sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2017 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72 };
  73
  74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  75 static struct hammer2_mntlist hammer2_mntlist;
  76
  77 struct hammer2_pfslist hammer2_pfslist;
  78 struct lock hammer2_mntlk;
  79
  80 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
  81 int hammer2_debug;
  82 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
  83 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
  84 int hammer2_dedup_enable = 1;
  85 int hammer2_always_compress = 0;        /* always try to compress */
  86 int hammer2_inval_enable = 0;
  87 int hammer2_flush_pipe = 100;
  88 int hammer2_synchronous_flush = 1;
  89 int hammer2_dio_count;
  90 int hammer2_limit_dio = 256;
  91 int hammer2_bulkfree_tps = 5000;
  92 long hammer2_chain_allocs;
  93 long hammer2_chain_frees;
  94 long hammer2_limit_dirty_chains;
  95 long hammer2_count_modified_chains;
  96 long hammer2_iod_invals;
  97 long hammer2_iod_file_read;
  98 long hammer2_iod_meta_read;
  99 long hammer2_iod_indr_read;
 100 long hammer2_iod_fmap_read;
 101 long hammer2_iod_volu_read;
 102 long hammer2_iod_file_write;
 103 long hammer2_iod_file_wembed;
 104 long hammer2_iod_file_wzero;
 105 long hammer2_iod_file_wdedup;
 106 long hammer2_iod_meta_write;
 107 long hammer2_iod_indr_write;
 108 long hammer2_iod_fmap_write;
 109 long hammer2_iod_volu_write;
 110
 111 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 112 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 113                 "Buffer used for compression.");
 114
 115 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 116 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 117                 "Buffer used for decompression.");
 118
 119 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 120
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
 122            &hammer2_supported_version, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 124            &hammer2_debug, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 126            &hammer2_cluster_meta_read, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 128            &hammer2_cluster_data_read, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 130            &hammer2_dedup_enable, 0, "");
 131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 132            &hammer2_always_compress, 0, "");
 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 134            &hammer2_inval_enable, 0, "");
 135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 136            &hammer2_flush_pipe, 0, "");
 137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 138            &hammer2_synchronous_flush, 0, "");
 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
 140            &hammer2_bulkfree_tps, 0, "");
 141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 142            &hammer2_chain_allocs, 0, "");
 143 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 144            &hammer2_chain_frees, 0, "");
 145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 146            &hammer2_limit_dirty_chains, 0, "");
 147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 148            &hammer2_count_modified_chains, 0, "");
 149 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 150            &hammer2_dio_count, 0, "");
 151 SYSCTL_INT(_vfs_hammer2, OID_AUTO, limit_dio, CTLFLAG_RW,
 152            &hammer2_limit_dio, 0, "");
 153
 154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 155            &hammer2_iod_invals, 0, "");
 156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 157            &hammer2_iod_file_read, 0, "");
 158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 159            &hammer2_iod_meta_read, 0, "");
 160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 161            &hammer2_iod_indr_read, 0, "");
 162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 163            &hammer2_iod_fmap_read, 0, "");
 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 165            &hammer2_iod_volu_read, 0, "");
 166
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 168            &hammer2_iod_file_write, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 170            &hammer2_iod_file_wembed, 0, "");
 171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 172            &hammer2_iod_file_wzero, 0, "");
 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 174            &hammer2_iod_file_wdedup, 0, "");
 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 176            &hammer2_iod_meta_write, 0, "");
 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 178            &hammer2_iod_indr_write, 0, "");
 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 180            &hammer2_iod_fmap_write, 0, "");
 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 182            &hammer2_iod_volu_write, 0, "");
 183
 184 long hammer2_check_icrc32;
 185 long hammer2_check_xxhash64;
 186 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 187            &hammer2_check_icrc32, 0, "");
 188 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 189            &hammer2_check_xxhash64, 0, "");
 190
 191 static int hammer2_vfs_init(struct vfsconf *conf);
 192 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 193 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 194                                 struct ucred *cred);
 195 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 196                                 struct vnode *, struct ucred *);
 197 static int hammer2_recovery(hammer2_dev_t *hmp);
 198 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 199 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 200 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 201                                 struct ucred *cred);
 202 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 203                                 struct ucred *cred);
 204 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 205                                 struct fid *fhp, struct vnode **vpp);
 206 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 207 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 208                                 int *exflagsp, struct ucred **credanonp);
 209
 210 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 211 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 212
 213 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 214
 215 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 216 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 217                                 hammer2_dev_t *hmp);
 218
 219 /*
 220  * HAMMER2 vfs operations.
 221  */
 222 static struct vfsops hammer2_vfsops = {
 223         .vfs_init       = hammer2_vfs_init,
 224         .vfs_uninit     = hammer2_vfs_uninit,
 225         .vfs_sync       = hammer2_vfs_sync,
 226         .vfs_mount      = hammer2_vfs_mount,
 227         .vfs_unmount    = hammer2_vfs_unmount,
 228         .vfs_root       = hammer2_vfs_root,
 229         .vfs_statfs     = hammer2_vfs_statfs,
 230         .vfs_statvfs    = hammer2_vfs_statvfs,
 231         .vfs_vget       = hammer2_vfs_vget,
 232         .vfs_vptofh     = hammer2_vfs_vptofh,
 233         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 234         .vfs_checkexp   = hammer2_vfs_checkexp
 235 };
 236
 237 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 238
 239 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 240 MODULE_VERSION(hammer2, 1);
 241
 242 static
 243 int
 244 hammer2_vfs_init(struct vfsconf *conf)
 245 {
 246         static struct objcache_malloc_args margs_read;
 247         static struct objcache_malloc_args margs_write;
 248         static struct objcache_malloc_args margs_vop;
 249
 250         int error;
 251
 252         error = 0;
 253
 254         /*
 255          * A large DIO cache is needed to retain dedup enablement masks.
 256          * The bulkfree code clears related masks as part of the disk block
 257          * recycling algorithm, preventing it from being used for a later
 258          * dedup.
 259          *
 260          * NOTE: A large buffer cache can actually interfere with dedup
 261          *       operation because we dedup based on media physical buffers
 262          *       and not logical buffers.  Try to make the DIO chace large
 263          *       enough to avoid this problem, but also cap it.
 264          */
 265         hammer2_limit_dio = nbuf * 2;
 266         if (hammer2_limit_dio > 100000)
 267                 hammer2_limit_dio = 100000;
 268
 269         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 270                 error = EINVAL;
 271         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 272                 error = EINVAL;
 273         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 274                 error = EINVAL;
 275
 276         if (error)
 277                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 278
 279         margs_read.objsize = 65536;
 280         margs_read.mtype = M_HAMMER2_DEBUFFER;
 281
 282         margs_write.objsize = 32768;
 283         margs_write.mtype = M_HAMMER2_CBUFFER;
 284
 285         margs_vop.objsize = sizeof(hammer2_xop_t);
 286         margs_vop.mtype = M_HAMMER2;
 287
 288         /*
 289          * Note thaht for the XOPS cache we want backing store allocations
 290          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 291          * confusion), so use the backing store function that does it.  This
 292          * means that initial XOPS objects are zerod but REUSED objects are
 293          * not.  So we are responsible for cleaning the object up sufficiently
 294          * for our needs before objcache_put()ing it back (typically just the
 295          * FIFO indices).
 296          */
 297         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 298                                 0, 1, NULL, NULL, NULL,
 299                                 objcache_malloc_alloc,
 300                                 objcache_malloc_free,
 301                                 &margs_read);
 302         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 303                                 0, 1, NULL, NULL, NULL,
 304                                 objcache_malloc_alloc,
 305                                 objcache_malloc_free,
 306                                 &margs_write);
 307         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 308                                 0, 1, NULL, NULL, NULL,
 309                                 objcache_malloc_alloc_zero,
 310                                 objcache_malloc_free,
 311                                 &margs_vop);
 312
 313
 314         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 315         TAILQ_INIT(&hammer2_mntlist);
 316         TAILQ_INIT(&hammer2_pfslist);
 317
 318         hammer2_limit_dirty_chains = maxvnodes / 10;
 319         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 320                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 321
 322         return (error);
 323 }
 324
 325 static
 326 int
 327 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 328 {
 329         objcache_destroy(cache_buffer_read);
 330         objcache_destroy(cache_buffer_write);
 331         objcache_destroy(cache_xops);
 332         return 0;
 333 }
 334
 335 /*
 336  * Core PFS allocator.  Used to allocate or reference the pmp structure
 337  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 338  * The pmp can be passed in or loaded by this function using the chain and
 339  * inode data.
 340  *
 341  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 342  * transactions.  Note that synchronization does not use this field.
 343  * (typically frontend operations and synchronization cannot run on the
 344  * same PFS node at the same time).
 345  *
 346  * XXX check locking
 347  */
 348 hammer2_pfs_t *
 349 hammer2_pfsalloc(hammer2_chain_t *chain,
 350                  const hammer2_inode_data_t *ripdata,
 351                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 352 {
 353         hammer2_pfs_t *pmp;
 354         hammer2_inode_t *iroot;
 355         int count;
 356         int i;
 357         int j;
 358
 359         pmp = NULL;
 360
 361         /*
 362          * Locate or create the PFS based on the cluster id.  If ripdata
 363          * is NULL this is a spmp which is unique and is always allocated.
 364          *
 365          * If the device is mounted in local mode all PFSs are considered
 366          * independent and not part of any cluster (for debugging only).
 367          */
 368         if (ripdata) {
 369                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 370                         if (force_local != pmp->force_local)
 371                                 continue;
 372                         if (force_local == NULL &&
 373                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 374                                  sizeof(pmp->pfs_clid)) == 0) {
 375                                         break;
 376                         } else if (force_local && pmp->pfs_names[0] &&
 377                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 378                                         break;
 379                         }
 380                 }
 381         }
 382
 383         if (pmp == NULL) {
 384                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 385                 pmp->force_local = force_local;
 386                 hammer2_trans_manage_init(pmp);
 387                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 388                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 389                 lockinit(&pmp->lock, "pfslk", 0, 0);
 390                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 391                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 392                 spin_init(&pmp->xop_spin, "h2xop");
 393                 spin_init(&pmp->lru_spin, "h2lru");
 394                 RB_INIT(&pmp->inum_tree);
 395                 TAILQ_INIT(&pmp->sideq);
 396                 TAILQ_INIT(&pmp->lru_list);
 397                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 398
 399                 /*
 400                  * Distribute backend operations to threads
 401                  */
 402                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 403                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 404
 405                 /*
 406                  * Save the last media transaction id for the flusher.  Set
 407                  * initial
 408                  */
 409                 if (ripdata)
 410                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 411                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 412
 413                 /*
 414                  * The synchronization thread may start too early, make
 415                  * sure it stays frozen until we are ready to let it go.
