct hammer2_mount_info info;
 793         hammer2_pfs_t *pmp;
 794         hammer2_pfs_t *spmp;
 795         hammer2_dev_t *hmp;
 796         hammer2_dev_t *force_local;
 797         hammer2_key_t key_next;
 798         hammer2_key_t key_dummy;
 799         hammer2_key_t lhc;
 800         struct vnode *devvp;
 801         struct nlookupdata nd;
 802         hammer2_chain_t *parent;
 803         hammer2_chain_t *chain;
 804         hammer2_cluster_t *cluster;
 805         const hammer2_inode_data_t *ripdata;
 806         hammer2_blockref_t bref;
 807         struct file *fp;
 808         char devstr[MNAMELEN];
 809         size_t size;
 810         size_t done;
 811         char *dev;
 812         char *label;
 813         int ronly = 1;
 814         int error;
 815         int cache_index;
 816         int i;
 817
 818         hmp = NULL;
 819         pmp = NULL;
 820         dev = NULL;
 821         label = NULL;
 822         devvp = NULL;
 823         cache_index = -1;
 824
 825         kprintf("hammer2_mount\n");
 826
 827         if (path == NULL) {
 828                 /*
 829                  * Root mount
 830                  */
 831                 bzero(&info, sizeof(info));
 832                 info.cluster_fd = -1;
 833                 ksnprintf(devstr, sizeof(devstr), "%s",
 834                           mp->mnt_stat.f_mntfromname);
 835                 kprintf("hammer2_mount: root '%s'\n", devstr);
 836         } else {
 837                 /*
 838                  * Non-root mount or updating a mount
 839                  */
 840                 error = copyin(data, &info, sizeof(info));
 841                 if (error)
 842                         return (error);
 843
 844                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 845                 if (error)
 846                         return (error);
 847         }
 848
 849         /* Extract device and label */
 850         dev = devstr;
 851         label = strchr(devstr, '@');
 852         if (label == NULL ||
 853             ((label + 1) - dev) > done) {
 854                 return (EINVAL);
 855         }
 856         *label = '\0';
 857         label++;
 858         if (*label == '\0')
 859                 return (EINVAL);
 860
 861         if (mp->mnt_flag & MNT_UPDATE) {
 862                 /*
 863                  * Update mount.  Note that pmp->iroot->cluster is
 864                  * an inode-embedded cluster and thus cannot be
 865                  * directly locked.
 866                  *
 867                  * XXX HAMMER2 needs to implement NFS export via
 868                  *     mountctl.
 869                  */
 870                 pmp = MPTOPMP(mp);
 871                 pmp->hflags = info.hflags;
 872                 cluster = &pmp->iroot->cluster;
 873                 for (i = 0; i < cluster->nchains; ++i) {
 874                         if (cluster->array[i].chain == NULL)
 875                                 continue;
 876                         hmp = cluster->array[i].chain->hmp;
 877                         devvp = hmp->devvp;
 878                         error = hammer2_remount(hmp, mp, path,
 879                                                 devvp, cred);
 880                         if (error)
 881                                 break;
 882                 }
 883
 884                 return error;
 885         }
 886
 887         /*
 888          * HMP device mount
 889          *
 890          * Lookup name and verify it refers to a block device.
 891          */
 892         if (path) {
 893                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 894                 if (error == 0)
 895                         error = nlookup(&nd);
 896                 if (error == 0)
 897                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 898                 nlookup_done(&nd);
 899         } else {
 900                 /* root mount */
 901                 cdev_t cdev = kgetdiskbyname(dev);
 902                 error = bdevvp(cdev, &devvp);
 903                 if (error)
 904                         kprintf("hammer2: cannot find '%s'\n", dev);
 905         }
 906
 907         if (error == 0) {
 908                 if (vn_isdisk(devvp, &error))
 909                         error = vfs_mountedon(devvp);
 910         }
 911
 912         /*
 913          * Determine if the device has already been mounted.  After this
 914          * check hmp will be non-NULL if we are doing the second or more
 915          * hammer2 mounts from the same device.
 916          */
 917         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 918         TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 919                 if (hmp->devvp == devvp)
 920                         break;
 921         }
 922
 923         /*
 924          * Open the device if this isn't a secondary mount and construct
 925          * the H2 device mount (hmp).
 926          */
 927         if (hmp == NULL) {
 928                 hammer2_chain_t *schain;
 929                 hammer2_xid_t xid;
 930
 931                 if (error == 0 && vcount(devvp) > 0)
 932                         error = EBUSY;
 933
 934                 /*
 935                  * Now open the device
 936                  */
 937                 if (error == 0) {
 938                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
 939                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
 940                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
 941                         if (error == 0) {
 942                                 error = VOP_OPEN(devvp,
 943                                                  ronly ? FREAD : FREAD | FWRITE,
 944                                                  FSCRED, NULL);
 945                         }
 946                         vn_unlock(devvp);
 947                 }
 948                 if (error && devvp) {
 949                         vrele(devvp);
 950                         devvp = NULL;
 951                 }
 952                 if (error) {
 953                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 954                         return error;
 955                 }
 956                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
 957                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
 958                 hmp->ronly = ronly;
 959                 hmp->devvp = devvp;
 960                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
 961                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
 962                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
 963                 RB_INIT(&hmp->iotree);
 964                 spin_init(&hmp->io_spin, "hm2mount_io");
 965                 spin_init(&hmp->list_spin, "hm2mount_list");
 966                 TAILQ_INIT(&hmp->flushq);
 967
 968                 lockinit(&hmp->vollk, "h2vol", 0, 0);
 969                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
 970
 971                 /*
 972                  * vchain setup. vchain.data is embedded.
 973                  * vchain.refs is initialized and will never drop to 0.
 974                  *
 975                  * NOTE! voldata is not yet loaded.
 976                  */
 977                 hmp->vchain.hmp = hmp;
 978                 hmp->vchain.refs = 1;
 979                 hmp->vchain.data = (void *)&hmp->voldata;
 980                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
 981                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 982                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 983
 984                 hammer2_chain_core_init(&hmp->vchain);
 985                 /* hmp->vchain.u.xxx is left NULL */
 986
 987                 /*
 988                  * fchain setup.  fchain.data is embedded.
 989                  * fchain.refs is initialized and will never drop to 0.
 990                  *
 991                  * The data is not used but needs to be initialized to
 992                  * pass assertion muster.  We use this chain primarily
 993                  * as a placeholder for the freemap's top-level RBTREE
 994                  * so it does not interfere with the volume's topology
 995                  * RBTREE.
