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 = 0;          /* bdwrite() so later inval works */
  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 long hammer2_chain_allocs;
  92 long hammer2_chain_frees;
  93 long hammer2_limit_dirty_chains;
  94 long hammer2_count_modified_chains;
  95 long hammer2_iod_invals;
  96 long hammer2_iod_file_read;
  97 long hammer2_iod_meta_read;
  98 long hammer2_iod_indr_read;
  99 long hammer2_iod_fmap_read;
 100 long hammer2_iod_volu_read;
 101 long hammer2_iod_file_write;
 102 long hammer2_iod_file_wembed;
 103 long hammer2_iod_file_wzero;
 104 long hammer2_iod_file_wdedup;
 105 long hammer2_iod_meta_write;
 106 long hammer2_iod_indr_write;
 107 long hammer2_iod_fmap_write;
 108 long hammer2_iod_volu_write;
 109
 110 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 111 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 112                 "Buffer used for compression.");
 113
 114 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 115 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 116                 "Buffer used for decompression.");
 117
 118 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 119
 120 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 121            &hammer2_debug, 0, "");
 122 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 123            &hammer2_cluster_meta_read, 0, "");
 124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 125            &hammer2_cluster_data_read, 0, "");
 126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 127            &hammer2_cluster_write, 0, "");
 128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 129            &hammer2_dedup_enable, 0, "");
 130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 131            &hammer2_always_compress, 0, "");
 132 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 133            &hammer2_inval_enable, 0, "");
 134 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 135            &hammer2_flush_pipe, 0, "");
 136 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 137            &hammer2_synchronous_flush, 0, "");
 138 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 139            &hammer2_chain_allocs, 0, "");
 140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 141            &hammer2_chain_frees, 0, "");
 142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 143            &hammer2_limit_dirty_chains, 0, "");
 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 145            &hammer2_count_modified_chains, 0, "");
 146 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 147            &hammer2_dio_count, 0, "");
 148 SYSCTL_INT(_vfs_hammer2, OID_AUTO, limit_dio, CTLFLAG_RW,
 149            &hammer2_limit_dio, 0, "");
 150
 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 152            &hammer2_iod_invals, 0, "");
 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 154            &hammer2_iod_file_read, 0, "");
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 156            &hammer2_iod_meta_read, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 158            &hammer2_iod_indr_read, 0, "");
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 160            &hammer2_iod_fmap_read, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 162            &hammer2_iod_volu_read, 0, "");
 163
 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 165            &hammer2_iod_file_write, 0, "");
 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 167            &hammer2_iod_file_wembed, 0, "");
 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 169            &hammer2_iod_file_wzero, 0, "");
 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 171            &hammer2_iod_file_wdedup, 0, "");
 172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 173            &hammer2_iod_meta_write, 0, "");
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 175            &hammer2_iod_indr_write, 0, "");
 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 177            &hammer2_iod_fmap_write, 0, "");
 178 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 179            &hammer2_iod_volu_write, 0, "");
 180
 181 long hammer2_check_icrc32;
 182 long hammer2_check_xxhash64;
 183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 184            &hammer2_check_icrc32, 0, "");
 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 186            &hammer2_check_xxhash64, 0, "");
 187
 188 static int hammer2_vfs_init(struct vfsconf *conf);
 189 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 190 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 191                                 struct ucred *cred);
 192 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 193                                 struct vnode *, struct ucred *);
 194 static int hammer2_recovery(hammer2_dev_t *hmp);
 195 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 196 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 197 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 198                                 struct ucred *cred);
 199 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 200                                 struct ucred *cred);
 201 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 202                                 struct fid *fhp, struct vnode **vpp);
 203 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 204 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 205                                 int *exflagsp, struct ucred **credanonp);
 206
 207 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 208 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 209
 210 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 211
 212 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 213 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 214                                 hammer2_dev_t *hmp);
 215
 216 /*
 217  * HAMMER2 vfs operations.
 218  */
 219 static struct vfsops hammer2_vfsops = {
 220         .vfs_init       = hammer2_vfs_init,
 221         .vfs_uninit     = hammer2_vfs_uninit,
 222         .vfs_sync       = hammer2_vfs_sync,
 223         .vfs_mount      = hammer2_vfs_mount,
 224         .vfs_unmount    = hammer2_vfs_unmount,
 225         .vfs_root       = hammer2_vfs_root,
 226         .vfs_statfs     = hammer2_vfs_statfs,
 227         .vfs_statvfs    = hammer2_vfs_statvfs,
 228         .vfs_vget       = hammer2_vfs_vget,
 229         .vfs_vptofh     = hammer2_vfs_vptofh,
 230         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 231         .vfs_checkexp   = hammer2_vfs_checkexp
 232 };
 233
 234 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 235
 236 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 237 MODULE_VERSION(hammer2, 1);
 238
 239 static
 240 int
 241 hammer2_vfs_init(struct vfsconf *conf)
 242 {
 243         static struct objcache_malloc_args margs_read;
 244         static struct objcache_malloc_args margs_write;
 245         static struct objcache_malloc_args margs_vop;
 246
 247         int error;
 248
 249         error = 0;
 250
 251         /*
 252          * A large DIO cache is needed to retain dedup enablement masks.
 253          * The bulkfree code clears related masks as part of the disk block
 254          * recycling algorithm, preventing it from being used for a later
 255          * dedup.
 256          *
 257          * NOTE: A large buffer cache can actually interfere with dedup
 258          *       operation because we dedup based on media physical buffers
 259          *       and not logical buffers.  Try to make the DIO chace large
 260          *       enough to avoid this problem, but also cap it.
 261          */
 262         hammer2_limit_dio = nbuf * 2;
 263         if (hammer2_limit_dio > 100000)
 264                 hammer2_limit_dio = 100000;
 265
 266         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 267                 error = EINVAL;
 268         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 269                 error = EINVAL;
 270         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 271                 error = EINVAL;
 272
 273         if (error)
 274                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 275
 276         margs_read.objsize = 65536;
 277         margs_read.mtype = M_HAMMER2_DEBUFFER;
 278
 279         margs_write.objsize = 32768;
 280         margs_write.mtype = M_HAMMER2_CBUFFER;
 281
 282         margs_vop.objsize = sizeof(hammer2_xop_t);
 283         margs_vop.mtype = M_HAMMER2;
 284
 285         /*
 286          * Note thaht for the XOPS cache we want backing store allocations
 287          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 288          * confusion), so use the backing store function that does it.  This
 289          * means that initial XOPS objects are zerod but REUSED objects are
 290          * not.  So we are responsible for cleaning the object up sufficiently
 291          * for our needs before objcache_put()ing it back (typically just the
 292          * FIFO indices).
 293          */
 294         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 295                                 0, 1, NULL, NULL, NULL,
 296                                 objcache_malloc_alloc,
 297                                 objcache_malloc_free,
 298                                 &margs_read);
 299         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 300                                 0, 1, NULL, NULL, NULL,
 301                                 objcache_malloc_alloc,
 302                                 objcache_malloc_free,
 303                                 &margs_write);
 304         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 305                                 0, 1, NULL, NULL, NULL,
 306                                 objcache_malloc_alloc_zero,
 307                                 objcache_malloc_free,
 308                                 &margs_vop);
 309
 310
 311         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 312         TAILQ_INIT(&hammer2_mntlist);
 313         TAILQ_INIT(&hammer2_pfslist);
 314
 315         hammer2_limit_dirty_chains = maxvnodes / 10;
 316         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 317                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 318
 319         return (error);
 320 }
 321
 322 static
 323 int
 324 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 325 {
 326         objcache_destroy(cache_buffer_read);
 327         objcache_destroy(cache_buffer_write);
 328         objcache_destroy(cache_xops);
 329         return 0;
 330 }
 331
 332 /*
 333  * Core PFS allocator.  Used to allocate or reference the pmp structure
 334  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 335  * The pmp can be passed in or loaded by this function using the chain and
 336  * inode data.
 337  *
 338  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 339  * transactions.  Note that synchronization does not use this field.
 340  * (typically frontend operations and synchronization cannot run on the
 341  * same PFS node at the same time).
 342  *
 343  * XXX check locking
 344  */
 345 hammer2_pfs_t *
 346 hammer2_pfsalloc(hammer2_chain_t *chain,
 347                  const hammer2_inode_data_t *ripdata,
 348                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 349 {
 350         hammer2_pfs_t *pmp;
 351         hammer2_inode_t *iroot;
 352         int count;
 353         int i;
 354         int j;
 355
 356         pmp = NULL;
 357
 358         /*
 359          * Locate or create the PFS based on the cluster id.  If ripdata
 360          * is NULL this is a spmp which is unique and is always allocated.
 361          *
 362          * If the device is mounted in local mode all PFSs are considered
 363          * independent and not part of any cluster (for debugging only).
 364          */
 365         if (ripdata) {
 366                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 367                         if (force_local != pmp->force_local)
 368                                 continue;
 369                         if (force_local == NULL &&
 370                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 371                                  sizeof(pmp->pfs_clid)) == 0) {
 372                                         break;
 373                         } else if (force_local && pmp->pfs_names[0] &&
 374                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 375                                         break;
 376                         }
 377                 }
 378         }
 379
 380         if (pmp == NULL) {
 381                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 382                 pmp->force_local = force_local;
 383                 hammer2_trans_manage_init(pmp);
 384                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 385                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 386                 lockinit(&pmp->lock, "pfslk", 0, 0);
 387                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 388                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 389                 spin_init(&pmp->xop_spin, "h2xop");
 390                 spin_init(&pmp->lru_spin, "h2lru");
 391                 RB_INIT(&pmp->inum_tree);
 392                 TAILQ_INIT(&pmp->sideq);
 393                 TAILQ_INIT(&pmp->lru_list);
 394                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 395
 396                 /*
 397                  * Distribute backend operations to threads
 398                  */
 399                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 400                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 401
 402                 /*
 403                  * Save the last media transaction id for the flusher.  Set
 404                  * initial
 405                  */
 406                 if (ripdata)
 407                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 408                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 409
 410                 /*
 411                  * The synchronization thread may start too early, make
 412                  * sure it stays frozen until we are ready to let it go.
