sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2018 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72         int pass;
  73 };
  74
  75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  76 static struct hammer2_mntlist hammer2_mntlist;
  77
  78 struct hammer2_pfslist hammer2_pfslist;
  79 struct hammer2_pfslist hammer2_spmplist;
  80 struct lock hammer2_mntlk;
  81
  82 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
  83 int hammer2_debug;
  84 long hammer2_debug_inode;
  85 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
  86 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
  87 int hammer2_cluster_write = 0;          /* physical write clustering */
  88 int hammer2_dedup_enable = 1;
  89 int hammer2_always_compress = 0;        /* always try to compress */
  90 int hammer2_inval_enable = 0;
  91 int hammer2_flush_pipe = 100;
  92 int hammer2_dio_count;
  93 int hammer2_dio_limit = 256;
  94 int hammer2_bulkfree_tps = 5000;
  95 long hammer2_chain_allocs;
  96 long hammer2_chain_frees;
  97 long hammer2_limit_dirty_chains;
  98 long hammer2_limit_dirty_inodes;
  99 long hammer2_count_modified_chains;
 100 long hammer2_iod_invals;
 101 long hammer2_iod_file_read;
 102 long hammer2_iod_meta_read;
 103 long hammer2_iod_indr_read;
 104 long hammer2_iod_fmap_read;
 105 long hammer2_iod_volu_read;
 106 long hammer2_iod_file_write;
 107 long hammer2_iod_file_wembed;
 108 long hammer2_iod_file_wzero;
 109 long hammer2_iod_file_wdedup;
 110 long hammer2_iod_meta_write;
 111 long hammer2_iod_indr_write;
 112 long hammer2_iod_fmap_write;
 113 long hammer2_iod_volu_write;
 114 long hammer2_iod_inode_creates;
 115 long hammer2_iod_inode_deletes;
 116
 117 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 118 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 119                 "Buffer used for compression.");
 120
 121 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 122 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 123                 "Buffer used for decompression.");
 124
 125 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 126
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
 128            &hammer2_supported_version, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 130            &hammer2_debug, 0, "");
 131 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW,
 132            &hammer2_debug_inode, 0, "");
 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 134            &hammer2_cluster_meta_read, 0, "");
 135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 136            &hammer2_cluster_data_read, 0, "");
 137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 138            &hammer2_cluster_write, 0, "");
 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 140            &hammer2_dedup_enable, 0, "");
 141 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 142            &hammer2_always_compress, 0, "");
 143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 144            &hammer2_inval_enable, 0, "");
 145 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 146            &hammer2_flush_pipe, 0, "");
 147 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
 148            &hammer2_bulkfree_tps, 0, "");
 149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 150            &hammer2_chain_allocs, 0, "");
 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 152            &hammer2_chain_frees, 0, "");
 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 154            &hammer2_limit_dirty_chains, 0, "");
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
 156            &hammer2_limit_dirty_inodes, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 158            &hammer2_count_modified_chains, 0, "");
 159 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 160            &hammer2_dio_count, 0, "");
 161 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
 162            &hammer2_dio_limit, 0, "");
 163
 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 165            &hammer2_iod_invals, 0, "");
 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 167            &hammer2_iod_file_read, 0, "");
 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 169            &hammer2_iod_meta_read, 0, "");
 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 171            &hammer2_iod_indr_read, 0, "");
 172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 173            &hammer2_iod_fmap_read, 0, "");
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 175            &hammer2_iod_volu_read, 0, "");
 176
 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 178            &hammer2_iod_file_write, 0, "");
 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 180            &hammer2_iod_file_wembed, 0, "");
 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 182            &hammer2_iod_file_wzero, 0, "");
 183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 184            &hammer2_iod_file_wdedup, 0, "");
 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 186            &hammer2_iod_meta_write, 0, "");
 187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 188            &hammer2_iod_indr_write, 0, "");
 189 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 190            &hammer2_iod_fmap_write, 0, "");
 191 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 192            &hammer2_iod_volu_write, 0, "");
 193 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW,
 194            &hammer2_iod_inode_creates, 0, "");
 195 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW,
 196            &hammer2_iod_inode_deletes, 0, "");
 197
 198 long hammer2_process_icrc32;
 199 long hammer2_process_xxhash64;
 200 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
 201            &hammer2_process_icrc32, 0, "");
 202 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
 203            &hammer2_process_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 static void hammer2_vfs_modifying(struct mount *mp);
 224
 225 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 226 #if 0
 227 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 228 #endif
 229
 230 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 231
 232 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 233 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 234                                 hammer2_dev_t *hmp);
 235 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
 236
 237 /*
 238  * HAMMER2 vfs operations.
 239  */
 240 static struct vfsops hammer2_vfsops = {
 241         .vfs_init       = hammer2_vfs_init,
 242         .vfs_uninit     = hammer2_vfs_uninit,
 243         .vfs_sync       = hammer2_vfs_sync,
 244         .vfs_mount      = hammer2_vfs_mount,
 245         .vfs_unmount    = hammer2_vfs_unmount,
 246         .vfs_root       = hammer2_vfs_root,
 247         .vfs_statfs     = hammer2_vfs_statfs,
 248         .vfs_statvfs    = hammer2_vfs_statvfs,
 249         .vfs_vget       = hammer2_vfs_vget,
 250         .vfs_vptofh     = hammer2_vfs_vptofh,
 251         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 252         .vfs_checkexp   = hammer2_vfs_checkexp,
 253         .vfs_modifying  = hammer2_vfs_modifying
 254 };
 255
 256 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 257
 258 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 259 MODULE_VERSION(hammer2, 1);
 260
 261 static
 262 int
 263 hammer2_vfs_init(struct vfsconf *conf)
 264 {
 265         static struct objcache_malloc_args margs_read;
 266         static struct objcache_malloc_args margs_write;
 267         static struct objcache_malloc_args margs_vop;
 268
 269         int error;
 270
 271         error = 0;
 272         kmalloc_raise_limit(M_HAMMER2, 0);      /* unlimited */
 273
 274         /*
 275          * A large DIO cache is needed to retain dedup enablement masks.
 276          * The bulkfree code clears related masks as part of the disk block
 277          * recycling algorithm, preventing it from being used for a later
 278          * dedup.
 279          *
 280          * NOTE: A large buffer cache can actually interfere with dedup
 281          *       operation because we dedup based on media physical buffers
 282          *       and not logical buffers.  Try to make the DIO case large
 283          *       enough to avoid this problem, but also cap it.
 284          */
 285         hammer2_dio_limit = nbuf * 2;
 286         if (hammer2_dio_limit > 100000)
 287                 hammer2_dio_limit = 100000;
 288
 289         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 290                 error = EINVAL;
 291         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 292                 error = EINVAL;
 293         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 294                 error = EINVAL;
 295
 296         if (error)
 297                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 298
 299         margs_read.objsize = 65536;
 300         margs_read.mtype = M_HAMMER2_DEBUFFER;
 301
 302         margs_write.objsize = 32768;
 303         margs_write.mtype = M_HAMMER2_CBUFFER;
 304
 305         margs_vop.objsize = sizeof(hammer2_xop_t);
 306         margs_vop.mtype = M_HAMMER2;
 307
 308         /*
 309          * Note thaht for the XOPS cache we want backing store allocations
 310          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 311          * confusion), so use the backing store function that does it.  This
 312          * means that initial XOPS objects are zerod but REUSED objects are
 313          * not.  So we are responsible for cleaning the object up sufficiently
 314          * for our needs before objcache_put()ing it back (typically just the
 315          * FIFO indices).
 316          */
 317         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 318                                 0, 1, NULL, NULL, NULL,
 319                                 objcache_malloc_alloc,
 320                                 objcache_malloc_free,
 321                                 &margs_read);
 322         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 323                                 0, 1, NULL, NULL, NULL,
 324                                 objcache_malloc_alloc,
 325                                 objcache_malloc_free,
 326                                 &margs_write);
 327         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 328                                 0, 1, NULL, NULL, NULL,
 329                                 objcache_malloc_alloc_zero,
 330                                 objcache_malloc_free,
 331                                 &margs_vop);
 332
 333
 334         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 335         TAILQ_INIT(&hammer2_mntlist);
 336         TAILQ_INIT(&hammer2_pfslist);
 337         TAILQ_INIT(&hammer2_spmplist);
 338
 339         hammer2_limit_dirty_chains = maxvnodes / 10;
 340         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 341                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 342         if (hammer2_limit_dirty_chains < 1000)
 343                 hammer2_limit_dirty_chains = 1000;
 344
 345         hammer2_limit_dirty_inodes = maxvnodes / 25;
 346         if (hammer2_limit_dirty_inodes < 100)
 347                 hammer2_limit_dirty_inodes = 100;
 348         if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
 349                 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
 350
 351         return (error);
 352 }
 353
 354 static
 355 int
 356 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 357 {
 358         objcache_destroy(cache_buffer_read);
 359         objcache_destroy(cache_buffer_write);
 360         objcache_destroy(cache_xops);
 361         return 0;
 362 }
 363
 364 /*
 365  * Core PFS allocator.  Used to allocate or reference the pmp structure
 366  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 367  * The pmp can be passed in or loaded by this function using the chain and
 368  * inode data.
 369  *
 370  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 371  * transactions.  Note that synchronization does not use this field.
 372  * (typically frontend operations and synchronization cannot run on the
 373  * same PFS node at the same time).
 374  *
 375  * XXX check locking
 376  */
 377 hammer2_pfs_t *
 378 hammer2_pfsalloc(hammer2_chain_t *chain,
 379                  const hammer2_inode_data_t *ripdata,
 380                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 381 {
 382         hammer2_pfs_t *pmp;
 383         hammer2_inode_t *iroot;
 384         int count;
 385         int i;
 386         int j;
 387
 388         pmp = NULL;
 389
 390         /*
 391          * Locate or create the PFS based on the cluster id.  If ripdata
 392          * is NULL this is a spmp which is unique and is always allocated.
 393          *
 394          * If the device is mounted in local mode all PFSs are considered
 395          * independent and not part of any cluster (for debugging only).
 396          */
 397         if (ripdata) {
 398                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 399                         if (force_local != pmp->force_local)
 400                                 continue;
 401                         if (force_local == NULL &&
 402                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 403                                  sizeof(pmp->pfs_clid)) == 0) {
 404                                         break;
 405                         } else if (force_local && pmp->pfs_names[0] &&
 406                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 407                                         break;
 408                         }
 409                 }
 410         }
 411
 412         if (pmp == NULL) {
 413                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 414                 pmp->force_local = force_local;
 415                 hammer2_trans_manage_init(pmp);
 416                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 417                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 418                 lockinit(&pmp->lock, "pfslk", 0, 0);
 419                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 420                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 421                 spin_init(&pmp->xop_spin, "h2xop");
 422                 spin_init(&pmp->lru_spin, "h2lru");
 423                 RB_INIT(&pmp->inum_tree);
 424                 TAILQ_INIT(&pmp->syncq);
 425                 TAILQ_INIT(&pmp->depq);
 426                 TAILQ_INIT(&pmp->lru_list);
 427                 spin_init(&pmp->list_spin, "h2pfsalloc_list");
 428
 429                 /*
 430                  * Distribute backend operations to threads
 431                  */
 432                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 433                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 434
 435                 /*
 436                  * Save the last media transaction id for the flusher.  Set
 437                  * initial
 438                  */
 439                 if (ripdata) {
 440                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 441                         TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 442                 } else {
 443                         pmp->flags |= HAMMER2_PMPF_SPMP;
 444                         TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
 445                 }
 446
 447                 /*
 448                  * The synchronization thread may start too early, make
 449                  * sure it stays frozen until we are ready to let it go.
 450                  * XXX
 451                  */
 452                 /*
 453                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 454                                          HAMMER2_THREAD_REMASTER;
 455                 */
 456         }
 457
 458         /*
 459          * Create the PFS's root inode and any missing XOP helper threads.
 460          */
 461         if ((iroot = pmp->iroot) == NULL) {
 462                 iroot = hammer2_inode_get(pmp, NULL, 1, -1);
 463                 if (ripdata)
 464                         iroot->meta = ripdata->meta;
 465                 pmp->iroot = iroot;
 466                 hammer2_inode_ref(iroot);
 467                 hammer2_inode_unlock(iroot);
 468         }
 469
 470         /*
 471          * Stop here if no chain is passed in.
 472          */
 473         if (chain == NULL)
 474                 goto done;
 475
 476         /*
 477          * When a chain is passed in we must add it to the PFS's root
 478          * inode, update pmp->pfs_types[], and update the syncronization
 479          * threads.