 996                  */
 997                 hmp->fchain.hmp = hmp;
 998                 hmp->fchain.refs = 1;
 999                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1000                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1001                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1002                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1003                 hmp->fchain.bref.methods =
1004                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1005                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1006
1007                 hammer2_chain_core_init(&hmp->fchain);
1008                 /* hmp->fchain.u.xxx is left NULL */
1009
1010                 /*
1011                  * Install the volume header and initialize fields from
1012                  * voldata.
1013                  */
1014                 error = hammer2_install_volume_header(hmp);
1015                 if (error) {
1016                         hammer2_unmount_helper(mp, NULL, hmp);
1017                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1018                         hammer2_vfs_unmount(mp, MNT_FORCE);
1019                         return error;
1020                 }
1021
1022                 /*
1023                  * Really important to get these right or flush will get
1024                  * confused.
1025                  */
1026                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1027                 kprintf("alloc spmp %p tid %016jx\n",
1028                         hmp->spmp, hmp->voldata.mirror_tid);
1029                 spmp = hmp->spmp;
1030
1031                 /*
1032                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1033                  * is inherited from the volume header.
1034                  */
1035                 xid = 0;
1036                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1037                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1038                 hmp->vchain.pmp = spmp;
1039                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1040                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1041                 hmp->fchain.pmp = spmp;
1042
1043                 /*
1044                  * First locate the super-root inode, which is key 0
1045                  * relative to the volume header's blockset.
1046                  *
1047                  * Then locate the root inode by scanning the directory keyspace
1048                  * represented by the label.
1049                  */
1050                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1051                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1052                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1053                                       &cache_index, 0);
1054                 hammer2_chain_lookup_done(parent);
1055                 if (schain == NULL) {
1056                         kprintf("hammer2_mount: invalid super-root\n");
1057                         hammer2_unmount_helper(mp, NULL, hmp);
1058                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1059                         hammer2_vfs_unmount(mp, MNT_FORCE);
1060                         return EINVAL;
1061                 }
1062                 if (schain->error) {
1063                         kprintf("hammer2_mount: error %s reading super-root\n",
1064                                 hammer2_error_str(schain->error));
1065                         hammer2_chain_unlock(schain);
1066                         hammer2_chain_drop(schain);
1067                         schain = NULL;
1068                         hammer2_unmount_helper(mp, NULL, hmp);
1069                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1070                         hammer2_vfs_unmount(mp, MNT_FORCE);
1071                         return EINVAL;
1072                 }
1073
1074                 /*
1075                  * The super-root always uses an inode_tid of 1 when
1076                  * creating PFSs.
1077                  */
1078                 spmp->inode_tid = 1;
1079                 spmp->modify_tid = schain->bref.modify_tid + 1;
1080
1081                 /*
1082                  * Sanity-check schain's pmp and finish initialization.
1083                  * Any chain belonging to the super-root topology should
1084                  * have a NULL pmp (not even set to spmp).
1085                  */
1086                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1087                 KKASSERT(schain->pmp == NULL);
1088                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1089
1090                 /*
1091                  * Replace the dummy spmp->iroot with a real one.  It's
1092                  * easier to just do a wholesale replacement than to try
1093                  * to update the chain and fixup the iroot fields.
1094                  *
1095                  * The returned inode is locked with the supplied cluster.
1096                  */
1097                 cluster = hammer2_cluster_from_chain(schain);
1098                 hammer2_inode_drop(spmp->iroot);
1099                 spmp->iroot = NULL;
1100                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1101                 spmp->spmp_hmp = hmp;
1102                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1103                 spmp->pfs_hmps[0] = hmp;
1104                 hammer2_inode_ref(spmp->iroot);
1105                 hammer2_inode_unlock(spmp->iroot);
1106                 hammer2_cluster_unlock(cluster);
1107                 hammer2_cluster_drop(cluster);
1108                 schain = NULL;
1109                 /* leave spmp->iroot with one ref */
1110
1111                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1112                         error = hammer2_recovery(hmp);
1113                         /* XXX do something with error */
1114                 }
1115                 hammer2_update_pmps(hmp);
1116                 hammer2_iocom_init(hmp);
1117
1118                 /*
1119                  * Ref the cluster management messaging descriptor.  The mount
1120                  * program deals with the other end of the communications pipe.
1121                  *
1122                  * Root mounts typically do not supply one.
1123                  */
1124                 if (info.cluster_fd >= 0) {
1125                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1126                         if (fp) {
1127                                 hammer2_cluster_reconnect(hmp, fp);
1128                         } else {
1129                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1130                         }
1131                 }
1132         } else {
1133                 spmp = hmp->spmp;
1134                 if (info.hflags & HMNT2_DEVFLAGS) {
1135                         kprintf("hammer2: Warning: mount flags pertaining "
1136                                 "to the whole device may only be specified "
1137                                 "on the first mount of the device: %08x\n",
1138                                 info.hflags & HMNT2_DEVFLAGS);
1139                 }
1140         }
1141
1142         /*
1143          * Force local mount (disassociate all PFSs from their clusters).
1144          * Used primarily for debugging.
1145          */
1146         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1147
1148         /*
1149          * Lookup the mount point under the media-localized super-root.
1150          * Scanning hammer2_pfslist doesn't help us because it represents
1151          * PFS cluster ids which can aggregate several named PFSs together.
1152          *
1153          * cluster->pmp will incorrectly point to spmp and must be fixed
1154          * up later on.
1155          */
1156         hammer2_inode_lock(spmp->iroot, 0);
1157         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1158         lhc = hammer2_dirhash(label, strlen(label));
1159         chain = hammer2_chain_lookup(&parent, &key_next,
1160                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1161                                      &cache_index, 0);
1162         while (chain) {
1163                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1164                     strcmp(label, chain->data->ipdata.filename) == 0) {
1165                         break;
1166                 }
1167                 chain = hammer2_chain_next(&parent, chain, &key_next,
1168                                             key_next,
1169                                             lhc + HAMMER2_DIRHASH_LOMASK,
1170                                             &cache_index, 0);
1171         }
1172         if (parent) {
1173                 hammer2_chain_unlock(parent);
1174                 hammer2_chain_drop(parent);
1175         }
1176         hammer2_inode_unlock(spmp->iroot);
1177
1178         /*
1179          * PFS could not be found?
1180          */
1181         if (chain == NULL) {
1182                 kprintf("hammer2_mount: PFS label not found\n");
1183                 hammer2_unmount_helper(mp, NULL, hmp);
1184                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1185                 hammer2_vfs_unmount(mp, MNT_FORCE);
1186
1187                 return EINVAL;
1188         }
1189
1190         /*
1191          * Acquire the pmp structure (it should have already been allocated
1192          * via hammer2_update_pmps() so do not pass cluster in to add to
1193          * available chains).