 413                  * XXX
 414                  */
 415                 /*
 416                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 417                                          HAMMER2_THREAD_REMASTER;
 418                 */
 419         }
 420
 421         /*
 422          * Create the PFS's root inode and any missing XOP helper threads.
 423          */
 424         if ((iroot = pmp->iroot) == NULL) {
 425                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 426                 if (ripdata)
 427                         iroot->meta = ripdata->meta;
 428                 pmp->iroot = iroot;
 429                 hammer2_inode_ref(iroot);
 430                 hammer2_inode_unlock(iroot);
 431         }
 432
 433         /*
 434          * Stop here if no chain is passed in.
 435          */
 436         if (chain == NULL)
 437                 goto done;
 438
 439         /*
 440          * When a chain is passed in we must add it to the PFS's root
 441          * inode, update pmp->pfs_types[], and update the syncronization
 442          * threads.
 443          *
 444          * When forcing local mode, mark the PFS as a MASTER regardless.
 445          *
 446          * At the moment empty spots can develop due to removals or failures.
 447          * Ultimately we want to re-fill these spots but doing so might
 448          * confused running code. XXX
 449          */
 450         hammer2_inode_ref(iroot);
 451         hammer2_mtx_ex(&iroot->lock);
 452         j = iroot->cluster.nchains;
 453
 454         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 455
 456         if (j == HAMMER2_MAXCLUSTER) {
 457                 kprintf("hammer2_mount: cluster full!\n");
 458                 /* XXX fatal error? */
 459         } else {
 460                 KKASSERT(chain->pmp == NULL);
 461                 chain->pmp = pmp;
 462                 hammer2_chain_ref(chain);
 463                 iroot->cluster.array[j].chain = chain;
 464                 if (force_local)
 465                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 466                 else
 467                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 468                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 469                 pmp->pfs_hmps[j] = chain->hmp;
 470
 471                 /*
 472                  * If the PFS is already mounted we must account
 473                  * for the mount_count here.
 474                  */
 475                 if (pmp->mp)
 476                         ++chain->hmp->mount_count;
 477
 478                 /*
 479                  * May have to fixup dirty chain tracking.  Previous
 480                  * pmp was NULL so nothing to undo.
 481                  */
 482                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 483                         hammer2_pfs_memory_inc(pmp);
 484                 ++j;
 485         }
 486         iroot->cluster.nchains = j;
 487
 488         /*
 489          * Update nmasters from any PFS inode which is part of the cluster.
 490          * It is possible that this will result in a value which is too
 491          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 492          * override this value later on.
 493          *
 494          * (This informs us of masters that might not currently be
 495          *  discoverable by this mount).
 496          */
 497         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 498                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 499         }
 500
 501         /*
 502          * Count visible masters.  Masters are usually added with
 503          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 504          * are more (XXX and must update the master inodes).
 505          */
 506         count = 0;
 507         for (i = 0; i < iroot->cluster.nchains; ++i) {
 508                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 509                         ++count;
 510         }
 511         if (pmp->pfs_nmasters < count)
 512                 pmp->pfs_nmasters = count;
 513
 514         /*
 515          * Create missing synchronization and support threads.
 516          *
 517          * Single-node masters (including snapshots) have nothing to
 518          * synchronize and do not require this thread.
 519          *
 520          * Multi-node masters or any number of soft masters, slaves, copy,
 521          * or other PFS types need the thread.
 522          *
 523          * Each thread is responsible for its particular cluster index.
 524          * We use independent threads so stalls or mismatches related to
 525          * any given target do not affect other targets.
 526          */
 527         for (i = 0; i < iroot->cluster.nchains; ++i) {
 528                 /*
 529                  * Single-node masters (including snapshots) have nothing
 530                  * to synchronize and will make direct xops support calls,
 531                  * thus they do not require this thread.
 532                  *
 533                  * Note that there can be thousands of snapshots.  We do not
 534                  * want to create thousands of threads.
 535                  */
 536                 if (pmp->pfs_nmasters <= 1 &&
 537                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 538                         continue;
 539                 }
 540
 541                 /*
 542                  * Sync support thread
 543                  */
 544                 if (pmp->sync_thrs[i].td == NULL) {
 545                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 546                                            "h2nod", i, -1,
 547                                            hammer2_primary_sync_thread);
 548                 }
 549         }
 550
 551         /*
 552          * Create missing Xop threads
 553          *
 554          * NOTE: We create helper threads for all mounted PFSs or any
 555          *       PFSs with 2+ nodes (so the sync thread can update them,
 556          *       even if not mounted).
 557          */
 558         if (pmp->mp || iroot->cluster.nchains >= 2)
 559                 hammer2_xop_helper_create(pmp);
 560
 561         hammer2_mtx_unlock(&iroot->lock);
 562         hammer2_inode_drop(iroot);
 563 done:
 564         return pmp;
 565 }
 566
 567 /*
 568  * Deallocate an element of a probed PFS.  If destroying and this is a
 569  * MASTER, adjust nmasters.
 570  *
 571  * This function does not physically destroy the PFS element in its device
 572  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 573  */
 574 void
 575 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 576 {
 577         hammer2_inode_t *iroot;
 578         hammer2_chain_t *chain;
 579         int j;
 580
 581         /*
 582          * Cleanup our reference on iroot.  iroot is (should) not be needed
 583          * by the flush code.
 584          */
 585         iroot = pmp->iroot;
 586         if (iroot) {
 587                 /*
 588                  * Stop synchronizing
 589                  *
 590                  * XXX flush after acquiring the iroot lock.
 591                  * XXX clean out the cluster index from all inode structures.
 592                  */
 593                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 594
 595                 /*
 596                  * Remove the cluster index from the group.  If destroying
 597                  * the PFS and this is a master, adjust pfs_nmasters.
 598                  */
 599                 hammer2_mtx_ex(&iroot->lock);
 600                 chain = iroot->cluster.array[clindex].chain;
 601                 iroot->cluster.array[clindex].chain = NULL;
 602
 603                 switch(pmp->pfs_types[clindex]) {
 604                 case HAMMER2_PFSTYPE_MASTER:
 605                         if (destroying && pmp->pfs_nmasters > 0)
 606                                 --pmp->pfs_nmasters;
 607                         /* XXX adjust ripdata->meta.pfs_nmasters */
 608                         break;
 609                 default:
 610                         break;
 611                 }
 612                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 613
 614                 hammer2_mtx_unlock(&iroot->lock);
 615
 616                 /*
 617                  * Release the chain.
 618                  */
 619                 if (chain) {
 620                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 621                         hammer2_chain_drop(chain);
 622                 }
 623
 624                 /*
 625                  * Terminate all XOP threads for the cluster index.
 626                  */
 627                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 628                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 629         }
 630 }
 631
 632 /*
 633  * Destroy a PFS, typically only occurs after the last mount on a device
 634  * has gone away.
 635  */
 636 static void
 637 hammer2_pfsfree(hammer2_pfs_t *pmp)
 638 {
 639         hammer2_inode_t *iroot;
 640         hammer2_chain_t *chain;
 641         int i;
 642         int j;
 643
 644         /*
 645          * Cleanup our reference on iroot.  iroot is (should) not be needed
 646          * by the flush code.
 647          */
 648         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 649
 650         iroot = pmp->iroot;
 651         if (iroot) {
 652                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 653                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 654                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 655                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 656                 }
 657 #if REPORT_REFS_ERRORS
 658                 if (pmp->iroot->refs != 1)
 659                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 660                                 pmp->iroot, pmp->iroot->refs);
 661 #else
 662                 KKASSERT(pmp->iroot->refs == 1);
 663 #endif
 664                 /* ref for pmp->iroot */
 665                 hammer2_inode_drop(pmp->iroot);
 666                 pmp->iroot = NULL;
 667         }
 668
 669         /*
 670          * Cleanup chains remaining on LRU list.
 671          */
 672         kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
 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                         kprintf("unmount hmp %p remove spmp %p\n",
 710                                 hmp, pmp);
 711                         hmp->spmp = NULL;
 712                 }
 713
 714                 /*
 715                  * Determine if this PFS is affected.  If it is we must
 716                  * freeze all management threads and lock its iroot.
 717                  *
 718                  * Freezing a management thread forces it idle, operations
 719                  * in-progress will be aborted and it will have to start
 720                  * over again when unfrozen, or exit if told to exit.
 721                  */
 722                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 723                         if (pmp->pfs_hmps[i] == hmp)
 724                                 break;
 725                 }
 726                 if (i != HAMMER2_MAXCLUSTER) {
 727                         /*
 728                          * Make sure all synchronization threads are locked
 729                          * down.
 730                          */
 731                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 732                                 if (pmp->pfs_hmps[i] == NULL)
 733                                         continue;
 734                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 735                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 736                                         hammer2_thr_freeze_async(
 737                                                 &pmp->xop_groups[j].thrs[i]);
 738                                 }
 739                         }
 740                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 741                                 if (pmp->pfs_hmps[i] == NULL)
 742                                         continue;
 743                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 744                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 745                                         hammer2_thr_freeze(
 746                                                 &pmp->xop_groups[j].thrs[i]);
 747                                 }
 748                         }
 749
 750                         /*
 751                          * Lock the inode and clean out matching chains.
 752                          * Note that we cannot use hammer2_inode_lock_*()
 753                          * here because that would attempt to validate the
 754                          * cluster that we are in the middle of ripping
 755                          * apart.
 756                          *
 757                          * WARNING! We are working directly on the inodes
 758                          *          embedded cluster.
 759                          */
 760                         hammer2_mtx_ex(&iroot->lock);
 761
 762                         /*
 763                          * Remove the chain from matching elements of the PFS.