 480          *
 481          * When forcing local mode, mark the PFS as a MASTER regardless.
 482          *
 483          * At the moment empty spots can develop due to removals or failures.
 484          * Ultimately we want to re-fill these spots but doing so might
 485          * confused running code. XXX
 486          */
 487         hammer2_inode_ref(iroot);
 488         hammer2_mtx_ex(&iroot->lock);
 489         j = iroot->cluster.nchains;
 490
 491         if (j == HAMMER2_MAXCLUSTER) {
 492                 kprintf("hammer2_mount: cluster full!\n");
 493                 /* XXX fatal error? */
 494         } else {
 495                 KKASSERT(chain->pmp == NULL);
 496                 chain->pmp = pmp;
 497                 hammer2_chain_ref(chain);
 498                 iroot->cluster.array[j].chain = chain;
 499                 if (force_local)
 500                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 501                 else
 502                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 503                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 504                 pmp->pfs_hmps[j] = chain->hmp;
 505                 hammer2_spin_ex(&pmp->inum_spin);
 506                 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset;
 507                 hammer2_spin_unex(&pmp->inum_spin);
 508
 509                 /*
 510                  * If the PFS is already mounted we must account
 511                  * for the mount_count here.
 512                  */
 513                 if (pmp->mp)
 514                         ++chain->hmp->mount_count;
 515
 516                 /*
 517                  * May have to fixup dirty chain tracking.  Previous
 518                  * pmp was NULL so nothing to undo.
 519                  */
 520                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 521                         hammer2_pfs_memory_inc(pmp);
 522                 ++j;
 523         }
 524         iroot->cluster.nchains = j;
 525
 526         /*
 527          * Update nmasters from any PFS inode which is part of the cluster.
 528          * It is possible that this will result in a value which is too
 529          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 530          * override this value later on.
 531          *
 532          * (This informs us of masters that might not currently be
 533          *  discoverable by this mount).
 534          */
 535         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 536                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 537         }
 538
 539         /*
 540          * Count visible masters.  Masters are usually added with
 541          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 542          * are more (XXX and must update the master inodes).
 543          */
 544         count = 0;
 545         for (i = 0; i < iroot->cluster.nchains; ++i) {
 546                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 547                         ++count;
 548         }
 549         if (pmp->pfs_nmasters < count)
 550                 pmp->pfs_nmasters = count;
 551
 552         /*
 553          * Create missing synchronization and support threads.
 554          *
 555          * Single-node masters (including snapshots) have nothing to
 556          * synchronize and do not require this thread.
 557          *
 558          * Multi-node masters or any number of soft masters, slaves, copy,
 559          * or other PFS types need the thread.
 560          *
 561          * Each thread is responsible for its particular cluster index.
 562          * We use independent threads so stalls or mismatches related to
 563          * any given target do not affect other targets.
 564          */
 565         for (i = 0; i < iroot->cluster.nchains; ++i) {
 566                 /*
 567                  * Single-node masters (including snapshots) have nothing
 568                  * to synchronize and will make direct xops support calls,
 569                  * thus they do not require this thread.
 570                  *
 571                  * Note that there can be thousands of snapshots.  We do not
 572                  * want to create thousands of threads.
 573                  */
 574                 if (pmp->pfs_nmasters <= 1 &&
 575                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 576                         continue;
 577                 }
 578
 579                 /*
 580                  * Sync support thread
 581                  */
 582                 if (pmp->sync_thrs[i].td == NULL) {
 583                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 584                                            "h2nod", i, -1,
 585                                            hammer2_primary_sync_thread);
 586                 }
 587         }
 588
 589         /*
 590          * Create missing Xop threads
 591          *
 592          * NOTE: We create helper threads for all mounted PFSs or any
 593          *       PFSs with 2+ nodes (so the sync thread can update them,
 594          *       even if not mounted).
 595          */
 596         if (pmp->mp || iroot->cluster.nchains >= 2)
 597                 hammer2_xop_helper_create(pmp);
 598
 599         hammer2_mtx_unlock(&iroot->lock);
 600         hammer2_inode_drop(iroot);
 601 done:
 602         return pmp;
 603 }
 604
 605 /*
 606  * Deallocate an element of a probed PFS.  If destroying and this is a
 607  * MASTER, adjust nmasters.
 608  *
 609  * This function does not physically destroy the PFS element in its device
 610  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 611  */
 612 void
 613 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 614 {
 615         hammer2_inode_t *iroot;
 616         hammer2_chain_t *chain;
 617         int j;
 618
 619         /*
 620          * Cleanup our reference on iroot.  iroot is (should) not be needed
 621          * by the flush code.
 622          */
 623         iroot = pmp->iroot;
 624         if (iroot) {
 625                 /*
 626                  * Stop synchronizing
 627                  *
 628                  * XXX flush after acquiring the iroot lock.
 629                  * XXX clean out the cluster index from all inode structures.
 630                  */
 631                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 632
 633                 /*
 634                  * Remove the cluster index from the group.  If destroying
 635                  * the PFS and this is a master, adjust pfs_nmasters.
 636                  */
 637                 hammer2_mtx_ex(&iroot->lock);
 638                 chain = iroot->cluster.array[clindex].chain;
 639                 iroot->cluster.array[clindex].chain = NULL;
 640
 641                 switch(pmp->pfs_types[clindex]) {
 642                 case HAMMER2_PFSTYPE_MASTER:
 643                         if (destroying && pmp->pfs_nmasters > 0)
 644                                 --pmp->pfs_nmasters;
 645                         /* XXX adjust ripdata->meta.pfs_nmasters */
 646                         break;
 647                 default:
 648                         break;
 649                 }
 650                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 651
 652                 hammer2_mtx_unlock(&iroot->lock);
 653
 654                 /*
 655                  * Release the chain.
 656                  */
 657                 if (chain) {
 658                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 659                         hammer2_chain_drop(chain);
 660                 }
 661
 662                 /*
 663                  * Terminate all XOP threads for the cluster index.
 664                  */
 665                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 666                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 667         }
 668 }
 669
 670 /*
 671  * Destroy a PFS, typically only occurs after the last mount on a device
 672  * has gone away.
 673  */
 674 static void
 675 hammer2_pfsfree(hammer2_pfs_t *pmp)
 676 {
 677         hammer2_inode_t *iroot;
 678         hammer2_chain_t *chain;
 679         int chains_still_present = 0;
 680         int i;
 681         int j;
 682
 683         /*
 684          * Cleanup our reference on iroot.  iroot is (should) not be needed
 685          * by the flush code.
 686          */
 687         if (pmp->flags & HAMMER2_PMPF_SPMP)
 688                 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
 689         else
 690                 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 691
 692         /*
 693          * Cleanup chains remaining on LRU list.
 694          */
 695         hammer2_spin_ex(&pmp->lru_spin);
 696         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 697                 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
 698                 atomic_add_int(&pmp->lru_count, -1);
 699                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
 700                 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
 701                 hammer2_chain_ref(chain);
 702                 hammer2_spin_unex(&pmp->lru_spin);
 703                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 704                 hammer2_chain_drop(chain);
 705                 hammer2_spin_ex(&pmp->lru_spin);
 706         }
 707         hammer2_spin_unex(&pmp->lru_spin);
 708
 709         /*
 710          * Clean up iroot
 711          */
 712         iroot = pmp->iroot;
 713         if (iroot) {
 714                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 715                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 716                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 717                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 718                         chain = iroot->cluster.array[i].chain;
 719                         if (chain && !RB_EMPTY(&chain->core.rbtree)) {
 720                                 kprintf("hammer2: Warning pmp %p still "
 721                                         "has active chains\n", pmp);
 722                                 chains_still_present = 1;
 723                         }
 724                 }
 725 #if REPORT_REFS_ERRORS
 726                 if (iroot->refs != 1)
 727                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 728                                 iroot, iroot->refs);
 729 #else
 730                 KKASSERT(iroot->refs == 1);
 731 #endif
 732                 /* ref for iroot */
 733                 hammer2_inode_drop(iroot);
 734                 pmp->iroot = NULL;
 735         }
 736
 737         /*
 738          * Free remaining pmp resources
 739          */
 740         if (chains_still_present) {
 741                 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp);
 742         } else {
 743                 kmalloc_destroy(&pmp->mmsg);
 744                 kmalloc_destroy(&pmp->minode);
 745                 kfree(pmp, M_HAMMER2);
 746         }
 747 }
 748
 749 /*
 750  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 751  * empty is terminated and freed.
 752  *
 753  * XXX inefficient.
 754  */
 755 static void
 756 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
 757 {
 758         hammer2_pfs_t *pmp;
 759         hammer2_inode_t *iroot;
 760         hammer2_chain_t *rchain;
 761         int i;
 762         int j;
 763         struct hammer2_pfslist *wlist;
 764
 765         if (which == 0)
 766                 wlist = &hammer2_pfslist;
 767         else
 768                 wlist = &hammer2_spmplist;
 769 again:
 770         TAILQ_FOREACH(pmp, wlist, mntentry) {
 771                 if ((iroot = pmp->iroot) == NULL)
 772                         continue;
 773
 774                 /*
 775                  * Determine if this PFS is affected.  If it is we must
 776                  * freeze all management threads and lock its iroot.
 777                  *
 778                  * Freezing a management thread forces it idle, operations
 779                  * in-progress will be aborted and it will have to start
 780                  * over again when unfrozen, or exit if told to exit.
 781                  */
 782                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 783                         if (pmp->pfs_hmps[i] == hmp)
 784                                 break;
 785                 }
 786                 if (i == HAMMER2_MAXCLUSTER)
 787                         continue;
 788
 789                 hammer2_vfs_sync_pmp(pmp, MNT_WAIT);
 790
 791                 /*
 792                  * Make sure all synchronization threads are locked
 793                  * down.
 794                  */
 795                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 796                         if (pmp->pfs_hmps[i] == NULL)
 797                                 continue;
 798                         hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 799                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 800                                 hammer2_thr_freeze_async(
 801                                         &pmp->xop_groups[j].thrs[i]);
 802                         }
 803                 }
 804                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 805                         if (pmp->pfs_hmps[i] == NULL)
 806                                 continue;
 807                         hammer2_thr_freeze(&pmp->sync_thrs[i]);
 808                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 809                                 hammer2_thr_freeze(
 810                                         &pmp->xop_groups[j].thrs[i]);
 811                         }
 812                 }
 813
 814                 /*
 815                  * Lock the inode and clean out matching chains.
 816                  * Note that we cannot use hammer2_inode_lock_*()
 817                  * here because that would attempt to validate the
 818                  * cluster that we are in the middle of ripping
 819                  * apart.
 820                  *
 821                  * WARNING! We are working directly on the inodes
 822                  *          embedded cluster.
 823                  */
 824                 hammer2_mtx_ex(&iroot->lock);
 825
 826                 /*
 827                  * Remove the chain from matching elements of the PFS.
 828                  */
 829                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 830                         if (pmp->pfs_hmps[i] != hmp)
 831                                 continue;
 832                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 833                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 834                                 hammer2_thr_delete(
 835                                         &pmp->xop_groups[j].thrs[i]);
 836                         }
 837                         rchain = iroot->cluster.array[i].chain;
 838                         iroot->cluster.array[i].chain = NULL;
 839                         pmp->pfs_types[i] = 0;
 840                         if (pmp->pfs_names[i]) {
 841                                 kfree(pmp->pfs_names[i], M_HAMMER2);
 842                                 pmp->pfs_names[i] = NULL;
 843                         }
 844                         if (rchain) {
 845                                 hammer2_chain_drop(rchain);
 846                                 /* focus hint */
 847                                 if (iroot->cluster.focus == rchain)
 848                                         iroot->cluster.focus = NULL;
 849                         }
 850                         pmp->pfs_hmps[i] = NULL;
 851                 }
 852                 hammer2_mtx_unlock(&iroot->lock);
 853
 854                 /*
 855                  * Cleanup trailing chains.  Gaps may remain.
 856                  */
 857                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 858                         if (pmp->pfs_hmps[i])
 859                                 break;
 860                 }
 861                 iroot->cluster.nchains = i + 1;
 862
 863                 /*
 864                  * If the PMP has no elements remaining we can destroy it.
 865                  * (this will transition management threads from frozen->exit).
 866                  */
 867                 if (iroot->cluster.nchains == 0) {
 868                         /*
 869                          * If this was the hmp's spmp, we need to clean
 870                          * a little more stuff out.
