sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72 };
  73
  74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  75 static struct hammer2_mntlist hammer2_mntlist;
  76
  77 struct hammer2_pfslist hammer2_pfslist;
  78 struct lock hammer2_mntlk;
  79
  80 int hammer2_debug;
  81 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
  82 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
  83 int hammer2_cluster_write = 4;          /* physical clustering (not file vp) */
  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, debug, CTLFLAG_RW,
 122            &hammer2_debug, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 124            &hammer2_cluster_meta_read, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 126            &hammer2_cluster_data_read, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 128            &hammer2_cluster_write, 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                 if (hmp->spmp == pmp)
 709                         hmp->spmp = NULL;
 710
 711                 /*
 712                  * Determine if this PFS is affected.  If it is we must
 713                  * freeze all management threads and lock its iroot.
 714                  *
 715                  * Freezing a management thread forces it idle, operations
 716                  * in-progress will be aborted and it will have to start
 717                  * over again when unfrozen, or exit if told to exit.
 718                  */
 719                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 720                         if (pmp->pfs_hmps[i] == hmp)
 721                                 break;
 722                 }
 723                 if (i != HAMMER2_MAXCLUSTER) {
 724                         /*
 725                          * Make sure all synchronization threads are locked
 726                          * down.
 727                          */
 728                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 729                                 if (pmp->pfs_hmps[i] == NULL)
 730                                         continue;
 731                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 732                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 733                                         hammer2_thr_freeze_async(
 734                                                 &pmp->xop_groups[j].thrs[i]);
 735                                 }
 736                         }
 737                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 738                                 if (pmp->pfs_hmps[i] == NULL)
 739                                         continue;
 740                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 741                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 742                                         hammer2_thr_freeze(
 743                                                 &pmp->xop_groups[j].thrs[i]);
 744                                 }
 745                         }
 746
 747                         /*
 748                          * Lock the inode and clean out matching chains.
 749                          * Note that we cannot use hammer2_inode_lock_*()
 750                          * here because that would attempt to validate the
 751                          * cluster that we are in the middle of ripping
 752                          * apart.
 753                          *
 754                          * WARNING! We are working directly on the inodes
 755                          *          embedded cluster.
 756                          */
 757                         hammer2_mtx_ex(&iroot->lock);
 758
 759                         /*
 760                          * Remove the chain from matching elements of the PFS.
 761                          */
 762                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 763                                 if (pmp->pfs_hmps[i] != hmp)
 764                                         continue;
 765                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 766                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 767                                         hammer2_thr_delete(
 768                                                 &pmp->xop_groups[j].thrs[i]);
 769                                 }
 770                                 rchain = iroot->cluster.array[i].chain;
 771                                 iroot->cluster.array[i].chain = NULL;
 772                                 pmp->pfs_types[i] = 0;
 773                                 if (pmp->pfs_names[i]) {
 774                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 775                                         pmp->pfs_names[i] = NULL;
 776                                 }
 777                                 if (rchain) {
 778                                         hammer2_chain_drop(rchain);
 779                                         /* focus hint */
 780                                         if (iroot->cluster.focus == rchain)
 781                                                 iroot->cluster.focus = NULL;
 782                                 }
 783                                 pmp->pfs_hmps[i] = NULL;
 784                         }
 785                         hammer2_mtx_unlock(&iroot->lock);
 786                         didfreeze = 1;  /* remaster, unfreeze down below */
 787                 } else {
 788                         didfreeze = 0;
 789                 }
 790
 791                 /*
 792                  * Cleanup trailing chains.  Gaps may remain.
 793                  */
 794                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 795                         if (pmp->pfs_hmps[i])
 796                                 break;
 797                 }
 798                 iroot->cluster.nchains = i + 1;
 799
 800                 /*
 801                  * If the PMP has no elements remaining we can destroy it.
 802                  * (this will transition management threads from frozen->exit).
 803                  */
 804                 if (iroot->cluster.nchains == 0) {
 805                         hammer2_pfsfree(pmp);
 806                         goto again;
 807                 }
 808
 809                 /*
 810                  * If elements still remain we need to set the REMASTER
 811                  * flag and unfreeze it.
 812                  */
 813                 if (didfreeze) {
 814                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 815                                 if (pmp->pfs_hmps[i] == NULL)
 816                                         continue;
 817                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 818                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 819                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 820                                         hammer2_thr_remaster(
 821                                                 &pmp->xop_groups[j].thrs[i]);
 822                                         hammer2_thr_unfreeze(
 823                                                 &pmp->xop_groups[j].thrs[i]);
 824                                 }
 825                         }
 826                 }
 827         }
 828 }
 829
 830 /*
 831  * Mount or remount HAMMER2 fileystem from physical media
 832  *
 833  *      mountroot
 834  *              mp              mount point structure
 835  *              path            NULL
 836  *              data            <unused>
 837  *              cred            <unused>
 838  *
 839  *      mount
 840  *              mp              mount point structure
 841  *              path            path to mount point
 842  *              data            pointer to argument structure in user space
 843  *                      volume  volume path (device@LABEL form)
 844  *                      hflags  user mount flags
 845  *              cred            user credentials
 846  *
 847  * RETURNS:     0       Success
 848  *              !0      error number
 849  */
 850 static
 851 int
 852 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 853                   struct ucred *cred)
 854 {
 855         struct hammer2_mount_info info;
 856         hammer2_pfs_t *pmp;
 857         hammer2_pfs_t *spmp;
 858         hammer2_dev_t *hmp;
 859         hammer2_dev_t *force_local;
 860         hammer2_key_t key_next;
 861         hammer2_key_t key_dummy;
 862         hammer2_key_t lhc;
 863         struct vnode *devvp;
 864         struct nlookupdata nd;
 865         hammer2_chain_t *parent;
 866         hammer2_chain_t *chain;
 867         hammer2_cluster_t *cluster;
 868         const hammer2_inode_data_t *ripdata;
 869         hammer2_blockref_t bref;
 870         struct file *fp;
 871         char devstr[MNAMELEN];
 872         size_t size;
 873         size_t done;
 874         char *dev;
 875         char *label;
 876         int ronly = 1;
 877         int error;
 878         int i;
 879
 880         hmp = NULL;
 881         pmp = NULL;
 882         dev = NULL;
 883         label = NULL;
 884         devvp = NULL;
 885
 886         kprintf("hammer2_mount\n");
 887
 888         if (path == NULL) {
 889                 /*
 890                  * Root mount
 891                  */
 892                 bzero(&info, sizeof(info));
 893                 info.cluster_fd = -1;
 894                 ksnprintf(devstr, sizeof(devstr), "%s",
 895                           mp->mnt_stat.f_mntfromname);
 896                 kprintf("hammer2_mount: root '%s'\n", devstr);
 897         } else {
 898                 /*
 899                  * Non-root mount or updating a mount
 900                  */
 901                 error = copyin(data, &info, sizeof(info));
 902                 if (error)
 903                         return (error);
 904
 905                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 906                 if (error)
 907                         return (error);
 908         }
 909
 910         /*
 911          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 912          * if no label specified, based on the partition id.  Error out if no
 913          * label or device (with partition id) is specified.  This is strictly
 914          * a convenience to match the default label created by newfs_hammer2,
 915          * our preference is that a label always be specified.
 916          *
 917          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
 918          *       that does not specify a device, as long as some H2 label
 919          *       has already been mounted from that device.  This makes
 920          *       mounting snapshots a lot easier.
 921          */
 922         dev = devstr;
 923         label = strchr(devstr, '@');
 924         if (label && ((label + 1) - dev) > done)
 925                 return (EINVAL);
 926         if (label == NULL || label[1] == 0) {
 927                 char slice;
 928
 929                 if (label == NULL)
 930                         label = devstr + strlen(devstr);
 931                 slice = label[-1];
 932                 switch(slice) {
 933                 case 'a':
 934                         label = "BOOT";
 935                         break;
 936                 case 'd':
 937                         label = "ROOT";
 938                         break;
 939                 default:
 940                         label = "DATA";
 941                         break;
 942                 }
 943         } else {
 944                 *label = '\0';
 945                 label++;
 946         }
 947
 948         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
 949                 dev, label, (mp->mnt_flag & MNT_RDONLY));
 950
 951         if (mp->mnt_flag & MNT_UPDATE) {
 952                 /*
 953                  * Update mount.  Note that pmp->iroot->cluster is
 954                  * an inode-embedded cluster and thus cannot be
 955                  * directly locked.