 764                          */
 765                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 766                                 if (pmp->pfs_hmps[i] != hmp)
 767                                         continue;
 768                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 769                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 770                                         hammer2_thr_delete(
 771                                                 &pmp->xop_groups[j].thrs[i]);
 772                                 }
 773                                 rchain = iroot->cluster.array[i].chain;
 774                                 iroot->cluster.array[i].chain = NULL;
 775                                 pmp->pfs_types[i] = 0;
 776                                 if (pmp->pfs_names[i]) {
 777                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 778                                         pmp->pfs_names[i] = NULL;
 779                                 }
 780                                 if (rchain) {
 781                                         hammer2_chain_drop(rchain);
 782                                         /* focus hint */
 783                                         if (iroot->cluster.focus == rchain)
 784                                                 iroot->cluster.focus = NULL;
 785                                 }
 786                                 pmp->pfs_hmps[i] = NULL;
 787                         }
 788                         hammer2_mtx_unlock(&iroot->lock);
 789                         didfreeze = 1;  /* remaster, unfreeze down below */
 790                 } else {
 791                         didfreeze = 0;
 792                 }
 793
 794                 /*
 795                  * Cleanup trailing chains.  Gaps may remain.
 796                  */
 797                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 798                         if (pmp->pfs_hmps[i])
 799                                 break;
 800                 }
 801                 iroot->cluster.nchains = i + 1;
 802
 803                 /*
 804                  * If the PMP has no elements remaining we can destroy it.
 805                  * (this will transition management threads from frozen->exit).
 806                  */
 807                 if (iroot->cluster.nchains == 0) {
 808                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 809                                 hmp, pmp);
 810                         hammer2_pfsfree(pmp);
 811                         goto again;
 812                 }
 813
 814                 /*
 815                  * If elements still remain we need to set the REMASTER
 816                  * flag and unfreeze it.
 817                  */
 818                 if (didfreeze) {
 819                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 820                                 if (pmp->pfs_hmps[i] == NULL)
 821                                         continue;
 822                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 823                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 824                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 825                                         hammer2_thr_remaster(
 826                                                 &pmp->xop_groups[j].thrs[i]);
 827                                         hammer2_thr_unfreeze(
 828                                                 &pmp->xop_groups[j].thrs[i]);
 829                                 }
 830                         }
 831                 }
 832         }
 833 }
 834
 835 /*
 836  * Mount or remount HAMMER2 fileystem from physical media
 837  *
 838  *      mountroot
 839  *              mp              mount point structure
 840  *              path            NULL
 841  *              data            <unused>
 842  *              cred            <unused>
 843  *
 844  *      mount
 845  *              mp              mount point structure
 846  *              path            path to mount point
 847  *              data            pointer to argument structure in user space
 848  *                      volume  volume path (device@LABEL form)
 849  *                      hflags  user mount flags
 850  *              cred            user credentials
 851  *
 852  * RETURNS:     0       Success
 853  *              !0      error number
 854  */
 855 static
 856 int
 857 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 858                   struct ucred *cred)
 859 {
 860         struct hammer2_mount_info info;
 861         hammer2_pfs_t *pmp;
 862         hammer2_pfs_t *spmp;
 863         hammer2_dev_t *hmp;
 864         hammer2_dev_t *force_local;
 865         hammer2_key_t key_next;
 866         hammer2_key_t key_dummy;
 867         hammer2_key_t lhc;
 868         struct vnode *devvp;
 869         struct nlookupdata nd;
 870         hammer2_chain_t *parent;
 871         hammer2_chain_t *chain;
 872         hammer2_cluster_t *cluster;
 873         const hammer2_inode_data_t *ripdata;
 874         hammer2_blockref_t bref;
 875         struct file *fp;
 876         char devstr[MNAMELEN];
 877         size_t size;
 878         size_t done;
 879         char *dev;
 880         char *label;
 881         int ronly = 1;
 882         int error;
 883         int i;
 884
 885         hmp = NULL;
 886         pmp = NULL;
 887         dev = NULL;
 888         label = NULL;
 889         devvp = NULL;
 890
 891         kprintf("hammer2_mount\n");
 892
 893         if (path == NULL) {
 894                 /*
 895                  * Root mount
 896                  */
 897                 bzero(&info, sizeof(info));
 898                 info.cluster_fd = -1;
 899                 ksnprintf(devstr, sizeof(devstr), "%s",
 900                           mp->mnt_stat.f_mntfromname);
 901                 kprintf("hammer2_mount: root '%s'\n", devstr);
 902         } else {
 903                 /*
 904                  * Non-root mount or updating a mount
 905                  */
 906                 error = copyin(data, &info, sizeof(info));
 907                 if (error)
 908                         return (error);
 909
 910                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 911                 if (error)
 912                         return (error);
 913         }
 914
 915         /*
 916          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 917          * if no label specified, based on the partition id.  Error out if no
 918          * partition id.  This is strictly a convenience to match the
 919          * default label created by newfs_hammer2, our preference is
 920          * that a label always be specified.
 921          */
 922         dev = devstr;
 923         label = strchr(devstr, '@');
 924         if (label && ((label + 1) - dev) > done)
 925                 return (EINVAL);
 926         if (label && label == devstr)
 927                 return (EINVAL);
 928         if (label == NULL || label[1] == 0) {
 929                 char slice;
 930
 931                 if (label == NULL)
 932                         label = devstr + strlen(devstr);
 933                 slice = label[-1];
 934                 switch(slice) {
 935                 case 'a':
 936                         label = "BOOT";
 937                         break;
 938                 case 'd':
 939                         label = "ROOT";
 940                         break;
 941                 default:
 942                         label = "DATA";
 943                         break;
 944                 }
 945         } else {
 946                 *label = '\0';
 947                 label++;
 948         }
 949
 950         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
 951                 dev, label, (mp->mnt_flag & MNT_RDONLY));
 952
 953         if (mp->mnt_flag & MNT_UPDATE) {
 954                 /*
 955                  * Update mount.  Note that pmp->iroot->cluster is
 956                  * an inode-embedded cluster and thus cannot be
 957                  * directly locked.
 958                  *
 959                  * XXX HAMMER2 needs to implement NFS export via
 960                  *     mountctl.
 961                  */
 962                 pmp = MPTOPMP(mp);
 963                 pmp->hflags = info.hflags;
 964                 cluster = &pmp->iroot->cluster;
 965                 for (i = 0; i < cluster->nchains; ++i) {
 966                         if (cluster->array[i].chain == NULL)
 967                                 continue;
 968                         hmp = cluster->array[i].chain->hmp;
 969                         devvp = hmp->devvp;
 970                         error = hammer2_remount(hmp, mp, path,
 971                                                 devvp, cred);
 972                         if (error)
 973                                 break;
 974                 }
 975
 976                 return error;
 977         }
 978
 979         /*
 980          * HMP device mount
 981          *
 982          * If a path is specified and dev is not an empty string, lookup the
 983          * name and verify that it referes to a block device.
 984          *
 985          * If a path is specified and dev is an empty string we fall through
 986          * and locate the label in the hmp search.
 987          */
 988         if (path && *dev != 0) {
 989                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 990                 if (error == 0)
 991                         error = nlookup(&nd);
 992                 if (error == 0)
 993                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 994                 nlookup_done(&nd);
 995         } else if (path == NULL) {
 996                 /* root mount */
 997                 cdev_t cdev = kgetdiskbyname(dev);
 998                 error = bdevvp(cdev, &devvp);
 999                 if (error)
1000                         kprintf("hammer2: cannot find '%s'\n", dev);
1001         } else {
1002                 /*
1003                  * We will locate the hmp using the label in the hmp loop.
1004                  */
1005                 error = 0;
1006         }
1007
1008         /*
1009          * Make sure its a block device.  Do not check to see if it is
1010          * already mounted until we determine that its a fresh H2 device.
1011          */
1012         if (error == 0 && devvp) {
1013                 vn_isdisk(devvp, &error);
1014         }
1015
1016         /*
1017          * Determine if the device has already been mounted.  After this
1018          * check hmp will be non-NULL if we are doing the second or more
1019          * hammer2 mounts from the same device.
1020          */
1021         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1022         if (devvp) {
1023                 /*
1024                  * Match the device.  Due to the way devfs works,
1025                  * we may not be able to directly match the vnode pointer,
1026                  * so also check to see if the underlying device matches.
1027                  */
1028                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1029                         if (hmp->devvp == devvp)
1030                                 break;
1031                         if (devvp->v_rdev &&
1032                             hmp->devvp->v_rdev == devvp->v_rdev) {
1033                                 break;
1034                         }
1035                 }
1036
1037                 /*
1038                  * If no match this may be a fresh H2 mount, make sure
1039                  * the device is not mounted on anything else.
1040                  */
1041                 if (hmp == NULL)
1042                         error = vfs_mountedon(devvp);
1043         } else if (error == 0) {
1044                 /*
1045                  * Match the label to a pmp already probed.
1046                  */
1047                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1048                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1049                                 if (pmp->pfs_names[i] &&
1050                                     strcmp(pmp->pfs_names[i], label) == 0) {
1051                                         hmp = pmp->pfs_hmps[i];
1052                                         break;
1053                                 }
1054                         }
1055                         if (hmp)
1056                                 break;
1057                 }
1058                 if (hmp == NULL)
1059                         error = ENOENT;
1060         }
1061
1062         /*
1063          * Open the device if this isn't a secondary mount and construct
1064          * the H2 device mount (hmp).
1065          */
1066         if (hmp == NULL) {
1067                 hammer2_chain_t *schain;
1068                 hammer2_xid_t xid;
1069
1070                 if (error == 0 && vcount(devvp) > 0) {
1071                         kprintf("Primary device already has references\n");
1072                         error = EBUSY;
1073                 }
1074
1075                 /*
1076                  * Now open the device
1077                  */
1078                 if (error == 0) {
1079                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1080                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1081                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1082                         if (error == 0) {
1083                                 error = VOP_OPEN(devvp,
1084                                              (ronly ? FREAD : FREAD | FWRITE),
1085                                              FSCRED, NULL);
1086                         }
1087                         vn_unlock(devvp);
1088                 }
1089                 if (error && devvp) {
1090                         vrele(devvp);
1091                         devvp = NULL;
1092                 }
1093                 if (error) {
1094                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1095                         return error;
1096                 }
1097                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1098                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1099                 hmp->ronly = ronly;
1100                 hmp->devvp = devvp;
1101                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1102                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1103                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1104                 RB_INIT(&hmp->iotree);
1105                 spin_init(&hmp->io_spin, "hm2mount_io");
1106                 spin_init(&hmp->list_spin, "hm2mount_list");
1107                 TAILQ_INIT(&hmp->flushq);
1108
1109                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1110                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1111                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1112
1113                 /*
1114                  * vchain setup. vchain.data is embedded.