 416                  * XXX
 417                  */
 418                 /*
 419                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 420                                          HAMMER2_THREAD_REMASTER;
 421                 */
 422         }
 423
 424         /*
 425          * Create the PFS's root inode and any missing XOP helper threads.
 426          */
 427         if ((iroot = pmp->iroot) == NULL) {
 428                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 429                 if (ripdata)
 430                         iroot->meta = ripdata->meta;
 431                 pmp->iroot = iroot;
 432                 hammer2_inode_ref(iroot);
 433                 hammer2_inode_unlock(iroot);
 434         }
 435
 436         /*
 437          * Stop here if no chain is passed in.
 438          */
 439         if (chain == NULL)
 440                 goto done;
 441
 442         /*
 443          * When a chain is passed in we must add it to the PFS's root
 444          * inode, update pmp->pfs_types[], and update the syncronization
 445          * threads.
 446          *
 447          * When forcing local mode, mark the PFS as a MASTER regardless.
 448          *
 449          * At the moment empty spots can develop due to removals or failures.
 450          * Ultimately we want to re-fill these spots but doing so might
 451          * confused running code. XXX
 452          */
 453         hammer2_inode_ref(iroot);
 454         hammer2_mtx_ex(&iroot->lock);
 455         j = iroot->cluster.nchains;
 456
 457         if (j == HAMMER2_MAXCLUSTER) {
 458                 kprintf("hammer2_mount: cluster full!\n");
 459                 /* XXX fatal error? */
 460         } else {
 461                 KKASSERT(chain->pmp == NULL);
 462                 chain->pmp = pmp;
 463                 hammer2_chain_ref(chain);
 464                 iroot->cluster.array[j].chain = chain;
 465                 if (force_local)
 466                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 467                 else
 468                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 469                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 470                 pmp->pfs_hmps[j] = chain->hmp;
 471
 472                 /*
 473                  * If the PFS is already mounted we must account
 474                  * for the mount_count here.
 475                  */
 476                 if (pmp->mp)
 477                         ++chain->hmp->mount_count;
 478
 479                 /*
 480                  * May have to fixup dirty chain tracking.  Previous
 481                  * pmp was NULL so nothing to undo.
 482                  */
 483                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 484                         hammer2_pfs_memory_inc(pmp);
 485                 ++j;
 486         }
 487         iroot->cluster.nchains = j;
 488
 489         /*
 490          * Update nmasters from any PFS inode which is part of the cluster.
 491          * It is possible that this will result in a value which is too
 492          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 493          * override this value later on.
 494          *
 495          * (This informs us of masters that might not currently be
 496          *  discoverable by this mount).
 497          */
 498         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 499                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 500         }
 501
 502         /*
 503          * Count visible masters.  Masters are usually added with
 504          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 505          * are more (XXX and must update the master inodes).
 506          */
 507         count = 0;
 508         for (i = 0; i < iroot->cluster.nchains; ++i) {
 509                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 510                         ++count;
 511         }
 512         if (pmp->pfs_nmasters < count)
 513                 pmp->pfs_nmasters = count;
 514
 515         /*
 516          * Create missing synchronization and support threads.
 517          *
 518          * Single-node masters (including snapshots) have nothing to
 519          * synchronize and do not require this thread.
 520          *
 521          * Multi-node masters or any number of soft masters, slaves, copy,
 522          * or other PFS types need the thread.
 523          *
 524          * Each thread is responsible for its particular cluster index.
 525          * We use independent threads so stalls or mismatches related to
 526          * any given target do not affect other targets.
 527          */
 528         for (i = 0; i < iroot->cluster.nchains; ++i) {
 529                 /*
 530                  * Single-node masters (including snapshots) have nothing
 531                  * to synchronize and will make direct xops support calls,
 532                  * thus they do not require this thread.
 533                  *
 534                  * Note that there can be thousands of snapshots.  We do not
 535                  * want to create thousands of threads.
 536                  */
 537                 if (pmp->pfs_nmasters <= 1 &&
 538                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 539                         continue;
 540                 }
 541
 542                 /*
 543                  * Sync support thread
 544                  */
 545                 if (pmp->sync_thrs[i].td == NULL) {
 546                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 547                                            "h2nod", i, -1,
 548                                            hammer2_primary_sync_thread);
 549                 }
 550         }
 551
 552         /*
 553          * Create missing Xop threads
 554          *
 555          * NOTE: We create helper threads for all mounted PFSs or any
 556          *       PFSs with 2+ nodes (so the sync thread can update them,
 557          *       even if not mounted).
 558          */
 559         if (pmp->mp || iroot->cluster.nchains >= 2)
 560                 hammer2_xop_helper_create(pmp);
 561
 562         hammer2_mtx_unlock(&iroot->lock);
 563         hammer2_inode_drop(iroot);
 564 done:
 565         return pmp;
 566 }
 567
 568 /*
 569  * Deallocate an element of a probed PFS.  If destroying and this is a
 570  * MASTER, adjust nmasters.
 571  *
 572  * This function does not physically destroy the PFS element in its device
 573  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 574  */
 575 void
 576 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 577 {
 578         hammer2_inode_t *iroot;
 579         hammer2_chain_t *chain;
 580         int j;
 581
 582         /*
 583          * Cleanup our reference on iroot.  iroot is (should) not be needed
 584          * by the flush code.
 585          */
 586         iroot = pmp->iroot;
 587         if (iroot) {
 588                 /*
 589                  * Stop synchronizing
 590                  *
 591                  * XXX flush after acquiring the iroot lock.
 592                  * XXX clean out the cluster index from all inode structures.
 593                  */
 594                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 595
 596                 /*
 597                  * Remove the cluster index from the group.  If destroying
 598                  * the PFS and this is a master, adjust pfs_nmasters.
 599                  */
 600                 hammer2_mtx_ex(&iroot->lock);
 601                 chain = iroot->cluster.array[clindex].chain;
 602                 iroot->cluster.array[clindex].chain = NULL;
 603
 604                 switch(pmp->pfs_types[clindex]) {
 605                 case HAMMER2_PFSTYPE_MASTER:
 606                         if (destroying && pmp->pfs_nmasters > 0)
 607                                 --pmp->pfs_nmasters;
 608                         /* XXX adjust ripdata->meta.pfs_nmasters */
 609                         break;
 610                 default:
 611                         break;
 612                 }
 613                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 614
 615                 hammer2_mtx_unlock(&iroot->lock);
 616
 617                 /*
 618                  * Release the chain.
 619                  */
 620                 if (chain) {
 621                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 622                         hammer2_chain_drop(chain);
 623                 }
 624
 625                 /*
 626                  * Terminate all XOP threads for the cluster index.
 627                  */
 628                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 629                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 630         }
 631 }
 632
 633 /*
 634  * Destroy a PFS, typically only occurs after the last mount on a device
 635  * has gone away.
 636  */
 637 static void
 638 hammer2_pfsfree(hammer2_pfs_t *pmp)
 639 {
 640         hammer2_inode_t *iroot;
 641         hammer2_chain_t *chain;
 642         int i;
 643         int j;
 644
 645         /*
 646          * Cleanup our reference on iroot.  iroot is (should) not be needed
 647          * by the flush code.
 648          */
 649         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 650
 651         iroot = pmp->iroot;
 652         if (iroot) {
 653                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 654                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 655                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 656                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 657                 }
 658 #if REPORT_REFS_ERRORS
 659                 if (pmp->iroot->refs != 1)
 660                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 661                                 pmp->iroot, pmp->iroot->refs);
 662 #else
 663                 KKASSERT(pmp->iroot->refs == 1);
 664 #endif
 665                 /* ref for pmp->iroot */
 666                 hammer2_inode_drop(pmp->iroot);
 667                 pmp->iroot = NULL;
 668         }
 669
 670         /*
 671          * Cleanup chains remaining on LRU list.
 672          */
 673         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 674                 hammer2_chain_ref(chain);
 675                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 676                 hammer2_chain_drop(chain);
 677         }
 678
 679         /*
 680          * Free remaining pmp resources
 681          */
 682         kmalloc_destroy(&pmp->mmsg);
 683         kmalloc_destroy(&pmp->minode);
 684
 685         kfree(pmp, M_HAMMER2);
 686 }
 687
 688 /*
 689  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 690  * empty is terminated and freed.
 691  *
 692  * XXX inefficient.
 693  */
 694 static void
 695 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 696 {
 697         hammer2_pfs_t *pmp;
 698         hammer2_inode_t *iroot;
 699         hammer2_chain_t *rchain;
 700         int didfreeze;
 701         int i;
 702         int j;
 703
 704 again:
 705         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 706                 if ((iroot = pmp->iroot) == NULL)
 707                         continue;
 708                 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
 709                 hammer2_inode_run_sideq(pmp, 1);
 710                 hammer2_bioq_sync(pmp);
 711                 hammer2_trans_done(pmp);
 712                 if (hmp->spmp == pmp) {
 713                         hmp->spmp = NULL;
 714                         hmp->vchain.pmp = NULL;
 715                         hmp->fchain.pmp = NULL;
 716                 }
 717
 718                 /*
 719                  * Determine if this PFS is affected.  If it is we must
 720                  * freeze all management threads and lock its iroot.
 721                  *
 722                  * Freezing a management thread forces it idle, operations
 723                  * in-progress will be aborted and it will have to start
 724                  * over again when unfrozen, or exit if told to exit.
 725                  */
 726                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 727                         if (pmp->pfs_hmps[i] == hmp)
 728                                 break;
 729                 }
 730                 if (i != HAMMER2_MAXCLUSTER) {
 731                         /*
 732                          * Make sure all synchronization threads are locked
 733                          * down.
 734                          */
 735                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 736                                 if (pmp->pfs_hmps[i] == NULL)
 737                                         continue;
 738                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 739                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 740                                         hammer2_thr_freeze_async(
 741                                                 &pmp->xop_groups[j].thrs[i]);
 742                                 }
 743                         }
 744                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 745                                 if (pmp->pfs_hmps[i] == NULL)
 746                                         continue;
 747                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 748                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 749                                         hammer2_thr_freeze(
 750                                                 &pmp->xop_groups[j].thrs[i]);
 751                                 }
 752                         }
 753
 754                         /*
 755                          * Lock the inode and clean out matching chains.