 956                  *
 957                  * XXX HAMMER2 needs to implement NFS export via
 958                  *     mountctl.
 959                  */
 960                 pmp = MPTOPMP(mp);
 961                 pmp->hflags = info.hflags;
 962                 cluster = &pmp->iroot->cluster;
 963                 for (i = 0; i < cluster->nchains; ++i) {
 964                         if (cluster->array[i].chain == NULL)
 965                                 continue;
 966                         hmp = cluster->array[i].chain->hmp;
 967                         devvp = hmp->devvp;
 968                         error = hammer2_remount(hmp, mp, path,
 969                                                 devvp, cred);
 970                         if (error)
 971                                 break;
 972                 }
 973
 974                 return error;
 975         }
 976
 977         /*
 978          * HMP device mount
 979          *
 980          * If a path is specified and dev is not an empty string, lookup the
 981          * name and verify that it referes to a block device.
 982          *
 983          * If a path is specified and dev is an empty string we fall through
 984          * and locate the label in the hmp search.
 985          */
 986         if (path && *dev != 0) {
 987                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 988                 if (error == 0)
 989                         error = nlookup(&nd);
 990                 if (error == 0)
 991                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 992                 nlookup_done(&nd);
 993         } else if (path == NULL) {
 994                 /* root mount */
 995                 cdev_t cdev = kgetdiskbyname(dev);
 996                 error = bdevvp(cdev, &devvp);
 997                 if (error)
 998                         kprintf("hammer2: cannot find '%s'\n", dev);
 999         } else {
1000                 /*
1001                  * We will locate the hmp using the label in the hmp loop.
1002                  */
1003                 error = 0;
1004         }
1005
1006         /*
1007          * Make sure its a block device.  Do not check to see if it is
1008          * already mounted until we determine that its a fresh H2 device.
1009          */
1010         if (error == 0 && devvp) {
1011                 vn_isdisk(devvp, &error);
1012         }
1013
1014         /*
1015          * Determine if the device has already been mounted.  After this
1016          * check hmp will be non-NULL if we are doing the second or more
1017          * hammer2 mounts from the same device.
1018          */
1019         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1020         if (devvp) {
1021                 /*
1022                  * Match the device.  Due to the way devfs works,
1023                  * we may not be able to directly match the vnode pointer,
1024                  * so also check to see if the underlying device matches.
1025                  */
1026                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1027                         if (hmp->devvp == devvp)
1028                                 break;
1029                         if (devvp->v_rdev &&
1030                             hmp->devvp->v_rdev == devvp->v_rdev) {
1031                                 break;
1032                         }
1033                 }
1034
1035                 /*
1036                  * If no match this may be a fresh H2 mount, make sure
1037                  * the device is not mounted on anything else.
1038                  */
1039                 if (hmp == NULL)
1040                         error = vfs_mountedon(devvp);
1041         } else if (error == 0) {
1042                 /*
1043                  * Match the label to a pmp already probed.
1044                  */
1045                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1046                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1047                                 if (pmp->pfs_names[i] &&
1048                                     strcmp(pmp->pfs_names[i], label) == 0) {
1049                                         hmp = pmp->pfs_hmps[i];
1050                                         break;
1051                                 }
1052                         }
1053                         if (hmp)
1054                                 break;
1055                 }
1056                 if (hmp == NULL)
1057                         error = ENOENT;
1058         }
1059
1060         /*
1061          * Open the device if this isn't a secondary mount and construct
1062          * the H2 device mount (hmp).
1063          */
1064         if (hmp == NULL) {
1065                 hammer2_chain_t *schain;
1066                 hammer2_xid_t xid;
1067
1068                 if (error == 0 && vcount(devvp) > 0) {
1069                         kprintf("Primary device already has references\n");
1070                         error = EBUSY;
1071                 }
1072
1073                 /*
1074                  * Now open the device
1075                  */
1076                 if (error == 0) {
1077                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1078                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1079                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1080                         if (error == 0) {
1081                                 error = VOP_OPEN(devvp,
1082                                              (ronly ? FREAD : FREAD | FWRITE),
1083                                              FSCRED, NULL);
1084                         }
1085                         vn_unlock(devvp);
1086                 }
1087                 if (error && devvp) {
1088                         vrele(devvp);
1089                         devvp = NULL;
1090                 }
1091                 if (error) {
1092                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1093                         return error;
1094                 }
1095                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1096                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1097                 hmp->ronly = ronly;
1098                 hmp->devvp = devvp;
1099                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1100                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1101                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1102                 RB_INIT(&hmp->iotree);
1103                 spin_init(&hmp->io_spin, "hm2mount_io");
1104                 spin_init(&hmp->list_spin, "hm2mount_list");
1105                 TAILQ_INIT(&hmp->flushq);
1106
1107                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1108                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1109                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1110
1111                 /*
1112                  * vchain setup. vchain.data is embedded.
1113                  * vchain.refs is initialized and will never drop to 0.
1114                  *
1115                  * NOTE! voldata is not yet loaded.
1116                  */
1117                 hmp->vchain.hmp = hmp;
1118                 hmp->vchain.refs = 1;
1119                 hmp->vchain.data = (void *)&hmp->voldata;
1120                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1121                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1122                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1123
1124                 hammer2_chain_core_init(&hmp->vchain);
1125                 /* hmp->vchain.u.xxx is left NULL */
1126
1127                 /*
1128                  * fchain setup.  fchain.data is embedded.
1129                  * fchain.refs is initialized and will never drop to 0.
1130                  *
1131                  * The data is not used but needs to be initialized to
1132                  * pass assertion muster.  We use this chain primarily
1133                  * as a placeholder for the freemap's top-level RBTREE
1134                  * so it does not interfere with the volume's topology
1135                  * RBTREE.
1136                  */
1137                 hmp->fchain.hmp = hmp;
1138                 hmp->fchain.refs = 1;
1139                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1140                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1141                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1142                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1143                 hmp->fchain.bref.methods =
1144                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1145                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1146
1147                 hammer2_chain_core_init(&hmp->fchain);
1148                 /* hmp->fchain.u.xxx is left NULL */
1149
1150                 /*
1151                  * Install the volume header and initialize fields from
1152                  * voldata.
1153                  */
1154                 error = hammer2_install_volume_header(hmp);
1155                 if (error) {
1156                         hammer2_unmount_helper(mp, NULL, hmp);
1157                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1158                         hammer2_vfs_unmount(mp, MNT_FORCE);
1159                         return error;
1160                 }
1161
1162                 /*
1163                  * Really important to get these right or flush will get
1164                  * confused.
1165                  */
1166                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1167                 kprintf("alloc spmp %p tid %016jx\n",
1168                         hmp->spmp, hmp->voldata.mirror_tid);
1169                 spmp = hmp->spmp;
1170
1171                 /*
1172                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1173                  * is inherited from the volume header.
1174                  */
1175                 xid = 0;
1176                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1177                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1178                 hmp->vchain.pmp = spmp;
1179                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1180                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1181                 hmp->fchain.pmp = spmp;
1182
1183                 /*
1184                  * First locate the super-root inode, which is key 0
1185                  * relative to the volume header's blockset.
1186                  *
1187                  * Then locate the root inode by scanning the directory keyspace
1188                  * represented by the label.