 871                          */
 872                         if (hmp->spmp == pmp) {
 873                                 hmp->spmp = NULL;
 874                                 hmp->vchain.pmp = NULL;
 875                                 hmp->fchain.pmp = NULL;
 876                         }
 877
 878                         /*
 879                          * Free the pmp and restart the loop
 880                          */
 881                         KKASSERT(TAILQ_EMPTY(&pmp->syncq));
 882                         KKASSERT(TAILQ_EMPTY(&pmp->depq));
 883                         hammer2_pfsfree(pmp);
 884                         goto again;
 885                 }
 886
 887                 /*
 888                  * If elements still remain we need to set the REMASTER
 889                  * flag and unfreeze it.
 890                  */
 891                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 892                         if (pmp->pfs_hmps[i] == NULL)
 893                                 continue;
 894                         hammer2_thr_remaster(&pmp->sync_thrs[i]);
 895                         hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 896                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 897                                 hammer2_thr_remaster(
 898                                         &pmp->xop_groups[j].thrs[i]);
 899                                 hammer2_thr_unfreeze(
 900                                         &pmp->xop_groups[j].thrs[i]);
 901                         }
 902                 }
 903         }
 904 }
 905
 906 /*
 907  * Mount or remount HAMMER2 fileystem from physical media
 908  *
 909  *      mountroot
 910  *              mp              mount point structure
 911  *              path            NULL
 912  *              data            <unused>
 913  *              cred            <unused>
 914  *
 915  *      mount
 916  *              mp              mount point structure
 917  *              path            path to mount point
 918  *              data            pointer to argument structure in user space
 919  *                      volume  volume path (device@LABEL form)
 920  *                      hflags  user mount flags
 921  *              cred            user credentials
 922  *
 923  * RETURNS:     0       Success
 924  *              !0      error number
 925  */
 926 static
 927 int
 928 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 929                   struct ucred *cred)
 930 {
 931         struct hammer2_mount_info info;
 932         hammer2_pfs_t *pmp;
 933         hammer2_pfs_t *spmp;
 934         hammer2_dev_t *hmp;
 935         hammer2_dev_t *force_local;
 936         hammer2_key_t key_next;
 937         hammer2_key_t key_dummy;
 938         hammer2_key_t lhc;
 939         struct vnode *devvp;
 940         struct nlookupdata nd;
 941         hammer2_chain_t *parent;
 942         hammer2_chain_t *chain;
 943         const hammer2_inode_data_t *ripdata;
 944         hammer2_blockref_t bref;
 945         struct file *fp;
 946         char devstr[MNAMELEN];
 947         size_t size;
 948         size_t done;
 949         char *dev;
 950         char *label;
 951         int ronly = 1;
 952         int error;
 953         int i;
 954
 955         hmp = NULL;
 956         pmp = NULL;
 957         dev = NULL;
 958         label = NULL;
 959         devvp = NULL;
 960
 961         if (path == NULL) {
 962                 /*
 963                  * Root mount
 964                  */
 965                 bzero(&info, sizeof(info));
 966                 info.cluster_fd = -1;
 967                 ksnprintf(devstr, sizeof(devstr), "%s",
 968                           mp->mnt_stat.f_mntfromname);
 969                 kprintf("hammer2_mount: root '%s'\n", devstr);
 970                 done = strlen(devstr) + 1;
 971         } else {
 972                 /*
 973                  * Non-root mount or updating a mount
 974                  */
 975                 error = copyin(data, &info, sizeof(info));
 976                 if (error)
 977                         return (error);
 978
 979                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 980                 if (error)
 981                         return (error);
 982                 kprintf("hammer2_mount: '%s'\n", devstr);
 983         }
 984
 985         /*
 986          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 987          * if no label specified, based on the partition id.  Error out if no
 988          * label or device (with partition id) is specified.  This is strictly
 989          * a convenience to match the default label created by newfs_hammer2,
 990          * our preference is that a label always be specified.
 991          *
 992          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
 993          *       that does not specify a device, as long as some H2 label
 994          *       has already been mounted from that device.  This makes
 995          *       mounting snapshots a lot easier.
 996          */
 997         dev = devstr;
 998         label = strchr(devstr, '@');
 999         if (label && ((label + 1) - dev) > done) {
1000                 kprintf("hammer2: mount: bad label %s/%zd\n",
1001                         devstr, done);
1002                 return (EINVAL);
1003         }
1004         if (label == NULL || label[1] == 0) {
1005                 char slice;
1006
1007                 if (label == NULL)
1008                         label = devstr + strlen(devstr);
1009                 else
1010                         *label = '\0';          /* clean up trailing @ */
1011
1012                 slice = label[-1];
1013                 switch(slice) {
1014                 case 'a':
1015                         label = "BOOT";
1016                         break;
1017                 case 'd':
1018                         label = "ROOT";
1019                         break;
1020                 default:
1021                         label = "DATA";
1022                         break;
1023                 }
1024         } else {
1025                 *label = '\0';
1026                 label++;
1027         }
1028
1029         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
1030                 dev, label, (mp->mnt_flag & MNT_RDONLY));
1031
1032         if (mp->mnt_flag & MNT_UPDATE) {
1033                 /*
1034                  * Update mount.  Note that pmp->iroot->cluster is
1035                  * an inode-embedded cluster and thus cannot be
1036                  * directly locked.
1037                  *
1038                  * XXX HAMMER2 needs to implement NFS export via
1039                  *     mountctl.
1040                  */
1041                 hammer2_cluster_t *cluster;
1042
1043                 pmp = MPTOPMP(mp);
1044                 pmp->hflags = info.hflags;
1045                 cluster = &pmp->iroot->cluster;
1046                 for (i = 0; i < cluster->nchains; ++i) {
1047                         if (cluster->array[i].chain == NULL)
1048                                 continue;
1049                         hmp = cluster->array[i].chain->hmp;
1050                         devvp = hmp->devvp;
1051                         error = hammer2_remount(hmp, mp, path,
1052                                                 devvp, cred);
1053                         if (error)
1054                                 break;
1055                 }
1056
1057                 return error;
1058         }
1059
1060         /*
1061          * HMP device mount
1062          *
1063          * If a path is specified and dev is not an empty string, lookup the
1064          * name and verify that it referes to a block device.
1065          *
1066          * If a path is specified and dev is an empty string we fall through
1067          * and locate the label in the hmp search.
1068          */
1069         if (path && *dev != 0) {
1070                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1071                 if (error == 0)
1072                         error = nlookup(&nd);
1073                 if (error == 0)
1074                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1075                 nlookup_done(&nd);
1076         } else if (path == NULL) {
1077                 /* root mount */
1078                 cdev_t cdev = kgetdiskbyname(dev);
1079                 error = bdevvp(cdev, &devvp);
1080                 if (error)
1081                         kprintf("hammer2: cannot find '%s'\n", dev);
1082         } else {
1083                 /*
1084                  * We will locate the hmp using the label in the hmp loop.
1085                  */
1086                 error = 0;
1087         }
1088
1089         /*
1090          * Make sure its a block device.  Do not check to see if it is
1091          * already mounted until we determine that its a fresh H2 device.
1092          */
1093         if (error == 0 && devvp) {
1094                 vn_isdisk(devvp, &error);
1095         }
1096
1097         /*
1098          * Determine if the device has already been mounted.  After this
1099          * check hmp will be non-NULL if we are doing the second or more
1100          * hammer2 mounts from the same device.
1101          */
1102         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1103         if (devvp) {
1104                 /*
1105                  * Match the device.  Due to the way devfs works,
1106                  * we may not be able to directly match the vnode pointer,
1107                  * so also check to see if the underlying device matches.
1108                  */
1109                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1110                         if (hmp->devvp == devvp)
1111                                 break;
1112                         if (devvp->v_rdev &&
1113                             hmp->devvp->v_rdev == devvp->v_rdev) {
1114                                 break;
1115                         }
1116                 }
1117
1118                 /*
1119                  * If no match this may be a fresh H2 mount, make sure
1120                  * the device is not mounted on anything else.
1121                  */
1122                 if (hmp == NULL)
1123                         error = vfs_mountedon(devvp);
1124         } else if (error == 0) {
1125                 /*
1126                  * Match the label to a pmp already probed.
1127                  */
1128                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1129                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1130                                 if (pmp->pfs_names[i] &&
1131                                     strcmp(pmp->pfs_names[i], label) == 0) {
1132                                         hmp = pmp->pfs_hmps[i];
1133                                         break;
1134                                 }
1135                         }
1136                         if (hmp)
1137                                 break;
1138                 }
1139                 if (hmp == NULL)
1140                         error = ENOENT;
1141         }
1142
1143         /*
1144          * Open the device if this isn't a secondary mount and construct
1145          * the H2 device mount (hmp).
1146          */
1147         if (hmp == NULL) {
1148                 hammer2_chain_t *schain;
1149                 hammer2_xid_t xid;
1150                 hammer2_xop_head_t xop;
1151
1152                 if (error == 0 && vcount(devvp) > 0) {
1153                         kprintf("Primary device already has references\n");
1154                         error = EBUSY;
1155                 }
1156
1157                 /*
1158                  * Now open the device
1159                  */
1160                 if (error == 0) {
1161                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1162                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1163                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1164                         if (error == 0) {
1165                                 error = VOP_OPEN(devvp,
1166                                              (ronly ? FREAD : FREAD | FWRITE),
1167                                              FSCRED, NULL);
1168                         }
1169                         vn_unlock(devvp);
1170                 }
1171                 if (error && devvp) {
1172                         vrele(devvp);
1173                         devvp = NULL;
1174                 }
1175                 if (error) {
1176                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1177                         return error;
1178                 }
1179                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1180                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1181                 hmp->ronly = ronly;
1182                 hmp->devvp = devvp;
1183                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1184                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1185                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1186                 RB_INIT(&hmp->iotree);
1187                 spin_init(&hmp->io_spin, "h2mount_io");
1188                 spin_init(&hmp->list_spin, "h2mount_list");
1189
1190                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1191                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1192                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1193
1194                 /*
1195                  * vchain setup. vchain.data is embedded.
1196                  * vchain.refs is initialized and will never drop to 0.
1197                  *
1198                  * NOTE! voldata is not yet loaded.
1199                  */
1200                 hmp->vchain.hmp = hmp;
1201                 hmp->vchain.refs = 1;
1202                 hmp->vchain.data = (void *)&hmp->voldata;
1203                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1204                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1205                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1206
1207                 hammer2_chain_core_init(&hmp->vchain);
1208                 /* hmp->vchain.u.xxx is left NULL */
1209
1210                 /*
1211                  * fchain setup.  fchain.data is embedded.
1212                  * fchain.refs is initialized and will never drop to 0.
1213                  *
1214                  * The data is not used but needs to be initialized to
1215                  * pass assertion muster.  We use this chain primarily
1216                  * as a placeholder for the freemap's top-level RBTREE
1217                  * so it does not interfere with the volume's topology
1218                  * RBTREE.
1219                  */
1220                 hmp->fchain.hmp = hmp;
1221                 hmp->fchain.refs = 1;
1222                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1223                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1224                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1225                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1226                 hmp->fchain.bref.methods =
1227                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1228                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1229
1230                 hammer2_chain_core_init(&hmp->fchain);
1231                 /* hmp->fchain.u.xxx is left NULL */
1232
1233                 /*
1234                  * Install the volume header and initialize fields from
1235                  * voldata.
1236                  */
1237                 error = hammer2_install_volume_header(hmp);
1238                 if (error) {
1239                         hammer2_unmount_helper(mp, NULL, hmp);
1240                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1241                         hammer2_vfs_unmount(mp, MNT_FORCE);
1242                         return error;
1243                 }
1244
1245                 /*
1246                  * Really important to get these right or flush will get
1247                  * confused.
1248                  */
1249                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1250                 spmp = hmp->spmp;
1251
1252                 /*
1253                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1254                  * is inherited from the volume header.
1255                  */
1256                 xid = 0;
1257                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1258                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1259                 hmp->vchain.pmp = spmp;
1260                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1261                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1262                 hmp->fchain.pmp = spmp;
1263
1264                 /*
1265                  * First locate the super-root inode, which is key 0
1266                  * relative to the volume header's blockset.
1267                  *
1268                  * Then locate the root inode by scanning the directory keyspace
1269                  * represented by the label.