1194          *
1195          * Check if the cluster has already been mounted.  A cluster can
1196          * only be mounted once, use null mounts to mount additional copies.
1197          */
1198         ripdata = &chain->data->ipdata;
1199         bref = chain->bref;
1200         pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid, force_local);
1201         hammer2_chain_unlock(chain);
1202         hammer2_chain_drop(chain);
1203
1204         if (pmp->mp) {
1205                 kprintf("hammer2_mount: PFS already mounted!\n");
1206                 hammer2_unmount_helper(mp, NULL, hmp);
1207                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1208                 hammer2_vfs_unmount(mp, MNT_FORCE);
1209
1210                 return EBUSY;
1211         }
1212
1213         /*
1214          * Finish the mount
1215          */
1216         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1217
1218         pmp->hflags = info.hflags;
1219         mp->mnt_flag = MNT_LOCAL;
1220         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1221         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1222
1223         /*
1224          * required mount structure initializations
1225          */
1226         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1227         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1228
1229         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1230         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1231
1232         /*
1233          * Optional fields
1234          */
1235         mp->mnt_iosize_max = MAXPHYS;
1236
1237         /*
1238          * Connect up mount pointers.
1239          */
1240         hammer2_mount_helper(mp, pmp);
1241
1242         lockmgr(&hammer2_mntlk, LK_RELEASE);
1243
1244         /*
1245          * Finish setup
1246          */
1247         vfs_getnewfsid(mp);
1248         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1249         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1250         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1251
1252         if (path) {
1253                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1254                           MNAMELEN - 1, &size);
1255                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1256         } /* else root mount, already in there */
1257
1258         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1259         if (path) {
1260                 copyinstr(path, mp->mnt_stat.f_mntonname,
1261                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1262                           &size);
1263         } else {
1264                 /* root mount */
1265                 mp->mnt_stat.f_mntonname[0] = '/';
1266         }
1267
1268         /*
1269          * Initial statfs to prime mnt_stat.
1270          */
1271         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1272
1273         return 0;
1274 }
1275
1276 /*
1277  * Scan PFSs under the super-root and create hammer2_pfs structures.
1278  */
1279 static
1280 void
1281 hammer2_update_pmps(hammer2_dev_t *hmp)
1282 {
1283         const hammer2_inode_data_t *ripdata;
1284         hammer2_chain_t *parent;
1285         hammer2_chain_t *chain;
1286         hammer2_blockref_t bref;
1287         hammer2_dev_t *force_local;
1288         hammer2_pfs_t *spmp;
1289         hammer2_pfs_t *pmp;
1290         hammer2_key_t key_next;
1291         int cache_index = -1;
1292
1293         /*
1294          * Force local mount (disassociate all PFSs from their clusters).
1295          * Used primarily for debugging.
1296          */
1297         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1298
1299         /*
1300          * Lookup mount point under the media-localized super-root.
1301          *
1302          * cluster->pmp will incorrectly point to spmp and must be fixed
1303          * up later on.
1304          */
1305         spmp = hmp->spmp;
1306         hammer2_inode_lock(spmp->iroot, 0);
1307         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1308         chain = hammer2_chain_lookup(&parent, &key_next,
1309                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1310                                          &cache_index, 0);
1311         while (chain) {
1312                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1313                         continue;
1314                 ripdata = &chain->data->ipdata;
1315                 bref = chain->bref;
1316                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1317
1318                 pmp = hammer2_pfsalloc(chain, ripdata,
1319                                        bref.modify_tid, force_local);
1320                 chain = hammer2_chain_next(&parent, chain, &key_next,
1321                                            key_next, HAMMER2_KEY_MAX,
1322                                            &cache_index, 0);
1323         }
1324         if (parent) {
1325                 hammer2_chain_unlock(parent);
1326                 hammer2_chain_drop(parent);
1327         }
1328         hammer2_inode_unlock(spmp->iroot);
1329 }
1330
1331 static
1332 int
1333 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1334                 struct vnode *devvp, struct ucred *cred)
1335 {
1336         int error;
1337
1338         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1339                 error = hammer2_recovery(hmp);
1340         } else {
1341                 error = 0;
1342         }
1343         return error;
1344 }
1345
1346 static
1347 int
1348 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1349 {
1350         hammer2_pfs_t *pmp;
1351         int flags;
1352         int error = 0;
1353
1354         pmp = MPTOPMP(mp);
1355
1356         if (pmp == NULL)
1357                 return(0);
1358
1359         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1360
1361         /*
1362          * If mount initialization proceeded far enough we must flush
1363          * its vnodes and sync the underlying mount points.  Three syncs
1364          * are required to fully flush the filesystem (freemap updates lag
1365          * by one flush, and one extra for safety).
1366          */
1367         if (mntflags & MNT_FORCE)
1368                 flags = FORCECLOSE;
1369         else
1370                 flags = 0;
1371         if (pmp->iroot) {
1372                 error = vflush(mp, 0, flags);
1373                 if (error)
1374                         goto failed;
1375                 hammer2_vfs_sync(mp, MNT_WAIT);
1376                 hammer2_vfs_sync(mp, MNT_WAIT);
1377                 hammer2_vfs_sync(mp, MNT_WAIT);
1378         }
1379
1380         /*
1381          * Cleanup the frontend support XOPS threads
1382          */
1383         hammer2_xop_helper_cleanup(pmp);
1384
1385         /*
1386          * Cleanup our reference on ihidden.
1387          */
1388         if (pmp->ihidden) {
1389                 hammer2_inode_drop(pmp->ihidden);
1390                 pmp->ihidden = NULL;
1391         }
1392         if (pmp->mp)
1393                 hammer2_unmount_helper(mp, pmp, NULL);
1394
1395         error = 0;
1396 failed:
1397         lockmgr(&hammer2_mntlk, LK_RELEASE);
1398
1399         return (error);
1400 }
1401
1402 /*
1403  * Mount helper, hook the system mount into our PFS.
1404  * The mount lock is held.
1405  *
1406  * We must bump the mount_count on related devices for any
1407  * mounted PFSs.
1408  */
1409 static
1410 void
1411 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1412 {
1413         hammer2_cluster_t *cluster;
1414         hammer2_chain_t *rchain;
1415         int i;
1416
1417         mp->mnt_data = (qaddr_t)pmp;
1418         pmp->mp = mp;
1419
1420         /*
1421          * After pmp->mp is set we have to adjust hmp->mount_count.