1189                  */
1190                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1191                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1192                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1193                                       &error, 0);
1194                 hammer2_chain_lookup_done(parent);
1195                 if (schain == NULL) {
1196                         kprintf("hammer2_mount: invalid super-root\n");
1197                         hammer2_unmount_helper(mp, NULL, hmp);
1198                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1199                         hammer2_vfs_unmount(mp, MNT_FORCE);
1200                         return EINVAL;
1201                 }
1202                 if (schain->error) {
1203                         kprintf("hammer2_mount: error %s reading super-root\n",
1204                                 hammer2_error_str(schain->error));
1205                         hammer2_chain_unlock(schain);
1206                         hammer2_chain_drop(schain);
1207                         schain = NULL;
1208                         hammer2_unmount_helper(mp, NULL, hmp);
1209                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1210                         hammer2_vfs_unmount(mp, MNT_FORCE);
1211                         return EINVAL;
1212                 }
1213
1214                 /*
1215                  * The super-root always uses an inode_tid of 1 when
1216                  * creating PFSs.
1217                  */
1218                 spmp->inode_tid = 1;
1219                 spmp->modify_tid = schain->bref.modify_tid + 1;
1220
1221                 /*
1222                  * Sanity-check schain's pmp and finish initialization.
1223                  * Any chain belonging to the super-root topology should
1224                  * have a NULL pmp (not even set to spmp).
1225                  */
1226                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1227                 KKASSERT(schain->pmp == NULL);
1228                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1229
1230                 /*
1231                  * Replace the dummy spmp->iroot with a real one.  It's
1232                  * easier to just do a wholesale replacement than to try
1233                  * to update the chain and fixup the iroot fields.
1234                  *
1235                  * The returned inode is locked with the supplied cluster.
1236                  */
1237                 cluster = hammer2_cluster_from_chain(schain);
1238                 hammer2_inode_drop(spmp->iroot);
1239                 spmp->iroot = NULL;
1240                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1241                 spmp->spmp_hmp = hmp;
1242                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1243                 spmp->pfs_hmps[0] = hmp;
1244                 hammer2_inode_ref(spmp->iroot);
1245                 hammer2_inode_unlock(spmp->iroot);
1246                 hammer2_cluster_unlock(cluster);
1247                 hammer2_cluster_drop(cluster);
1248                 schain = NULL;
1249                 /* leave spmp->iroot with one ref */
1250
1251                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1252                         error = hammer2_recovery(hmp);
1253                         /* XXX do something with error */
1254                 }
1255                 hammer2_update_pmps(hmp);
1256                 hammer2_iocom_init(hmp);
1257                 hammer2_bulkfree_init(hmp);
1258
1259                 /*
1260                  * Ref the cluster management messaging descriptor.  The mount
1261                  * program deals with the other end of the communications pipe.
1262                  *
1263                  * Root mounts typically do not supply one.
1264                  */
1265                 if (info.cluster_fd >= 0) {
1266                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1267                         if (fp) {
1268                                 hammer2_cluster_reconnect(hmp, fp);
1269                         } else {
1270                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1271                         }
1272                 }
1273         } else {
1274                 spmp = hmp->spmp;
1275                 if (info.hflags & HMNT2_DEVFLAGS) {
1276                         kprintf("hammer2: Warning: mount flags pertaining "
1277                                 "to the whole device may only be specified "
1278                                 "on the first mount of the device: %08x\n",
1279                                 info.hflags & HMNT2_DEVFLAGS);
1280                 }
1281         }
1282
1283         /*
1284          * Force local mount (disassociate all PFSs from their clusters).
1285          * Used primarily for debugging.
1286          */
1287         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1288
1289         /*
1290          * Lookup the mount point under the media-localized super-root.
1291          * Scanning hammer2_pfslist doesn't help us because it represents
1292          * PFS cluster ids which can aggregate several named PFSs together.
1293          *
1294          * cluster->pmp will incorrectly point to spmp and must be fixed
1295          * up later on.
1296          */
1297         hammer2_inode_lock(spmp->iroot, 0);
1298         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1299         lhc = hammer2_dirhash(label, strlen(label));
1300         chain = hammer2_chain_lookup(&parent, &key_next,
1301                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1302                                      &error, 0);
1303         while (chain) {
1304                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1305                     strcmp(label, chain->data->ipdata.filename) == 0) {
1306                         break;
1307                 }
1308                 chain = hammer2_chain_next(&parent, chain, &key_next,
1309                                             key_next,
1310                                             lhc + HAMMER2_DIRHASH_LOMASK,
1311                                             &error, 0);
1312         }
1313         if (parent) {
1314                 hammer2_chain_unlock(parent);
1315                 hammer2_chain_drop(parent);
1316         }
1317         hammer2_inode_unlock(spmp->iroot);
1318
1319         /*
1320          * PFS could not be found?
1321          */
1322         if (chain == NULL) {
1323                 if (error)
1324                         kprintf("hammer2_mount: PFS label I/O error\n");
1325                 else
1326                         kprintf("hammer2_mount: PFS label not found\n");
1327                 hammer2_unmount_helper(mp, NULL, hmp);
1328                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1329                 hammer2_vfs_unmount(mp, MNT_FORCE);
1330
1331                 return EINVAL;
1332         }
1333
1334         /*
1335          * Acquire the pmp structure (it should have already been allocated
1336          * via hammer2_update_pmps() so do not pass cluster in to add to
1337          * available chains).
1338          *
1339          * Check if the cluster has already been mounted.  A cluster can
1340          * only be mounted once, use null mounts to mount additional copies.
1341          */
1342         if (chain->error) {
1343                 kprintf("hammer2_mount: PFS label I/O error\n");
1344         } else {
1345                 ripdata = &chain->data->ipdata;
1346                 bref = chain->bref;
1347                 pmp = hammer2_pfsalloc(NULL, ripdata,
1348                                        bref.modify_tid, force_local);
1349         }
1350         hammer2_chain_unlock(chain);
1351         hammer2_chain_drop(chain);
1352
1353         /*
1354          * Finish the mount
1355          */
1356         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1357
1358         if (pmp->mp) {
1359                 kprintf("hammer2_mount: PFS already mounted!\n");
1360                 hammer2_unmount_helper(mp, NULL, hmp);
1361                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1362                 hammer2_vfs_unmount(mp, MNT_FORCE);
1363
1364                 return EBUSY;
1365         }
1366
1367         pmp->hflags = info.hflags;
1368         mp->mnt_flag |= MNT_LOCAL;
1369         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1370         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1371
1372         /*
1373          * required mount structure initializations
1374          */
1375         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1376         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1377
1378         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1379         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1380
1381         /*
1382          * Optional fields
1383          */
1384         mp->mnt_iosize_max = MAXPHYS;
1385
1386         /*
1387          * Connect up mount pointers.
1388          */
1389         hammer2_mount_helper(mp, pmp);
1390
1391         lockmgr(&hammer2_mntlk, LK_RELEASE);
1392
1393         /*
1394          * Finish setup
1395          */
1396         vfs_getnewfsid(mp);
1397         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1398         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1399         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1400
1401         if (path) {
1402                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1403                           MNAMELEN - 1, &size);
1404                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1405         } /* else root mount, already in there */
1406
1407         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1408         if (path) {
1409                 copyinstr(path, mp->mnt_stat.f_mntonname,
1410                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1411                           &size);
1412         } else {
1413                 /* root mount */
1414                 mp->mnt_stat.f_mntonname[0] = '/';
1415         }
1416
1417         /*
1418          * Initial statfs to prime mnt_stat.
1419          */
1420         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1421
1422         return 0;
1423 }
1424
1425 /*
1426  * Scan PFSs under the super-root and create hammer2_pfs structures.
1427  */
1428 static
1429 void
1430 hammer2_update_pmps(hammer2_dev_t *hmp)
1431 {
1432         const hammer2_inode_data_t *ripdata;
1433         hammer2_chain_t *parent;
1434         hammer2_chain_t *chain;
1435         hammer2_blockref_t bref;
1436         hammer2_dev_t *force_local;
1437         hammer2_pfs_t *spmp;
1438         hammer2_pfs_t *pmp;
1439         hammer2_key_t key_next;
1440         int error;
1441
1442         /*
1443          * Force local mount (disassociate all PFSs from their clusters).
1444          * Used primarily for debugging.
1445          */
1446         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1447
1448         /*
1449          * Lookup mount point under the media-localized super-root.
1450          *
1451          * cluster->pmp will incorrectly point to spmp and must be fixed
1452          * up later on.