1115                  * vchain.refs is initialized and will never drop to 0.
1116                  *
1117                  * NOTE! voldata is not yet loaded.
1118                  */
1119                 hmp->vchain.hmp = hmp;
1120                 hmp->vchain.refs = 1;
1121                 hmp->vchain.data = (void *)&hmp->voldata;
1122                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1123                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1124                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1125
1126                 hammer2_chain_core_init(&hmp->vchain);
1127                 /* hmp->vchain.u.xxx is left NULL */
1128
1129                 /*
1130                  * fchain setup.  fchain.data is embedded.
1131                  * fchain.refs is initialized and will never drop to 0.
1132                  *
1133                  * The data is not used but needs to be initialized to
1134                  * pass assertion muster.  We use this chain primarily
1135                  * as a placeholder for the freemap's top-level RBTREE
1136                  * so it does not interfere with the volume's topology
1137                  * RBTREE.
1138                  */
1139                 hmp->fchain.hmp = hmp;
1140                 hmp->fchain.refs = 1;
1141                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1142                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1143                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1144                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1145                 hmp->fchain.bref.methods =
1146                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1147                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1148
1149                 hammer2_chain_core_init(&hmp->fchain);
1150                 /* hmp->fchain.u.xxx is left NULL */
1151
1152                 /*
1153                  * Install the volume header and initialize fields from
1154                  * voldata.
1155                  */
1156                 error = hammer2_install_volume_header(hmp);
1157                 if (error) {
1158                         hammer2_unmount_helper(mp, NULL, hmp);
1159                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1160                         hammer2_vfs_unmount(mp, MNT_FORCE);
1161                         return error;
1162                 }
1163
1164                 /*
1165                  * Really important to get these right or flush will get
1166                  * confused.
1167                  */
1168                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1169                 kprintf("alloc spmp %p tid %016jx\n",
1170                         hmp->spmp, hmp->voldata.mirror_tid);
1171                 spmp = hmp->spmp;
1172
1173                 /*
1174                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1175                  * is inherited from the volume header.
1176                  */
1177                 xid = 0;
1178                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1179                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1180                 hmp->vchain.pmp = spmp;
1181                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1182                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1183                 hmp->fchain.pmp = spmp;
1184
1185                 /*
1186                  * First locate the super-root inode, which is key 0
1187                  * relative to the volume header's blockset.
1188                  *
1189                  * Then locate the root inode by scanning the directory keyspace
1190                  * represented by the label.
1191                  */
1192                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1193                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1194                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1195                                       &error, 0);
1196                 hammer2_chain_lookup_done(parent);
1197                 if (schain == NULL) {
1198                         kprintf("hammer2_mount: invalid super-root\n");
1199                         hammer2_unmount_helper(mp, NULL, hmp);
1200                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1201                         hammer2_vfs_unmount(mp, MNT_FORCE);
1202                         return EINVAL;
1203                 }
1204                 if (schain->error) {
1205                         kprintf("hammer2_mount: error %s reading super-root\n",
1206                                 hammer2_error_str(schain->error));
1207                         hammer2_chain_unlock(schain);
1208                         hammer2_chain_drop(schain);
1209                         schain = NULL;
1210                         hammer2_unmount_helper(mp, NULL, hmp);
1211                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1212                         hammer2_vfs_unmount(mp, MNT_FORCE);
1213                         return EINVAL;
1214                 }
1215
1216                 /*
1217                  * The super-root always uses an inode_tid of 1 when
1218                  * creating PFSs.
1219                  */
1220                 spmp->inode_tid = 1;
1221                 spmp->modify_tid = schain->bref.modify_tid + 1;
1222
1223                 /*
1224                  * Sanity-check schain's pmp and finish initialization.
1225                  * Any chain belonging to the super-root topology should
1226                  * have a NULL pmp (not even set to spmp).
1227                  */
1228                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1229                 KKASSERT(schain->pmp == NULL);
1230                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1231
1232                 /*
1233                  * Replace the dummy spmp->iroot with a real one.  It's
1234                  * easier to just do a wholesale replacement than to try
1235                  * to update the chain and fixup the iroot fields.
1236                  *
1237                  * The returned inode is locked with the supplied cluster.
1238                  */
1239                 cluster = hammer2_cluster_from_chain(schain);
1240                 hammer2_inode_drop(spmp->iroot);
1241                 spmp->iroot = NULL;
1242                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1243                 spmp->spmp_hmp = hmp;
1244                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1245                 spmp->pfs_hmps[0] = hmp;
1246                 hammer2_inode_ref(spmp->iroot);
1247                 hammer2_inode_unlock(spmp->iroot);
1248                 hammer2_cluster_unlock(cluster);
1249                 hammer2_cluster_drop(cluster);
1250                 schain = NULL;
1251                 /* leave spmp->iroot with one ref */
1252
1253                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1254                         error = hammer2_recovery(hmp);
1255                         /* XXX do something with error */
1256                 }
1257                 hammer2_update_pmps(hmp);
1258                 hammer2_iocom_init(hmp);
1259                 hammer2_bulkfree_init(hmp);
1260
1261                 /*
1262                  * Ref the cluster management messaging descriptor.  The mount
1263                  * program deals with the other end of the communications pipe.
1264                  *
1265                  * Root mounts typically do not supply one.
1266                  */
1267                 if (info.cluster_fd >= 0) {
1268                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1269                         if (fp) {
1270                                 hammer2_cluster_reconnect(hmp, fp);
1271                         } else {
1272                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1273                         }
1274                 }
1275         } else {
1276                 spmp = hmp->spmp;
1277                 if (info.hflags & HMNT2_DEVFLAGS) {
1278                         kprintf("hammer2: Warning: mount flags pertaining "
1279                                 "to the whole device may only be specified "
1280                                 "on the first mount of the device: %08x\n",
1281                                 info.hflags & HMNT2_DEVFLAGS);
1282                 }
1283         }
1284
1285         /*
1286          * Force local mount (disassociate all PFSs from their clusters).
1287          * Used primarily for debugging.
1288          */
1289         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1290
1291         /*
1292          * Lookup the mount point under the media-localized super-root.
1293          * Scanning hammer2_pfslist doesn't help us because it represents
1294          * PFS cluster ids which can aggregate several named PFSs together.
1295          *
1296          * cluster->pmp will incorrectly point to spmp and must be fixed
1297          * up later on.
1298          */
1299         hammer2_inode_lock(spmp->iroot, 0);
1300         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1301         lhc = hammer2_dirhash(label, strlen(label));
1302         chain = hammer2_chain_lookup(&parent, &key_next,
1303                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1304                                      &error, 0);
1305         while (chain) {
1306                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1307                     strcmp(label, chain->data->ipdata.filename) == 0) {
1308                         break;
1309                 }
1310                 chain = hammer2_chain_next(&parent, chain, &key_next,
1311                                             key_next,
1312                                             lhc + HAMMER2_DIRHASH_LOMASK,
1313                                             &error, 0);
1314         }
1315         if (parent) {
1316                 hammer2_chain_unlock(parent);
1317                 hammer2_chain_drop(parent);
1318         }
1319         hammer2_inode_unlock(spmp->iroot);
1320
1321         /*
1322          * PFS could not be found?
1323          */
1324         if (chain == NULL) {
1325                 if (error)
1326                         kprintf("hammer2_mount: PFS label I/O error\n");
1327                 else
1328                         kprintf("hammer2_mount: PFS label not found\n");
1329                 hammer2_unmount_helper(mp, NULL, hmp);
1330                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1331                 hammer2_vfs_unmount(mp, MNT_FORCE);
1332
1333                 return EINVAL;
1334         }
1335
1336         /*
1337          * Acquire the pmp structure (it should have already been allocated
1338          * via hammer2_update_pmps() so do not pass cluster in to add to
1339          * available chains).
1340          *
1341          * Check if the cluster has already been mounted.  A cluster can
1342          * only be mounted once, use null mounts to mount additional copies.
1343          */
1344         if (chain->error) {
1345                 kprintf("hammer2_mount: PFS label I/O error\n");
1346         } else {
1347                 ripdata = &chain->data->ipdata;
1348                 bref = chain->bref;
1349                 pmp = hammer2_pfsalloc(NULL, ripdata,
1350                                        bref.modify_tid, force_local);
1351         }
1352         hammer2_chain_unlock(chain);
1353         hammer2_chain_drop(chain);
1354
1355         /*
1356          * Finish the mount
1357          */
1358         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1359
1360         if (pmp->mp) {
1361                 kprintf("hammer2_mount: PFS already mounted!\n");
1362                 hammer2_unmount_helper(mp, NULL, hmp);
1363                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1364                 hammer2_vfs_unmount(mp, MNT_FORCE);
1365
1366                 return EBUSY;
1367         }
1368
1369         pmp->hflags = info.hflags;
1370         mp->mnt_flag |= MNT_LOCAL;
1371         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1372         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1373
1374         /*
1375          * required mount structure initializations
1376          */
1377         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1378         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1379
1380         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1381         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1382
1383         /*
1384          * Optional fields
1385          */
1386         mp->mnt_iosize_max = MAXPHYS;
1387
1388         /*
1389          * Connect up mount pointers.
1390          */
1391         hammer2_mount_helper(mp, pmp);
1392
1393         lockmgr(&hammer2_mntlk, LK_RELEASE);
1394
1395         /*
1396          * Finish setup
1397          */
1398         vfs_getnewfsid(mp);
1399         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1400         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1401         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1402
1403         if (path) {
1404                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1405                           MNAMELEN - 1, &size);
1406                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1407         } /* else root mount, already in there */
1408
1409         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1410         if (path) {
1411                 copyinstr(path, mp->mnt_stat.f_mntonname,
1412                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1413                           &size);
1414         } else {
1415                 /* root mount */
1416                 mp->mnt_stat.f_mntonname[0] = '/';
1417         }
1418
1419         /*
1420          * Initial statfs to prime mnt_stat.
1421          */
1422         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1423
1424         return 0;
1425 }
1426
1427 /*
1428  * Scan PFSs under the super-root and create hammer2_pfs structures.