1422          */
1423         cluster = &pmp->iroot->cluster;
1424         for (i = 0; i < cluster->nchains; ++i) {
1425                 rchain = cluster->array[i].chain;
1426                 if (rchain == NULL)
1427                         continue;
1428                 ++rchain->hmp->mount_count;
1429                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1430                         rchain->hmp, rchain->hmp->mount_count);
1431         }
1432
1433         /*
1434          * Create missing Xop threads
1435          */
1436         hammer2_xop_helper_create(pmp);
1437 }
1438
1439 /*
1440  * Mount helper, unhook the system mount from our PFS.
1441  * The mount lock is held.
1442  *
1443  * If hmp is supplied a mount responsible for being the first to open
1444  * the block device failed and the block device and all PFSs using the
1445  * block device must be cleaned up.
1446  *
1447  * If pmp is supplied multiple devices might be backing the PFS and each
1448  * must be disconnected.  This might not be the last PFS using some of the
1449  * underlying devices.  Also, we have to adjust our hmp->mount_count
1450  * accounting for the devices backing the pmp which is now undergoing an
1451  * unmount.
1452  */
1453 static
1454 void
1455 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1456 {
1457         hammer2_cluster_t *cluster;
1458         hammer2_chain_t *rchain;
1459         struct vnode *devvp;
1460         int dumpcnt;
1461         int ronly = 0;
1462         int i;
1463
1464         /*
1465          * If no device supplied this is a high-level unmount and we have to
1466          * to disconnect the mount, adjust mount_count, and locate devices
1467          * that might now have no mounts.
1468          */
1469         if (pmp) {
1470                 KKASSERT(hmp == NULL);
1471                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1472                 pmp->mp = NULL;
1473                 mp->mnt_data = NULL;
1474
1475                 /*
1476                  * After pmp->mp is cleared we have to account for
1477                  * mount_count.
1478                  */
1479                 cluster = &pmp->iroot->cluster;
1480                 for (i = 0; i < cluster->nchains; ++i) {
1481                         rchain = cluster->array[i].chain;
1482                         if (rchain == NULL)
1483                                 continue;
1484                         --rchain->hmp->mount_count;
1485                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1486                                 rchain->hmp, rchain->hmp->mount_count);
1487                         /* scrapping hmp now may invalidate the pmp */
1488                 }
1489 again:
1490                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1491                         if (hmp->mount_count == 0) {
1492                                 hammer2_unmount_helper(NULL, NULL, hmp);
1493                                 goto again;
1494                         }
1495                 }
1496                 return;
1497         }
1498
1499         /*
1500          * Try to terminate the block device.  We can't terminate it if
1501          * there are still PFSs referencing it.
1502          */
1503         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1504                 hmp, hmp->mount_count);
1505         if (hmp->mount_count)
1506                 return;
1507
1508         hammer2_pfsfree_scan(hmp);
1509         hammer2_dev_exlock(hmp);        /* XXX order */
1510
1511         /*
1512          * Cycle the volume data lock as a safety (probably not needed any
1513          * more).  To ensure everything is out we need to flush at least
1514          * three times.  (1) The running of the sideq can dirty the
1515          * filesystem, (2) A normal flush can dirty the freemap, and
1516          * (3) ensure that the freemap is fully synchronized.
1517          *
1518          * The next mount's recovery scan can clean everything up but we want
1519          * to leave the filesystem in a 100% clean state on a normal unmount.
1520          */
1521 #if 0
1522         hammer2_voldata_lock(hmp);
1523         hammer2_voldata_unlock(hmp);
1524 #endif
1525         hammer2_iocom_uninit(hmp);
1526
1527         if ((hmp->vchain.flags | hmp->fchain.flags) &
1528             HAMMER2_CHAIN_FLUSH_MASK) {
1529                 kprintf("hammer2_unmount: chains left over "
1530                         "after final sync\n");
1531                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1532                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1533
1534                 if (hammer2_debug & 0x0010)
1535                         Debugger("entered debugger");
1536         }
1537
1538         KKASSERT(hmp->spmp == NULL);
1539
1540         /*
1541          * Finish up with the device vnode
1542          */
1543         if ((devvp = hmp->devvp) != NULL) {
1544                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1545                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1546                 hmp->devvp = NULL;
1547                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1548                 vn_unlock(devvp);
1549                 vrele(devvp);
1550                 devvp = NULL;
1551         }
1552
1553         /*
1554          * Clear vchain/fchain flags that might prevent final cleanup
1555          * of these chains.
1556          */
1557         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1558                 atomic_add_long(&hammer2_count_modified_chains, -1);
1559                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1560                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1561         }
1562         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1563                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1564         }
1565
1566         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1567                 atomic_add_long(&hammer2_count_modified_chains, -1);
1568                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1569                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1570         }
1571         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1572                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1573         }
1574
1575         /*
1576          * Final drop of embedded freemap root chain to
1577          * clean up fchain.core (fchain structure is not
1578          * flagged ALLOCATED so it is cleaned out and then
1579          * left to rot).
1580          */
1581         hammer2_chain_drop(&hmp->fchain);
1582
1583         /*
1584          * Final drop of embedded volume root chain to clean
1585          * up vchain.core (vchain structure is not flagged
1586          * ALLOCATED so it is cleaned out and then left to
1587          * rot).
1588          */
1589         dumpcnt = 50;
1590         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1591         dumpcnt = 50;
1592         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1593         hammer2_dev_unlock(hmp);
1594         hammer2_chain_drop(&hmp->vchain);
1595
1596         hammer2_io_cleanup(hmp, &hmp->iotree);
1597         if (hmp->iofree_count) {
1598                 kprintf("io_cleanup: %d I/O's left hanging\n",
1599                         hmp->iofree_count);
1600         }
1601
1602         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1603         kmalloc_destroy(&hmp->mchain);
1604         kfree(hmp, M_HAMMER2);
1605 }
1606
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                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1645                         kprintf("vget: no collect error but also no rdata\n");
1646                         kprintf("xop %p\n", xop);
1647                         while ((hammer2_debug & 0x80000) == 0) {
1648                                 tsleep(xop, PCATCH, "wait", hz * 10);
1649                         }
1650                         ip = NULL;
1651                 } else {
1652                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1653                 }
1654         }
1655         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1656
1657         if (ip) {
1658                 *vpp = hammer2_igetv(ip, &error);
1659                 hammer2_inode_unlock(ip);
1660         } else {
1661                 *vpp = NULL;
1662                 error = ENOENT;
1663         }
1664         return (error);
1665 }
1666
1667 static
1668 int
1669 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1670 {
1671         hammer2_pfs_t *pmp;
1672         struct vnode *vp;
1673         int error;
1674
1675         pmp = MPTOPMP(mp);
1676         if (pmp->iroot == NULL) {
1677                 *vpp = NULL;
1678                 return EINVAL;
1679         }
1680
1681         error = 0;
1682         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1683
1684         while (pmp->inode_tid == 0) {
1685                 hammer2_xop_ipcluster_t *xop;
1686                 hammer2_inode_meta_t *meta;
1687
1688                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1689                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1690                 error = hammer2_xop_collect(&xop->head, 0);
1691
1692                 if (error == 0) {
1693                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1694                         pmp->iroot->meta = *meta;
1695                         pmp->inode_tid = meta->pfs_inum + 1;
1696                         if (pmp->inode_tid < HAMMER2_INODE_START)
1697                                 pmp->inode_tid = HAMMER2_INODE_START;
1698                         pmp->modify_tid =
1699                                 xop->head.cluster.focus->bref.modify_tid + 1;
1700                         kprintf("PFS: Starting inode %jd\n",
1701                                 (intmax_t)pmp->inode_tid);
1702                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1703                                 pmp->inode_tid, pmp->modify_tid);
1704                         wakeup(&pmp->iroot);
1705
1706                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1707
1708                         /*
1709                          * Prime the mount info.