1453          */
1454         spmp = hmp->spmp;
1455         hammer2_inode_lock(spmp->iroot, 0);
1456         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1457         chain = hammer2_chain_lookup(&parent, &key_next,
1458                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1459                                          &error, 0);
1460         while (chain) {
1461                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1462                         continue;
1463                 if (chain->error) {
1464                         kprintf("I/O error scanning PFS labels\n");
1465                 } else {
1466                         ripdata = &chain->data->ipdata;
1467                         bref = chain->bref;
1468
1469                         pmp = hammer2_pfsalloc(chain, ripdata,
1470                                                bref.modify_tid, force_local);
1471                 }
1472                 chain = hammer2_chain_next(&parent, chain, &key_next,
1473                                            key_next, HAMMER2_KEY_MAX,
1474                                            &error, 0);
1475         }
1476         if (parent) {
1477                 hammer2_chain_unlock(parent);
1478                 hammer2_chain_drop(parent);
1479         }
1480         hammer2_inode_unlock(spmp->iroot);
1481 }
1482
1483 static
1484 int
1485 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1486                 struct vnode *devvp, struct ucred *cred)
1487 {
1488         int error;
1489
1490         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1491                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1492                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1493                 vn_unlock(devvp);
1494                 error = hammer2_recovery(hmp);
1495                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1496                 if (error == 0) {
1497                         VOP_CLOSE(devvp, FREAD, NULL);
1498                         hmp->ronly = 0;
1499                 } else {
1500                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1501                 }
1502                 vn_unlock(devvp);
1503         } else {
1504                 error = 0;
1505         }
1506         return error;
1507 }
1508
1509 static
1510 int
1511 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1512 {
1513         hammer2_pfs_t *pmp;
1514         int flags;
1515         int error = 0;
1516
1517         pmp = MPTOPMP(mp);
1518
1519         if (pmp == NULL)
1520                 return(0);
1521
1522         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1523
1524         /*
1525          * If mount initialization proceeded far enough we must flush
1526          * its vnodes and sync the underlying mount points.  Three syncs
1527          * are required to fully flush the filesystem (freemap updates lag
1528          * by one flush, and one extra for safety).
1529          */
1530         if (mntflags & MNT_FORCE)
1531                 flags = FORCECLOSE;
1532         else
1533                 flags = 0;
1534         if (pmp->iroot) {
1535                 error = vflush(mp, 0, flags);
1536                 if (error)
1537                         goto failed;
1538                 hammer2_vfs_sync(mp, MNT_WAIT);
1539                 hammer2_vfs_sync(mp, MNT_WAIT);
1540                 hammer2_vfs_sync(mp, MNT_WAIT);
1541         }
1542
1543         /*
1544          * Cleanup the frontend support XOPS threads
1545          */
1546         hammer2_xop_helper_cleanup(pmp);
1547
1548         if (pmp->mp)
1549                 hammer2_unmount_helper(mp, pmp, NULL);
1550
1551         error = 0;
1552 failed:
1553         lockmgr(&hammer2_mntlk, LK_RELEASE);
1554
1555         return (error);
1556 }
1557
1558 /*
1559  * Mount helper, hook the system mount into our PFS.
1560  * The mount lock is held.
1561  *
1562  * We must bump the mount_count on related devices for any
1563  * mounted PFSs.
1564  */
1565 static
1566 void
1567 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1568 {
1569         hammer2_cluster_t *cluster;
1570         hammer2_chain_t *rchain;
1571         int i;
1572
1573         mp->mnt_data = (qaddr_t)pmp;
1574         pmp->mp = mp;
1575
1576         /*
1577          * After pmp->mp is set we have to adjust hmp->mount_count.
1578          */
1579         cluster = &pmp->iroot->cluster;
1580         for (i = 0; i < cluster->nchains; ++i) {
1581                 rchain = cluster->array[i].chain;
1582                 if (rchain == NULL)
1583                         continue;
1584                 ++rchain->hmp->mount_count;
1585                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1586                         rchain->hmp, rchain->hmp->mount_count);
1587         }
1588
1589         /*
1590          * Create missing Xop threads
1591          */
1592         hammer2_xop_helper_create(pmp);
1593 }
1594
1595 /*
1596  * Mount helper, unhook the system mount from our PFS.
1597  * The mount lock is held.
1598  *
1599  * If hmp is supplied a mount responsible for being the first to open
1600  * the block device failed and the block device and all PFSs using the
1601  * block device must be cleaned up.
1602  *
1603  * If pmp is supplied multiple devices might be backing the PFS and each
1604  * must be disconnected.  This might not be the last PFS using some of the
1605  * underlying devices.  Also, we have to adjust our hmp->mount_count
1606  * accounting for the devices backing the pmp which is now undergoing an
1607  * unmount.
1608  */
1609 static
1610 void
1611 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1612 {
1613         hammer2_cluster_t *cluster;
1614         hammer2_chain_t *rchain;
1615         struct vnode *devvp;
1616         int dumpcnt;
1617         int ronly;
1618         int i;
1619
1620         /*
1621          * If no device supplied this is a high-level unmount and we have to
1622          * to disconnect the mount, adjust mount_count, and locate devices
1623          * that might now have no mounts.
1624          */
1625         if (pmp) {
1626                 KKASSERT(hmp == NULL);
1627                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1628                 pmp->mp = NULL;
1629                 mp->mnt_data = NULL;
1630
1631                 /*
1632                  * After pmp->mp is cleared we have to account for
1633                  * mount_count.
1634                  */
1635                 cluster = &pmp->iroot->cluster;
1636                 for (i = 0; i < cluster->nchains; ++i) {
1637                         rchain = cluster->array[i].chain;
1638                         if (rchain == NULL)
1639                                 continue;
1640                         --rchain->hmp->mount_count;
1641                         /* scrapping hmp now may invalidate the pmp */
1642                 }
1643 again:
1644                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1645                         if (hmp->mount_count == 0) {
1646                                 hammer2_unmount_helper(NULL, NULL, hmp);
1647                                 goto again;
1648                         }
1649                 }
1650                 return;
1651         }
1652
1653         /*
1654          * Try to terminate the block device.  We can't terminate it if
1655          * there are still PFSs referencing it.
1656          */
1657         if (hmp->mount_count)
1658                 return;
1659
1660         /*
1661          * Decomission the network before we start messing with the
1662          * device and PFS.
1663          */
1664         hammer2_iocom_uninit(hmp);
1665
1666         hammer2_bulkfree_uninit(hmp);
1667         hammer2_pfsfree_scan(hmp);
1668         hammer2_dev_exlock(hmp);        /* XXX order */
1669
1670         /*
1671          * Cycle the volume data lock as a safety (probably not needed any
1672          * more).  To ensure everything is out we need to flush at least
1673          * three times.  (1) The running of the sideq can dirty the
1674          * filesystem, (2) A normal flush can dirty the freemap, and
1675          * (3) ensure that the freemap is fully synchronized.
1676          *
1677          * The next mount's recovery scan can clean everything up but we want
1678          * to leave the filesystem in a 100% clean state on a normal unmount.
1679          */
1680 #if 0
1681         hammer2_voldata_lock(hmp);
1682         hammer2_voldata_unlock(hmp);
1683 #endif
1684
1685         if ((hmp->vchain.flags | hmp->fchain.flags) &
1686             HAMMER2_CHAIN_FLUSH_MASK) {
1687                 kprintf("hammer2_unmount: chains left over "
1688                         "after final sync\n");
1689                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1690                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1691
1692                 if (hammer2_debug & 0x0010)
1693                         Debugger("entered debugger");
1694         }
1695
1696         KKASSERT(hmp->spmp == NULL);
1697
1698         /*
1699          * Finish up with the device vnode
1700          */
1701         if ((devvp = hmp->devvp) != NULL) {
1702                 ronly = hmp->ronly;
1703                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1704                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1705                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1706                         ronly);
1707                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1708                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1709                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1710                 hmp->devvp = NULL;
1711                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1712                 vn_unlock(devvp);
1713                 vrele(devvp);
1714                 devvp = NULL;
1715         }
1716
1717         /*
1718          * Clear vchain/fchain flags that might prevent final cleanup
1719          * of these chains.