 756                          * Note that we cannot use hammer2_inode_lock_*()
 757                          * here because that would attempt to validate the
 758                          * cluster that we are in the middle of ripping
 759                          * apart.
 760                          *
 761                          * WARNING! We are working directly on the inodes
 762                          *          embedded cluster.
 763                          */
 764                         hammer2_mtx_ex(&iroot->lock);
 765
 766                         /*
 767                          * Remove the chain from matching elements of the PFS.
 768                          */
 769                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 770                                 if (pmp->pfs_hmps[i] != hmp)
 771                                         continue;
 772                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 773                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 774                                         hammer2_thr_delete(
 775                                                 &pmp->xop_groups[j].thrs[i]);
 776                                 }
 777                                 rchain = iroot->cluster.array[i].chain;
 778                                 iroot->cluster.array[i].chain = NULL;
 779                                 pmp->pfs_types[i] = 0;
 780                                 if (pmp->pfs_names[i]) {
 781                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 782                                         pmp->pfs_names[i] = NULL;
 783                                 }
 784                                 if (rchain) {
 785                                         hammer2_chain_drop(rchain);
 786                                         /* focus hint */
 787                                         if (iroot->cluster.focus == rchain)
 788                                                 iroot->cluster.focus = NULL;
 789                                 }
 790                                 pmp->pfs_hmps[i] = NULL;
 791                         }
 792                         hammer2_mtx_unlock(&iroot->lock);
 793                         didfreeze = 1;  /* remaster, unfreeze down below */
 794                 } else {
 795                         didfreeze = 0;
 796                 }
 797
 798                 /*
 799                  * Cleanup trailing chains.  Gaps may remain.
 800                  */
 801                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 802                         if (pmp->pfs_hmps[i])
 803                                 break;
 804                 }
 805                 iroot->cluster.nchains = i + 1;
 806
 807                 /*
 808                  * If the PMP has no elements remaining we can destroy it.
 809                  * (this will transition management threads from frozen->exit).
 810                  */
 811                 if (iroot->cluster.nchains == 0) {
 812                         hammer2_pfsfree(pmp);
 813                         goto again;
 814                 }
 815
 816                 /*
 817                  * If elements still remain we need to set the REMASTER
 818                  * flag and unfreeze it.
 819                  */
 820                 if (didfreeze) {
 821                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 822                                 if (pmp->pfs_hmps[i] == NULL)
 823                                         continue;
 824                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 825                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 826                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 827                                         hammer2_thr_remaster(
 828                                                 &pmp->xop_groups[j].thrs[i]);
 829                                         hammer2_thr_unfreeze(
 830                                                 &pmp->xop_groups[j].thrs[i]);
 831                                 }
 832                         }
 833                 }
 834         }
 835 }
 836
 837 /*
 838  * Mount or remount HAMMER2 fileystem from physical media
 839  *
 840  *      mountroot
 841  *              mp              mount point structure
 842  *              path            NULL
 843  *              data            <unused>
 844  *              cred            <unused>
 845  *
 846  *      mount
 847  *              mp              mount point structure
 848  *              path            path to mount point
 849  *              data            pointer to argument structure in user space
 850  *                      volume  volume path (device@LABEL form)
 851  *                      hflags  user mount flags
 852  *              cred            user credentials
 853  *
 854  * RETURNS:     0       Success
 855  *              !0      error number
 856  */
 857 static
 858 int
 859 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 860                   struct ucred *cred)
 861 {
 862         struct hammer2_mount_info info;
 863         hammer2_pfs_t *pmp;
 864         hammer2_pfs_t *spmp;
 865         hammer2_dev_t *hmp;
 866         hammer2_dev_t *force_local;
 867         hammer2_key_t key_next;
 868         hammer2_key_t key_dummy;
 869         hammer2_key_t lhc;
 870         struct vnode *devvp;
 871         struct nlookupdata nd;
 872         hammer2_chain_t *parent;
 873         hammer2_chain_t *chain;
 874         hammer2_cluster_t *cluster;
 875         const hammer2_inode_data_t *ripdata;
 876         hammer2_blockref_t bref;
 877         struct file *fp;
 878         char devstr[MNAMELEN];
 879         size_t size;
 880         size_t done;
 881         char *dev;
 882         char *label;
 883         int ronly = 1;
 884         int error;
 885         int i;
 886
 887         hmp = NULL;
 888         pmp = NULL;
 889         dev = NULL;
 890         label = NULL;
 891         devvp = NULL;
 892
 893         kprintf("hammer2_mount\n");
 894
 895         if (path == NULL) {
 896                 /*
 897                  * Root mount
 898                  */
 899                 bzero(&info, sizeof(info));
 900                 info.cluster_fd = -1;
 901                 ksnprintf(devstr, sizeof(devstr), "%s",
 902                           mp->mnt_stat.f_mntfromname);
 903                 kprintf("hammer2_mount: root '%s'\n", devstr);
 904         } else {
 905                 /*
 906                  * Non-root mount or updating a mount
 907                  */
 908                 error = copyin(data, &info, sizeof(info));
 909                 if (error)
 910                         return (error);
 911
 912                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 913                 if (error)
 914                         return (error);
 915         }
 916
 917         /*
 918          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 919          * if no label specified, based on the partition id.  Error out if no
 920          * label or device (with partition id) is specified.  This is strictly
 921          * a convenience to match the default label created by newfs_hammer2,
 922          * our preference is that a label always be specified.
 923          *
 924          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
 925          *       that does not specify a device, as long as some H2 label
 926          *       has already been mounted from that device.  This makes
 927          *       mounting snapshots a lot easier.
 928          */
 929         dev = devstr;
 930         label = strchr(devstr, '@');
 931         if (label && ((label + 1) - dev) > done)
 932                 return (EINVAL);
 933         if (label == NULL || label[1] == 0) {
 934                 char slice;
 935
 936                 if (label == NULL)
 937                         label = devstr + strlen(devstr);
 938                 slice = label[-1];
 939                 switch(slice) {
 940                 case 'a':
 941                         label = "BOOT";
 942                         break;
 943                 case 'd':
 944                         label = "ROOT";
 945                         break;
 946                 default:
 947                         label = "DATA";
 948                         break;
 949                 }
 950         } else {
 951                 *label = '\0';
 952                 label++;
 953         }
 954
 955         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
 956                 dev, label, (mp->mnt_flag & MNT_RDONLY));
 957
 958         if (mp->mnt_flag & MNT_UPDATE) {
 959                 /*
 960                  * Update mount.  Note that pmp->iroot->cluster is
 961                  * an inode-embedded cluster and thus cannot be
 962                  * directly locked.
 963                  *
 964                  * XXX HAMMER2 needs to implement NFS export via
 965                  *     mountctl.
 966                  */
 967                 pmp = MPTOPMP(mp);
 968                 pmp->hflags = info.hflags;
 969                 cluster = &pmp->iroot->cluster;
 970                 for (i = 0; i < cluster->nchains; ++i) {
 971                         if (cluster->array[i].chain == NULL)
 972                                 continue;
 973                         hmp = cluster->array[i].chain->hmp;
 974                         devvp = hmp->devvp;
 975                         error = hammer2_remount(hmp, mp, path,
 976                                                 devvp, cred);
 977                         if (error)
 978                                 break;
 979                 }
 980
 981                 return error;
 982         }
 983
 984         /*
 985          * HMP device mount
 986          *
 987          * If a path is specified and dev is not an empty string, lookup the
 988          * name and verify that it referes to a block device.
 989          *
 990          * If a path is specified and dev is an empty string we fall through
 991          * and locate the label in the hmp search.
 992          */
 993         if (path && *dev != 0) {
 994                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 995                 if (error == 0)
 996                         error = nlookup(&nd);
 997                 if (error == 0)
 998                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 999                 nlookup_done(&nd);
1000         } else if (path == NULL) {
1001                 /* root mount */
1002                 cdev_t cdev = kgetdiskbyname(dev);
1003                 error = bdevvp(cdev, &devvp);
1004                 if (error)
1005                         kprintf("hammer2: cannot find '%s'\n", dev);
1006         } else {
1007                 /*
1008                  * We will locate the hmp using the label in the hmp loop.
1009                  */
1010                 error = 0;
1011         }
1012
1013         /*
1014          * Make sure its a block device.  Do not check to see if it is
1015          * already mounted until we determine that its a fresh H2 device.
1016          */
1017         if (error == 0 && devvp) {
1018                 vn_isdisk(devvp, &error);
1019         }
1020
1021         /*
1022          * Determine if the device has already been mounted.  After this
1023          * check hmp will be non-NULL if we are doing the second or more
1024          * hammer2 mounts from the same device.
1025          */
1026         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1027         if (devvp) {
1028                 /*
1029                  * Match the device.  Due to the way devfs works,
1030                  * we may not be able to directly match the vnode pointer,
1031                  * so also check to see if the underlying device matches.
1032                  */
1033                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1034                         if (hmp->devvp == devvp)
1035                                 break;
1036                         if (devvp->v_rdev &&
1037                             hmp->devvp->v_rdev == devvp->v_rdev) {
1038                                 break;
1039                         }
1040                 }
1041
1042                 /*
1043                  * If no match this may be a fresh H2 mount, make sure
1044                  * the device is not mounted on anything else.
1045                  */
1046                 if (hmp == NULL)
1047                         error = vfs_mountedon(devvp);
1048         } else if (error == 0) {
1049                 /*
1050                  * Match the label to a pmp already probed.
1051                  */
1052                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1053                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1054                                 if (pmp->pfs_names[i] &&
1055                                     strcmp(pmp->pfs_names[i], label) == 0) {
1056                                         hmp = pmp->pfs_hmps[i];
1057                                         break;
1058                                 }
1059                         }
1060                         if (hmp)
1061                                 break;
1062                 }
1063                 if (hmp == NULL)
1064                         error = ENOENT;
1065         }
1066
1067         /*
1068          * Open the device if this isn't a secondary mount and construct
1069          * the H2 device mount (hmp).