1270                  */
1271                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1272                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1273                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1274                                       &error, 0);
1275                 hammer2_chain_lookup_done(parent);
1276                 if (schain == NULL) {
1277                         kprintf("hammer2_mount: invalid super-root\n");
1278                         hammer2_unmount_helper(mp, NULL, hmp);
1279                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1280                         hammer2_vfs_unmount(mp, MNT_FORCE);
1281                         return EINVAL;
1282                 }
1283                 if (schain->error) {
1284                         kprintf("hammer2_mount: error %s reading super-root\n",
1285                                 hammer2_error_str(schain->error));
1286                         hammer2_chain_unlock(schain);
1287                         hammer2_chain_drop(schain);
1288                         schain = NULL;
1289                         hammer2_unmount_helper(mp, NULL, hmp);
1290                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1291                         hammer2_vfs_unmount(mp, MNT_FORCE);
1292                         return EINVAL;
1293                 }
1294
1295                 /*
1296                  * The super-root always uses an inode_tid of 1 when
1297                  * creating PFSs.
1298                  */
1299                 spmp->inode_tid = 1;
1300                 spmp->modify_tid = schain->bref.modify_tid + 1;
1301
1302                 /*
1303                  * Sanity-check schain's pmp and finish initialization.
1304                  * Any chain belonging to the super-root topology should
1305                  * have a NULL pmp (not even set to spmp).
1306                  */
1307                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1308                 KKASSERT(schain->pmp == NULL);
1309                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1310
1311                 /*
1312                  * Replace the dummy spmp->iroot with a real one.  It's
1313                  * easier to just do a wholesale replacement than to try
1314                  * to update the chain and fixup the iroot fields.
1315                  *
1316                  * The returned inode is locked with the supplied cluster.
1317                  */
1318                 hammer2_dummy_xop_from_chain(&xop, schain);
1319                 hammer2_inode_drop(spmp->iroot);
1320                 spmp->iroot = NULL;
1321                 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1);
1322                 spmp->spmp_hmp = hmp;
1323                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1324                 spmp->pfs_hmps[0] = hmp;
1325                 hammer2_inode_ref(spmp->iroot);
1326                 hammer2_inode_unlock(spmp->iroot);
1327                 hammer2_cluster_unlock(&xop.cluster);
1328                 hammer2_chain_drop(schain);
1329                 /* do not call hammer2_cluster_drop() on an embedded cluster */
1330                 schain = NULL;  /* now invalid */
1331                 /* leave spmp->iroot with one ref */
1332
1333                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1334                         error = hammer2_recovery(hmp);
1335                         if (error == 0)
1336                                 error |= hammer2_fixup_pfses(hmp);
1337                         /* XXX do something with error */
1338                 }
1339                 hammer2_update_pmps(hmp);
1340                 hammer2_iocom_init(hmp);
1341                 hammer2_bulkfree_init(hmp);
1342
1343                 /*
1344                  * Ref the cluster management messaging descriptor.  The mount
1345                  * program deals with the other end of the communications pipe.
1346                  *
1347                  * Root mounts typically do not supply one.
1348                  */
1349                 if (info.cluster_fd >= 0) {
1350                         fp = holdfp(curthread, info.cluster_fd, -1);
1351                         if (fp) {
1352                                 hammer2_cluster_reconnect(hmp, fp);
1353                         } else {
1354                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1355                         }
1356                 }
1357         } else {
1358                 spmp = hmp->spmp;
1359                 if (info.hflags & HMNT2_DEVFLAGS) {
1360                         kprintf("hammer2: Warning: mount flags pertaining "
1361                                 "to the whole device may only be specified "
1362                                 "on the first mount of the device: %08x\n",
1363                                 info.hflags & HMNT2_DEVFLAGS);
1364                 }
1365         }
1366
1367         /*
1368          * Force local mount (disassociate all PFSs from their clusters).
1369          * Used primarily for debugging.
1370          */
1371         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1372
1373         /*
1374          * Lookup the mount point under the media-localized super-root.
1375          * Scanning hammer2_pfslist doesn't help us because it represents
1376          * PFS cluster ids which can aggregate several named PFSs together.
1377          *
1378          * cluster->pmp will incorrectly point to spmp and must be fixed
1379          * up later on.
1380          */
1381         hammer2_inode_lock(spmp->iroot, 0);
1382         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1383         lhc = hammer2_dirhash(label, strlen(label));
1384         chain = hammer2_chain_lookup(&parent, &key_next,
1385                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1386                                      &error, 0);
1387         while (chain) {
1388                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1389                     strcmp(label, chain->data->ipdata.filename) == 0) {
1390                         break;
1391                 }
1392                 chain = hammer2_chain_next(&parent, chain, &key_next,
1393                                             key_next,
1394                                             lhc + HAMMER2_DIRHASH_LOMASK,
1395                                             &error, 0);
1396         }
1397         if (parent) {
1398                 hammer2_chain_unlock(parent);
1399                 hammer2_chain_drop(parent);
1400         }
1401         hammer2_inode_unlock(spmp->iroot);
1402
1403         /*
1404          * PFS could not be found?
1405          */
1406         if (chain == NULL) {
1407                 if (error)
1408                         kprintf("hammer2_mount: PFS label I/O error\n");
1409                 else
1410                         kprintf("hammer2_mount: PFS label not found\n");
1411                 hammer2_unmount_helper(mp, NULL, hmp);
1412                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1413                 hammer2_vfs_unmount(mp, MNT_FORCE);
1414
1415                 return EINVAL;
1416         }
1417
1418         /*
1419          * Acquire the pmp structure (it should have already been allocated
1420          * via hammer2_update_pmps() so do not pass cluster in to add to
1421          * available chains).
1422          *
1423          * Check if the cluster has already been mounted.  A cluster can
1424          * only be mounted once, use null mounts to mount additional copies.
1425          */
1426         if (chain->error) {
1427                 kprintf("hammer2_mount: PFS label I/O error\n");
1428         } else {
1429                 ripdata = &chain->data->ipdata;
1430                 bref = chain->bref;
1431                 pmp = hammer2_pfsalloc(NULL, ripdata,
1432                                        bref.modify_tid, force_local);
1433         }
1434         hammer2_chain_unlock(chain);
1435         hammer2_chain_drop(chain);
1436
1437         /*
1438          * Finish the mount
1439          */
1440         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1441
1442         if (pmp->mp) {
1443                 kprintf("hammer2_mount: PFS already mounted!\n");
1444                 hammer2_unmount_helper(mp, NULL, hmp);
1445                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1446                 hammer2_vfs_unmount(mp, MNT_FORCE);
1447
1448                 return EBUSY;
1449         }
1450
1451         pmp->hflags = info.hflags;
1452         mp->mnt_flag |= MNT_LOCAL;
1453         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1454         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1455
1456         /*
1457          * required mount structure initializations
1458          */
1459         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1460         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1461
1462         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1463         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1464
1465         /*
1466          * Optional fields
1467          */
1468         mp->mnt_iosize_max = MAXPHYS;
1469
1470         /*
1471          * Connect up mount pointers.
1472          */
1473         hammer2_mount_helper(mp, pmp);
1474
1475         lockmgr(&hammer2_mntlk, LK_RELEASE);
1476
1477         /*
1478          * Finish setup
1479          */
1480         vfs_getnewfsid(mp);
1481         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1482         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1483         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1484
1485         if (path) {
1486                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1487                           MNAMELEN - 1, &size);
1488                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1489         } /* else root mount, already in there */
1490
1491         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1492         if (path) {
1493                 copyinstr(path, mp->mnt_stat.f_mntonname,
1494                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1495                           &size);
1496         } else {
1497                 /* root mount */
1498                 mp->mnt_stat.f_mntonname[0] = '/';
1499         }
1500
1501         /*
1502          * Initial statfs to prime mnt_stat.
1503          */
1504         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1505
1506         return 0;
1507 }
1508
1509 /*
1510  * Scan PFSs under the super-root and create hammer2_pfs structures.
1511  */
1512 static
1513 void
1514 hammer2_update_pmps(hammer2_dev_t *hmp)
1515 {
1516         const hammer2_inode_data_t *ripdata;
1517         hammer2_chain_t *parent;
1518         hammer2_chain_t *chain;
1519         hammer2_blockref_t bref;
1520         hammer2_dev_t *force_local;
1521         hammer2_pfs_t *spmp;
1522         hammer2_pfs_t *pmp;
1523         hammer2_key_t key_next;
1524         int error;
1525
1526         /*
1527          * Force local mount (disassociate all PFSs from their clusters).
1528          * Used primarily for debugging.
1529          */
1530         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1531
1532         /*
1533          * Lookup mount point under the media-localized super-root.
1534          *
1535          * cluster->pmp will incorrectly point to spmp and must be fixed
1536          * up later on.
1537          */
1538         spmp = hmp->spmp;
1539         hammer2_inode_lock(spmp->iroot, 0);
1540         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1541         chain = hammer2_chain_lookup(&parent, &key_next,
1542                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1543                                          &error, 0);
1544         while (chain) {
1545                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1546                         continue;
1547                 if (chain->error) {
1548                         kprintf("I/O error scanning PFS labels\n");
1549                 } else {
1550                         ripdata = &chain->data->ipdata;
1551                         bref = chain->bref;
1552
1553                         pmp = hammer2_pfsalloc(chain, ripdata,
1554                                                bref.modify_tid, force_local);
1555                 }
1556                 chain = hammer2_chain_next(&parent, chain, &key_next,
1557                                            key_next, HAMMER2_KEY_MAX,
1558                                            &error, 0);
1559         }
1560         if (parent) {
1561                 hammer2_chain_unlock(parent);
1562                 hammer2_chain_drop(parent);
1563         }
1564         hammer2_inode_unlock(spmp->iroot);
1565 }
1566
1567 static
1568 int
1569 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1570                 struct vnode *devvp, struct ucred *cred)
1571 {
1572         int error;
1573
1574         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1575                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1576                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1577                 vn_unlock(devvp);
1578                 error = hammer2_recovery(hmp);
1579                 if (error == 0)
1580                         error |= hammer2_fixup_pfses(hmp);
1581                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1582                 if (error == 0) {
1583                         VOP_CLOSE(devvp, FREAD, NULL);
1584                         hmp->ronly = 0;
1585                 } else {
1586                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1587                 }
1588                 vn_unlock(devvp);
1589         } else {
1590                 error = 0;
1591         }
1592         return error;
1593 }
1594
1595 static
1596 int
1597 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1598 {
1599         hammer2_pfs_t *pmp;
1600         int flags;
1601         int error = 0;
1602
1603         pmp = MPTOPMP(mp);
1604
1605         if (pmp == NULL)
1606                 return(0);
1607
1608         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1609
1610         /*
1611          * If mount initialization proceeded far enough we must flush
1612          * its vnodes and sync the underlying mount points.  Three syncs
1613          * are required to fully flush the filesystem (freemap updates lag
1614          * by one flush, and one extra for safety).
1615          */
1616         if (mntflags & MNT_FORCE)
1617                 flags = FORCECLOSE;
1618         else
1619                 flags = 0;
1620         if (pmp->iroot) {
1621                 error = vflush(mp, 0, flags);
1622                 if (error)
1623                         goto failed;
1624                 hammer2_vfs_sync(mp, MNT_WAIT);
1625                 hammer2_vfs_sync(mp, MNT_WAIT);
1626                 hammer2_vfs_sync(mp, MNT_WAIT);
1627         }
1628
1629         /*
1630          * Cleanup the frontend support XOPS threads
1631          */
1632         hammer2_xop_helper_cleanup(pmp);
1633
1634         if (pmp->mp)
1635                 hammer2_unmount_helper(mp, pmp, NULL);
1636
1637         error = 0;
1638 failed:
1639         lockmgr(&hammer2_mntlk, LK_RELEASE);
1640
1641         return (error);
1642 }
1643
1644 /*
1645  * Mount helper, hook the system mount into our PFS.
1646  * The mount lock is held.
1647  *
1648  * We must bump the mount_count on related devices for any
1649  * mounted PFSs.
1650  */
1651 static
1652 void
1653 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1654 {
1655         hammer2_cluster_t *cluster;
1656         hammer2_chain_t *rchain;
1657         int i;
1658
1659         mp->mnt_data = (qaddr_t)pmp;
1660         pmp->mp = mp;
1661
1662         /*
1663          * After pmp->mp is set we have to adjust hmp->mount_count.
1664          */
1665         cluster = &pmp->iroot->cluster;
1666         for (i = 0; i < cluster->nchains; ++i) {
1667                 rchain = cluster->array[i].chain;
1668                 if (rchain == NULL)
1669                         continue;
1670                 ++rchain->hmp->mount_count;
1671         }
1672
1673         /*
1674          * Create missing Xop threads
1675          */
1676         hammer2_xop_helper_create(pmp);
1677 }
1678
1679 /*
1680  * Mount helper, unhook the system mount from our PFS.
1681  * The mount lock is held.
1682  *
1683  * If hmp is supplied a mount responsible for being the first to open
1684  * the block device failed and the block device and all PFSs using the
1685  * block device must be cleaned up.