1720          */
1721         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1722                 atomic_add_long(&hammer2_count_modified_chains, -1);
1723                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1724                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1725         }
1726         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1727                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1728         }
1729
1730         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1731                 atomic_add_long(&hammer2_count_modified_chains, -1);
1732                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1733                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1734         }
1735         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1736                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1737         }
1738
1739         /*
1740          * Final drop of embedded freemap root chain to
1741          * clean up fchain.core (fchain structure is not
1742          * flagged ALLOCATED so it is cleaned out and then
1743          * left to rot).
1744          */
1745         hammer2_chain_drop(&hmp->fchain);
1746
1747         /*
1748          * Final drop of embedded volume root chain to clean
1749          * up vchain.core (vchain structure is not flagged
1750          * ALLOCATED so it is cleaned out and then left to
1751          * rot).
1752          */
1753         dumpcnt = 50;
1754         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1755         dumpcnt = 50;
1756         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1757         hammer2_dev_unlock(hmp);
1758         hammer2_chain_drop(&hmp->vchain);
1759
1760         hammer2_io_cleanup(hmp, &hmp->iotree);
1761         if (hmp->iofree_count) {
1762                 kprintf("io_cleanup: %d I/O's left hanging\n",
1763                         hmp->iofree_count);
1764         }
1765
1766         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1767         kmalloc_destroy(&hmp->mchain);
1768         kfree(hmp, M_HAMMER2);
1769 }
1770
1771 int
1772 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1773                  ino_t ino, struct vnode **vpp)
1774 {
1775         hammer2_xop_lookup_t *xop;
1776         hammer2_pfs_t *pmp;
1777         hammer2_inode_t *ip;
1778         hammer2_tid_t inum;
1779         int error;
1780
1781         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1782
1783         error = 0;
1784         pmp = MPTOPMP(mp);
1785
1786         /*
1787          * Easy if we already have it cached
1788          */
1789         ip = hammer2_inode_lookup(pmp, inum);
1790         if (ip) {
1791                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1792                 *vpp = hammer2_igetv(ip, &error);
1793                 hammer2_inode_unlock(ip);
1794                 hammer2_inode_drop(ip);         /* from lookup */
1795
1796                 return error;
1797         }
1798
1799         /*
1800          * Otherwise we have to find the inode
1801          */
1802         xop = hammer2_xop_alloc(pmp->iroot, 0);
1803         xop->lhc = inum;
1804         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1805         error = hammer2_xop_collect(&xop->head, 0);
1806
1807         if (error == 0) {
1808                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1809                         kprintf("vget: no collect error but also no rdata\n");
1810                         kprintf("xop %p\n", xop);
1811                         while ((hammer2_debug & 0x80000) == 0) {
1812                                 tsleep(xop, PCATCH, "wait", hz * 10);
1813                         }
1814                         ip = NULL;
1815                 } else {
1816                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1817                 }
1818         }
1819         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1820
1821         if (ip) {
1822                 *vpp = hammer2_igetv(ip, &error);
1823                 hammer2_inode_unlock(ip);
1824         } else {
1825                 *vpp = NULL;
1826                 error = ENOENT;
1827         }
1828         return (error);
1829 }
1830
1831 static
1832 int
1833 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1834 {
1835         hammer2_pfs_t *pmp;
1836         struct vnode *vp;
1837         int error;
1838
1839         pmp = MPTOPMP(mp);
1840         if (pmp->iroot == NULL) {
1841                 *vpp = NULL;
1842                 return EINVAL;
1843         }
1844
1845         error = 0;
1846         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1847
1848         while (pmp->inode_tid == 0) {
1849                 hammer2_xop_ipcluster_t *xop;
1850                 hammer2_inode_meta_t *meta;
1851
1852                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1853                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1854                 error = hammer2_xop_collect(&xop->head, 0);
1855
1856                 if (error == 0) {
1857                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1858                         pmp->iroot->meta = *meta;
1859                         pmp->inode_tid = meta->pfs_inum + 1;
1860                         if (pmp->inode_tid < HAMMER2_INODE_START)
1861                                 pmp->inode_tid = HAMMER2_INODE_START;
1862                         pmp->modify_tid =
1863                                 xop->head.cluster.focus->bref.modify_tid + 1;
1864                         kprintf("PFS: Starting inode %jd\n",
1865                                 (intmax_t)pmp->inode_tid);
1866                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1867                                 pmp->inode_tid, pmp->modify_tid);
1868                         wakeup(&pmp->iroot);
1869
1870                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1871
1872                         /*
1873                          * Prime the mount info.
1874                          */
1875                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1876                         break;
1877                 }
1878
1879                 /*
1880                  * Loop, try again
1881                  */
1882                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1883                 hammer2_inode_unlock(pmp->iroot);
1884                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1885                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1886                 if (error == EINTR)
1887                         break;
1888         }
1889
1890         if (error) {
1891                 hammer2_inode_unlock(pmp->iroot);
1892                 *vpp = NULL;
1893         } else {
1894                 vp = hammer2_igetv(pmp->iroot, &error);
1895                 hammer2_inode_unlock(pmp->iroot);
1896                 *vpp = vp;
1897         }
1898
1899         return (error);
1900 }
1901
1902 /*
1903  * Filesystem status
1904  *
1905  * XXX incorporate ipdata->meta.inode_quota and data_quota
1906  */
1907 static
1908 int
1909 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1910 {
1911         hammer2_pfs_t *pmp;
1912         hammer2_dev_t *hmp;
1913         hammer2_blockref_t bref;
1914         int i;
1915
1916         /*
1917          * NOTE: iroot might not have validated the cluster yet.
1918          */
1919         pmp = MPTOPMP(mp);
1920
1921         mp->mnt_stat.f_files = 0;
1922         mp->mnt_stat.f_ffree = 0;
1923         mp->mnt_stat.f_blocks = 0;
1924         mp->mnt_stat.f_bfree = 0;
1925         mp->mnt_stat.f_bavail = 0;
1926
1927         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1928                 hmp = pmp->pfs_hmps[i];
1929                 if (hmp == NULL)
1930                         continue;
1931                 if (pmp->iroot->cluster.array[i].chain)
1932                         bref = pmp->iroot->cluster.array[i].chain->bref;
1933                 else
1934                         bzero(&bref, sizeof(bref));
1935
1936                 mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1937                 mp->mnt_stat.f_ffree = 0;
1938                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1939                                         mp->mnt_vstat.f_bsize;
1940                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1941                                         mp->mnt_vstat.f_bsize;
1942                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1943
1944                 if (cred && cred->cr_uid != 0) {
1945                         uint64_t adj;
1946
1947                         /* 5% */
1948                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
1949                         mp->mnt_stat.f_blocks -= adj;
1950                         mp->mnt_stat.f_bfree -= adj;
1951                         mp->mnt_stat.f_bavail -= adj;
1952                 }
1953
1954                 *sbp = mp->mnt_stat;
1955         }
1956         return (0);
1957 }
1958
1959 static
1960 int
1961 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1962 {
1963         hammer2_pfs_t *pmp;
1964         hammer2_dev_t *hmp;
1965         hammer2_blockref_t bref;
1966         int i;
1967
1968         /*
1969          * NOTE: iroot might not have validated the cluster yet.