1429  */
1430 static
1431 void
1432 hammer2_update_pmps(hammer2_dev_t *hmp)
1433 {
1434         const hammer2_inode_data_t *ripdata;
1435         hammer2_chain_t *parent;
1436         hammer2_chain_t *chain;
1437         hammer2_blockref_t bref;
1438         hammer2_dev_t *force_local;
1439         hammer2_pfs_t *spmp;
1440         hammer2_pfs_t *pmp;
1441         hammer2_key_t key_next;
1442         int error;
1443
1444         /*
1445          * Force local mount (disassociate all PFSs from their clusters).
1446          * Used primarily for debugging.
1447          */
1448         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1449
1450         /*
1451          * Lookup mount point under the media-localized super-root.
1452          *
1453          * cluster->pmp will incorrectly point to spmp and must be fixed
1454          * up later on.
1455          */
1456         spmp = hmp->spmp;
1457         hammer2_inode_lock(spmp->iroot, 0);
1458         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1459         chain = hammer2_chain_lookup(&parent, &key_next,
1460                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1461                                          &error, 0);
1462         while (chain) {
1463                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1464                         continue;
1465                 if (chain->error) {
1466                         kprintf("I/O error scanning PFS labels\n");
1467                 } else {
1468                         ripdata = &chain->data->ipdata;
1469                         bref = chain->bref;
1470                         kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1471
1472                         pmp = hammer2_pfsalloc(chain, ripdata,
1473                                                bref.modify_tid, force_local);
1474                 }
1475                 chain = hammer2_chain_next(&parent, chain, &key_next,
1476                                            key_next, HAMMER2_KEY_MAX,
1477                                            &error, 0);
1478         }
1479         if (parent) {
1480                 hammer2_chain_unlock(parent);
1481                 hammer2_chain_drop(parent);
1482         }
1483         hammer2_inode_unlock(spmp->iroot);
1484 }
1485
1486 static
1487 int
1488 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1489                 struct vnode *devvp, struct ucred *cred)
1490 {
1491         int error;
1492
1493         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1494                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1495                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1496                 vn_unlock(devvp);
1497                 error = hammer2_recovery(hmp);
1498                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1499                 if (error == 0) {
1500                         VOP_CLOSE(devvp, FREAD, NULL);
1501                         hmp->ronly = 0;
1502                 } else {
1503                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1504                 }
1505                 vn_unlock(devvp);
1506         } else {
1507                 error = 0;
1508         }
1509         return error;
1510 }
1511
1512 static
1513 int
1514 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1515 {
1516         hammer2_pfs_t *pmp;
1517         int flags;
1518         int error = 0;
1519
1520         pmp = MPTOPMP(mp);
1521
1522         if (pmp == NULL)
1523                 return(0);
1524
1525         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1526
1527         /*
1528          * If mount initialization proceeded far enough we must flush
1529          * its vnodes and sync the underlying mount points.  Three syncs
1530          * are required to fully flush the filesystem (freemap updates lag
1531          * by one flush, and one extra for safety).
1532          */
1533         if (mntflags & MNT_FORCE)
1534                 flags = FORCECLOSE;
1535         else
1536                 flags = 0;
1537         if (pmp->iroot) {
1538                 error = vflush(mp, 0, flags);
1539                 if (error)
1540                         goto failed;
1541                 hammer2_vfs_sync(mp, MNT_WAIT);
1542                 hammer2_vfs_sync(mp, MNT_WAIT);
1543                 hammer2_vfs_sync(mp, MNT_WAIT);
1544         }
1545
1546         /*
1547          * Cleanup the frontend support XOPS threads
1548          */
1549         hammer2_xop_helper_cleanup(pmp);
1550
1551         if (pmp->mp)
1552                 hammer2_unmount_helper(mp, pmp, NULL);
1553
1554         error = 0;
1555 failed:
1556         lockmgr(&hammer2_mntlk, LK_RELEASE);
1557
1558         return (error);
1559 }
1560
1561 /*
1562  * Mount helper, hook the system mount into our PFS.
1563  * The mount lock is held.
1564  *
1565  * We must bump the mount_count on related devices for any
1566  * mounted PFSs.
1567  */
1568 static
1569 void
1570 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1571 {
1572         hammer2_cluster_t *cluster;
1573         hammer2_chain_t *rchain;
1574         int i;
1575
1576         mp->mnt_data = (qaddr_t)pmp;
1577         pmp->mp = mp;
1578
1579         /*
1580          * After pmp->mp is set we have to adjust hmp->mount_count.
1581          */
1582         cluster = &pmp->iroot->cluster;
1583         for (i = 0; i < cluster->nchains; ++i) {
1584                 rchain = cluster->array[i].chain;
1585                 if (rchain == NULL)
1586                         continue;
1587                 ++rchain->hmp->mount_count;
1588                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1589                         rchain->hmp, rchain->hmp->mount_count);
1590         }
1591
1592         /*
1593          * Create missing Xop threads
1594          */
1595         hammer2_xop_helper_create(pmp);
1596 }
1597
1598 /*
1599  * Mount helper, unhook the system mount from our PFS.
1600  * The mount lock is held.
1601  *
1602  * If hmp is supplied a mount responsible for being the first to open
1603  * the block device failed and the block device and all PFSs using the
1604  * block device must be cleaned up.
1605  *
1606  * If pmp is supplied multiple devices might be backing the PFS and each
1607  * must be disconnected.  This might not be the last PFS using some of the
1608  * underlying devices.  Also, we have to adjust our hmp->mount_count
1609  * accounting for the devices backing the pmp which is now undergoing an
1610  * unmount.
1611  */
1612 static
1613 void
1614 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1615 {
1616         hammer2_cluster_t *cluster;
1617         hammer2_chain_t *rchain;
1618         struct vnode *devvp;
1619         int dumpcnt;
1620         int ronly;
1621         int i;
1622
1623         /*
1624          * If no device supplied this is a high-level unmount and we have to
1625          * to disconnect the mount, adjust mount_count, and locate devices
1626          * that might now have no mounts.
1627          */
1628         if (pmp) {
1629                 KKASSERT(hmp == NULL);
1630                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1631                 pmp->mp = NULL;
1632                 mp->mnt_data = NULL;
1633
1634                 /*
1635                  * After pmp->mp is cleared we have to account for
1636                  * mount_count.
1637                  */
1638                 cluster = &pmp->iroot->cluster;
1639                 for (i = 0; i < cluster->nchains; ++i) {
1640                         rchain = cluster->array[i].chain;
1641                         if (rchain == NULL)
1642                                 continue;
1643                         --rchain->hmp->mount_count;
1644                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1645                                 rchain->hmp, rchain->hmp->mount_count);
1646                         /* scrapping hmp now may invalidate the pmp */
1647                 }
1648 again:
1649                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1650                         if (hmp->mount_count == 0) {
1651                                 hammer2_unmount_helper(NULL, NULL, hmp);
1652                                 goto again;
1653                         }
1654                 }
1655                 return;
1656         }
1657
1658         /*
1659          * Try to terminate the block device.  We can't terminate it if
1660          * there are still PFSs referencing it.
1661          */
1662         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1663                 hmp, hmp->mount_count);
1664         if (hmp->mount_count)
1665                 return;
1666
1667         hammer2_bulkfree_uninit(hmp);
1668         hammer2_pfsfree_scan(hmp);
1669         hammer2_dev_exlock(hmp);        /* XXX order */
1670
1671         /*
1672          * Cycle the volume data lock as a safety (probably not needed any
1673          * more).  To ensure everything is out we need to flush at least
1674          * three times.  (1) The running of the sideq can dirty the
1675          * filesystem, (2) A normal flush can dirty the freemap, and
1676          * (3) ensure that the freemap is fully synchronized.
1677          *
1678          * The next mount's recovery scan can clean everything up but we want
1679          * to leave the filesystem in a 100% clean state on a normal unmount.
1680          */
1681 #if 0
1682         hammer2_voldata_lock(hmp);
1683         hammer2_voldata_unlock(hmp);
1684 #endif
1685         hammer2_iocom_uninit(hmp);
1686
1687         if ((hmp->vchain.flags | hmp->fchain.flags) &
1688             HAMMER2_CHAIN_FLUSH_MASK) {
1689                 kprintf("hammer2_unmount: chains left over "
1690                         "after final sync\n");
1691                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1692                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1693
1694                 if (hammer2_debug & 0x0010)
1695                         Debugger("entered debugger");
1696         }
1697
1698         KKASSERT(hmp->spmp == NULL);
1699
1700         /*
1701          * Finish up with the device vnode
1702          */
1703         if ((devvp = hmp->devvp) != NULL) {
1704                 ronly = hmp->ronly;
1705                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1706                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1707                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1708                         ronly);
1709                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1710                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1711                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1712                 hmp->devvp = NULL;
1713                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1714                 vn_unlock(devvp);
1715                 vrele(devvp);
1716                 devvp = NULL;
1717         }
1718
1719         /*
1720          * Clear vchain/fchain flags that might prevent final cleanup
1721          * of these chains.
1722          */
1723         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1724                 atomic_add_long(&hammer2_count_modified_chains, -1);
1725                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1726                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1727         }
1728         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1729                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1730         }
1731
1732         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1733                 atomic_add_long(&hammer2_count_modified_chains, -1);
1734                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1735                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1736         }
1737         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1738                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1739         }
1740
1741         /*
1742          * Final drop of embedded freemap root chain to
1743          * clean up fchain.core (fchain structure is not
1744          * flagged ALLOCATED so it is cleaned out and then
1745          * left to rot).
1746          */
1747         hammer2_chain_drop(&hmp->fchain);
1748
1749         /*
1750          * Final drop of embedded volume root chain to clean
1751          * up vchain.core (vchain structure is not flagged
1752          * ALLOCATED so it is cleaned out and then left to
1753          * rot).