1686  *
1687  * If pmp is supplied multiple devices might be backing the PFS and each
1688  * must be disconnected.  This might not be the last PFS using some of the
1689  * underlying devices.  Also, we have to adjust our hmp->mount_count
1690  * accounting for the devices backing the pmp which is now undergoing an
1691  * unmount.
1692  */
1693 static
1694 void
1695 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1696 {
1697         hammer2_cluster_t *cluster;
1698         hammer2_chain_t *rchain;
1699         struct vnode *devvp;
1700         int dumpcnt;
1701         int ronly;
1702         int i;
1703
1704         /*
1705          * If no device supplied this is a high-level unmount and we have to
1706          * to disconnect the mount, adjust mount_count, and locate devices
1707          * that might now have no mounts.
1708          */
1709         if (pmp) {
1710                 KKASSERT(hmp == NULL);
1711                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1712                 pmp->mp = NULL;
1713                 mp->mnt_data = NULL;
1714
1715                 /*
1716                  * After pmp->mp is cleared we have to account for
1717                  * mount_count.
1718                  */
1719                 cluster = &pmp->iroot->cluster;
1720                 for (i = 0; i < cluster->nchains; ++i) {
1721                         rchain = cluster->array[i].chain;
1722                         if (rchain == NULL)
1723                                 continue;
1724                         --rchain->hmp->mount_count;
1725                         /* scrapping hmp now may invalidate the pmp */
1726                 }
1727 again:
1728                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1729                         if (hmp->mount_count == 0) {
1730                                 hammer2_unmount_helper(NULL, NULL, hmp);
1731                                 goto again;
1732                         }
1733                 }
1734                 return;
1735         }
1736
1737         /*
1738          * Try to terminate the block device.  We can't terminate it if
1739          * there are still PFSs referencing it.
1740          */
1741         if (hmp->mount_count)
1742                 return;
1743
1744         /*
1745          * Decomission the network before we start messing with the
1746          * device and PFS.
1747          */
1748         hammer2_iocom_uninit(hmp);
1749
1750         hammer2_bulkfree_uninit(hmp);
1751         hammer2_pfsfree_scan(hmp, 0);
1752 #if 0
1753         hammer2_dev_exlock(hmp);        /* XXX order */
1754 #endif
1755
1756         /*
1757          * Cycle the volume data lock as a safety (probably not needed any
1758          * more).  To ensure everything is out we need to flush at least
1759          * three times.  (1) The running of the sideq can dirty the
1760          * filesystem, (2) A normal flush can dirty the freemap, and
1761          * (3) ensure that the freemap is fully synchronized.
1762          *
1763          * The next mount's recovery scan can clean everything up but we want
1764          * to leave the filesystem in a 100% clean state on a normal unmount.
1765          */
1766 #if 0
1767         hammer2_voldata_lock(hmp);
1768         hammer2_voldata_unlock(hmp);
1769 #endif
1770
1771         /*
1772          * Flush whatever is left.  Unmounted but modified PFS's might still
1773          * have some dirty chains on them.
1774          */
1775         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1776         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1777
1778         if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1779                 hammer2_voldata_modify(hmp);
1780                 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1781                                             HAMMER2_FLUSH_ALL);
1782         }
1783         hammer2_chain_unlock(&hmp->fchain);
1784
1785         if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1786                 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1787                                             HAMMER2_FLUSH_ALL);
1788         }
1789         hammer2_chain_unlock(&hmp->vchain);
1790
1791         if ((hmp->vchain.flags | hmp->fchain.flags) &
1792             HAMMER2_CHAIN_FLUSH_MASK) {
1793                 kprintf("hammer2_unmount: chains left over "
1794                         "after final sync\n");
1795                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1796                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1797
1798                 if (hammer2_debug & 0x0010)
1799                         Debugger("entered debugger");
1800         }
1801
1802         hammer2_pfsfree_scan(hmp, 1);
1803
1804         KKASSERT(hmp->spmp == NULL);
1805
1806         /*
1807          * Finish up with the device vnode
1808          */
1809         if ((devvp = hmp->devvp) != NULL) {
1810                 ronly = hmp->ronly;
1811                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1812                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1813                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1814                         ronly);
1815                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1816                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1817                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1818                 hmp->devvp = NULL;
1819                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1820                 vn_unlock(devvp);
1821                 vrele(devvp);
1822                 devvp = NULL;
1823         }
1824
1825         /*
1826          * Clear vchain/fchain flags that might prevent final cleanup
1827          * of these chains.
1828          */
1829         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1830                 atomic_add_long(&hammer2_count_modified_chains, -1);
1831                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1832                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1833         }
1834         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1835                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1836         }
1837
1838         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1839                 atomic_add_long(&hammer2_count_modified_chains, -1);
1840                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1841                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1842         }
1843         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1844                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1845         }
1846
1847         /*
1848          * Final drop of embedded freemap root chain to
1849          * clean up fchain.core (fchain structure is not
1850          * flagged ALLOCATED so it is cleaned out and then
1851          * left to rot).
1852          */
1853         hammer2_chain_drop(&hmp->fchain);
1854
1855         /*
1856          * Final drop of embedded volume root chain to clean
1857          * up vchain.core (vchain structure is not flagged
1858          * ALLOCATED so it is cleaned out and then left to
1859          * rot).
1860          */
1861         dumpcnt = 50;
1862         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1863         dumpcnt = 50;
1864         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1865 #if 0
1866         hammer2_dev_unlock(hmp);
1867 #endif
1868         hammer2_chain_drop(&hmp->vchain);
1869
1870         hammer2_io_cleanup(hmp, &hmp->iotree);
1871         if (hmp->iofree_count) {
1872                 kprintf("io_cleanup: %d I/O's left hanging\n",
1873                         hmp->iofree_count);
1874         }
1875
1876         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1877         kmalloc_destroy(&hmp->mchain);
1878         kfree(hmp, M_HAMMER2);
1879 }
1880
1881 int
1882 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1883                  ino_t ino, struct vnode **vpp)
1884 {
1885         hammer2_xop_lookup_t *xop;
1886         hammer2_pfs_t *pmp;
1887         hammer2_inode_t *ip;
1888         hammer2_tid_t inum;
1889         int error;
1890
1891         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1892
1893         error = 0;
1894         pmp = MPTOPMP(mp);
1895
1896         /*
1897          * Easy if we already have it cached
1898          */
1899         ip = hammer2_inode_lookup(pmp, inum);
1900         if (ip) {
1901                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1902                 *vpp = hammer2_igetv(ip, &error);
1903                 hammer2_inode_unlock(ip);
1904                 hammer2_inode_drop(ip);         /* from lookup */
1905
1906                 return error;
1907         }
1908
1909         /*
1910          * Otherwise we have to find the inode
1911          */
1912         xop = hammer2_xop_alloc(pmp->iroot, 0);
1913         xop->lhc = inum;
1914         hammer2_xop_start(&xop->head, &hammer2_lookup_desc);
1915         error = hammer2_xop_collect(&xop->head, 0);
1916
1917         if (error == 0)
1918                 ip = hammer2_inode_get(pmp, &xop->head, -1, -1);
1919         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1920
1921         if (ip) {
1922                 *vpp = hammer2_igetv(ip, &error);
1923                 hammer2_inode_unlock(ip);
1924         } else {
1925                 *vpp = NULL;
1926                 error = ENOENT;
1927         }
1928         return (error);
1929 }
1930
1931 static
1932 int
1933 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1934 {
1935         hammer2_pfs_t *pmp;
1936         struct vnode *vp;
1937         int error;
1938
1939         pmp = MPTOPMP(mp);
1940         if (pmp->iroot == NULL) {
1941                 kprintf("hammer2 (%s): no root inode\n",
1942                         mp->mnt_stat.f_mntfromname);
1943                 *vpp = NULL;
1944                 return EINVAL;
1945         }
1946
1947         error = 0;
1948         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1949
1950         while (pmp->inode_tid == 0) {
1951                 hammer2_xop_ipcluster_t *xop;
1952                 const hammer2_inode_meta_t *meta;
1953
1954                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1955                 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc);
1956                 error = hammer2_xop_collect(&xop->head, 0);
1957
1958                 if (error == 0) {
1959                         meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1960                         pmp->iroot->meta = *meta;
1961                         pmp->inode_tid = meta->pfs_inum + 1;
1962                         hammer2_xop_pdata(&xop->head);
1963                         /* meta invalid */
1964
1965                         if (pmp->inode_tid < HAMMER2_INODE_START)
1966                                 pmp->inode_tid = HAMMER2_INODE_START;
1967                         pmp->modify_tid =
1968                                 xop->head.cluster.focus->bref.modify_tid + 1;
1969 #if 0
1970                         kprintf("PFS: Starting inode %jd\n",
1971                                 (intmax_t)pmp->inode_tid);
1972                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1973                                 pmp->inode_tid, pmp->modify_tid);
1974 #endif
1975                         wakeup(&pmp->iroot);
1976
1977                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1978
1979                         /*
1980                          * Prime the mount info.
1981                          */
1982                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1983                         break;
1984                 }
1985
1986                 /*
1987                  * Loop, try again
1988                  */
1989                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1990                 hammer2_inode_unlock(pmp->iroot);
1991                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1992                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1993                 if (error == EINTR)
1994                         break;
1995         }
1996
1997         if (error) {
1998                 hammer2_inode_unlock(pmp->iroot);
1999                 *vpp = NULL;
2000         } else {
2001                 vp = hammer2_igetv(pmp->iroot, &error);
2002                 hammer2_inode_unlock(pmp->iroot);
2003                 *vpp = vp;
2004         }
2005
2006         return (error);
2007 }
2008
2009 /*
2010  * Filesystem status
2011  *
2012  * XXX incorporate ipdata->meta.inode_quota and data_quota
2013  */
2014 static
2015 int
2016 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
2017 {
2018         hammer2_pfs_t *pmp;
2019         hammer2_dev_t *hmp;
2020         hammer2_blockref_t bref;
2021         struct statfs tmp;
2022         int i;
2023
2024         /*
2025          * NOTE: iroot might not have validated the cluster yet.
2026          */
2027         pmp = MPTOPMP(mp);
2028
2029         bzero(&tmp, sizeof(tmp));
2030
2031         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2032                 hmp = pmp->pfs_hmps[i];
2033                 if (hmp == NULL)
2034                         continue;
2035                 if (pmp->iroot->cluster.array[i].chain)
2036                         bref = pmp->iroot->cluster.array[i].chain->bref;
2037                 else
2038                         bzero(&bref, sizeof(bref));
2039
2040                 tmp.f_files = bref.embed.stats.inode_count;
2041                 tmp.f_ffree = 0;
2042                 tmp.f_blocks = hmp->voldata.allocator_size /
2043                                mp->mnt_vstat.f_bsize;
2044                 tmp.f_bfree = hmp->voldata.allocator_free /
2045                               mp->mnt_vstat.f_bsize;
2046                 tmp.f_bavail = tmp.f_bfree;
2047
2048                 if (cred && cred->cr_uid != 0) {
2049                         uint64_t adj;
2050
2051                         /* 5% */
2052                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2053                         tmp.f_blocks -= adj;
2054                         tmp.f_bfree -= adj;
2055                         tmp.f_bavail -= adj;
2056                 }
2057
2058                 mp->mnt_stat.f_blocks = tmp.f_blocks;
2059                 mp->mnt_stat.f_bfree = tmp.f_bfree;
2060                 mp->mnt_stat.f_bavail = tmp.f_bavail;
2061                 mp->mnt_stat.f_files = tmp.f_files;
2062                 mp->mnt_stat.f_ffree = tmp.f_ffree;
2063
2064                 *sbp = mp->mnt_stat;
2065         }
2066         return (0);
2067 }
2068
2069 static
2070 int
2071 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2072 {
2073         hammer2_pfs_t *pmp;
2074         hammer2_dev_t *hmp;
2075         hammer2_blockref_t bref;
2076         struct statvfs tmp;
2077         int i;
2078
2079         /*
2080          * NOTE: iroot might not have validated the cluster yet.