1970          */
1971         pmp = MPTOPMP(mp);
1972
1973         mp->mnt_vstat.f_bsize = 0;
1974         mp->mnt_vstat.f_files = 0;
1975         mp->mnt_vstat.f_ffree = 0;
1976         mp->mnt_vstat.f_blocks = 0;
1977         mp->mnt_vstat.f_bfree = 0;
1978         mp->mnt_vstat.f_bavail = 0;
1979
1980         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1981                 hmp = pmp->pfs_hmps[i];
1982                 if (hmp == NULL)
1983                         continue;
1984                 if (pmp->iroot->cluster.array[i].chain)
1985                         bref = pmp->iroot->cluster.array[i].chain->bref;
1986                 else
1987                         bzero(&bref, sizeof(bref));
1988
1989                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1990                 mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
1991                 mp->mnt_vstat.f_ffree = 0;
1992                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1993                                         mp->mnt_vstat.f_bsize;
1994                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1995                                         mp->mnt_vstat.f_bsize;
1996                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1997
1998                 if (cred && cred->cr_uid != 0) {
1999                         uint64_t adj;
2000
2001                         /* 5% */
2002                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2003                         mp->mnt_vstat.f_blocks -= adj;
2004                         mp->mnt_vstat.f_bfree -= adj;
2005                         mp->mnt_vstat.f_bavail -= adj;
2006                 }
2007
2008                 *sbp = mp->mnt_vstat;
2009         }
2010         return (0);
2011 }
2012
2013 /*
2014  * Mount-time recovery (RW mounts)
2015  *
2016  * Updates to the free block table are allowed to lag flushes by one
2017  * transaction.  In case of a crash, then on a fresh mount we must do an
2018  * incremental scan of the last committed transaction id and make sure that
2019  * all related blocks have been marked allocated.
2020  *
2021  * The super-root topology and each PFS has its own transaction id domain,
2022  * so we must track PFS boundary transitions.
2023  */
2024 struct hammer2_recovery_elm {
2025         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2026         hammer2_chain_t *chain;
2027         hammer2_tid_t sync_tid;
2028 };
2029
2030 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2031
2032 struct hammer2_recovery_info {
2033         struct hammer2_recovery_list list;
2034         hammer2_tid_t   mtid;
2035         int     depth;
2036 };
2037
2038 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2039                         hammer2_chain_t *parent,
2040                         struct hammer2_recovery_info *info,
2041                         hammer2_tid_t sync_tid);
2042
2043 #define HAMMER2_RECOVERY_MAXDEPTH       10
2044
2045 static
2046 int
2047 hammer2_recovery(hammer2_dev_t *hmp)
2048 {
2049         struct hammer2_recovery_info info;
2050         struct hammer2_recovery_elm *elm;
2051         hammer2_chain_t *parent;
2052         hammer2_tid_t sync_tid;
2053         hammer2_tid_t mirror_tid;
2054         int error;
2055
2056         hammer2_trans_init(hmp->spmp, 0);
2057
2058         sync_tid = hmp->voldata.freemap_tid;
2059         mirror_tid = hmp->voldata.mirror_tid;
2060
2061         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2062         if (sync_tid >= mirror_tid) {
2063                 kprintf(" no recovery needed\n");
2064         } else {
2065                 kprintf(" freemap recovery %016jx-%016jx\n",
2066                         sync_tid + 1, mirror_tid);
2067         }
2068
2069         TAILQ_INIT(&info.list);
2070         info.depth = 0;
2071         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2072         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2073         hammer2_chain_lookup_done(parent);
2074
2075         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2076                 TAILQ_REMOVE(&info.list, elm, entry);
2077                 parent = elm->chain;
2078                 sync_tid = elm->sync_tid;
2079                 kfree(elm, M_HAMMER2);
2080
2081                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2082                 error |= hammer2_recovery_scan(hmp, parent, &info,
2083                                               hmp->voldata.freemap_tid);
2084                 hammer2_chain_unlock(parent);
2085                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2086         }
2087         hammer2_trans_done(hmp->spmp);
2088
2089         return error;
2090 }
2091
2092 static
2093 int
2094 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2095                       struct hammer2_recovery_info *info,
2096                       hammer2_tid_t sync_tid)
2097 {
2098         const hammer2_inode_data_t *ripdata;
2099         hammer2_chain_t *chain;
2100         hammer2_blockref_t bref;
2101         int tmp_error;
2102         int rup_error;
2103         int error;
2104         int first;
2105
2106         /*
2107          * Adjust freemap to ensure that the block(s) are marked allocated.
2108          */
2109         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2110                 hammer2_freemap_adjust(hmp, &parent->bref,
2111                                        HAMMER2_FREEMAP_DORECOVER);
2112         }
2113
2114         /*
2115          * Check type for recursive scan
2116          */
2117         switch(parent->bref.type) {
2118         case HAMMER2_BREF_TYPE_VOLUME:
2119                 /* data already instantiated */
2120                 break;
2121         case HAMMER2_BREF_TYPE_INODE:
2122                 /*
2123                  * Must instantiate data for DIRECTDATA test and also
2124                  * for recursion.
2125                  */
2126                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2127                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2128                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2129                         /* not applicable to recovery scan */
2130                         hammer2_chain_unlock(parent);
2131                         return 0;
2132                 }
2133                 hammer2_chain_unlock(parent);
2134                 break;
2135         case HAMMER2_BREF_TYPE_INDIRECT:
2136                 /*
2137                  * Must instantiate data for recursion
2138                  */
2139                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2140                 hammer2_chain_unlock(parent);
2141                 break;
2142         case HAMMER2_BREF_TYPE_DIRENT:
2143         case HAMMER2_BREF_TYPE_DATA:
2144         case HAMMER2_BREF_TYPE_FREEMAP:
2145         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2146         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2147                 /* not applicable to recovery scan */
2148                 return 0;
2149                 break;
2150         default:
2151                 return HAMMER2_ERROR_BADBREF;
2152         }
2153
2154         /*
2155          * Defer operation if depth limit reached or if we are crossing a
2156          * PFS boundary.
2157          */
2158         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2159                 struct hammer2_recovery_elm *elm;
2160
2161                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2162                 elm->chain = parent;
2163                 elm->sync_tid = sync_tid;
2164                 hammer2_chain_ref(parent);
2165                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2166                 /* unlocked by caller */
2167
2168                 return(0);
2169         }
2170
2171
2172         /*
2173          * Recursive scan of the last flushed transaction only.  We are
2174          * doing this without pmp assignments so don't leave the chains
2175          * hanging around after we are done with them.
2176          *
2177          * error        Cumulative error this level only
2178          * rup_error    Cumulative error for recursion
2179          * tmp_error    Specific non-cumulative recursion error
2180          */
2181         chain = NULL;
2182         first = 1;
2183         rup_error = 0;
2184         error = 0;
2185
2186         for (;;) {
2187                 error |= hammer2_chain_scan(parent, &chain, &bref,
2188                                             &first,
2189                                             HAMMER2_LOOKUP_NODATA);
2190
2191                 /*
2192                  * Problem during scan or EOF
2193                  */
2194                 if (error)
2195                         break;
2196
2197                 /*
2198                  * If this is a leaf
2199                  */
2200                 if (chain == NULL) {
2201                         if (bref.mirror_tid > sync_tid) {
2202                                 hammer2_freemap_adjust(hmp, &bref,
2203                                                      HAMMER2_FREEMAP_DORECOVER);
2204                         }
2205                         continue;
2206                 }
2207
2208                 /*
2209                  * This may or may not be a recursive node.
2210                  */
2211                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2212                 if (bref.mirror_tid > sync_tid) {
2213                         ++info->depth;
2214                         tmp_error = hammer2_recovery_scan(hmp, chain,
2215                                                            info, sync_tid);
2216                         --info->depth;
2217                 } else {
2218                         tmp_error = 0;
2219                 }
2220
2221                 /*
2222                  * Flush the recovery at the PFS boundary to stage it for
2223                  * the final flush of the super-root topology.
2224                  */
2225                 if (tmp_error == 0 &&
2226                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2227                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2228                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2229                 }
2230                 rup_error |= tmp_error;
2231         }
2232         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2233 }
2234
2235 /*
2236  * Sync a mount point; this is called on a per-mount basis from the
2237  * filesystem syncer process periodically and whenever a user issues
2238  * a sync.
2239  */
2240 int
2241 hammer2_vfs_sync(struct mount *mp, int waitfor)
2242 {
2243         hammer2_xop_flush_t *xop;
2244         struct hammer2_sync_info info;
2245         hammer2_inode_t *iroot;
2246         hammer2_pfs_t *pmp;
2247         int flags;
2248         int error;
2249
2250         pmp = MPTOPMP(mp);
2251         iroot = pmp->iroot;
2252         KKASSERT(iroot);
2253         KKASSERT(iroot->pmp == pmp);
2254
2255         /*
2256          * We can't acquire locks on existing vnodes while in a transaction
2257          * without risking a deadlock.  This assumes that vfsync() can be
2258          * called without the vnode locked (which it can in DragonFly).