1710                          */
1711                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1712
1713                         /*
1714                          * With the cluster operational, check for and
1715                          * install ihidden if needed.  The install_hidden
1716                          * code needs to get a transaction so we must unlock
1717                          * iroot around it.
1718                          *
1719                          * This is only applicable PFS mounts, there is no
1720                          * hidden directory in the spmp.
1721                          */
1722                         hammer2_inode_unlock(pmp->iroot);
1723                         hammer2_inode_install_hidden(pmp);
1724                         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1725
1726                         break;
1727                 }
1728
1729                 /*
1730                  * Loop, try again
1731                  */
1732                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1733                 hammer2_inode_unlock(pmp->iroot);
1734                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1735                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1736                 if (error == EINTR)
1737                         break;
1738         }
1739
1740         if (error) {
1741                 hammer2_inode_unlock(pmp->iroot);
1742                 *vpp = NULL;
1743         } else {
1744                 vp = hammer2_igetv(pmp->iroot, &error);
1745                 hammer2_inode_unlock(pmp->iroot);
1746                 *vpp = vp;
1747         }
1748
1749         return (error);
1750 }
1751
1752 /*
1753  * Filesystem status
1754  *
1755  * XXX incorporate ipdata->meta.inode_quota and data_quota
1756  */
1757 static
1758 int
1759 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1760 {
1761         hammer2_pfs_t *pmp;
1762         hammer2_dev_t *hmp;
1763         hammer2_blockref_t bref;
1764         int i;
1765
1766         /*
1767          * NOTE: iroot might not have validated the cluster yet.
1768          */
1769         pmp = MPTOPMP(mp);
1770
1771         mp->mnt_stat.f_files = 0;
1772         mp->mnt_stat.f_ffree = 0;
1773         mp->mnt_stat.f_blocks = 0;
1774         mp->mnt_stat.f_bfree = 0;
1775         mp->mnt_stat.f_bavail = 0;
1776
1777         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1778                 hmp = pmp->pfs_hmps[i];
1779                 if (hmp == NULL)
1780                         continue;
1781                 if (pmp->iroot->cluster.array[i].chain)
1782                         bref = pmp->iroot->cluster.array[i].chain->bref;
1783                 else
1784                         bzero(&bref, sizeof(bref));
1785
1786                 mp->mnt_stat.f_files = bref.inode_count;
1787                 mp->mnt_stat.f_ffree = 0;
1788                 mp->mnt_stat.f_blocks = (bref.data_count +
1789                                          hmp->voldata.allocator_free) /
1790                                         mp->mnt_vstat.f_bsize;
1791                 mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free /
1792                                         mp->mnt_vstat.f_bsize;
1793                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1794
1795                 *sbp = mp->mnt_stat;
1796         }
1797         return (0);
1798 }
1799
1800 static
1801 int
1802 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1803 {
1804         hammer2_pfs_t *pmp;
1805         hammer2_dev_t *hmp;
1806         hammer2_blockref_t bref;
1807         int i;
1808
1809         /*
1810          * NOTE: iroot might not have validated the cluster yet.
1811          */
1812         pmp = MPTOPMP(mp);
1813
1814         mp->mnt_vstat.f_bsize = 0;
1815         mp->mnt_vstat.f_files = 0;
1816         mp->mnt_vstat.f_ffree = 0;
1817         mp->mnt_vstat.f_blocks = 0;
1818         mp->mnt_vstat.f_bfree = 0;
1819         mp->mnt_vstat.f_bavail = 0;
1820
1821         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1822                 hmp = pmp->pfs_hmps[i];
1823                 if (hmp == NULL)
1824                         continue;
1825                 if (pmp->iroot->cluster.array[i].chain)
1826                         bref = pmp->iroot->cluster.array[i].chain->bref;
1827                 else
1828                         bzero(&bref, sizeof(bref));
1829
1830                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1831                 mp->mnt_vstat.f_files = bref.inode_count;
1832                 mp->mnt_vstat.f_ffree = 0;
1833                 mp->mnt_vstat.f_blocks = (bref.data_count +
1834                                          hmp->voldata.allocator_free) /
1835                                         mp->mnt_vstat.f_bsize;
1836                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1837                                         mp->mnt_vstat.f_bsize;
1838                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1839
1840                 *sbp = mp->mnt_vstat;
1841         }
1842         return (0);
1843 }
1844
1845 /*
1846  * Mount-time recovery (RW mounts)
1847  *
1848  * Updates to the free block table are allowed to lag flushes by one
1849  * transaction.  In case of a crash, then on a fresh mount we must do an
1850  * incremental scan of the last committed transaction id and make sure that
1851  * all related blocks have been marked allocated.
1852  *
1853  * The super-root topology and each PFS has its own transaction id domain,
1854  * so we must track PFS boundary transitions.