2081          */
2082         pmp = MPTOPMP(mp);
2083         bzero(&tmp, sizeof(tmp));
2084
2085         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2086                 hmp = pmp->pfs_hmps[i];
2087                 if (hmp == NULL)
2088                         continue;
2089                 if (pmp->iroot->cluster.array[i].chain)
2090                         bref = pmp->iroot->cluster.array[i].chain->bref;
2091                 else
2092                         bzero(&bref, sizeof(bref));
2093
2094                 tmp.f_files = bref.embed.stats.inode_count;
2095                 tmp.f_ffree = 0;
2096                 tmp.f_blocks = hmp->voldata.allocator_size /
2097                                mp->mnt_vstat.f_bsize;
2098                 tmp.f_bfree = hmp->voldata.allocator_free /
2099                               mp->mnt_vstat.f_bsize;
2100                 tmp.f_bavail = tmp.f_bfree;
2101
2102                 if (cred && cred->cr_uid != 0) {
2103                         uint64_t adj;
2104
2105                         /* 5% */
2106                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2107                         tmp.f_blocks -= adj;
2108                         tmp.f_bfree -= adj;
2109                         tmp.f_bavail -= adj;
2110                 }
2111
2112                 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2113                 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2114                 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2115                 mp->mnt_vstat.f_files = tmp.f_files;
2116                 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2117
2118                 *sbp = mp->mnt_vstat;
2119         }
2120         return (0);
2121 }
2122
2123 /*
2124  * Mount-time recovery (RW mounts)
2125  *
2126  * Updates to the free block table are allowed to lag flushes by one
2127  * transaction.  In case of a crash, then on a fresh mount we must do an
2128  * incremental scan of the last committed transaction id and make sure that
2129  * all related blocks have been marked allocated.
2130  *
2131  * The super-root topology and each PFS has its own transaction id domain,
2132  * so we must track PFS boundary transitions.
2133  */
2134 struct hammer2_recovery_elm {
2135         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2136         hammer2_chain_t *chain;
2137         hammer2_tid_t sync_tid;
2138 };
2139
2140 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2141
2142 struct hammer2_recovery_info {
2143         struct hammer2_recovery_list list;
2144         hammer2_tid_t   mtid;
2145         int     depth;
2146 };
2147
2148 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2149                         hammer2_chain_t *parent,
2150                         struct hammer2_recovery_info *info,
2151                         hammer2_tid_t sync_tid);
2152
2153 #define HAMMER2_RECOVERY_MAXDEPTH       10
2154
2155 static
2156 int
2157 hammer2_recovery(hammer2_dev_t *hmp)
2158 {
2159         struct hammer2_recovery_info info;
2160         struct hammer2_recovery_elm *elm;
2161         hammer2_chain_t *parent;
2162         hammer2_tid_t sync_tid;
2163         hammer2_tid_t mirror_tid;
2164         int error;
2165
2166         hammer2_trans_init(hmp->spmp, 0);
2167
2168         sync_tid = hmp->voldata.freemap_tid;
2169         mirror_tid = hmp->voldata.mirror_tid;
2170
2171         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2172         if (sync_tid >= mirror_tid) {
2173                 kprintf(" no recovery needed\n");
2174         } else {
2175                 kprintf(" freemap recovery %016jx-%016jx\n",
2176                         sync_tid + 1, mirror_tid);
2177         }
2178
2179         TAILQ_INIT(&info.list);
2180         info.depth = 0;
2181         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2182         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2183         hammer2_chain_lookup_done(parent);
2184
2185         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2186                 TAILQ_REMOVE(&info.list, elm, entry);
2187                 parent = elm->chain;
2188                 sync_tid = elm->sync_tid;
2189                 kfree(elm, M_HAMMER2);
2190
2191                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2192                 error |= hammer2_recovery_scan(hmp, parent, &info,
2193                                               hmp->voldata.freemap_tid);
2194                 hammer2_chain_unlock(parent);
2195                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2196         }
2197
2198         hammer2_trans_done(hmp->spmp, 0);
2199
2200         return error;
2201 }
2202
2203 static
2204 int
2205 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2206                       struct hammer2_recovery_info *info,
2207                       hammer2_tid_t sync_tid)
2208 {
2209         const hammer2_inode_data_t *ripdata;
2210         hammer2_chain_t *chain;
2211         hammer2_blockref_t bref;
2212         int tmp_error;
2213         int rup_error;
2214         int error;
2215         int first;
2216
2217         /*
2218          * Adjust freemap to ensure that the block(s) are marked allocated.
2219          */
2220         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2221                 hammer2_freemap_adjust(hmp, &parent->bref,
2222                                        HAMMER2_FREEMAP_DORECOVER);
2223         }
2224
2225         /*
2226          * Check type for recursive scan
2227          */
2228         switch(parent->bref.type) {
2229         case HAMMER2_BREF_TYPE_VOLUME:
2230                 /* data already instantiated */
2231                 break;
2232         case HAMMER2_BREF_TYPE_INODE:
2233                 /*
2234                  * Must instantiate data for DIRECTDATA test and also
2235                  * for recursion.
2236                  */
2237                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2238                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2239                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2240                         /* not applicable to recovery scan */
2241                         hammer2_chain_unlock(parent);
2242                         return 0;
2243                 }
2244                 hammer2_chain_unlock(parent);
2245                 break;
2246         case HAMMER2_BREF_TYPE_INDIRECT:
2247                 /*
2248                  * Must instantiate data for recursion
2249                  */
2250                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2251                 hammer2_chain_unlock(parent);
2252                 break;
2253         case HAMMER2_BREF_TYPE_DIRENT:
2254         case HAMMER2_BREF_TYPE_DATA:
2255         case HAMMER2_BREF_TYPE_FREEMAP:
2256         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2257         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2258                 /* not applicable to recovery scan */
2259                 return 0;
2260                 break;
2261         default:
2262                 return HAMMER2_ERROR_BADBREF;
2263         }
2264
2265         /*
2266          * Defer operation if depth limit reached or if we are crossing a
2267          * PFS boundary.
2268          */
2269         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2270                 struct hammer2_recovery_elm *elm;
2271
2272                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2273                 elm->chain = parent;
2274                 elm->sync_tid = sync_tid;
2275                 hammer2_chain_ref(parent);
2276                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2277                 /* unlocked by caller */
2278
2279                 return(0);
2280         }
2281
2282
2283         /*
2284          * Recursive scan of the last flushed transaction only.  We are
2285          * doing this without pmp assignments so don't leave the chains
2286          * hanging around after we are done with them.
2287          *
2288          * error        Cumulative error this level only
2289          * rup_error    Cumulative error for recursion
2290          * tmp_error    Specific non-cumulative recursion error
2291          */
2292         chain = NULL;
2293         first = 1;
2294         rup_error = 0;
2295         error = 0;
2296
2297         for (;;) {
2298                 error |= hammer2_chain_scan(parent, &chain, &bref,
2299                                             &first,
2300                                             HAMMER2_LOOKUP_NODATA);
2301
2302                 /*
2303                  * Problem during scan or EOF
2304                  */
2305                 if (error)
2306                         break;
2307
2308                 /*
2309                  * If this is a leaf
2310                  */
2311                 if (chain == NULL) {
2312                         if (bref.mirror_tid > sync_tid) {
2313                                 hammer2_freemap_adjust(hmp, &bref,
2314                                                      HAMMER2_FREEMAP_DORECOVER);
2315                         }
2316                         continue;
2317                 }
2318
2319                 /*
2320                  * This may or may not be a recursive node.
2321                  */
2322                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2323                 if (bref.mirror_tid > sync_tid) {
2324                         ++info->depth;
2325                         tmp_error = hammer2_recovery_scan(hmp, chain,
2326                                                            info, sync_tid);
2327                         --info->depth;
2328                 } else {
2329                         tmp_error = 0;
2330                 }
2331
2332                 /*
2333                  * Flush the recovery at the PFS boundary to stage it for
2334                  * the final flush of the super-root topology.
2335                  */
2336                 if (tmp_error == 0 &&
2337                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2338                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2339                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2340                                              HAMMER2_FLUSH_ALL);
2341                 }
2342                 rup_error |= tmp_error;
2343         }
2344         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2345 }
2346
2347 /*
2348  * This fixes up an error introduced in earlier H2 implementations where
2349  * moving a PFS inode into an indirect block wound up causing the
2350  * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2351  */
2352 static
2353 int
2354 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2355 {
2356         const hammer2_inode_data_t *ripdata;
2357         hammer2_chain_t *parent;
2358         hammer2_chain_t *chain;
2359         hammer2_key_t key_next;
2360         hammer2_pfs_t *spmp;
2361         int error;
2362
2363         error = 0;
2364
2365         /*
2366          * Lookup mount point under the media-localized super-root.
2367          *
2368          * cluster->pmp will incorrectly point to spmp and must be fixed
2369          * up later on.
2370          */
2371         spmp = hmp->spmp;
2372         hammer2_inode_lock(spmp->iroot, 0);
2373         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2374         chain = hammer2_chain_lookup(&parent, &key_next,
2375                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2376                                          &error, 0);
2377         while (chain) {
2378                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2379                         continue;
2380                 if (chain->error) {
2381                         kprintf("I/O error scanning PFS labels\n");
2382                         error |= chain->error;
2383                 } else if ((chain->bref.flags &
2384                             HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2385                         int error2;
2386
2387                         ripdata = &chain->data->ipdata;
2388                         hammer2_trans_init(hmp->spmp, 0);
2389                         error2 = hammer2_chain_modify(chain,
2390                                                       chain->bref.modify_tid,
2391                                                       0, 0);
2392                         if (error2 == 0) {
2393                                 kprintf("hammer2: Correct mis-flagged PFS %s\n",
2394                                         ripdata->filename);
2395                                 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2396                         } else {
2397                                 error |= error2;
2398                         }
2399                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2400                                              HAMMER2_FLUSH_ALL);
2401                         hammer2_trans_done(hmp->spmp, 0);
2402                 }
2403                 chain = hammer2_chain_next(&parent, chain, &key_next,
2404                                            key_next, HAMMER2_KEY_MAX,
2405                                            &error, 0);
2406         }
2407         if (parent) {
2408                 hammer2_chain_unlock(parent);
2409                 hammer2_chain_drop(parent);
2410         }
2411         hammer2_inode_unlock(spmp->iroot);
2412
2413         return error;
2414 }
2415
2416 /*
2417  * Sync a mount point; this is called periodically on a per-mount basis from
2418  * the filesystem syncer, and whenever a user issues a sync.
2419  */
2420 int
2421 hammer2_vfs_sync(struct mount *mp, int waitfor)
2422 {
2423         int error;
2424
2425         error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor);
2426
2427         return error;
2428 }
2429
2430 /*
2431  * Because frontend operations lock vnodes before we get a chance to
2432  * lock the related inode, we can't just acquire a vnode lock without
2433  * risking a deadlock.  The frontend may be holding a vnode lock while
2434  * also blocked on our SYNCQ flag while trying to get the inode lock.
2435  *
2436  * To deal with this situation we can check the vnode lock situation
2437  * after locking the inode and perform a work-around.
2438  */
2439 int
2440 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor)
2441 {
2442         struct mount *mp;
2443         /*hammer2_xop_flush_t *xop;*/
2444         /*struct hammer2_sync_info info;*/
2445         hammer2_inode_t *ip;
2446         hammer2_depend_t *depend;
2447         hammer2_depend_t *depend_next;
2448         struct vnode *vp;
2449         uint32_t pass2;
2450         int error;
2451         int dorestart;
2452
2453         mp = pmp->mp;
2454
2455         /*
2456          * Move all inodes on sideq to syncq.  This will clear sideq.
2457          * This should represent all flushable inodes.  These inodes
2458          * will already have refs due to being on syncq or sideq.  We
2459          * must do this all at once with the spinlock held to ensure that
2460          * all inode dependencies are part of the same flush.
2461          *
2462          * We should be able to do this asynchronously from frontend
2463          * operations because we will be locking the inodes later on
2464          * to actually flush them, and that will partition any frontend
2465          * op using the same inode.  Either it has already locked the
2466          * inode and we will block, or it has not yet locked the inode
2467          * and it will block until we are finished flushing that inode.
2468          *
2469          * When restarting, only move the inodes flagged as PASS2 from
2470          * SIDEQ to SYNCQ.  PASS2 propagation by inode_lock4() and
2471          * inode_depend() are atomic with the spin-lock.
2472          */
2473         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2474 #ifdef HAMMER2_DEBUG_SYNC
2475         kprintf("FILESYSTEM SYNC BOUNDARY\n");
2476 #endif
2477         dorestart = 0;
2478
2479         /*
2480          * Move inodes from depq to syncq, releasing the related
2481          * depend structures.
2482          */
2483 restart:
2484 #ifdef HAMMER2_DEBUG_SYNC
2485         kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart);
2486 #endif
2487         hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/);
2488         hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN);
2489
2490         /*
2491          * Move inodes from depq to syncq.  When restarting, only depq's
2492          * marked pass2 are moved.