1754          */
1755         dumpcnt = 50;
1756         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1757         dumpcnt = 50;
1758         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1759         hammer2_dev_unlock(hmp);
1760         hammer2_chain_drop(&hmp->vchain);
1761
1762         hammer2_io_cleanup(hmp, &hmp->iotree);
1763         if (hmp->iofree_count) {
1764                 kprintf("io_cleanup: %d I/O's left hanging\n",
1765                         hmp->iofree_count);
1766         }
1767
1768         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1769         kmalloc_destroy(&hmp->mchain);
1770         kfree(hmp, M_HAMMER2);
1771 }
1772
1773 int
1774 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1775                  ino_t ino, struct vnode **vpp)
1776 {
1777         hammer2_xop_lookup_t *xop;
1778         hammer2_pfs_t *pmp;
1779         hammer2_inode_t *ip;
1780         hammer2_tid_t inum;
1781         int error;
1782
1783         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1784
1785         error = 0;
1786         pmp = MPTOPMP(mp);
1787
1788         /*
1789          * Easy if we already have it cached
1790          */
1791         ip = hammer2_inode_lookup(pmp, inum);
1792         if (ip) {
1793                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1794                 *vpp = hammer2_igetv(ip, &error);
1795                 hammer2_inode_unlock(ip);
1796                 hammer2_inode_drop(ip);         /* from lookup */
1797
1798                 return error;
1799         }
1800
1801         /*
1802          * Otherwise we have to find the inode
1803          */
1804         xop = hammer2_xop_alloc(pmp->iroot, 0);
1805         xop->lhc = inum;
1806         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1807         error = hammer2_xop_collect(&xop->head, 0);
1808
1809         if (error == 0) {
1810                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1811                         kprintf("vget: no collect error but also no rdata\n");
1812                         kprintf("xop %p\n", xop);
1813                         while ((hammer2_debug & 0x80000) == 0) {
1814                                 tsleep(xop, PCATCH, "wait", hz * 10);
1815                         }
1816                         ip = NULL;
1817                 } else {
1818                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1819                 }
1820         }
1821         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1822
1823         if (ip) {
1824                 *vpp = hammer2_igetv(ip, &error);
1825                 hammer2_inode_unlock(ip);
1826         } else {
1827                 *vpp = NULL;
1828                 error = ENOENT;
1829         }
1830         return (error);
1831 }
1832
1833 static
1834 int
1835 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1836 {
1837         hammer2_pfs_t *pmp;
1838         struct vnode *vp;
1839         int error;
1840
1841         pmp = MPTOPMP(mp);
1842         if (pmp->iroot == NULL) {
1843                 *vpp = NULL;
1844                 return EINVAL;
1845         }
1846
1847         error = 0;
1848         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1849
1850         while (pmp->inode_tid == 0) {
1851                 hammer2_xop_ipcluster_t *xop;
1852                 hammer2_inode_meta_t *meta;
1853
1854                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1855                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1856                 error = hammer2_xop_collect(&xop->head, 0);
1857
1858                 if (error == 0) {
1859                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1860                         pmp->iroot->meta = *meta;
1861                         pmp->inode_tid = meta->pfs_inum + 1;
1862                         if (pmp->inode_tid < HAMMER2_INODE_START)
1863                                 pmp->inode_tid = HAMMER2_INODE_START;
1864                         pmp->modify_tid =
1865                                 xop->head.cluster.focus->bref.modify_tid + 1;
1866                         kprintf("PFS: Starting inode %jd\n",
1867                                 (intmax_t)pmp->inode_tid);
1868                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1869                                 pmp->inode_tid, pmp->modify_tid);
1870                         wakeup(&pmp->iroot);
1871
1872                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1873
1874                         /*
1875                          * Prime the mount info.
1876                          */
1877                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1878                         break;
1879                 }
1880
1881                 /*
1882                  * Loop, try again
1883                  */
1884                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1885                 hammer2_inode_unlock(pmp->iroot);
1886                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1887                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1888                 if (error == EINTR)
1889                         break;
1890         }
1891
1892         if (error) {
1893                 hammer2_inode_unlock(pmp->iroot);
1894                 *vpp = NULL;
1895         } else {
1896                 vp = hammer2_igetv(pmp->iroot, &error);
1897                 hammer2_inode_unlock(pmp->iroot);
1898                 *vpp = vp;
1899         }
1900
1901         return (error);
1902 }
1903
1904 /*
1905  * Filesystem status
1906  *
1907  * XXX incorporate ipdata->meta.inode_quota and data_quota
1908  */
1909 static
1910 int
1911 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1912 {
1913         hammer2_pfs_t *pmp;
1914         hammer2_dev_t *hmp;
1915         hammer2_blockref_t bref;
1916         int i;
1917
1918         /*
1919          * NOTE: iroot might not have validated the cluster yet.
1920          */
1921         pmp = MPTOPMP(mp);
1922
1923         mp->mnt_stat.f_files = 0;
1924         mp->mnt_stat.f_ffree = 0;
1925         mp->mnt_stat.f_blocks = 0;
1926         mp->mnt_stat.f_bfree = 0;
1927         mp->mnt_stat.f_bavail = 0;
1928
1929         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1930                 hmp = pmp->pfs_hmps[i];
1931                 if (hmp == NULL)
1932                         continue;
1933                 if (pmp->iroot->cluster.array[i].chain)
1934                         bref = pmp->iroot->cluster.array[i].chain->bref;
1935                 else
1936                         bzero(&bref, sizeof(bref));
1937
1938                 mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1939                 mp->mnt_stat.f_ffree = 0;
1940                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1941                                         mp->mnt_vstat.f_bsize;
1942                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1943                                         mp->mnt_vstat.f_bsize;
1944                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1945
1946                 if (cred && cred->cr_uid != 0) {
1947                         uint64_t adj;
1948
1949                         /* 5% */
1950                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
1951                         mp->mnt_stat.f_blocks -= adj;
1952                         mp->mnt_stat.f_bfree -= adj;
1953                         mp->mnt_stat.f_bavail -= adj;
1954                 }
1955
1956                 *sbp = mp->mnt_stat;
1957         }
1958         return (0);
1959 }
1960
1961 static
1962 int
1963 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1964 {
1965         hammer2_pfs_t *pmp;
1966         hammer2_dev_t *hmp;
1967         hammer2_blockref_t bref;
1968         int i;
1969
1970         /*
1971          * NOTE: iroot might not have validated the cluster yet.
1972          */
1973         pmp = MPTOPMP(mp);
1974
1975         mp->mnt_vstat.f_bsize = 0;
1976         mp->mnt_vstat.f_files = 0;
1977         mp->mnt_vstat.f_ffree = 0;
1978         mp->mnt_vstat.f_blocks = 0;
1979         mp->mnt_vstat.f_bfree = 0;
1980         mp->mnt_vstat.f_bavail = 0;
1981
1982         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1983                 hmp = pmp->pfs_hmps[i];
1984                 if (hmp == NULL)
1985                         continue;
1986                 if (pmp->iroot->cluster.array[i].chain)
1987                         bref = pmp->iroot->cluster.array[i].chain->bref;
1988                 else
1989                         bzero(&bref, sizeof(bref));
1990
1991                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1992                 mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
1993                 mp->mnt_vstat.f_ffree = 0;
1994                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1995                                         mp->mnt_vstat.f_bsize;
1996                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1997                                         mp->mnt_vstat.f_bsize;
1998                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1999
2000                 if (cred && cred->cr_uid != 0) {
2001                         uint64_t adj;
2002
2003                         /* 5% */
2004                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2005                         mp->mnt_vstat.f_blocks -= adj;
2006                         mp->mnt_vstat.f_bfree -= adj;
2007                         mp->mnt_vstat.f_bavail -= adj;
2008                 }
2009
2010                 *sbp = mp->mnt_vstat;
2011         }
2012         return (0);
2013 }
2014
2015 /*
2016  * Mount-time recovery (RW mounts)
2017  *
2018  * Updates to the free block table are allowed to lag flushes by one
2019  * transaction.  In case of a crash, then on a fresh mount we must do an
2020  * incremental scan of the last committed transaction id and make sure that
2021  * all related blocks have been marked allocated.
2022  *
2023  * The super-root topology and each PFS has its own transaction id domain,
2024  * so we must track PFS boundary transitions.
2025  */
2026 struct hammer2_recovery_elm {
2027         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2028         hammer2_chain_t *chain;
2029         hammer2_tid_t sync_tid;
2030 };
2031
2032 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2033
2034 struct hammer2_recovery_info {
2035         struct hammer2_recovery_list list;
2036         hammer2_tid_t   mtid;
2037         int     depth;
2038 };
2039
2040 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2041                         hammer2_chain_t *parent,
2042                         struct hammer2_recovery_info *info,
2043                         hammer2_tid_t sync_tid);
2044
2045 #define HAMMER2_RECOVERY_MAXDEPTH       10
2046
2047 static
2048 int
2049 hammer2_recovery(hammer2_dev_t *hmp)
2050 {
2051         struct hammer2_recovery_info info;
2052         struct hammer2_recovery_elm *elm;
2053         hammer2_chain_t *parent;
2054         hammer2_tid_t sync_tid;
2055         hammer2_tid_t mirror_tid;
2056         int error;
2057
2058         hammer2_trans_init(hmp->spmp, 0);
2059
2060         sync_tid = hmp->voldata.freemap_tid;
2061         mirror_tid = hmp->voldata.mirror_tid;
2062
2063         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2064         if (sync_tid >= mirror_tid) {
2065                 kprintf(" no recovery needed\n");
2066         } else {
2067                 kprintf(" freemap recovery %016jx-%016jx\n",
2068                         sync_tid + 1, mirror_tid);
2069         }
2070
2071         TAILQ_INIT(&info.list);
2072         info.depth = 0;
2073         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2074         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2075         hammer2_chain_lookup_done(parent);
2076
2077         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2078                 TAILQ_REMOVE(&info.list, elm, entry);
2079                 parent = elm->chain;
2080                 sync_tid = elm->sync_tid;
2081                 kfree(elm, M_HAMMER2);
2082
2083                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2084                 error |= hammer2_recovery_scan(hmp, parent, &info,
2085                                               hmp->voldata.freemap_tid);
2086                 hammer2_chain_unlock(parent);
2087                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2088         }
2089         hammer2_trans_done(hmp->spmp);
2090
2091         return error;
2092 }
2093
2094 static
2095 int
2096 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2097                       struct hammer2_recovery_info *info,
2098                       hammer2_tid_t sync_tid)
2099 {
2100         const hammer2_inode_data_t *ripdata;
2101         hammer2_chain_t *chain;
2102         hammer2_blockref_t bref;
2103         int tmp_error;
2104         int rup_error;
2105         int error;
2106         int first;
2107
2108         /*
2109          * Adjust freemap to ensure that the block(s) are marked allocated.