1855  */
1856 struct hammer2_recovery_elm {
1857         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1858         hammer2_chain_t *chain;
1859         hammer2_tid_t sync_tid;
1860 };
1861
1862 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1863
1864 struct hammer2_recovery_info {
1865         struct hammer2_recovery_list list;
1866         hammer2_tid_t   mtid;
1867         int     depth;
1868 };
1869
1870 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1871                         hammer2_chain_t *parent,
1872                         struct hammer2_recovery_info *info,
1873                         hammer2_tid_t sync_tid);
1874
1875 #define HAMMER2_RECOVERY_MAXDEPTH       10
1876
1877 static
1878 int
1879 hammer2_recovery(hammer2_dev_t *hmp)
1880 {
1881         struct hammer2_recovery_info info;
1882         struct hammer2_recovery_elm *elm;
1883         hammer2_chain_t *parent;
1884         hammer2_tid_t sync_tid;
1885         hammer2_tid_t mirror_tid;
1886         int error;
1887         int cumulative_error = 0;
1888
1889         hammer2_trans_init(hmp->spmp, 0);
1890
1891         sync_tid = hmp->voldata.freemap_tid;
1892         mirror_tid = hmp->voldata.mirror_tid;
1893
1894         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1895         if (sync_tid >= mirror_tid) {
1896                 kprintf(" no recovery needed\n");
1897         } else {
1898                 kprintf(" freemap recovery %016jx-%016jx\n",
1899                         sync_tid + 1, mirror_tid);
1900         }
1901
1902         TAILQ_INIT(&info.list);
1903         info.depth = 0;
1904         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1905         cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
1906         hammer2_chain_lookup_done(parent);
1907
1908         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1909                 TAILQ_REMOVE(&info.list, elm, entry);
1910                 parent = elm->chain;
1911                 sync_tid = elm->sync_tid;
1912                 kfree(elm, M_HAMMER2);
1913
1914                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1915                 error = hammer2_recovery_scan(hmp, parent, &info,
1916                                               hmp->voldata.freemap_tid);
1917                 hammer2_chain_unlock(parent);
1918                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
1919                 if (error)
1920                         cumulative_error = error;
1921         }
1922         hammer2_trans_done(hmp->spmp);
1923
1924         return cumulative_error;
1925 }
1926
1927 static
1928 int
1929 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1930                       struct hammer2_recovery_info *info,
1931                       hammer2_tid_t sync_tid)
1932 {
1933         const hammer2_inode_data_t *ripdata;
1934         hammer2_chain_t *chain;
1935         hammer2_blockref_t bref;
1936         int cache_index;
1937         int cumulative_error = 0;
1938         int error;
1939         int first;
1940
1941         /*
1942          * Adjust freemap to ensure that the block(s) are marked allocated.
1943          */
1944         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1945                 hammer2_freemap_adjust(hmp, &parent->bref,
1946                                        HAMMER2_FREEMAP_DORECOVER);
1947         }
1948
1949         /*
1950          * Check type for recursive scan
1951          */
1952         switch(parent->bref.type) {
1953         case HAMMER2_BREF_TYPE_VOLUME:
1954                 /* data already instantiated */
1955                 break;
1956         case HAMMER2_BREF_TYPE_INODE:
1957                 /*
1958                  * Must instantiate data for DIRECTDATA test and also
1959                  * for recursion.
1960                  */
1961                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1962                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
1963                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1964                         /* not applicable to recovery scan */
1965                         hammer2_chain_unlock(parent);
1966                         return 0;
1967                 }
1968                 hammer2_chain_unlock(parent);
1969                 break;
1970         case HAMMER2_BREF_TYPE_INDIRECT:
1971                 /*
1972                  * Must instantiate data for recursion
1973                  */
1974                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1975                 hammer2_chain_unlock(parent);
1976                 break;
1977         case HAMMER2_BREF_TYPE_DATA:
1978         case HAMMER2_BREF_TYPE_FREEMAP:
1979         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1980         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1981                 /* not applicable to recovery scan */
1982                 return 0;
1983                 break;
1984         default:
1985                 return EDOM;
1986         }
1987
1988         /*
1989          * Defer operation if depth limit reached or if we are crossing a
1990          * PFS boundary.
1991          */
1992         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1993                 struct hammer2_recovery_elm *elm;
1994
1995                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1996                 elm->chain = parent;
1997                 elm->sync_tid = sync_tid;
1998                 hammer2_chain_ref(parent);
1999                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2000                 /* unlocked by caller */
2001
2002                 return(0);
2003         }
2004
2005
2006         /*
2007          * Recursive scan of the last flushed transaction only.  We are
2008          * doing this without pmp assignments so don't leave the chains
2009          * hanging around after we are done with them.
2010          */
2011         cache_index = 0;
2012         chain = NULL;
2013         first = 1;
2014
2015         while (hammer2_chain_scan(parent, &chain, &bref,
2016                                   &first, &cache_index,
2017                                   HAMMER2_LOOKUP_NODATA) != NULL) {
2018                 /*
2019                  * If this is a leaf
2020                  */
2021                 if (chain == NULL) {
2022                         if (bref.mirror_tid > sync_tid) {
2023                                 hammer2_freemap_adjust(hmp, &bref,
2024                                                      HAMMER2_FREEMAP_DORECOVER);
2025                         }
2026                         continue;
2027                 }
2028
2029                 /*
2030                  * This may or may not be a recursive node.
2031                  */
2032                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2033                 if (bref.mirror_tid > sync_tid) {
2034                         ++info->depth;
2035                         error = hammer2_recovery_scan(hmp, chain,
2036                                                       info, sync_tid);
2037                         --info->depth;
2038                         if (error)
2039                                 cumulative_error = error;
2040                 }
2041
2042                 /*
2043                  * Flush the recovery at the PFS boundary to stage it for
2044                  * the final flush of the super-root topology.
2045                  */
2046                 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2047                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2048                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2049                 }
2050         }
2051
2052         return cumulative_error;
2053 }
2054
2055 /*
2056  * Sync a mount point; this is called on a per-mount basis from the
2057  * filesystem syncer process periodically and whenever a user issues
2058  * a sync.
2059  */
2060 int
2061 hammer2_vfs_sync(struct mount *mp, int waitfor)
2062 {
2063         hammer2_xop_flush_t *xop;
2064         struct hammer2_sync_info info;
2065         hammer2_inode_t *iroot;
2066         hammer2_pfs_t *pmp;
2067         int flags;
2068         int error;
2069
2070         pmp = MPTOPMP(mp);
2071         iroot = pmp->iroot;
2072         KKASSERT(iroot);
2073         KKASSERT(iroot->pmp == pmp);
2074
2075         /*
2076          * We can't acquire locks on existing vnodes while in a transaction
2077          * without risking a deadlock.  This assumes that vfsync() can be
2078          * called without the vnode locked (which it can in DragonFly).
2079          * Otherwise we'd have to implement a multi-pass or flag the lock
2080          * failures and retry.
2081          *
2082          * The reclamation code interlocks with the sync list's token
2083          * (by removing the vnode from the scan list) before unlocking
2084          * the inode, giving us time to ref the inode.
2085          */
2086         /*flags = VMSC_GETVP;*/
2087         flags = 0;
2088         if (waitfor & MNT_LAZY)
2089                 flags |= VMSC_ONEPASS;
2090
2091 #if 0
2092         /*
2093          * Preflush the vnodes using a normal transaction before interlocking
2094          * with a flush transaction.