2493          */
2494         hammer2_spin_ex(&pmp->list_spin);
2495         depend_next = TAILQ_FIRST(&pmp->depq);
2496
2497         while ((depend = depend_next) != NULL) {
2498                 depend_next = TAILQ_NEXT(depend, entry);
2499                 if (dorestart && depend->pass2 == 0)
2500                         continue;
2501                 TAILQ_FOREACH(ip, &depend->sideq, entry) {
2502                         KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ);
2503                         atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ);
2504                         atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ);
2505                         ip->depend = NULL;
2506                 }
2507                 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry);
2508                 pmp->sideq_count -= depend->count;
2509                 depend->count = 0;
2510                 depend->pass2 = 0;
2511                 TAILQ_REMOVE(&pmp->depq, depend, entry);
2512         }
2513
2514         hammer2_spin_unex(&pmp->list_spin);
2515         hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/
2516                                       HAMMER2_TRANS_WAITING);
2517         dorestart = 0;
2518
2519         /*
2520          * sideq_count may have dropped enough to allow us to unstall
2521          * the frontend.
2522          */
2523         hammer2_pfs_memory_inc(pmp);
2524         hammer2_pfs_memory_wakeup(pmp);
2525
2526         /*
2527          * Now run through all inodes on syncq.
2528          *
2529          * Flush transactions only interlock with other flush transactions.
2530          * Any conflicting frontend operations will block on the inode, but
2531          * may hold a vnode lock while doing so.
2532          */
2533         hammer2_spin_ex(&pmp->list_spin);
2534         while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) {
2535                 /*
2536                  * Remove the inode from the SYNCQ, transfer the syncq ref
2537                  * to us.  We must clear SYNCQ to allow any potential
2538                  * front-end deadlock to proceed.  We must set PASS2 so
2539                  * the dependency code knows what to do.
2540                  */
2541                 pass2 = ip->flags;
2542                 cpu_ccfence();
2543                 if (atomic_cmpset_int(&ip->flags,
2544                               pass2,
2545                               (pass2 & ~(HAMMER2_INODE_SYNCQ |
2546                                          HAMMER2_INODE_SYNCQ_WAKEUP)) |
2547                               HAMMER2_INODE_SYNCQ_PASS2) == 0) {
2548                         continue;
2549                 }
2550                 TAILQ_REMOVE(&pmp->syncq, ip, entry);
2551                 hammer2_spin_unex(&pmp->list_spin);
2552                 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP)
2553                         wakeup(&ip->flags);
2554
2555                 /*
2556                  * Relock the inode, and we inherit a ref from the above.
2557                  * We will check for a race after we acquire the vnode.
2558                  */
2559                 hammer2_mtx_ex(&ip->lock);
2560
2561                 /*
2562                  * We need the vp in order to vfsync() dirty buffers, so if
2563                  * one isn't attached we can skip it.
2564                  *
2565                  * Ordering the inode lock and then the vnode lock has the
2566                  * potential to deadlock.  If we had left SYNCQ set that could
2567                  * also deadlock us against the frontend even if we don't hold
2568                  * any locks, but the latter is not a problem now since we
2569                  * cleared it.  igetv will temporarily release the inode lock
2570                  * in a safe manner to work-around the deadlock.
2571                  *
2572                  * Unfortunately it is still possible to deadlock when the
2573                  * frontend obtains multiple inode locks, because all the
2574                  * related vnodes are already locked (nor can the vnode locks
2575                  * be released and reacquired without messing up RECLAIM and
2576                  * INACTIVE sequencing).
2577                  *
2578                  * The solution for now is to move the vp back onto SIDEQ
2579                  * and set dorestart, which will restart the flush after we
2580                  * exhaust the current SYNCQ.  Note that additional
2581                  * dependencies may build up, so we definitely need to move
2582                  * the whole SIDEQ back to SYNCQ when we restart.
2583                  */
2584                 vp = ip->vp;
2585                 if (vp) {
2586                         if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) {
2587                                 /*
2588                                  * Failed to get the vnode, requeue the inode
2589                                  * (PASS2 is already set so it will be found
2590                                  * again on the restart).
2591                                  *
2592                                  * Then unlock, possibly sleep, and retry
2593                                  * later.  We sleep if PASS2 was *previously*
2594                                  * set, before we set it again above.
2595                                  */
2596                                 vp = NULL;
2597                                 dorestart = 1;
2598 #ifdef HAMMER2_DEBUG_SYNC
2599                                 kprintf("inum %ld (sync delayed by vnode)\n",
2600                                         (long)ip->meta.inum);
2601 #endif
2602                                 hammer2_inode_delayed_sideq(ip);
2603
2604                                 hammer2_mtx_unlock(&ip->lock);
2605                                 hammer2_inode_drop(ip);
2606
2607                                 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) {
2608                                         tsleep(&dorestart, 0, "h2syndel", 2);
2609                                 }
2610                                 hammer2_spin_ex(&pmp->list_spin);
2611                                 continue;
2612                         }
2613                 } else {
2614                         vp = NULL;
2615                 }
2616
2617                 /*
2618                  * If the inode wound up on a SIDEQ again it will already be
2619                  * prepped for another PASS2.  In this situation if we flush
2620                  * it now we will just wind up flushing it again in the same
2621                  * syncer run, so we might as well not flush it now.
2622                  */
2623                 if (ip->flags & HAMMER2_INODE_SIDEQ) {
2624                         hammer2_mtx_unlock(&ip->lock);
2625                         hammer2_inode_drop(ip);
2626                         if (vp)
2627                                 vput(vp);
2628                         dorestart = 1;
2629                         hammer2_spin_ex(&pmp->list_spin);
2630                         continue;
2631                 }
2632
2633                 /*
2634                  * Ok we have the inode exclusively locked and if vp is
2635                  * not NULL that will also be exclusively locked.  Do the
2636                  * meat of the flush.
2637                  *
2638                  * vp token needed for v_rbdirty_tree check / vclrisdirty
2639                  * sequencing.  Though we hold the vnode exclusively so
2640                  * we shouldn't need to hold the token also in this case.
2641                  */
2642                 if (vp) {
2643                         vfsync(vp, MNT_WAIT, 1, NULL, NULL);
2644                         bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */
2645                 }
2646
2647                 /*
2648                  * If the inode has not yet been inserted into the tree
2649                  * we must do so.  Then sync and flush it.  The flush should
2650                  * update the parent.
2651                  */
2652                 if (ip->flags & HAMMER2_INODE_DELETING) {
2653 #ifdef HAMMER2_DEBUG_SYNC
2654                         kprintf("inum %ld destroy\n", (long)ip->meta.inum);
2655 #endif
2656                         hammer2_inode_chain_des(ip);
2657                         atomic_add_long(&hammer2_iod_inode_deletes, 1);
2658                 } else if (ip->flags & HAMMER2_INODE_CREATING) {
2659 #ifdef HAMMER2_DEBUG_SYNC
2660                         kprintf("inum %ld insert\n", (long)ip->meta.inum);
2661 #endif
2662                         hammer2_inode_chain_ins(ip);
2663                         atomic_add_long(&hammer2_iod_inode_creates, 1);
2664                 }
2665 #ifdef HAMMER2_DEBUG_SYNC
2666                 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum);
2667 #endif
2668
2669                 /*
2670                  * Because I kinda messed up the design and index the inodes
2671                  * under the root inode, along side the directory entries,
2672                  * we can't flush the inode index under the iroot until the
2673                  * end.  If we do it now we might miss effects created by
2674                  * other inodes on the SYNCQ.
2675                  *
2676                  * Do a normal (non-FSSYNC) flush instead, which allows the
2677                  * vnode code to work the same.  We don't want to force iroot
2678                  * back onto the SIDEQ, and we also don't want the flush code
2679                  * to update pfs_iroot_blocksets until the final flush later.
2680                  *
2681                  * XXX at the moment this will likely result in a double-flush
2682                  * of the iroot chain.
2683                  */
2684                 hammer2_inode_chain_sync(ip);
2685                 if (ip == pmp->iroot) {
2686                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP);
2687                 } else {
2688                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2689                                                       HAMMER2_XOP_FSSYNC);
2690                 }
2691                 if (vp) {
2692                         lwkt_gettoken(&vp->v_token);
2693                         if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2694                                           HAMMER2_INODE_RESIZED |
2695                                           HAMMER2_INODE_DIRTYDATA)) == 0 &&
2696                             RB_EMPTY(&vp->v_rbdirty_tree) &&
2697                             !bio_track_active(&vp->v_track_write)) {
2698                                 vclrisdirty(vp);
2699                         } else {
2700                                 hammer2_inode_delayed_sideq(ip);
2701                         }
2702                         lwkt_reltoken(&vp->v_token);
2703                         vput(vp);
2704                         vp = NULL;      /* safety */
2705                 }
2706                 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2);
2707                 hammer2_inode_unlock(ip);       /* unlock+drop */
2708                 /* ip pointer invalid */
2709
2710                 /*
2711                  * If the inode got dirted after we dropped our locks,
2712                  * it will have already been moved back to the SIDEQ.
2713                  */
2714                 hammer2_spin_ex(&pmp->list_spin);
2715         }
2716         hammer2_spin_unex(&pmp->list_spin);
2717         if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) {
2718 #ifdef HAMMER2_DEBUG_SYNC
2719                 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n");
2720                 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/
2721 #endif
2722                 dorestart = 1;
2723                 goto restart;
2724         }
2725 #ifdef HAMMER2_DEBUG_SYNC
2726         kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n");
2727         /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/
2728 #endif
2729
2730         /*
2731          * We have to flush the PFS root last, even if it does not appear to
2732          * be dirty, because all the inodes in the PFS are indexed under it.
2733          * The normal flushing of iroot above would only occur if directory
2734          * entries under the root were changed.
2735          *
2736          * Specifying VOLHDR will cause an additionl flush of hmp->spmp
2737          * for the media making up the cluster.
2738          */
2739         if ((ip = pmp->iroot) != NULL) {
2740                 hammer2_inode_ref(ip);
2741                 hammer2_mtx_ex(&ip->lock);
2742                 hammer2_inode_chain_sync(ip);
2743                 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2744                                               HAMMER2_XOP_FSSYNC |
2745                                               HAMMER2_XOP_VOLHDR);
2746                 hammer2_inode_unlock(ip);       /* unlock+drop */
2747         }
2748 #ifdef HAMMER2_DEBUG_SYNC
2749         kprintf("FILESYSTEM SYNC STAGE 2 DONE\n");
2750 #endif
2751
2752         /*
2753          * device bioq sync
2754          */
2755         hammer2_bioq_sync(pmp);
2756
2757 #if 0
2758         info.pass = 1;
2759         info.waitfor = MNT_WAIT;
2760         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2761
2762         info.pass = 2;
2763         info.waitfor = MNT_WAIT;
2764         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2765 #endif
2766 #if 0
2767         /*
2768          * Generally speaking we now want to flush the media topology from
2769          * the iroot through to the inodes.  The flush stops at any inode
2770          * boundary, which allows the frontend to continue running concurrent
2771          * modifying operations on inodes (including kernel flushes of
2772          * buffers) without interfering with the main sync.
2773          *
2774          * Use the XOP interface to concurrently flush all nodes to
2775          * synchronize the PFSROOT subtopology to the media.  A standard
2776          * end-of-scan ENOENT error indicates cluster sufficiency.
2777          *
2778          * Note that this flush will not be visible on crash recovery until
2779          * we flush the super-root topology in the next loop.
2780          *
2781          * XXX For now wait for all flushes to complete.
2782          */
2783         if (mp && (ip = pmp->iroot) != NULL) {
2784                 /*
2785                  * If unmounting try to flush everything including any
2786                  * sub-trees under inodes, just in case there is dangling
2787                  * modified data, as a safety.  Otherwise just flush up to
2788                  * the inodes in this stage.