2259          * Otherwise we'd have to implement a multi-pass or flag the lock
2260          * failures and retry.
2261          *
2262          * The reclamation code interlocks with the sync list's token
2263          * (by removing the vnode from the scan list) before unlocking
2264          * the inode, giving us time to ref the inode.
2265          */
2266         /*flags = VMSC_GETVP;*/
2267         flags = 0;
2268         if (waitfor & MNT_LAZY)
2269                 flags |= VMSC_ONEPASS;
2270
2271         /*
2272          * Preflush the vnodes using a normal transaction before interlocking
2273          * with a flush transaction.  We do this to try to run as much of
2274          * the compression as possible outside the flush transaction.
2275          *
2276          * For efficiency do an async pass before making sure with a
2277          * synchronous pass on all related buffer cache buffers.
2278          */
2279         hammer2_trans_init(pmp, 0);
2280         info.error = 0;
2281         info.waitfor = MNT_NOWAIT;
2282         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2283         info.waitfor = MNT_WAIT;
2284         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2285         hammer2_trans_done(pmp);
2286
2287         /*
2288          * Start our flush transaction.  This does not return until all
2289          * concurrent transactions have completed and will prevent any
2290          * new transactions from running concurrently, except for the
2291          * buffer cache transactions.
2292          *
2293          * (1) vfsync() all dirty vnodes via vfsyncscan().
2294          *
2295          * (2) Flush any remaining dirty inodes (the sideq), including any
2296          *     which may have been created during or raced against the
2297          *     vfsync().  To catch all cases this must be done after the
2298          *     vfsync().
2299          *
2300          * (3) Wait for any pending BIO I/O to complete (hammer2_bioq_sync()).
2301          *
2302          * NOTE! It is still possible for the paging code to push pages
2303          *       out via a UIO_NOCOPY hammer2_vop_write() during the main
2304          *       flush.
2305          */
2306         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2307
2308         info.error = 0;
2309         info.waitfor = MNT_NOWAIT;
2310         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2311         info.waitfor = MNT_WAIT;
2312         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2313         hammer2_inode_run_sideq(pmp, 1);
2314         hammer2_bioq_sync(pmp);
2315
2316         /*
2317          * Use the XOP interface to concurrently flush all nodes to
2318          * synchronize the PFSROOT subtopology to the media.  A standard
2319          * end-of-scan ENOENT error indicates cluster sufficiency.
2320          *
2321          * Note that this flush will not be visible on crash recovery until
2322          * we flush the super-root topology in the next loop.
2323          *
2324          * XXX For now wait for all flushes to complete.
2325          */
2326         if (iroot) {
2327                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2328                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2329                 error = hammer2_xop_collect(&xop->head,
2330                                             HAMMER2_XOP_COLLECT_WAITALL);
2331                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2332                 if (error == HAMMER2_ERROR_ENOENT)
2333                         error = 0;
2334                 else
2335                         error = hammer2_error_to_errno(error);
2336         } else {
2337                 error = 0;
2338         }
2339         hammer2_trans_done(pmp);
2340
2341         return (error);
2342 }
2343
2344 /*
2345  * Sync passes.
2346  *
2347  * Note that we ignore the tranasction mtid we got above.  Instead,
2348  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2349  * transactions.
2350  */
2351 static int
2352 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2353 {
2354         struct hammer2_sync_info *info = data;
2355         hammer2_inode_t *ip;
2356         int error;
2357
2358         /*
2359          * Degenerate cases.  Note that ip == NULL typically means the
2360          * syncer vnode itself and we don't want to vclrisdirty() in that
2361          * situation.
2362          */
2363         ip = VTOI(vp);
2364         if (ip == NULL) {
2365                 return(0);
2366         }
2367         if (vp->v_type == VNON || vp->v_type == VBAD) {
2368                 vclrisdirty(vp);
2369                 return(0);
2370         }
2371
2372         /*
2373          * VOP_FSYNC will start a new transaction so replicate some code
2374          * here to do it inline (see hammer2_vop_fsync()).
2375          *
2376          * WARNING: The vfsync interacts with the buffer cache and might
2377          *          block, we can't hold the inode lock at that time.
2378          *          However, we MUST ref ip before blocking to ensure that
2379          *          it isn't ripped out from under us (since we do not
2380          *          hold a lock on the vnode).
2381          */
2382         hammer2_inode_ref(ip);
2383         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2384             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2385                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2386                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2387                                  HAMMER2_INODE_MODIFIED)) {
2388                         hammer2_inode_lock(ip, 0);
2389                         if (ip->flags & (HAMMER2_INODE_RESIZED |
2390                                          HAMMER2_INODE_MODIFIED)) {
2391                                 hammer2_inode_chain_sync(ip);
2392                         }
2393                         hammer2_inode_unlock(ip);
2394                 }
2395         }
2396         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2397             RB_EMPTY(&vp->v_rbdirty_tree)) {
2398                 vclrisdirty(vp);
2399         }
2400
2401         hammer2_inode_drop(ip);
2402 #if 1
2403         error = 0;
2404         if (error)
2405                 info->error = error;
2406 #endif
2407         return(0);
2408 }
2409
2410 static
2411 int
2412 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2413 {
2414         hammer2_inode_t *ip;
2415
2416         KKASSERT(MAXFIDSZ >= 16);
2417         ip = VTOI(vp);
2418         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2419         fhp->fid_ext = 0;
2420         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2421         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2422
2423         return 0;
2424 }
2425
2426 static
2427 int
2428 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2429                struct fid *fhp, struct vnode **vpp)
2430 {
2431         hammer2_pfs_t *pmp;
2432         hammer2_tid_t inum;
2433         int error;
2434
2435         pmp = MPTOPMP(mp);
2436         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2437         if (vpp) {
2438                 if (inum == 1)
2439                         error = hammer2_vfs_root(mp, vpp);
2440                 else
2441                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2442         } else {
2443                 error = 0;
2444         }
2445         if (error)
2446                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2447         return error;
2448 }
2449
2450 static
2451 int
2452 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2453                  int *exflagsp, struct ucred **credanonp)
2454 {
2455         hammer2_pfs_t *pmp;
2456         struct netcred *np;
2457         int error;
2458
2459         pmp = MPTOPMP(mp);
2460         np = vfs_export_lookup(mp, &pmp->export, nam);
2461         if (np) {
2462                 *exflagsp = np->netc_exflags;
2463                 *credanonp = &np->netc_anon;
2464                 error = 0;
2465         } else {
2466                 error = EACCES;
2467         }
2468         return error;
2469 }
2470
2471 /*
2472  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2473  * header into the HMP
2474  *
2475  * XXX read four volhdrs and use the one with the highest TID whos CRC
2476  *     matches.
2477  *
2478  * XXX check iCRCs.
2479  *
2480  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2481  *     nonexistant locations.
2482  *
2483  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2484  */
2485 static
2486 int
2487 hammer2_install_volume_header(hammer2_dev_t *hmp)
2488 {
2489         hammer2_volume_data_t *vd;
2490         struct buf *bp;
2491         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2492         int error_reported;
2493         int error;
2494         int valid;
2495         int i;
2496
2497         error_reported = 0;
2498         error = 0;
2499         valid = 0;
2500         bp = NULL;
2501
2502         /*
2503          * There are up to 4 copies of the volume header (syncs iterate
2504          * between them so there is no single master).  We don't trust the
2505          * volu_size field so we don't know precisely how large the filesystem
2506          * is, so depend on the OS to return an error if we go beyond the
2507          * block device's EOF.