1070          */
1071         if (hmp == NULL) {
1072                 hammer2_chain_t *schain;
1073                 hammer2_xid_t xid;
1074
1075                 if (error == 0 && vcount(devvp) > 0) {
1076                         kprintf("Primary device already has references\n");
1077                         error = EBUSY;
1078                 }
1079
1080                 /*
1081                  * Now open the device
1082                  */
1083                 if (error == 0) {
1084                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1085                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1086                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1087                         if (error == 0) {
1088                                 error = VOP_OPEN(devvp,
1089                                              (ronly ? FREAD : FREAD | FWRITE),
1090                                              FSCRED, NULL);
1091                         }
1092                         vn_unlock(devvp);
1093                 }
1094                 if (error && devvp) {
1095                         vrele(devvp);
1096                         devvp = NULL;
1097                 }
1098                 if (error) {
1099                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1100                         return error;
1101                 }
1102                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1103                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1104                 hmp->ronly = ronly;
1105                 hmp->devvp = devvp;
1106                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1107                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1108                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1109                 RB_INIT(&hmp->iotree);
1110                 spin_init(&hmp->io_spin, "hm2mount_io");
1111                 spin_init(&hmp->list_spin, "hm2mount_list");
1112                 TAILQ_INIT(&hmp->flushq);
1113
1114                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1115                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1116                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1117
1118                 /*
1119                  * vchain setup. vchain.data is embedded.
1120                  * vchain.refs is initialized and will never drop to 0.
1121                  *
1122                  * NOTE! voldata is not yet loaded.
1123                  */
1124                 hmp->vchain.hmp = hmp;
1125                 hmp->vchain.refs = 1;
1126                 hmp->vchain.data = (void *)&hmp->voldata;
1127                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1128                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1129                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1130
1131                 hammer2_chain_core_init(&hmp->vchain);
1132                 /* hmp->vchain.u.xxx is left NULL */
1133
1134                 /*
1135                  * fchain setup.  fchain.data is embedded.
1136                  * fchain.refs is initialized and will never drop to 0.
1137                  *
1138                  * The data is not used but needs to be initialized to
1139                  * pass assertion muster.  We use this chain primarily
1140                  * as a placeholder for the freemap's top-level RBTREE
1141                  * so it does not interfere with the volume's topology
1142                  * RBTREE.
1143                  */
1144                 hmp->fchain.hmp = hmp;
1145                 hmp->fchain.refs = 1;
1146                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1147                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1148                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1149                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1150                 hmp->fchain.bref.methods =
1151                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1152                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1153
1154                 hammer2_chain_core_init(&hmp->fchain);
1155                 /* hmp->fchain.u.xxx is left NULL */
1156
1157                 /*
1158                  * Install the volume header and initialize fields from
1159                  * voldata.
1160                  */
1161                 error = hammer2_install_volume_header(hmp);
1162                 if (error) {
1163                         hammer2_unmount_helper(mp, NULL, hmp);
1164                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1165                         hammer2_vfs_unmount(mp, MNT_FORCE);
1166                         return error;
1167                 }
1168
1169                 /*
1170                  * Really important to get these right or flush will get
1171                  * confused.
1172                  */
1173                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1174                 kprintf("alloc spmp %p tid %016jx\n",
1175                         hmp->spmp, hmp->voldata.mirror_tid);
1176                 spmp = hmp->spmp;
1177
1178                 /*
1179                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1180                  * is inherited from the volume header.
1181                  */
1182                 xid = 0;
1183                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1184                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1185                 hmp->vchain.pmp = spmp;
1186                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1187                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1188                 hmp->fchain.pmp = spmp;
1189
1190                 /*
1191                  * First locate the super-root inode, which is key 0
1192                  * relative to the volume header's blockset.
1193                  *
1194                  * Then locate the root inode by scanning the directory keyspace
1195                  * represented by the label.
1196                  */
1197                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1198                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1199                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1200                                       &error, 0);
1201                 hammer2_chain_lookup_done(parent);
1202                 if (schain == NULL) {
1203                         kprintf("hammer2_mount: invalid super-root\n");
1204                         hammer2_unmount_helper(mp, NULL, hmp);
1205                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1206                         hammer2_vfs_unmount(mp, MNT_FORCE);
1207                         return EINVAL;
1208                 }
1209                 if (schain->error) {
1210                         kprintf("hammer2_mount: error %s reading super-root\n",
1211                                 hammer2_error_str(schain->error));
1212                         hammer2_chain_unlock(schain);
1213                         hammer2_chain_drop(schain);
1214                         schain = NULL;
1215                         hammer2_unmount_helper(mp, NULL, hmp);
1216                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1217                         hammer2_vfs_unmount(mp, MNT_FORCE);
1218                         return EINVAL;
1219                 }
1220
1221                 /*
1222                  * The super-root always uses an inode_tid of 1 when
1223                  * creating PFSs.
1224                  */
1225                 spmp->inode_tid = 1;
1226                 spmp->modify_tid = schain->bref.modify_tid + 1;
1227
1228                 /*
1229                  * Sanity-check schain's pmp and finish initialization.
1230                  * Any chain belonging to the super-root topology should
1231                  * have a NULL pmp (not even set to spmp).
1232                  */
1233                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1234                 KKASSERT(schain->pmp == NULL);
1235                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1236
1237                 /*
1238                  * Replace the dummy spmp->iroot with a real one.  It's
1239                  * easier to just do a wholesale replacement than to try
1240                  * to update the chain and fixup the iroot fields.
1241                  *
1242                  * The returned inode is locked with the supplied cluster.
1243                  */
1244                 cluster = hammer2_cluster_from_chain(schain);
1245                 hammer2_inode_drop(spmp->iroot);
1246                 spmp->iroot = NULL;
1247                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1248                 spmp->spmp_hmp = hmp;
1249                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1250                 spmp->pfs_hmps[0] = hmp;
1251                 hammer2_inode_ref(spmp->iroot);
1252                 hammer2_inode_unlock(spmp->iroot);
1253                 hammer2_cluster_unlock(cluster);
1254                 hammer2_cluster_drop(cluster);
1255                 schain = NULL;
1256                 /* leave spmp->iroot with one ref */
1257
1258                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1259                         error = hammer2_recovery(hmp);
1260                         /* XXX do something with error */
1261                 }
1262                 hammer2_update_pmps(hmp);
1263                 hammer2_iocom_init(hmp);
1264                 hammer2_bulkfree_init(hmp);
1265
1266                 /*
1267                  * Ref the cluster management messaging descriptor.  The mount
1268                  * program deals with the other end of the communications pipe.
1269                  *
1270                  * Root mounts typically do not supply one.
1271                  */
1272                 if (info.cluster_fd >= 0) {
1273                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1274                         if (fp) {
1275                                 hammer2_cluster_reconnect(hmp, fp);
1276                         } else {
1277                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1278                         }
1279                 }
1280         } else {
1281                 spmp = hmp->spmp;
1282                 if (info.hflags & HMNT2_DEVFLAGS) {
1283                         kprintf("hammer2: Warning: mount flags pertaining "
1284                                 "to the whole device may only be specified "
1285                                 "on the first mount of the device: %08x\n",
1286                                 info.hflags & HMNT2_DEVFLAGS);
1287                 }
1288         }
1289
1290         /*
1291          * Force local mount (disassociate all PFSs from their clusters).
1292          * Used primarily for debugging.
1293          */
1294         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1295
1296         /*
1297          * Lookup the mount point under the media-localized super-root.
1298          * Scanning hammer2_pfslist doesn't help us because it represents
1299          * PFS cluster ids which can aggregate several named PFSs together.
1300          *
1301          * cluster->pmp will incorrectly point to spmp and must be fixed
1302          * up later on.
1303          */
1304         hammer2_inode_lock(spmp->iroot, 0);
1305         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1306         lhc = hammer2_dirhash(label, strlen(label));
1307         chain = hammer2_chain_lookup(&parent, &key_next,
1308                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1309                                      &error, 0);
1310         while (chain) {
1311                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1312                     strcmp(label, chain->data->ipdata.filename) == 0) {
1313                         break;
1314                 }
1315                 chain = hammer2_chain_next(&parent, chain, &key_next,
1316                                             key_next,
1317                                             lhc + HAMMER2_DIRHASH_LOMASK,
1318                                             &error, 0);
1319         }
1320         if (parent) {
1321                 hammer2_chain_unlock(parent);
1322                 hammer2_chain_drop(parent);
1323         }
1324         hammer2_inode_unlock(spmp->iroot);
1325
1326         /*
1327          * PFS could not be found?
1328          */
1329         if (chain == NULL) {
1330                 if (error)
1331                         kprintf("hammer2_mount: PFS label I/O error\n");
1332                 else
1333                         kprintf("hammer2_mount: PFS label not found\n");
1334                 hammer2_unmount_helper(mp, NULL, hmp);
1335                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1336                 hammer2_vfs_unmount(mp, MNT_FORCE);
1337
1338                 return EINVAL;
1339         }
1340
1341         /*
1342          * Acquire the pmp structure (it should have already been allocated
1343          * via hammer2_update_pmps() so do not pass cluster in to add to
1344          * available chains).
1345          *
1346          * Check if the cluster has already been mounted.  A cluster can
1347          * only be mounted once, use null mounts to mount additional copies.
1348          */
1349         if (chain->error) {
1350                 kprintf("hammer2_mount: PFS label I/O error\n");
1351         } else {
1352                 ripdata = &chain->data->ipdata;
1353                 bref = chain->bref;
1354                 pmp = hammer2_pfsalloc(NULL, ripdata,
1355                                        bref.modify_tid, force_local);
1356         }
1357         hammer2_chain_unlock(chain);
1358         hammer2_chain_drop(chain);
1359
1360         /*
1361          * Finish the mount
1362          */
1363         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1364
1365         if (pmp->mp) {
1366                 kprintf("hammer2_mount: PFS already mounted!\n");
1367                 hammer2_unmount_helper(mp, NULL, hmp);
1368                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1369                 hammer2_vfs_unmount(mp, MNT_FORCE);
1370
1371                 return EBUSY;
1372         }
1373
1374         pmp->hflags = info.hflags;
1375         mp->mnt_flag |= MNT_LOCAL;
1376         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1377         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1378
1379         /*
1380          * required mount structure initializations
1381          */
1382         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1383         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1384
1385         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1386         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1387
1388         /*
1389          * Optional fields
1390          */
1391         mp->mnt_iosize_max = MAXPHYS;
1392
1393         /*
1394          * Connect up mount pointers.
1395          */
1396         hammer2_mount_helper(mp, pmp);
1397
1398         lockmgr(&hammer2_mntlk, LK_RELEASE);
1399
1400         /*
1401          * Finish setup
1402          */
1403         vfs_getnewfsid(mp);
1404         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1405         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1406         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1407
1408         if (path) {
1409                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1410                           MNAMELEN - 1, &size);
1411                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1412         } /* else root mount, already in there */
1413
1414         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1415         if (path) {
1416                 copyinstr(path, mp->mnt_stat.f_mntonname,
1417                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1418                           &size);
1419         } else {
1420                 /* root mount */
1421                 mp->mnt_stat.f_mntonname[0] = '/';
1422         }
1423
1424         /*
1425          * Initial statfs to prime mnt_stat.