2110          */
2111         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2112                 hammer2_freemap_adjust(hmp, &parent->bref,
2113                                        HAMMER2_FREEMAP_DORECOVER);
2114         }
2115
2116         /*
2117          * Check type for recursive scan
2118          */
2119         switch(parent->bref.type) {
2120         case HAMMER2_BREF_TYPE_VOLUME:
2121                 /* data already instantiated */
2122                 break;
2123         case HAMMER2_BREF_TYPE_INODE:
2124                 /*
2125                  * Must instantiate data for DIRECTDATA test and also
2126                  * for recursion.
2127                  */
2128                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2129                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2130                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2131                         /* not applicable to recovery scan */
2132                         hammer2_chain_unlock(parent);
2133                         return 0;
2134                 }
2135                 hammer2_chain_unlock(parent);
2136                 break;
2137         case HAMMER2_BREF_TYPE_INDIRECT:
2138                 /*
2139                  * Must instantiate data for recursion
2140                  */
2141                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2142                 hammer2_chain_unlock(parent);
2143                 break;
2144         case HAMMER2_BREF_TYPE_DIRENT:
2145         case HAMMER2_BREF_TYPE_DATA:
2146         case HAMMER2_BREF_TYPE_FREEMAP:
2147         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2148         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2149                 /* not applicable to recovery scan */
2150                 return 0;
2151                 break;
2152         default:
2153                 return HAMMER2_ERROR_BADBREF;
2154         }
2155
2156         /*
2157          * Defer operation if depth limit reached or if we are crossing a
2158          * PFS boundary.
2159          */
2160         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2161                 struct hammer2_recovery_elm *elm;
2162
2163                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2164                 elm->chain = parent;
2165                 elm->sync_tid = sync_tid;
2166                 hammer2_chain_ref(parent);
2167                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2168                 /* unlocked by caller */
2169
2170                 return(0);
2171         }
2172
2173
2174         /*
2175          * Recursive scan of the last flushed transaction only.  We are
2176          * doing this without pmp assignments so don't leave the chains
2177          * hanging around after we are done with them.
2178          *
2179          * error        Cumulative error this level only
2180          * rup_error    Cumulative error for recursion
2181          * tmp_error    Specific non-cumulative recursion error
2182          */
2183         chain = NULL;
2184         first = 1;
2185         rup_error = 0;
2186         error = 0;
2187
2188         for (;;) {
2189                 error |= hammer2_chain_scan(parent, &chain, &bref,
2190                                             &first,
2191                                             HAMMER2_LOOKUP_NODATA);
2192
2193                 /*
2194                  * Problem during scan or EOF
2195                  */
2196                 if (error)
2197                         break;
2198
2199                 /*
2200                  * If this is a leaf
2201                  */
2202                 if (chain == NULL) {
2203                         if (bref.mirror_tid > sync_tid) {
2204                                 hammer2_freemap_adjust(hmp, &bref,
2205                                                      HAMMER2_FREEMAP_DORECOVER);
2206                         }
2207                         continue;
2208                 }
2209
2210                 /*
2211                  * This may or may not be a recursive node.
2212                  */
2213                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2214                 if (bref.mirror_tid > sync_tid) {
2215                         ++info->depth;
2216                         tmp_error = hammer2_recovery_scan(hmp, chain,
2217                                                            info, sync_tid);
2218                         --info->depth;
2219                 } else {
2220                         tmp_error = 0;
2221                 }
2222
2223                 /*
2224                  * Flush the recovery at the PFS boundary to stage it for
2225                  * the final flush of the super-root topology.
2226                  */
2227                 if (tmp_error == 0 &&
2228                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2229                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2230                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2231                 }
2232                 rup_error |= tmp_error;
2233         }
2234         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2235 }
2236
2237 /*
2238  * Sync a mount point; this is called on a per-mount basis from the
2239  * filesystem syncer process periodically and whenever a user issues
2240  * a sync.
2241  */
2242 int
2243 hammer2_vfs_sync(struct mount *mp, int waitfor)
2244 {
2245         hammer2_xop_flush_t *xop;
2246         struct hammer2_sync_info info;
2247         hammer2_inode_t *iroot;
2248         hammer2_pfs_t *pmp;
2249         int flags;
2250         int error;
2251
2252         pmp = MPTOPMP(mp);
2253         iroot = pmp->iroot;
2254         KKASSERT(iroot);
2255         KKASSERT(iroot->pmp == pmp);
2256
2257         /*
2258          * We can't acquire locks on existing vnodes while in a transaction
2259          * without risking a deadlock.  This assumes that vfsync() can be
2260          * called without the vnode locked (which it can in DragonFly).
2261          * Otherwise we'd have to implement a multi-pass or flag the lock
2262          * failures and retry.
2263          *
2264          * The reclamation code interlocks with the sync list's token
2265          * (by removing the vnode from the scan list) before unlocking
2266          * the inode, giving us time to ref the inode.
2267          */
2268         /*flags = VMSC_GETVP;*/
2269         flags = 0;
2270         if (waitfor & MNT_LAZY)
2271                 flags |= VMSC_ONEPASS;
2272
2273         /*
2274          * Preflush the vnodes using a normal transaction before interlocking
2275          * with a flush transaction.  We do this to try to run as much of
2276          * the compression as possible outside the flush transaction.
2277          *
2278          * For efficiency do an async pass before making sure with a
2279          * synchronous pass on all related buffer cache buffers.
2280          */
2281         hammer2_trans_init(pmp, 0);
2282         info.error = 0;
2283         info.waitfor = MNT_NOWAIT;
2284         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2285         info.waitfor = MNT_WAIT;
2286         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2287         hammer2_trans_done(pmp);
2288
2289         /*
2290          * Start our flush transaction.  This does not return until all
2291          * concurrent transactions have completed and will prevent any
2292          * new transactions from running concurrently, except for the
2293          * buffer cache transactions.
2294          *
2295          * NOTE!  It is still possible for the paging code to push pages
2296          *        out via a UIO_NOCOPY hammer2_vop_write() during the main
2297          *        flush.
2298          */
2299         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2300         hammer2_inode_run_sideq(pmp);
2301
2302         info.error = 0;
2303         info.waitfor = MNT_NOWAIT;
2304         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2305         info.waitfor = MNT_WAIT;
2306         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2307         hammer2_bioq_sync(pmp);
2308
2309         /*
2310          * Use the XOP interface to concurrently flush all nodes to
2311          * synchronize the PFSROOT subtopology to the media.  A standard
2312          * end-of-scan ENOENT error indicates cluster sufficiency.
2313          *
2314          * Note that this flush will not be visible on crash recovery until
2315          * we flush the super-root topology in the next loop.
2316          *
2317          * XXX For now wait for all flushes to complete.
2318          */
2319         if (iroot) {
2320                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2321                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2322                 error = hammer2_xop_collect(&xop->head,
2323                                             HAMMER2_XOP_COLLECT_WAITALL);
2324                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2325                 if (error == HAMMER2_ERROR_ENOENT)
2326                         error = 0;
2327                 else
2328                         error = hammer2_error_to_errno(error);
2329         } else {
2330                 error = 0;
2331         }
2332         hammer2_trans_done(pmp);
2333
2334         return (error);
2335 }
2336
2337 /*
2338  * Sync passes.
2339  *
2340  * Note that we ignore the tranasction mtid we got above.  Instead,
2341  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2342  * transactions.
2343  */
2344 static int
2345 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2346 {
2347         struct hammer2_sync_info *info = data;
2348         hammer2_inode_t *ip;
2349         int error;
2350
2351         /*
2352          * Degenerate cases.  Note that ip == NULL typically means the
2353          * syncer vnode itself and we don't want to vclrisdirty() in that
2354          * situation.
2355          */
2356         ip = VTOI(vp);
2357         if (ip == NULL) {
2358                 return(0);
2359         }
2360         if (vp->v_type == VNON || vp->v_type == VBAD) {
2361                 vclrisdirty(vp);
2362                 return(0);
2363         }
2364
2365         /*
2366          * VOP_FSYNC will start a new transaction so replicate some code
2367          * here to do it inline (see hammer2_vop_fsync()).
2368          *
2369          * WARNING: The vfsync interacts with the buffer cache and might
2370          *          block, we can't hold the inode lock at that time.
2371          *          However, we MUST ref ip before blocking to ensure that
2372          *          it isn't ripped out from under us (since we do not
2373          *          hold a lock on the vnode).
2374          */
2375         hammer2_inode_ref(ip);
2376         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2377             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2378                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2379                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2380                                  HAMMER2_INODE_MODIFIED)) {
2381                         hammer2_inode_lock(ip, 0);
2382                         if (ip->flags & (HAMMER2_INODE_RESIZED |
2383                                          HAMMER2_INODE_MODIFIED)) {
2384                                 hammer2_inode_chain_sync(ip);
2385                         }
2386                         hammer2_inode_unlock(ip);
2387                 }
2388         }
2389         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2390             RB_EMPTY(&vp->v_rbdirty_tree)) {
2391                 vclrisdirty(vp);
2392         }
2393
2394         hammer2_inode_drop(ip);
2395 #if 1
2396         error = 0;
2397         if (error)
2398                 info->error = error;
2399 #endif
2400         return(0);
2401 }
2402
2403 static
2404 int
2405 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2406 {
2407         hammer2_inode_t *ip;
2408
2409         KKASSERT(MAXFIDSZ >= 16);
2410         ip = VTOI(vp);
2411         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2412         fhp->fid_ext = 0;
2413         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2414         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2415
2416         return 0;
2417 }
2418
2419 static
2420 int
2421 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2422                struct fid *fhp, struct vnode **vpp)
2423 {
2424         hammer2_pfs_t *pmp;
2425         hammer2_tid_t inum;
2426         int error;
2427
2428         pmp = MPTOPMP(mp);
2429         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2430         if (vpp) {
2431                 if (inum == 1)
2432                         error = hammer2_vfs_root(mp, vpp);
2433                 else
2434                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2435         } else {
2436                 error = 0;
2437         }
2438         if (error)
2439                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2440         return error;
2441 }
2442
2443 static
2444 int
2445 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2446                  int *exflagsp, struct ucred **credanonp)
2447 {
2448         hammer2_pfs_t *pmp;
2449         struct netcred *np;
2450         int error;
2451
2452         pmp = MPTOPMP(mp);
2453         np = vfs_export_lookup(mp, &pmp->export, nam);
2454         if (np) {
2455                 *exflagsp = np->netc_exflags;
2456                 *credanonp = &np->netc_anon;
2457                 error = 0;
2458         } else {
2459                 error = EACCES;
2460         }
2461         return error;
2462 }
2463
2464 /*
2465  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2466  * header into the HMP
2467  *
2468  * XXX read four volhdrs and use the one with the highest TID whos CRC
2469  *     matches.