2508          */
2509         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2510                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2511                               HAMMER2_VOLUME_BYTES, &bp);
2512                 if (error) {
2513                         brelse(bp);
2514                         bp = NULL;
2515                         continue;
2516                 }
2517
2518                 vd = (struct hammer2_volume_data *) bp->b_data;
2519                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2520                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2521                         brelse(bp);
2522                         bp = NULL;
2523                         continue;
2524                 }
2525
2526                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2527                         /* XXX: Reversed-endianness filesystem */
2528                         kprintf("hammer2: reverse-endian filesystem detected");
2529                         brelse(bp);
2530                         bp = NULL;
2531                         continue;
2532                 }
2533
2534                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2535                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2536                                       HAMMER2_VOLUME_ICRC0_SIZE);
2537                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2538                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2539                                        HAMMER2_VOLUME_ICRC1_SIZE);
2540                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2541                         kprintf("hammer2 volume header crc "
2542                                 "mismatch copy #%d %08x/%08x\n",
2543                                 i, crc0, crc);
2544                         error_reported = 1;
2545                         brelse(bp);
2546                         bp = NULL;
2547                         continue;
2548                 }
2549                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2550                         valid = 1;
2551                         hmp->voldata = *vd;
2552                         hmp->volhdrno = i;
2553                 }
2554                 brelse(bp);
2555                 bp = NULL;
2556         }
2557         if (valid) {
2558                 hmp->volsync = hmp->voldata;
2559                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2560                 error = 0;
2561                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2562                         kprintf("hammer2: using volume header #%d\n",
2563                                 hmp->volhdrno);
2564                 }
2565         } else {
2566                 error = EINVAL;
2567                 kprintf("hammer2: no valid volume headers found!\n");
2568         }
2569         return (error);
2570 }
2571
2572 /*
2573  * This handles hysteresis on regular file flushes.  Because the BIOs are
2574  * routed to a thread it is possible for an excessive number to build up
2575  * and cause long front-end stalls long before the runningbuffspace limit
2576  * is hit, so we implement hammer2_flush_pipe to control the
2577  * hysteresis.
2578  *
2579  * This is a particular problem when compression is used.
2580  */
2581 void
2582 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2583 {
2584         atomic_add_int(&pmp->count_lwinprog, 1);
2585 }
2586
2587 void
2588 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2589 {
2590         int lwinprog;
2591
2592         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2593         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2594             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2595                 atomic_clear_int(&pmp->count_lwinprog,
2596                                  HAMMER2_LWINPROG_WAITING);
2597                 wakeup(&pmp->count_lwinprog);
2598         }
2599         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2600             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2601                 atomic_clear_int(&pmp->count_lwinprog,
2602                                  HAMMER2_LWINPROG_WAITING0);
2603                 wakeup(&pmp->count_lwinprog);
2604         }
2605 }
2606
2607 void
2608 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2609 {
2610         int lwinprog;
2611         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2612                                     HAMMER2_LWINPROG_WAITING0;
2613
2614         for (;;) {
2615                 lwinprog = pmp->count_lwinprog;
2616                 cpu_ccfence();
2617                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2618                         break;
2619                 tsleep_interlock(&pmp->count_lwinprog, 0);
2620                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2621                 lwinprog = pmp->count_lwinprog;
2622                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2623                         break;
2624                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2625         }
2626 }
2627
2628 /*
2629  * Manage excessive memory resource use for chain and related
2630  * structures.
2631  */
2632 void
2633 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2634 {
2635         uint32_t waiting;
2636         uint32_t count;
2637         uint32_t limit;
2638 #if 0
2639         static int zzticks;
2640 #endif
2641
2642         /*
2643          * Atomic check condition and wait.  Also do an early speedup of
2644          * the syncer to try to avoid hitting the wait.
2645          */
2646         for (;;) {
2647                 waiting = pmp->inmem_dirty_chains;
2648                 cpu_ccfence();
2649                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2650
2651                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2652                 if (limit < hammer2_limit_dirty_chains)
2653                         limit = hammer2_limit_dirty_chains;
2654                 if (limit < 1000)
2655                         limit = 1000;
2656
2657 #if 0
2658                 if ((int)(ticks - zzticks) > hz) {
2659                         zzticks = ticks;
2660                         kprintf("count %ld %ld\n", count, limit);
2661                 }
2662 #endif
2663
2664                 /*
2665                  * Block if there are too many dirty chains present, wait
2666                  * for the flush to clean some out.
2667                  */
2668                 if (count > limit) {
2669                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2670                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2671                                                waiting,
2672                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2673                                 speedup_syncer(pmp->mp);
2674                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2675                                        "chnmem", hz);
2676                         }
2677                         continue;       /* loop on success or fail */
2678                 }
2679
2680                 /*
2681                  * Try to start an early flush before we are forced to block.
2682                  */
2683                 if (count > limit * 7 / 10)
2684                         speedup_syncer(pmp->mp);
2685                 break;
2686         }
2687 }
2688
2689 void
2690 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2691 {
2692         if (pmp) {
2693                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2694         }
2695 }
2696
2697 void
2698 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2699 {
2700         uint32_t waiting;
2701
2702         if (pmp == NULL)
2703                 return;
2704
2705         for (;;) {
2706                 waiting = pmp->inmem_dirty_chains;
2707                 cpu_ccfence();
2708                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2709                                        waiting,
2710                                        (waiting - 1) &
2711                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2712                         break;
2713                 }
2714         }
2715
2716         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2717                 wakeup(&pmp->inmem_dirty_chains);
2718 }
2719
2720 /*
2721  * Returns 0 if the filesystem has tons of free space
2722  * Returns 1 if the filesystem has less than 10% remaining
2723  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2724  */
2725 int
2726 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2727 {
2728         hammer2_pfs_t *pmp;
2729         hammer2_dev_t *hmp;
2730         hammer2_off_t free_reserved;
2731         hammer2_off_t free_nominal;
2732         int i;
2733
2734         pmp = ip->pmp;
2735
2736         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2737                 free_reserved = HAMMER2_SEGSIZE;
2738                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2739                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2740                         hmp = pmp->pfs_hmps[i];
2741                         if (hmp == NULL)
2742                                 continue;
2743                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2744                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2745                                 continue;
2746
2747                         if (free_nominal > hmp->voldata.allocator_free)
2748                                 free_nominal = hmp->voldata.allocator_free;
2749                         if (free_reserved < hmp->free_reserved)
2750                                 free_reserved = hmp->free_reserved;
2751                 }
2752
2753                 /*
2754                  * SMP races ok
2755                  */
2756                 pmp->free_reserved = free_reserved;
2757                 pmp->free_nominal = free_nominal;
2758                 pmp->free_ticks = ticks;
2759         } else {
2760                 free_reserved = pmp->free_reserved;
2761                 free_nominal = pmp->free_nominal;
2762         }
2763         if (cred && cred->cr_uid != 0) {
2764                 if ((int64_t)(free_nominal - bytes) <
2765                     (int64_t)free_reserved) {
2766                         return 2;
2767                 }
2768         } else {
2769                 if ((int64_t)(free_nominal - bytes) <
2770                     (int64_t)free_reserved / 2) {
2771                         return 2;
2772                 }
2773         }
2774         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2775                 return 1;
2776         return 0;
2777 }
2778
2779 /*
2780  * Debugging
2781  */
2782 void
2783 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2784 {
2785         hammer2_chain_t *scan;
2786         hammer2_chain_t *parent;
2787
2788         --*countp;
2789         if (*countp == 0) {
2790                 kprintf("%*.*s...\n", tab, tab, "");
2791                 return;
2792         }
2793         if (*countp < 0)
2794                 return;
2795         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2796                 tab, tab, "", pfx,
2797                 chain, chain->bref.type,
2798                 chain->bref.key, chain->bref.keybits,
2799                 chain->bref.mirror_tid);
2800
2801         kprintf("%*.*s      [%08x] (%s) refs=%d",
2802                 tab, tab, "",
2803                 chain->flags,
2804                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2805                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2806                 chain->refs);
2807
2808         parent = chain->parent;
2809         if (parent)
2810                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2811                         tab, tab, "",
2812                         parent, parent->flags, parent->refs);
2813         if (RB_EMPTY(&chain->core.rbtree)) {
2814                 kprintf("\n");
2815         } else {
2816                 kprintf(" {\n");
2817                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2818                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2819                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2820                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2821                                 chain->data->ipdata.filename);
2822                 else
2823                         kprintf("%*.*s}\n", tab, tab, "");
2824         }
2825 }