2789                  */
2790                 kprintf("MP & IROOT\n");
2791 #ifdef HAMMER2_DEBUG_SYNC
2792                 kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n");
2793 #endif
2794                 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2795                         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2796                                                     HAMMER2_XOP_VOLHDR |
2797                                                     HAMMER2_XOP_FSSYNC |
2798                                                     HAMMER2_XOP_INODE_STOP);
2799                 } else {
2800                         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2801                                                     HAMMER2_XOP_INODE_STOP |
2802                                                     HAMMER2_XOP_VOLHDR |
2803                                                     HAMMER2_XOP_FSSYNC |
2804                                                     HAMMER2_XOP_INODE_STOP);
2805                 }
2806                 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc);
2807                 error = hammer2_xop_collect(&xop->head,
2808                                             HAMMER2_XOP_COLLECT_WAITALL);
2809                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2810 #ifdef HAMMER2_DEBUG_SYNC
2811                 kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n");
2812 #endif
2813                 if (error == HAMMER2_ERROR_ENOENT)
2814                         error = 0;
2815                 else
2816                         error = hammer2_error_to_errno(error);
2817         } else {
2818                 error = 0;
2819         }
2820 #endif
2821         error = 0;      /* XXX */
2822         hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH);
2823
2824         return (error);
2825 }
2826
2827 static
2828 int
2829 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2830 {
2831         hammer2_inode_t *ip;
2832
2833         KKASSERT(MAXFIDSZ >= 16);
2834         ip = VTOI(vp);
2835         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2836         fhp->fid_ext = 0;
2837         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2838         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2839
2840         return 0;
2841 }
2842
2843 static
2844 int
2845 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2846                struct fid *fhp, struct vnode **vpp)
2847 {
2848         hammer2_pfs_t *pmp;
2849         hammer2_tid_t inum;
2850         int error;
2851
2852         pmp = MPTOPMP(mp);
2853         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2854         if (vpp) {
2855                 if (inum == 1)
2856                         error = hammer2_vfs_root(mp, vpp);
2857                 else
2858                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2859         } else {
2860                 error = 0;
2861         }
2862         if (error)
2863                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2864         return error;
2865 }
2866
2867 static
2868 int
2869 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2870                  int *exflagsp, struct ucred **credanonp)
2871 {
2872         hammer2_pfs_t *pmp;
2873         struct netcred *np;
2874         int error;
2875
2876         pmp = MPTOPMP(mp);
2877         np = vfs_export_lookup(mp, &pmp->export, nam);
2878         if (np) {
2879                 *exflagsp = np->netc_exflags;
2880                 *credanonp = &np->netc_anon;
2881                 error = 0;
2882         } else {
2883                 error = EACCES;
2884         }
2885         return error;
2886 }
2887
2888 /*
2889  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2890  * header into the HMP
2891  *
2892  * XXX read four volhdrs and use the one with the highest TID whos CRC
2893  *     matches.
2894  *
2895  * XXX check iCRCs.
2896  *
2897  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2898  *     nonexistant locations.
2899  *
2900  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2901  */
2902 static
2903 int
2904 hammer2_install_volume_header(hammer2_dev_t *hmp)
2905 {
2906         hammer2_volume_data_t *vd;
2907         struct buf *bp;
2908         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2909         int error_reported;
2910         int error;
2911         int valid;
2912         int i;
2913
2914         error_reported = 0;
2915         error = 0;
2916         valid = 0;
2917         bp = NULL;
2918
2919         /*
2920          * There are up to 4 copies of the volume header (syncs iterate
2921          * between them so there is no single master).  We don't trust the
2922          * volu_size field so we don't know precisely how large the filesystem
2923          * is, so depend on the OS to return an error if we go beyond the
2924          * block device's EOF.
2925          */
2926         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2927                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2928                               HAMMER2_VOLUME_BYTES, &bp);
2929                 if (error) {
2930                         brelse(bp);
2931                         bp = NULL;
2932                         continue;
2933                 }
2934
2935                 vd = (struct hammer2_volume_data *) bp->b_data;
2936                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2937                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2938                         brelse(bp);
2939                         bp = NULL;
2940                         continue;
2941                 }
2942
2943                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2944                         /* XXX: Reversed-endianness filesystem */
2945                         kprintf("hammer2: reverse-endian filesystem detected");
2946                         brelse(bp);
2947                         bp = NULL;
2948                         continue;
2949                 }
2950
2951                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2952                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2953                                       HAMMER2_VOLUME_ICRC0_SIZE);
2954                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2955                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2956                                        HAMMER2_VOLUME_ICRC1_SIZE);
2957                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2958                         kprintf("hammer2 volume header crc "
2959                                 "mismatch copy #%d %08x/%08x\n",
2960                                 i, crc0, crc);
2961                         error_reported = 1;
2962                         brelse(bp);
2963                         bp = NULL;
2964                         continue;
2965                 }
2966                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2967                         valid = 1;
2968                         hmp->voldata = *vd;
2969                         hmp->volhdrno = i;
2970                 }
2971                 brelse(bp);
2972                 bp = NULL;
2973         }
2974         if (valid) {
2975                 hmp->volsync = hmp->voldata;
2976                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2977                 error = 0;
2978                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2979                         kprintf("hammer2: using volume header #%d\n",
2980                                 hmp->volhdrno);
2981                 }
2982         } else {
2983                 error = EINVAL;
2984                 kprintf("hammer2: no valid volume headers found!\n");
2985         }
2986         return (error);
2987 }
2988
2989 /*
2990  * This handles hysteresis on regular file flushes.  Because the BIOs are
2991  * routed to a thread it is possible for an excessive number to build up
2992  * and cause long front-end stalls long before the runningbuffspace limit
2993  * is hit, so we implement hammer2_flush_pipe to control the
2994  * hysteresis.
2995  *
2996  * This is a particular problem when compression is used.
2997  */
2998 void
2999 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
3000 {
3001         atomic_add_int(&pmp->count_lwinprog, 1);
3002 }
3003
3004 void
3005 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
3006 {
3007         int lwinprog;
3008
3009         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
3010         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
3011             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
3012                 atomic_clear_int(&pmp->count_lwinprog,
3013                                  HAMMER2_LWINPROG_WAITING);
3014                 wakeup(&pmp->count_lwinprog);
3015         }
3016         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
3017             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
3018                 atomic_clear_int(&pmp->count_lwinprog,
3019                                  HAMMER2_LWINPROG_WAITING0);
3020                 wakeup(&pmp->count_lwinprog);
3021         }
3022 }
3023
3024 void
3025 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
3026 {
3027         int lwinprog;
3028         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
3029                                     HAMMER2_LWINPROG_WAITING0;
3030
3031         for (;;) {
3032                 lwinprog = pmp->count_lwinprog;
3033                 cpu_ccfence();
3034                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3035                         break;
3036                 tsleep_interlock(&pmp->count_lwinprog, 0);
3037                 atomic_set_int(&pmp->count_lwinprog, lwflag);
3038                 lwinprog = pmp->count_lwinprog;
3039                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3040                         break;
3041                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
3042         }
3043 }
3044
3045 /*
3046  * It is possible for an excessive number of dirty chains or dirty inodes
3047  * to build up.  When this occurs we start an asynchronous filesystem sync.
3048  * If the level continues to build up, we stall, waiting for it to drop,
3049  * with some hysteresis.
3050  *
3051  * We limit the stall to two seconds per call.
3052  *
3053  * This relies on the kernel calling hammer2_vfs_modifying() prior to
3054  * obtaining any vnode locks before making a modifying VOP call.
3055  */
3056 static void
3057 hammer2_vfs_modifying(struct mount *mp)
3058 {
3059         hammer2_pfs_memory_wait(MPTOPMP(mp));
3060 }
3061
3062 /*
3063  * Initiate an asynchronous filesystem sync and, with hysteresis,
3064  * stall if the internal data structure count becomes too bloated.
3065  */
3066 void
3067 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
3068 {
3069         uint32_t waiting;
3070         int loops;
3071
3072         if (pmp == NULL || pmp->mp == NULL)
3073                 return;
3074
3075         for (loops = 0; loops < 2; ++loops) {
3076                 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK;
3077                 cpu_ccfence();
3078
3079                 /*
3080                  * Start the syncer running at 1/2 the limit
3081                  */
3082                 if (waiting > hammer2_limit_dirty_chains / 2 ||
3083                     pmp->sideq_count > hammer2_limit_dirty_inodes / 2) {
3084                         trigger_syncer(pmp->mp);
3085                 }
3086
3087                 /*
3088                  * Stall at the limit waiting for the counts to drop.
3089                  * This code will typically be woken up once the count
3090                  * drops below 3/4 the limit, or in one second.
3091                  */
3092                 if (waiting < hammer2_limit_dirty_chains &&
3093                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
3094                         break;
3095                 }
3096                 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
3097                 atomic_set_int(&pmp->inmem_dirty_chains,
3098                                HAMMER2_DIRTYCHAIN_WAITING);
3099                 if (waiting < hammer2_limit_dirty_chains &&
3100                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
3101                         break;
3102                 }
3103                 trigger_syncer(pmp->mp);
3104                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED, "h2memw", hz);
3105 #if 0
3106                 limit = pmp->mp->mnt_nvnodelistsize / 10;
3107                 if (limit < hammer2_limit_dirty_chains)
3108                         limit = hammer2_limit_dirty_chains;
3109                 if (limit < 1000)
3110                         limit = 1000;
3111 #endif
3112         }
3113 }
3114
3115 void
3116 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
3117 {
3118         if (pmp) {
3119                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
3120         }
3121 }
3122
3123 void
3124 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
3125 {
3126         uint32_t waiting;
3127
3128         if (pmp) {
3129                 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
3130                 /* don't need --waiting to test flag */
3131
3132                 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) &&
3133                     (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <=
3134                     hammer2_limit_dirty_chains * 2 / 3 &&
3135                     pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) {
3136                         atomic_clear_int(&pmp->inmem_dirty_chains,
3137                                          HAMMER2_DIRTYCHAIN_WAITING);
3138                         wakeup(&pmp->inmem_dirty_chains);
3139                 }
3140         }
3141 }
3142
3143 /*
3144  * Returns 0 if the filesystem has tons of free space
3145  * Returns 1 if the filesystem has less than 10% remaining
3146  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
3147  */
3148 int
3149 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
3150 {
3151         hammer2_pfs_t *pmp;
3152         hammer2_dev_t *hmp;
3153         hammer2_off_t free_reserved;
3154         hammer2_off_t free_nominal;
3155         int i;
3156
3157         pmp = ip->pmp;
3158
3159         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
3160                 free_reserved = HAMMER2_SEGSIZE;
3161                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
3162                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
3163                         hmp = pmp->pfs_hmps[i];
3164                         if (hmp == NULL)
3165                                 continue;
3166                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
3167                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
3168                                 continue;
3169
3170                         if (free_nominal > hmp->voldata.allocator_free)
3171                                 free_nominal = hmp->voldata.allocator_free;
3172                         if (free_reserved < hmp->free_reserved)
3173                                 free_reserved = hmp->free_reserved;
3174                 }
3175
3176                 /*
3177                  * SMP races ok
3178                  */
3179                 pmp->free_reserved = free_reserved;
3180                 pmp->free_nominal = free_nominal;
3181                 pmp->free_ticks = ticks;
3182         } else {
3183                 free_reserved = pmp->free_reserved;
3184                 free_nominal = pmp->free_nominal;
3185         }
3186         if (cred && cred->cr_uid != 0) {
3187                 if ((int64_t)(free_nominal - bytes) <
3188                     (int64_t)free_reserved) {
3189                         return 2;
3190                 }
3191         } else {
3192                 if ((int64_t)(free_nominal - bytes) <
3193                     (int64_t)free_reserved / 2) {
3194                         return 2;
3195                 }
3196         }
3197         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
3198                 return 1;
3199         return 0;
3200 }
3201
3202 /*
3203  * Debugging
3204  */
3205 void
3206 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
3207                    u_int flags)
3208 {
3209         hammer2_chain_t *scan;
3210         hammer2_chain_t *parent;
3211
3212         --*countp;
3213         if (*countp == 0) {
3214                 kprintf("%*.*s...\n", tab, tab, "");
3215                 return;
3216         }
3217         if (*countp < 0)
3218                 return;
3219         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
3220                 tab, tab, "", pfx,
3221                 chain, chain->bref.type,
3222                 chain->bref.key, chain->bref.keybits,
3223                 chain->bref.mirror_tid);
3224
3225         kprintf("%*.*s      [%08x] (%s) refs=%d",
3226                 tab, tab, "",
3227                 chain->flags,
3228                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3229                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
3230                 chain->refs);
3231
3232         parent = chain->parent;
3233         if (parent)
3234                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
3235                         tab, tab, "",
3236                         parent, parent->flags, parent->refs);
3237         if (RB_EMPTY(&chain->core.rbtree)) {
3238                 kprintf("\n");
3239         } else {
3240                 kprintf(" {\n");
3241                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3242                         if ((scan->flags & flags) || flags == (u_int)-1) {
3243                                 hammer2_dump_chain(scan, tab + 4, countp, 'a',
3244                                                    flags);
3245                         }
3246                 }
3247                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3248                         kprintf("%*.*s}(%s)\n", tab, tab, "",
3249                                 chain->data->ipdata.filename);
3250                 else
3251                         kprintf("%*.*s}\n", tab, tab, "");
3252         }
3253 }