1426          */
1427         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1428
1429         return 0;
1430 }
1431
1432 /*
1433  * Scan PFSs under the super-root and create hammer2_pfs structures.
1434  */
1435 static
1436 void
1437 hammer2_update_pmps(hammer2_dev_t *hmp)
1438 {
1439         const hammer2_inode_data_t *ripdata;
1440         hammer2_chain_t *parent;
1441         hammer2_chain_t *chain;
1442         hammer2_blockref_t bref;
1443         hammer2_dev_t *force_local;
1444         hammer2_pfs_t *spmp;
1445         hammer2_pfs_t *pmp;
1446         hammer2_key_t key_next;
1447         int error;
1448
1449         /*
1450          * Force local mount (disassociate all PFSs from their clusters).
1451          * Used primarily for debugging.
1452          */
1453         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1454
1455         /*
1456          * Lookup mount point under the media-localized super-root.
1457          *
1458          * cluster->pmp will incorrectly point to spmp and must be fixed
1459          * up later on.
1460          */
1461         spmp = hmp->spmp;
1462         hammer2_inode_lock(spmp->iroot, 0);
1463         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1464         chain = hammer2_chain_lookup(&parent, &key_next,
1465                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1466                                          &error, 0);
1467         while (chain) {
1468                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1469                         continue;
1470                 if (chain->error) {
1471                         kprintf("I/O error scanning PFS labels\n");
1472                 } else {
1473                         ripdata = &chain->data->ipdata;
1474                         bref = chain->bref;
1475
1476                         pmp = hammer2_pfsalloc(chain, ripdata,
1477                                                bref.modify_tid, force_local);
1478                 }
1479                 chain = hammer2_chain_next(&parent, chain, &key_next,
1480                                            key_next, HAMMER2_KEY_MAX,
1481                                            &error, 0);
1482         }
1483         if (parent) {
1484                 hammer2_chain_unlock(parent);
1485                 hammer2_chain_drop(parent);
1486         }
1487         hammer2_inode_unlock(spmp->iroot);
1488 }
1489
1490 static
1491 int
1492 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1493                 struct vnode *devvp, struct ucred *cred)
1494 {
1495         int error;
1496
1497         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1498                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1499                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1500                 vn_unlock(devvp);
1501                 error = hammer2_recovery(hmp);
1502                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1503                 if (error == 0) {
1504                         VOP_CLOSE(devvp, FREAD, NULL);
1505                         hmp->ronly = 0;
1506                 } else {
1507                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1508                 }
1509                 vn_unlock(devvp);
1510         } else {
1511                 error = 0;
1512         }
1513         return error;
1514 }
1515
1516 static
1517 int
1518 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1519 {
1520         hammer2_pfs_t *pmp;
1521         int flags;
1522         int error = 0;
1523
1524         pmp = MPTOPMP(mp);
1525
1526         if (pmp == NULL)
1527                 return(0);
1528
1529         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1530
1531         /*
1532          * If mount initialization proceeded far enough we must flush
1533          * its vnodes and sync the underlying mount points.  Three syncs
1534          * are required to fully flush the filesystem (freemap updates lag
1535          * by one flush, and one extra for safety).
1536          */
1537         if (mntflags & MNT_FORCE)
1538                 flags = FORCECLOSE;
1539         else
1540                 flags = 0;
1541         if (pmp->iroot) {
1542                 error = vflush(mp, 0, flags);
1543                 if (error)
1544                         goto failed;
1545                 hammer2_vfs_sync(mp, MNT_WAIT);
1546                 hammer2_vfs_sync(mp, MNT_WAIT);
1547                 hammer2_vfs_sync(mp, MNT_WAIT);
1548         }
1549
1550         /*
1551          * Cleanup the frontend support XOPS threads
1552          */
1553         hammer2_xop_helper_cleanup(pmp);
1554
1555         if (pmp->mp)
1556                 hammer2_unmount_helper(mp, pmp, NULL);
1557
1558         error = 0;
1559 failed:
1560         lockmgr(&hammer2_mntlk, LK_RELEASE);
1561
1562         return (error);
1563 }
1564
1565 /*
1566  * Mount helper, hook the system mount into our PFS.
1567  * The mount lock is held.
1568  *
1569  * We must bump the mount_count on related devices for any
1570  * mounted PFSs.
1571  */
1572 static
1573 void
1574 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1575 {
1576         hammer2_cluster_t *cluster;
1577         hammer2_chain_t *rchain;
1578         int i;
1579
1580         mp->mnt_data = (qaddr_t)pmp;
1581         pmp->mp = mp;
1582
1583         /*
1584          * After pmp->mp is set we have to adjust hmp->mount_count.
1585          */
1586         cluster = &pmp->iroot->cluster;
1587         for (i = 0; i < cluster->nchains; ++i) {
1588                 rchain = cluster->array[i].chain;
1589                 if (rchain == NULL)
1590                         continue;
1591                 ++rchain->hmp->mount_count;
1592                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1593                         rchain->hmp, rchain->hmp->mount_count);
1594         }
1595
1596         /*
1597          * Create missing Xop threads
1598          */
1599         hammer2_xop_helper_create(pmp);
1600 }
1601
1602 /*
1603  * Mount helper, unhook the system mount from our PFS.
1604  * The mount lock is held.
1605  *
1606  * If hmp is supplied a mount responsible for being the first to open
1607  * the block device failed and the block device and all PFSs using the
1608  * block device must be cleaned up.
1609  *
1610  * If pmp is supplied multiple devices might be backing the PFS and each
1611  * must be disconnected.  This might not be the last PFS using some of the
1612  * underlying devices.  Also, we have to adjust our hmp->mount_count
1613  * accounting for the devices backing the pmp which is now undergoing an
1614  * unmount.
1615  */
1616 static
1617 void
1618 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1619 {
1620         hammer2_cluster_t *cluster;
1621         hammer2_chain_t *rchain;
1622         struct vnode *devvp;
1623         int dumpcnt;
1624         int ronly;
1625         int i;
1626
1627         /*
1628          * If no device supplied this is a high-level unmount and we have to
1629          * to disconnect the mount, adjust mount_count, and locate devices
1630          * that might now have no mounts.
1631          */
1632         if (pmp) {
1633                 KKASSERT(hmp == NULL);
1634                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1635                 pmp->mp = NULL;
1636                 mp->mnt_data = NULL;
1637
1638                 /*
1639                  * After pmp->mp is cleared we have to account for
1640                  * mount_count.
1641                  */
1642                 cluster = &pmp->iroot->cluster;
1643                 for (i = 0; i < cluster->nchains; ++i) {
1644                         rchain = cluster->array[i].chain;
1645                         if (rchain == NULL)
1646                                 continue;
1647                         --rchain->hmp->mount_count;
1648                         /* scrapping hmp now may invalidate the pmp */
1649                 }
1650 again:
1651                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1652                         if (hmp->mount_count == 0) {
1653                                 hammer2_unmount_helper(NULL, NULL, hmp);
1654                                 goto again;
1655                         }
1656                 }
1657                 return;
1658         }
1659
1660         /*
1661          * Try to terminate the block device.  We can't terminate it if
1662          * there are still PFSs referencing it.
1663          */
1664         if (hmp->mount_count)
1665                 return;
1666
1667         /*
1668          * Decomission the network before we start messing with the
1669          * device and PFS.
1670          */
1671         hammer2_iocom_uninit(hmp);
1672
1673         hammer2_bulkfree_uninit(hmp);
1674         hammer2_pfsfree_scan(hmp);
1675         hammer2_dev_exlock(hmp);        /* XXX order */
1676
1677         /*
1678          * Cycle the volume data lock as a safety (probably not needed any
1679          * more).  To ensure everything is out we need to flush at least
1680          * three times.  (1) The running of the sideq can dirty the
1681          * filesystem, (2) A normal flush can dirty the freemap, and
1682          * (3) ensure that the freemap is fully synchronized.
1683          *
1684          * The next mount's recovery scan can clean everything up but we want
1685          * to leave the filesystem in a 100% clean state on a normal unmount.
1686          */
1687 #if 0
1688         hammer2_voldata_lock(hmp);
1689         hammer2_voldata_unlock(hmp);
1690 #endif
1691
1692         /*
1693          * Flush whatever is left.  Unmounted but modified PFS's might still
1694          * have some dirty chains on them.
1695          */
1696         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1697         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1698         hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP | HAMMER2_FLUSH_ALL);
1699         hammer2_chain_unlock(&hmp->fchain);
1700         hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP | HAMMER2_FLUSH_ALL);
1701         hammer2_chain_unlock(&hmp->vchain);
1702
1703         if ((hmp->vchain.flags | hmp->fchain.flags) &
1704             HAMMER2_CHAIN_FLUSH_MASK) {
1705                 kprintf("hammer2_unmount: chains left over "
1706                         "after final sync\n");
1707                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1708                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1709
1710                 if (hammer2_debug & 0x0010)
1711                         Debugger("entered debugger");
1712         }
1713
1714         KKASSERT(hmp->spmp == NULL);
1715
1716         /*
1717          * Finish up with the device vnode
1718          */
1719         if ((devvp = hmp->devvp) != NULL) {
1720                 ronly = hmp->ronly;
1721                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1722                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1723                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1724                         ronly);
1725                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1726                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1727                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1728                 hmp->devvp = NULL;
1729                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1730                 vn_unlock(devvp);
1731                 vrele(devvp);
1732                 devvp = NULL;
1733         }
1734
1735         /*
1736          * Clear vchain/fchain flags that might prevent final cleanup
1737          * of these chains.
1738          */
1739         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1740                 atomic_add_long(&hammer2_count_modified_chains, -1);
1741                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1742                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1743         }
1744         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1745                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1746         }
1747
1748         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1749                 atomic_add_long(&hammer2_count_modified_chains, -1);
1750                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1751                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1752         }
1753         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1754                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1755         }
1756
1757         /*
1758          * Final drop of embedded freemap root chain to
1759          * clean up fchain.core (fchain structure is not
1760          * flagged ALLOCATED so it is cleaned out and then
1761          * left to rot).
1762          */
1763         hammer2_chain_drop(&hmp->fchain);
1764
1765         /*
1766          * Final drop of embedded volume root chain to clean
1767          * up vchain.core (vchain structure is not flagged
1768          * ALLOCATED so it is cleaned out and then left to
1769          * rot).