2095          */
2096         hammer2_trans_init(pmp, 0);
2097         info.error = 0;
2098         info.waitfor = MNT_NOWAIT;
2099         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2100         hammer2_trans_done(pmp);
2101 #endif
2102
2103         /*
2104          * Start our flush transaction.  This does not return until all
2105          * concurrent transactions have completed and will prevent any
2106          * new transactions from running concurrently, except for the
2107          * buffer cache transactions.
2108          *
2109          * For efficiency do an async pass before making sure with a
2110          * synchronous pass on all related buffer cache buffers.  It
2111          * should theoretically not be possible for any new file buffers
2112          * to be instantiated during this sequence.
2113          */
2114         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH |
2115                                 HAMMER2_TRANS_PREFLUSH);
2116         hammer2_inode_run_sideq(pmp);
2117
2118         info.error = 0;
2119         info.waitfor = MNT_NOWAIT;
2120         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2121         info.waitfor = MNT_WAIT;
2122         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2123
2124         /*
2125          * Clear PREFLUSH.  This prevents (or asserts on) any new logical
2126          * buffer cache flushes which occur during the flush.  Device buffers
2127          * are not affected.
2128          */
2129         hammer2_bioq_sync(pmp);
2130         hammer2_trans_clear_preflush(pmp);
2131
2132         /*
2133          * Use the XOP interface to concurrently flush all nodes to
2134          * synchronize the PFSROOT subtopology to the media.  A standard
2135          * end-of-scan ENOENT error indicates cluster sufficiency.
2136          *
2137          * Note that this flush will not be visible on crash recovery until
2138          * we flush the super-root topology in the next loop.
2139          *
2140          * XXX For now wait for all flushes to complete.
2141          */
2142         if (iroot) {
2143                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2144                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2145                 error = hammer2_xop_collect(&xop->head,
2146                                             HAMMER2_XOP_COLLECT_WAITALL);
2147                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2148                 if (error == ENOENT)
2149                         error = 0;
2150         } else {
2151                 error = 0;
2152         }
2153         hammer2_trans_done(pmp);
2154
2155         return (error);
2156 }
2157
2158 /*
2159  * Sync passes.
2160  *
2161  * Note that we ignore the tranasction mtid we got above.  Instead,
2162  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2163  * transactions.
2164  */
2165 static int
2166 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2167 {
2168         struct hammer2_sync_info *info = data;
2169         hammer2_inode_t *ip;
2170         int error;
2171
2172         /*
2173          * Degenerate cases.  Note that ip == NULL typically means the
2174          * syncer vnode itself and we don't want to vclrisdirty() in that
2175          * situation.
2176          */
2177         ip = VTOI(vp);
2178         if (ip == NULL) {
2179                 return(0);
2180         }
2181         if (vp->v_type == VNON || vp->v_type == VBAD) {
2182                 vclrisdirty(vp);
2183                 return(0);
2184         }
2185
2186         /*
2187          * VOP_FSYNC will start a new transaction so replicate some code
2188          * here to do it inline (see hammer2_vop_fsync()).
2189          *
2190          * WARNING: The vfsync interacts with the buffer cache and might
2191          *          block, we can't hold the inode lock at that time.
2192          *          However, we MUST ref ip before blocking to ensure that
2193          *          it isn't ripped out from under us (since we do not
2194          *          hold a lock on the vnode).
2195          */
2196         hammer2_inode_ref(ip);
2197         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2198             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2199                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2200                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2201                                  HAMMER2_INODE_MODIFIED)) {
2202                         hammer2_inode_lock(ip, 0);
2203                         hammer2_inode_chain_sync(ip);
2204                         hammer2_inode_unlock(ip);
2205                 }
2206         }
2207         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2208             RB_EMPTY(&vp->v_rbdirty_tree)) {
2209                 vclrisdirty(vp);
2210         }
2211
2212         hammer2_inode_drop(ip);
2213 #if 1
2214         error = 0;
2215         if (error)
2216                 info->error = error;
2217 #endif
2218         return(0);
2219 }
2220
2221 static
2222 int
2223 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2224 {
2225         hammer2_inode_t *ip;
2226
2227         KKASSERT(MAXFIDSZ >= 16);
2228         ip = VTOI(vp);
2229         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2230         fhp->fid_ext = 0;
2231         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2232         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2233
2234         return 0;
2235 }
2236
2237 static
2238 int
2239 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2240                struct fid *fhp, struct vnode **vpp)
2241 {
2242         hammer2_pfs_t *pmp;
2243         hammer2_tid_t inum;
2244         int error;
2245
2246         pmp = MPTOPMP(mp);
2247         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2248         if (vpp) {
2249                 if (inum == 1)
2250                         error = hammer2_vfs_root(mp, vpp);
2251                 else
2252                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2253         } else {
2254                 error = 0;
2255         }
2256         if (error)
2257                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2258         return error;
2259 }
2260
2261 static
2262 int
2263 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2264                  int *exflagsp, struct ucred **credanonp)
2265 {
2266         hammer2_pfs_t *pmp;
2267         struct netcred *np;
2268         int error;
2269
2270         pmp = MPTOPMP(mp);
2271         np = vfs_export_lookup(mp, &pmp->export, nam);
2272         if (np) {
2273                 *exflagsp = np->netc_exflags;
2274                 *credanonp = &np->netc_anon;
2275                 error = 0;
2276         } else {
2277                 error = EACCES;
2278         }
2279         return error;
2280 }
2281
2282 /*
2283  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2284  * header into the HMP
2285  *
2286  * XXX read four volhdrs and use the one with the highest TID whos CRC
2287  *     matches.
2288  *
2289  * XXX check iCRCs.
2290  *
2291  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2292  *     nonexistant locations.
2293  *
2294  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2295  */
2296 static
2297 int
2298 hammer2_install_volume_header(hammer2_dev_t *hmp)
2299 {
2300         hammer2_volume_data_t *vd;
2301         struct buf *bp;
2302         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2303         int error_reported;
2304         int error;
2305         int valid;
2306         int i;
2307
2308         error_reported = 0;
2309         error = 0;
2310         valid = 0;
2311         bp = NULL;
2312
2313         /*
2314          * There are up to 4 copies of the volume header (syncs iterate
2315          * between them so there is no single master).  We don't trust the
2316          * volu_size field so we don't know precisely how large the filesystem
2317          * is, so depend on the OS to return an error if we go beyond the
2318          * block device's EOF.