2470  *
2471  * XXX check iCRCs.
2472  *
2473  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2474  *     nonexistant locations.
2475  *
2476  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2477  */
2478 static
2479 int
2480 hammer2_install_volume_header(hammer2_dev_t *hmp)
2481 {
2482         hammer2_volume_data_t *vd;
2483         struct buf *bp;
2484         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2485         int error_reported;
2486         int error;
2487         int valid;
2488         int i;
2489
2490         error_reported = 0;
2491         error = 0;
2492         valid = 0;
2493         bp = NULL;
2494
2495         /*
2496          * There are up to 4 copies of the volume header (syncs iterate
2497          * between them so there is no single master).  We don't trust the
2498          * volu_size field so we don't know precisely how large the filesystem
2499          * is, so depend on the OS to return an error if we go beyond the
2500          * block device's EOF.
2501          */
2502         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2503                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2504                               HAMMER2_VOLUME_BYTES, &bp);
2505                 if (error) {
2506                         brelse(bp);
2507                         bp = NULL;
2508                         continue;
2509                 }
2510
2511                 vd = (struct hammer2_volume_data *) bp->b_data;
2512                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2513                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2514                         brelse(bp);
2515                         bp = NULL;
2516                         continue;
2517                 }
2518
2519                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2520                         /* XXX: Reversed-endianness filesystem */
2521                         kprintf("hammer2: reverse-endian filesystem detected");
2522                         brelse(bp);
2523                         bp = NULL;
2524                         continue;
2525                 }
2526
2527                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2528                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2529                                       HAMMER2_VOLUME_ICRC0_SIZE);
2530                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2531                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2532                                        HAMMER2_VOLUME_ICRC1_SIZE);
2533                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2534                         kprintf("hammer2 volume header crc "
2535                                 "mismatch copy #%d %08x/%08x\n",
2536                                 i, crc0, crc);
2537                         error_reported = 1;
2538                         brelse(bp);
2539                         bp = NULL;
2540                         continue;
2541                 }
2542                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2543                         valid = 1;
2544                         hmp->voldata = *vd;
2545                         hmp->volhdrno = i;
2546                 }
2547                 brelse(bp);
2548                 bp = NULL;
2549         }
2550         if (valid) {
2551                 hmp->volsync = hmp->voldata;
2552                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2553                 error = 0;
2554                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2555                         kprintf("hammer2: using volume header #%d\n",
2556                                 hmp->volhdrno);
2557                 }
2558         } else {
2559                 error = EINVAL;
2560                 kprintf("hammer2: no valid volume headers found!\n");
2561         }
2562         return (error);
2563 }
2564
2565 /*
2566  * This handles hysteresis on regular file flushes.  Because the BIOs are
2567  * routed to a thread it is possible for an excessive number to build up
2568  * and cause long front-end stalls long before the runningbuffspace limit
2569  * is hit, so we implement hammer2_flush_pipe to control the
2570  * hysteresis.
2571  *
2572  * This is a particular problem when compression is used.
2573  */
2574 void
2575 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2576 {
2577         atomic_add_int(&pmp->count_lwinprog, 1);
2578 }
2579
2580 void
2581 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2582 {
2583         int lwinprog;
2584
2585         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2586         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2587             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2588                 atomic_clear_int(&pmp->count_lwinprog,
2589                                  HAMMER2_LWINPROG_WAITING);
2590                 wakeup(&pmp->count_lwinprog);
2591         }
2592         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2593             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2594                 atomic_clear_int(&pmp->count_lwinprog,
2595                                  HAMMER2_LWINPROG_WAITING0);
2596                 wakeup(&pmp->count_lwinprog);
2597         }
2598 }
2599
2600 void
2601 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2602 {
2603         int lwinprog;
2604         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2605                                     HAMMER2_LWINPROG_WAITING0;
2606
2607         for (;;) {
2608                 lwinprog = pmp->count_lwinprog;
2609                 cpu_ccfence();
2610                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2611                         break;
2612                 tsleep_interlock(&pmp->count_lwinprog, 0);
2613                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2614                 lwinprog = pmp->count_lwinprog;
2615                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2616                         break;
2617                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2618         }
2619 }
2620
2621 /*
2622  * Manage excessive memory resource use for chain and related
2623  * structures.
2624  */
2625 void
2626 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2627 {
2628         uint32_t waiting;
2629         uint32_t count;
2630         uint32_t limit;
2631 #if 0
2632         static int zzticks;
2633 #endif
2634
2635         /*
2636          * Atomic check condition and wait.  Also do an early speedup of
2637          * the syncer to try to avoid hitting the wait.
2638          */
2639         for (;;) {
2640                 waiting = pmp->inmem_dirty_chains;
2641                 cpu_ccfence();
2642                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2643
2644                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2645                 if (limit < hammer2_limit_dirty_chains)
2646                         limit = hammer2_limit_dirty_chains;
2647                 if (limit < 1000)
2648                         limit = 1000;
2649
2650 #if 0
2651                 if ((int)(ticks - zzticks) > hz) {
2652                         zzticks = ticks;
2653                         kprintf("count %ld %ld\n", count, limit);
2654                 }
2655 #endif
2656
2657                 /*
2658                  * Block if there are too many dirty chains present, wait
2659                  * for the flush to clean some out.
2660                  */
2661                 if (count > limit) {
2662                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2663                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2664                                                waiting,
2665                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2666                                 speedup_syncer(pmp->mp);
2667                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2668                                        "chnmem", hz);
2669                         }
2670                         continue;       /* loop on success or fail */
2671                 }
2672
2673                 /*
2674                  * Try to start an early flush before we are forced to block.
2675                  */
2676                 if (count > limit * 7 / 10)
2677                         speedup_syncer(pmp->mp);
2678                 break;
2679         }
2680 }
2681
2682 void
2683 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2684 {
2685         if (pmp) {
2686                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2687         }
2688 }
2689
2690 void
2691 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2692 {
2693         uint32_t waiting;
2694
2695         if (pmp == NULL)
2696                 return;
2697
2698         for (;;) {
2699                 waiting = pmp->inmem_dirty_chains;
2700                 cpu_ccfence();
2701                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2702                                        waiting,
2703                                        (waiting - 1) &
2704                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2705                         break;
2706                 }
2707         }
2708
2709         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2710                 wakeup(&pmp->inmem_dirty_chains);
2711 }
2712
2713 /*
2714  * Returns 0 if the filesystem has tons of free space
2715  * Returns 1 if the filesystem has less than 10% remaining
2716  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2717  */
2718 int
2719 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2720 {
2721         hammer2_pfs_t *pmp;
2722         hammer2_dev_t *hmp;
2723         hammer2_off_t free_reserved;
2724         hammer2_off_t free_nominal;
2725         int i;
2726
2727         pmp = ip->pmp;
2728
2729         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2730                 free_reserved = HAMMER2_SEGSIZE;
2731                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2732                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2733                         hmp = pmp->pfs_hmps[i];
2734                         if (hmp == NULL)
2735                                 continue;
2736                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2737                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2738                                 continue;
2739
2740                         if (free_nominal > hmp->voldata.allocator_free)
2741                                 free_nominal = hmp->voldata.allocator_free;
2742                         if (free_reserved < hmp->free_reserved)
2743                                 free_reserved = hmp->free_reserved;
2744                 }
2745
2746                 /*
2747                  * SMP races ok
2748                  */
2749                 pmp->free_reserved = free_reserved;
2750                 pmp->free_nominal = free_nominal;
2751                 pmp->free_ticks = ticks;
2752         } else {
2753                 free_reserved = pmp->free_reserved;
2754                 free_nominal = pmp->free_nominal;
2755         }
2756         if (cred && cred->cr_uid != 0) {
2757                 if ((int64_t)(free_nominal - bytes) <
2758                     (int64_t)free_reserved) {
2759                         return 2;
2760                 }
2761         } else {
2762                 if ((int64_t)(free_nominal - bytes) <
2763                     (int64_t)free_reserved / 2) {
2764                         return 2;
2765                 }
2766         }
2767         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2768                 return 1;
2769         return 0;
2770 }
2771
2772 /*
2773  * Debugging
2774  */
2775 void
2776 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2777 {
2778         hammer2_chain_t *scan;
2779         hammer2_chain_t *parent;
2780
2781         --*countp;
2782         if (*countp == 0) {
2783                 kprintf("%*.*s...\n", tab, tab, "");
2784                 return;
2785         }
2786         if (*countp < 0)
2787                 return;
2788         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2789                 tab, tab, "", pfx,
2790                 chain, chain->bref.type,
2791                 chain->bref.key, chain->bref.keybits,
2792                 chain->bref.mirror_tid);
2793
2794         kprintf("%*.*s      [%08x] (%s) refs=%d",
2795                 tab, tab, "",
2796                 chain->flags,
2797                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2798                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2799                 chain->refs);
2800
2801         parent = chain->parent;
2802         if (parent)
2803                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2804                         tab, tab, "",
2805                         parent, parent->flags, parent->refs);
2806         if (RB_EMPTY(&chain->core.rbtree)) {
2807                 kprintf("\n");
2808         } else {
2809                 kprintf(" {\n");
2810                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2811                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2812                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2813                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2814                                 chain->data->ipdata.filename);
2815                 else
2816                         kprintf("%*.*s}\n", tab, tab, "");
2817         }
2818 }