1770          */
1771         dumpcnt = 50;
1772         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1773         dumpcnt = 50;
1774         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1775         hammer2_dev_unlock(hmp);
1776         hammer2_chain_drop(&hmp->vchain);
1777
1778         hammer2_io_cleanup(hmp, &hmp->iotree);
1779         if (hmp->iofree_count) {
1780                 kprintf("io_cleanup: %d I/O's left hanging\n",
1781                         hmp->iofree_count);
1782         }
1783
1784         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1785         kmalloc_destroy(&hmp->mchain);
1786         kfree(hmp, M_HAMMER2);
1787 }
1788
1789 int
1790 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1791                  ino_t ino, struct vnode **vpp)
1792 {
1793         hammer2_xop_lookup_t *xop;
1794         hammer2_pfs_t *pmp;
1795         hammer2_inode_t *ip;
1796         hammer2_tid_t inum;
1797         int error;
1798
1799         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1800
1801         error = 0;
1802         pmp = MPTOPMP(mp);
1803
1804         /*
1805          * Easy if we already have it cached
1806          */
1807         ip = hammer2_inode_lookup(pmp, inum);
1808         if (ip) {
1809                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1810                 *vpp = hammer2_igetv(ip, &error);
1811                 hammer2_inode_unlock(ip);
1812                 hammer2_inode_drop(ip);         /* from lookup */
1813
1814                 return error;
1815         }
1816
1817         /*
1818          * Otherwise we have to find the inode
1819          */
1820         xop = hammer2_xop_alloc(pmp->iroot, 0);
1821         xop->lhc = inum;
1822         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1823         error = hammer2_xop_collect(&xop->head, 0);
1824
1825         if (error == 0) {
1826                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1827                         kprintf("vget: no collect error but also no rdata\n");
1828                         kprintf("xop %p\n", xop);
1829                         while ((hammer2_debug & 0x80000) == 0) {
1830                                 tsleep(xop, PCATCH, "wait", hz * 10);
1831                         }
1832                         ip = NULL;
1833                 } else {
1834                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1835                 }
1836         }
1837         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1838
1839         if (ip) {
1840                 *vpp = hammer2_igetv(ip, &error);
1841                 hammer2_inode_unlock(ip);
1842         } else {
1843                 *vpp = NULL;
1844                 error = ENOENT;
1845         }
1846         return (error);
1847 }
1848
1849 static
1850 int
1851 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1852 {
1853         hammer2_pfs_t *pmp;
1854         struct vnode *vp;
1855         int error;
1856
1857         pmp = MPTOPMP(mp);
1858         if (pmp->iroot == NULL) {
1859                 *vpp = NULL;
1860                 return EINVAL;
1861         }
1862
1863         error = 0;
1864         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1865
1866         while (pmp->inode_tid == 0) {
1867                 hammer2_xop_ipcluster_t *xop;
1868                 hammer2_inode_meta_t *meta;
1869
1870                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1871                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1872                 error = hammer2_xop_collect(&xop->head, 0);
1873
1874                 if (error == 0) {
1875                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1876                         pmp->iroot->meta = *meta;
1877                         pmp->inode_tid = meta->pfs_inum + 1;
1878                         if (pmp->inode_tid < HAMMER2_INODE_START)
1879                                 pmp->inode_tid = HAMMER2_INODE_START;
1880                         pmp->modify_tid =
1881                                 xop->head.cluster.focus->bref.modify_tid + 1;
1882                         kprintf("PFS: Starting inode %jd\n",
1883                                 (intmax_t)pmp->inode_tid);
1884                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1885                                 pmp->inode_tid, pmp->modify_tid);
1886                         wakeup(&pmp->iroot);
1887
1888                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1889
1890                         /*
1891                          * Prime the mount info.
1892                          */
1893                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1894                         break;
1895                 }
1896
1897                 /*
1898                  * Loop, try again
1899                  */
1900                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1901                 hammer2_inode_unlock(pmp->iroot);
1902                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1903                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1904                 if (error == EINTR)
1905                         break;
1906         }
1907
1908         if (error) {
1909                 hammer2_inode_unlock(pmp->iroot);
1910                 *vpp = NULL;
1911         } else {
1912                 vp = hammer2_igetv(pmp->iroot, &error);
1913                 hammer2_inode_unlock(pmp->iroot);
1914                 *vpp = vp;
1915         }
1916
1917         return (error);
1918 }
1919
1920 /*
1921  * Filesystem status
1922  *
1923  * XXX incorporate ipdata->meta.inode_quota and data_quota
1924  */
1925 static
1926 int
1927 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1928 {
1929         hammer2_pfs_t *pmp;
1930         hammer2_dev_t *hmp;
1931         hammer2_blockref_t bref;
1932         int i;
1933
1934         /*
1935          * NOTE: iroot might not have validated the cluster yet.
1936          */
1937         pmp = MPTOPMP(mp);
1938
1939         mp->mnt_stat.f_files = 0;
1940         mp->mnt_stat.f_ffree = 0;
1941         mp->mnt_stat.f_blocks = 0;
1942         mp->mnt_stat.f_bfree = 0;
1943         mp->mnt_stat.f_bavail = 0;
1944
1945         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1946                 hmp = pmp->pfs_hmps[i];
1947                 if (hmp == NULL)
1948                         continue;
1949                 if (pmp->iroot->cluster.array[i].chain)
1950                         bref = pmp->iroot->cluster.array[i].chain->bref;
1951                 else
1952                         bzero(&bref, sizeof(bref));
1953
1954                 mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1955                 mp->mnt_stat.f_ffree = 0;
1956                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1957                                         mp->mnt_vstat.f_bsize;
1958                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1959                                         mp->mnt_vstat.f_bsize;
1960                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1961
1962                 if (cred && cred->cr_uid != 0) {
1963                         uint64_t adj;
1964
1965                         /* 5% */
1966                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
1967                         mp->mnt_stat.f_blocks -= adj;
1968                         mp->mnt_stat.f_bfree -= adj;
1969                         mp->mnt_stat.f_bavail -= adj;
1970                 }
1971
1972                 *sbp = mp->mnt_stat;
1973         }
1974         return (0);
1975 }
1976
1977 static
1978 int
1979 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1980 {
1981         hammer2_pfs_t *pmp;
1982         hammer2_dev_t *hmp;
1983         hammer2_blockref_t bref;
1984         int i;
1985
1986         /*
1987          * NOTE: iroot might not have validated the cluster yet.
1988          */
1989         pmp = MPTOPMP(mp);
1990
1991         mp->mnt_vstat.f_bsize = 0;
1992         mp->mnt_vstat.f_files = 0;
1993         mp->mnt_vstat.f_ffree = 0;
1994         mp->mnt_vstat.f_blocks = 0;
1995         mp->mnt_vstat.f_bfree = 0;
1996         mp->mnt_vstat.f_bavail = 0;
1997
1998         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1999                 hmp = pmp->pfs_hmps[i];
2000                 if (hmp == NULL)
2001                         continue;
2002                 if (pmp->iroot->cluster.array[i].chain)
2003                         bref = pmp->iroot->cluster.array[i].chain->bref;
2004                 else
2005                         bzero(&bref, sizeof(bref));
2006
2007                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
2008                 mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
2009                 mp->mnt_vstat.f_ffree = 0;
2010                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
2011                                         mp->mnt_vstat.f_bsize;
2012                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
2013                                         mp->mnt_vstat.f_bsize;
2014                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
2015
2016                 if (cred && cred->cr_uid != 0) {
2017                         uint64_t adj;
2018
2019                         /* 5% */
2020                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2021                         mp->mnt_vstat.f_blocks -= adj;
2022                         mp->mnt_vstat.f_bfree -= adj;
2023                         mp->mnt_vstat.f_bavail -= adj;
2024                 }
2025
2026                 *sbp = mp->mnt_vstat;
2027         }
2028         return (0);
2029 }
2030
2031 /*
2032  * Mount-time recovery (RW mounts)
2033  *
2034  * Updates to the free block table are allowed to lag flushes by one
2035  * transaction.  In case of a crash, then on a fresh mount we must do an
2036  * incremental scan of the last committed transaction id and make sure that
2037  * all related blocks have been marked allocated.
2038  *
2039  * The super-root topology and each PFS has its own transaction id domain,
2040  * so we must track PFS boundary transitions.
2041  */
2042 struct hammer2_recovery_elm {
2043         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2044         hammer2_chain_t *chain;
2045         hammer2_tid_t sync_tid;
2046 };
2047
2048 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2049
2050 struct hammer2_recovery_info {
2051         struct hammer2_recovery_list list;
2052         hammer2_tid_t   mtid;
2053         int     depth;
2054 };
2055
2056 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2057                         hammer2_chain_t *parent,
2058                         struct hammer2_recovery_info *info,
2059                         hammer2_tid_t sync_tid);
2060
2061 #define HAMMER2_RECOVERY_MAXDEPTH       10
2062
2063 static
2064 int
2065 hammer2_recovery(hammer2_dev_t *hmp)
2066 {
2067         struct hammer2_recovery_info info;
2068         struct hammer2_recovery_elm *elm;
2069         hammer2_chain_t *parent;
2070         hammer2_tid_t sync_tid;
2071         hammer2_tid_t mirror_tid;
2072         int error;
2073
2074         hammer2_trans_init(hmp->spmp, 0);
2075
2076         sync_tid = hmp->voldata.freemap_tid;
2077         mirror_tid = hmp->voldata.mirror_tid;
2078
2079         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2080         if (sync_tid >= mirror_tid) {
2081                 kprintf(" no recovery needed\n");
2082         } else {
2083                 kprintf(" freemap recovery %016jx-%016jx\n",
2084                         sync_tid + 1, mirror_tid);
2085         }
2086
2087         TAILQ_INIT(&info.list);
2088         info.depth = 0;
2089         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2090         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2091         hammer2_chain_lookup_done(parent);
2092
2093         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2094                 TAILQ_REMOVE(&info.list, elm, entry);
2095                 parent = elm->chain;
2096                 sync_tid = elm->sync_tid;
2097                 kfree(elm, M_HAMMER2);
2098
2099                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2100                 error |= hammer2_recovery_scan(hmp, parent, &info,
2101                                               hmp->voldata.freemap_tid);
2102                 hammer2_chain_unlock(parent);
2103                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2104         }
2105         hammer2_trans_done(hmp->spmp);
2106
2107         return error;
2108 }
2109
2110 static
2111 int
2112 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2113                       struct hammer2_recovery_info *info,
2114                       hammer2_tid_t sync_tid)
2115 {
2116         const hammer2_inode_data_t *ripdata;
2117         hammer2_chain_t *chain;
2118         hammer2_blockref_t bref;
2119         int tmp_error;
2120         int rup_error;
2121         int error;
2122         int first;
2123
2124         /*
2125          * Adjust freemap to ensure that the block(s) are marked allocated.