2319          */
2320         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2321                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2322                               HAMMER2_VOLUME_BYTES, &bp);
2323                 if (error) {
2324                         brelse(bp);
2325                         bp = NULL;
2326                         continue;
2327                 }
2328
2329                 vd = (struct hammer2_volume_data *) bp->b_data;
2330                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2331                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2332                         brelse(bp);
2333                         bp = NULL;
2334                         continue;
2335                 }
2336
2337                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2338                         /* XXX: Reversed-endianness filesystem */
2339                         kprintf("hammer2: reverse-endian filesystem detected");
2340                         brelse(bp);
2341                         bp = NULL;
2342                         continue;
2343                 }
2344
2345                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2346                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2347                                       HAMMER2_VOLUME_ICRC0_SIZE);
2348                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2349                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2350                                        HAMMER2_VOLUME_ICRC1_SIZE);
2351                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2352                         kprintf("hammer2 volume header crc "
2353                                 "mismatch copy #%d %08x/%08x\n",
2354                                 i, crc0, crc);
2355                         error_reported = 1;
2356                         brelse(bp);
2357                         bp = NULL;
2358                         continue;
2359                 }
2360                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2361                         valid = 1;
2362                         hmp->voldata = *vd;
2363                         hmp->volhdrno = i;
2364                 }
2365                 brelse(bp);
2366                 bp = NULL;
2367         }
2368         if (valid) {
2369                 hmp->volsync = hmp->voldata;
2370                 error = 0;
2371                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2372                         kprintf("hammer2: using volume header #%d\n",
2373                                 hmp->volhdrno);
2374                 }
2375         } else {
2376                 error = EINVAL;
2377                 kprintf("hammer2: no valid volume headers found!\n");
2378         }
2379         return (error);
2380 }
2381
2382 /*
2383  * This handles hysteresis on regular file flushes.  Because the BIOs are
2384  * routed to a thread it is possible for an excessive number to build up
2385  * and cause long front-end stalls long before the runningbuffspace limit
2386  * is hit, so we implement hammer2_flush_pipe to control the
2387  * hysteresis.
2388  *
2389  * This is a particular problem when compression is used.
2390  */
2391 void
2392 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2393 {
2394         atomic_add_int(&pmp->count_lwinprog, 1);
2395 }
2396
2397 void
2398 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2399 {
2400         int lwinprog;
2401
2402         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2403         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2404             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2405                 atomic_clear_int(&pmp->count_lwinprog,
2406                                  HAMMER2_LWINPROG_WAITING);
2407                 wakeup(&pmp->count_lwinprog);
2408         }
2409         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2410             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2411                 atomic_clear_int(&pmp->count_lwinprog,
2412                                  HAMMER2_LWINPROG_WAITING0);
2413                 wakeup(&pmp->count_lwinprog);
2414         }
2415 }
2416
2417 void
2418 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2419 {
2420         int lwinprog;
2421         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2422                                     HAMMER2_LWINPROG_WAITING0;
2423
2424         for (;;) {
2425                 lwinprog = pmp->count_lwinprog;
2426                 cpu_ccfence();
2427                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2428                         break;
2429                 tsleep_interlock(&pmp->count_lwinprog, 0);
2430                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2431                 lwinprog = pmp->count_lwinprog;
2432                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2433                         break;
2434                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2435         }
2436 }
2437
2438 /*
2439  * Manage excessive memory resource use for chain and related
2440  * structures.
2441  */
2442 void
2443 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2444 {
2445         uint32_t waiting;
2446         uint32_t count;
2447         uint32_t limit;
2448 #if 0
2449         static int zzticks;
2450 #endif
2451
2452         /*
2453          * Atomic check condition and wait.  Also do an early speedup of
2454          * the syncer to try to avoid hitting the wait.
2455          */
2456         for (;;) {
2457                 waiting = pmp->inmem_dirty_chains;
2458                 cpu_ccfence();
2459                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2460
2461                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2462                 if (limit < hammer2_limit_dirty_chains)
2463                         limit = hammer2_limit_dirty_chains;
2464                 if (limit < 1000)
2465                         limit = 1000;
2466
2467 #if 0
2468                 if ((int)(ticks - zzticks) > hz) {
2469                         zzticks = ticks;
2470                         kprintf("count %ld %ld\n", count, limit);
2471                 }
2472 #endif
2473
2474                 /*
2475                  * Block if there are too many dirty chains present, wait
2476                  * for the flush to clean some out.
2477                  */
2478                 if (count > limit) {
2479                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2480                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2481                                                waiting,
2482                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2483                                 speedup_syncer(pmp->mp);
2484                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2485                                        "chnmem", hz);
2486                         }
2487                         continue;       /* loop on success or fail */
2488                 }
2489
2490                 /*
2491                  * Try to start an early flush before we are forced to block.
2492                  */
2493                 if (count > limit * 7 / 10)
2494                         speedup_syncer(pmp->mp);
2495                 break;
2496         }
2497 }
2498
2499 void
2500 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2501 {
2502         if (pmp) {
2503                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2504         }
2505 }
2506
2507 void
2508 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2509 {
2510         uint32_t waiting;
2511
2512         if (pmp == NULL)
2513                 return;
2514
2515         for (;;) {
2516                 waiting = pmp->inmem_dirty_chains;
2517                 cpu_ccfence();
2518                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2519                                        waiting,
2520                                        (waiting - 1) &
2521                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2522                         break;
2523                 }
2524         }
2525
2526         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2527                 wakeup(&pmp->inmem_dirty_chains);
2528 }
2529
2530 /*
2531  * Debugging
2532  */
2533 void
2534 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2535 {
2536         hammer2_chain_t *scan;
2537         hammer2_chain_t *parent;
2538
2539         --*countp;
2540         if (*countp == 0) {
2541                 kprintf("%*.*s...\n", tab, tab, "");
2542                 return;
2543         }
2544         if (*countp < 0)
2545                 return;
2546         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2547                 tab, tab, "", pfx,
2548                 chain, chain->bref.type,
2549                 chain->bref.key, chain->bref.keybits,
2550                 chain->bref.mirror_tid);
2551
2552         kprintf("%*.*s      [%08x] (%s) refs=%d",
2553                 tab, tab, "",
2554                 chain->flags,
2555                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2556                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2557                 chain->refs);
2558
2559         parent = chain->parent;
2560         if (parent)
2561                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2562                         tab, tab, "",
2563                         parent, parent->flags, parent->refs);
2564         if (RB_EMPTY(&chain->core.rbtree)) {
2565                 kprintf("\n");
2566         } else {
2567                 kprintf(" {\n");
2568                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2569                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2570                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2571                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2572                                 chain->data->ipdata.filename);
2573                 else
2574                         kprintf("%*.*s}\n", tab, tab, "");
2575         }
2576 }