2126          */
2127         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2128                 hammer2_freemap_adjust(hmp, &parent->bref,
2129                                        HAMMER2_FREEMAP_DORECOVER);
2130         }
2131
2132         /*
2133          * Check type for recursive scan
2134          */
2135         switch(parent->bref.type) {
2136         case HAMMER2_BREF_TYPE_VOLUME:
2137                 /* data already instantiated */
2138                 break;
2139         case HAMMER2_BREF_TYPE_INODE:
2140                 /*
2141                  * Must instantiate data for DIRECTDATA test and also
2142                  * for recursion.
2143                  */
2144                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2145                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2146                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2147                         /* not applicable to recovery scan */
2148                         hammer2_chain_unlock(parent);
2149                         return 0;
2150                 }
2151                 hammer2_chain_unlock(parent);
2152                 break;
2153         case HAMMER2_BREF_TYPE_INDIRECT:
2154                 /*
2155                  * Must instantiate data for recursion
2156                  */
2157                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2158                 hammer2_chain_unlock(parent);
2159                 break;
2160         case HAMMER2_BREF_TYPE_DIRENT:
2161         case HAMMER2_BREF_TYPE_DATA:
2162         case HAMMER2_BREF_TYPE_FREEMAP:
2163         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2164         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2165                 /* not applicable to recovery scan */
2166                 return 0;
2167                 break;
2168         default:
2169                 return HAMMER2_ERROR_BADBREF;
2170         }
2171
2172         /*
2173          * Defer operation if depth limit reached or if we are crossing a
2174          * PFS boundary.
2175          */
2176         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2177                 struct hammer2_recovery_elm *elm;
2178
2179                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2180                 elm->chain = parent;
2181                 elm->sync_tid = sync_tid;
2182                 hammer2_chain_ref(parent);
2183                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2184                 /* unlocked by caller */
2185
2186                 return(0);
2187         }
2188
2189
2190         /*
2191          * Recursive scan of the last flushed transaction only.  We are
2192          * doing this without pmp assignments so don't leave the chains
2193          * hanging around after we are done with them.
2194          *
2195          * error        Cumulative error this level only
2196          * rup_error    Cumulative error for recursion
2197          * tmp_error    Specific non-cumulative recursion error
2198          */
2199         chain = NULL;
2200         first = 1;
2201         rup_error = 0;
2202         error = 0;
2203
2204         for (;;) {
2205                 error |= hammer2_chain_scan(parent, &chain, &bref,
2206                                             &first,
2207                                             HAMMER2_LOOKUP_NODATA);
2208
2209                 /*
2210                  * Problem during scan or EOF
2211                  */
2212                 if (error)
2213                         break;
2214
2215                 /*
2216                  * If this is a leaf
2217                  */
2218                 if (chain == NULL) {
2219                         if (bref.mirror_tid > sync_tid) {
2220                                 hammer2_freemap_adjust(hmp, &bref,
2221                                                      HAMMER2_FREEMAP_DORECOVER);
2222                         }
2223                         continue;
2224                 }
2225
2226                 /*
2227                  * This may or may not be a recursive node.
2228                  */
2229                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2230                 if (bref.mirror_tid > sync_tid) {
2231                         ++info->depth;
2232                         tmp_error = hammer2_recovery_scan(hmp, chain,
2233                                                            info, sync_tid);
2234                         --info->depth;
2235                 } else {
2236                         tmp_error = 0;
2237                 }
2238
2239                 /*
2240                  * Flush the recovery at the PFS boundary to stage it for
2241                  * the final flush of the super-root topology.
2242                  */
2243                 if (tmp_error == 0 &&
2244                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2245                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2246                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2247                 }
2248                 rup_error |= tmp_error;
2249         }
2250         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2251 }
2252
2253 /*
2254  * Sync a mount point; this is called on a per-mount basis from the
2255  * filesystem syncer process periodically and whenever a user issues
2256  * a sync.
2257  */
2258 int
2259 hammer2_vfs_sync(struct mount *mp, int waitfor)
2260 {
2261         hammer2_xop_flush_t *xop;
2262         struct hammer2_sync_info info;
2263         hammer2_inode_t *iroot;
2264         hammer2_pfs_t *pmp;
2265         int flags;
2266         int error;
2267
2268         pmp = MPTOPMP(mp);
2269         iroot = pmp->iroot;
2270         KKASSERT(iroot);
2271         KKASSERT(iroot->pmp == pmp);
2272
2273         /*
2274          * We can't acquire locks on existing vnodes while in a transaction
2275          * without risking a deadlock.  This assumes that vfsync() can be
2276          * called without the vnode locked (which it can in DragonFly).
2277          * Otherwise we'd have to implement a multi-pass or flag the lock
2278          * failures and retry.
2279          *
2280          * The reclamation code interlocks with the sync list's token
2281          * (by removing the vnode from the scan list) before unlocking
2282          * the inode, giving us time to ref the inode.
2283          */
2284         /*flags = VMSC_GETVP;*/
2285         flags = 0;
2286         if (waitfor & MNT_LAZY)
2287                 flags |= VMSC_ONEPASS;
2288
2289         /*
2290          * Preflush the vnodes using a normal transaction before interlocking
2291          * with a flush transaction.  We do this to try to run as much of
2292          * the compression as possible outside the flush transaction.
2293          *
2294          * For efficiency do an async pass before making sure with a
2295          * synchronous pass on all related buffer cache buffers.
2296          */
2297         hammer2_trans_init(pmp, 0);
2298         info.error = 0;
2299         info.waitfor = MNT_NOWAIT;
2300         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2301         info.waitfor = MNT_WAIT;
2302         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2303         hammer2_trans_done(pmp);
2304
2305         /*
2306          * Start our flush transaction.  This does not return until all
2307          * concurrent transactions have completed and will prevent any
2308          * new transactions from running concurrently, except for the
2309          * buffer cache transactions.
2310          *
2311          * (1) vfsync() all dirty vnodes via vfsyncscan().
2312          *
2313          * (2) Flush any remaining dirty inodes (the sideq), including any
2314          *     which may have been created during or raced against the
2315          *     vfsync().  To catch all cases this must be done after the
2316          *     vfsync().
2317          *
2318          * (3) Wait for any pending BIO I/O to complete (hammer2_bioq_sync()).
2319          *
2320          * NOTE! It is still possible for the paging code to push pages
2321          *       out via a UIO_NOCOPY hammer2_vop_write() during the main
2322          *       flush.
2323          */
2324         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2325
2326         info.error = 0;
2327         info.waitfor = MNT_NOWAIT;
2328         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2329         info.waitfor = MNT_WAIT;
2330         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2331         hammer2_inode_run_sideq(pmp, 1);
2332         hammer2_bioq_sync(pmp);
2333
2334         /*
2335          * Use the XOP interface to concurrently flush all nodes to
2336          * synchronize the PFSROOT subtopology to the media.  A standard
2337          * end-of-scan ENOENT error indicates cluster sufficiency.
2338          *
2339          * Note that this flush will not be visible on crash recovery until
2340          * we flush the super-root topology in the next loop.
2341          *
2342          * XXX For now wait for all flushes to complete.
2343          */
2344         if (iroot) {
2345                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2346                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2347                 error = hammer2_xop_collect(&xop->head,
2348                                             HAMMER2_XOP_COLLECT_WAITALL);
2349                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2350                 if (error == HAMMER2_ERROR_ENOENT)
2351                         error = 0;
2352                 else
2353                         error = hammer2_error_to_errno(error);
2354         } else {
2355                 error = 0;
2356         }
2357         hammer2_trans_done(pmp);
2358
2359         return (error);
2360 }
2361
2362 /*
2363  * Sync passes.
2364  *
2365  * Note that we ignore the tranasction mtid we got above.  Instead,
2366  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2367  * transactions.
2368  */
2369 static int
2370 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2371 {
2372         struct hammer2_sync_info *info = data;
2373         hammer2_inode_t *ip;
2374         int error;
2375
2376         /*
2377          * Degenerate cases.  Note that ip == NULL typically means the
2378          * syncer vnode itself and we don't want to vclrisdirty() in that
2379          * situation.
2380          */
2381         ip = VTOI(vp);
2382         if (ip == NULL) {
2383                 return(0);
2384         }
2385         if (vp->v_type == VNON || vp->v_type == VBAD) {
2386                 vclrisdirty(vp);
2387                 return(0);
2388         }
2389
2390         /*
2391          * VOP_FSYNC will start a new transaction so replicate some code
2392          * here to do it inline (see hammer2_vop_fsync()).
2393          *
2394          * WARNING: The vfsync interacts with the buffer cache and might
2395          *          block, we can't hold the inode lock at that time.
2396          *          However, we MUST ref ip before blocking to ensure that
2397          *          it isn't ripped out from under us (since we do not
2398          *          hold a lock on the vnode).
2399          */
2400         hammer2_inode_ref(ip);
2401         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2402             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2403                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2404                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2405                                  HAMMER2_INODE_MODIFIED)) {
2406                         hammer2_inode_lock(ip, 0);
2407                         if (ip->flags & (HAMMER2_INODE_RESIZED |
2408                                          HAMMER2_INODE_MODIFIED)) {
2409                                 hammer2_inode_chain_sync(ip);
2410                         }
2411                         hammer2_inode_unlock(ip);
2412                 }
2413         }
2414         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2415             RB_EMPTY(&vp->v_rbdirty_tree)) {
2416                 vclrisdirty(vp);
2417         }
2418
2419         hammer2_inode_drop(ip);
2420 #if 1
2421         error = 0;
2422         if (error)
2423                 info->error = error;
2424 #endif
2425         return(0);
2426 }
2427
2428 static
2429 int
2430 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2431 {
2432         hammer2_inode_t *ip;
2433
2434         KKASSERT(MAXFIDSZ >= 16);
2435         ip = VTOI(vp);
2436         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2437         fhp->fid_ext = 0;
2438         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2439         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2440
2441         return 0;
2442 }
2443
2444 static
2445 int
2446 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2447                struct fid *fhp, struct vnode **vpp)
2448 {
2449         hammer2_pfs_t *pmp;
2450         hammer2_tid_t inum;
2451         int error;
2452
2453         pmp = MPTOPMP(mp);
2454         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2455         if (vpp) {
2456                 if (inum == 1)
2457                         error = hammer2_vfs_root(mp, vpp);
2458                 else
2459                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2460         } else {
2461                 error = 0;
2462         }
2463         if (error)
2464                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2465         return error;
2466 }
2467
2468 static
2469 int
2470 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2471                  int *exflagsp, struct ucred **credanonp)
2472 {
2473         hammer2_pfs_t *pmp;
2474         struct netcred *np;
2475         int error;
2476
2477         pmp = MPTOPMP(mp);
2478         np = vfs_export_lookup(mp, &pmp->export, nam);
2479         if (np) {
2480                 *exflagsp = np->netc_exflags;
2481                 *credanonp = &np->netc_anon;
2482                 error = 0;
2483         } else {
2484                 error = EACCES;
2485         }
2486         return error;
2487 }
2488
2489 /*
2490  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2491  * header into the HMP
2492  *
2493  * XXX read four volhdrs and use the one with the highest TID whos CRC
2494  *     matches.
2495  *
2496  * XXX check iCRCs.
2497  *
2498  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2499  *     nonexistant locations.
2500  *
2501  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2502  */
2503 static
2504 int
2505 hammer2_install_volume_header(hammer2_dev_t *hmp)
2506 {
2507         hammer2_volume_data_t *vd;
2508         struct buf *bp;
2509         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2510         int error_reported;
2511         int error;
2512         int valid;
2513         int i;
2514
2515         error_reported = 0;
2516         error = 0;
2517         valid = 0;
2518         bp = NULL;
2519
2520         /*
2521          * There are up to 4 copies of the volume header (syncs iterate
2522          * between them so there is no single master).  We don't trust the
2523          * volu_size field so we don't know precisely how large the filesystem
2524          * is, so depend on the OS to return an error if we go beyond the
2525          * block device's EOF.
2526          */
2527         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2528                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2529                               HAMMER2_VOLUME_BYTES, &bp);
2530                 if (error) {
2531                         brelse(bp);
2532                         bp = NULL;
2533                         continue;
2534                 }
2535
2536                 vd = (struct hammer2_volume_data *) bp->b_data;
2537                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2538                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2539                         brelse(bp);
2540                         bp = NULL;
2541                         continue;
2542                 }
2543
2544                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2545                         /* XXX: Reversed-endianness filesystem */
2546                         kprintf("hammer2: reverse-endian filesystem detected");
2547                         brelse(bp);
2548                         bp = NULL;
2549                         continue;
2550                 }
2551
2552                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2553                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2554                                       HAMMER2_VOLUME_ICRC0_SIZE);
2555                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2556                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2557                                        HAMMER2_VOLUME_ICRC1_SIZE);
2558                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2559                         kprintf("hammer2 volume header crc "
2560                                 "mismatch copy #%d %08x/%08x\n",
2561                                 i, crc0, crc);
2562                         error_reported = 1;
2563                         brelse(bp);
2564                         bp = NULL;
2565                         continue;
2566                 }
2567                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2568                         valid = 1;
2569                         hmp->voldata = *vd;
2570                         hmp->volhdrno = i;
2571                 }
2572                 brelse(bp);
2573                 bp = NULL;
2574         }
2575         if (valid) {
2576                 hmp->volsync = hmp->voldata;
2577                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2578                 error = 0;
2579                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2580                         kprintf("hammer2: using volume header #%d\n",
2581                                 hmp->volhdrno);
2582                 }
2583         } else {
2584                 error = EINVAL;
2585                 kprintf("hammer2: no valid volume headers found!\n");
2586         }
2587         return (error);
2588 }
2589
2590 /*
2591  * This handles hysteresis on regular file flushes.  Because the BIOs are
2592  * routed to a thread it is possible for an excessive number to build up
2593  * and cause long front-end stalls long before the runningbuffspace limit
2594  * is hit, so we implement hammer2_flush_pipe to control the
2595  * hysteresis.
2596  *
2597  * This is a particular problem when compression is used.
2598  */
2599 void
2600 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2601 {
2602         atomic_add_int(&pmp->count_lwinprog, 1);
2603 }
2604
2605 void
2606 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2607 {
2608         int lwinprog;
2609
2610         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2611         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2612             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2613                 atomic_clear_int(&pmp->count_lwinprog,
2614                                  HAMMER2_LWINPROG_WAITING);
2615                 wakeup(&pmp->count_lwinprog);
2616         }
2617         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2618             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2619                 atomic_clear_int(&pmp->count_lwinprog,
2620                                  HAMMER2_LWINPROG_WAITING0);
2621                 wakeup(&pmp->count_lwinprog);
2622         }
2623 }
2624
2625 void
2626 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2627 {
2628         int lwinprog;
2629         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2630                                     HAMMER2_LWINPROG_WAITING0;
2631
2632         for (;;) {
2633                 lwinprog = pmp->count_lwinprog;
2634                 cpu_ccfence();
2635                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2636                         break;
2637                 tsleep_interlock(&pmp->count_lwinprog, 0);
2638                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2639                 lwinprog = pmp->count_lwinprog;
2640                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2641                         break;
2642                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2643         }
2644 }
2645
2646 /*
2647  * Manage excessive memory resource use for chain and related
2648  * structures.
2649  */
2650 void
2651 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2652 {
2653         uint32_t waiting;
2654         uint32_t count;
2655         uint32_t limit;
2656 #if 0
2657         static int zzticks;
2658 #endif
2659
2660         /*
2661          * Atomic check condition and wait.  Also do an early speedup of
2662          * the syncer to try to avoid hitting the wait.
2663          */
2664         for (;;) {
2665                 waiting = pmp->inmem_dirty_chains;
2666                 cpu_ccfence();
2667                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2668
2669                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2670                 if (limit < hammer2_limit_dirty_chains)
2671                         limit = hammer2_limit_dirty_chains;
2672                 if (limit < 1000)
2673                         limit = 1000;
2674
2675 #if 0
2676                 if ((int)(ticks - zzticks) > hz) {
2677                         zzticks = ticks;
2678                         kprintf("count %ld %ld\n", count, limit);
2679                 }
2680 #endif
2681
2682                 /*
2683                  * Block if there are too many dirty chains present, wait
2684                  * for the flush to clean some out.
2685                  */
2686                 if (count > limit) {
2687                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2688                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2689                                                waiting,
2690                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2691                                 speedup_syncer(pmp->mp);
2692                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2693                                        "chnmem", hz);
2694                         }
2695                         continue;       /* loop on success or fail */
2696                 }
2697
2698                 /*
2699                  * Try to start an early flush before we are forced to block.
2700                  */
2701                 if (count > limit * 7 / 10)
2702                         speedup_syncer(pmp->mp);
2703                 break;
2704         }
2705 }
2706
2707 void
2708 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2709 {
2710         if (pmp) {
2711                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2712         }
2713 }
2714
2715 void
2716 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2717 {
2718         uint32_t waiting;
2719
2720         if (pmp == NULL)
2721                 return;
2722
2723         for (;;) {
2724                 waiting = pmp->inmem_dirty_chains;
2725                 cpu_ccfence();
2726                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2727                                        waiting,
2728                                        (waiting - 1) &
2729                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2730                         break;
2731                 }
2732         }
2733
2734         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2735                 wakeup(&pmp->inmem_dirty_chains);
2736 }
2737
2738 /*
2739  * Returns 0 if the filesystem has tons of free space
2740  * Returns 1 if the filesystem has less than 10% remaining
2741  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2742  */
2743 int
2744 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2745 {
2746         hammer2_pfs_t *pmp;
2747         hammer2_dev_t *hmp;
2748         hammer2_off_t free_reserved;
2749         hammer2_off_t free_nominal;
2750         int i;
2751
2752         pmp = ip->pmp;
2753
2754         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2755                 free_reserved = HAMMER2_SEGSIZE;
2756                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2757                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2758                         hmp = pmp->pfs_hmps[i];
2759                         if (hmp == NULL)
2760                                 continue;
2761                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2762                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2763                                 continue;
2764
2765                         if (free_nominal > hmp->voldata.allocator_free)
2766                                 free_nominal = hmp->voldata.allocator_free;
2767                         if (free_reserved < hmp->free_reserved)
2768                                 free_reserved = hmp->free_reserved;
2769                 }
2770
2771                 /*
2772                  * SMP races ok
2773                  */
2774                 pmp->free_reserved = free_reserved;
2775                 pmp->free_nominal = free_nominal;
2776                 pmp->free_ticks = ticks;
2777         } else {
2778                 free_reserved = pmp->free_reserved;
2779                 free_nominal = pmp->free_nominal;
2780         }
2781         if (cred && cred->cr_uid != 0) {
2782                 if ((int64_t)(free_nominal - bytes) <
2783                     (int64_t)free_reserved) {
2784                         return 2;
2785                 }
2786         } else {
2787                 if ((int64_t)(free_nominal - bytes) <
2788                     (int64_t)free_reserved / 2) {
2789                         return 2;
2790                 }
2791         }
2792         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2793                 return 1;
2794         return 0;
2795 }
2796
2797 /*
2798  * Debugging
2799  */
2800 void
2801 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2802 {
2803         hammer2_chain_t *scan;
2804         hammer2_chain_t *parent;
2805
2806         --*countp;
2807         if (*countp == 0) {
2808                 kprintf("%*.*s...\n", tab, tab, "");
2809                 return;
2810         }
2811         if (*countp < 0)
2812                 return;
2813         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2814                 tab, tab, "", pfx,
2815                 chain, chain->bref.type,
2816                 chain->bref.key, chain->bref.keybits,
2817                 chain->bref.mirror_tid);
2818
2819         kprintf("%*.*s      [%08x] (%s) refs=%d",
2820                 tab, tab, "",
2821                 chain->flags,
2822                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2823                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2824                 chain->refs);
2825
2826         parent = chain->parent;
2827         if (parent)
2828                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2829                         tab, tab, "",
2830                         parent, parent->flags, parent->refs);
2831         if (RB_EMPTY(&chain->core.rbtree)) {
2832                 kprintf("\n");
2833         } else {
2834                 kprintf(" {\n");
2835                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2836                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2837                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2838                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2839                                 chain->data->ipdata.filename);
2840                 else
2841                         kprintf("%*.*s}\n", tab, tab, "");
2842         }
2843 }