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_read = 4;           /* physical read-ahead */
  82 int hammer2_cluster_write = 0;          /* bdwrite() so later inval works */
  83 int hammer2_dedup_enable = 1;
  84 int hammer2_always_compress = 0;        /* always try to compress */
  85 int hammer2_inval_enable = 0;
  86 int hammer2_flush_pipe = 100;
  87 int hammer2_synchronous_flush = 1;
  88 int hammer2_dio_count;
  89 int hammer2_limit_dio = 256;
  90 long hammer2_chain_allocs;
  91 long hammer2_chain_frees;
  92 long hammer2_limit_dirty_chains;
  93 long hammer2_count_modified_chains;
  94 long hammer2_iod_invals;
  95 long hammer2_iod_file_read;
  96 long hammer2_iod_meta_read;
  97 long hammer2_iod_indr_read;
  98 long hammer2_iod_fmap_read;
  99 long hammer2_iod_volu_read;
 100 long hammer2_iod_file_write;
 101 long hammer2_iod_file_wembed;
 102 long hammer2_iod_file_wzero;
 103 long hammer2_iod_file_wdedup;
 104 long hammer2_iod_meta_write;
 105 long hammer2_iod_indr_write;
 106 long hammer2_iod_fmap_write;
 107 long hammer2_iod_volu_write;
 108
 109 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 110 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 111                 "Buffer used for compression.");
 112
 113 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 114 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 115                 "Buffer used for decompression.");
 116
 117 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 118
 119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 120            &hammer2_debug, 0, "");
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW,
 122            &hammer2_cluster_read, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 124            &hammer2_cluster_write, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 126            &hammer2_dedup_enable, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 128            &hammer2_always_compress, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 130            &hammer2_inval_enable, 0, "");
 131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 132            &hammer2_flush_pipe, 0, "");
 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 134            &hammer2_synchronous_flush, 0, "");
 135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 136            &hammer2_chain_allocs, 0, "");
 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 138            &hammer2_chain_frees, 0, "");
 139 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 140            &hammer2_limit_dirty_chains, 0, "");
 141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 142            &hammer2_count_modified_chains, 0, "");
 143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 144            &hammer2_dio_count, 0, "");
 145 SYSCTL_INT(_vfs_hammer2, OID_AUTO, limit_dio, CTLFLAG_RW,
 146            &hammer2_limit_dio, 0, "");
 147
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 149            &hammer2_iod_invals, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 151            &hammer2_iod_file_read, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 153            &hammer2_iod_meta_read, 0, "");
 154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 155            &hammer2_iod_indr_read, 0, "");
 156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 157            &hammer2_iod_fmap_read, 0, "");
 158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 159            &hammer2_iod_volu_read, 0, "");
 160
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 162            &hammer2_iod_file_write, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 164            &hammer2_iod_file_wembed, 0, "");
 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 166            &hammer2_iod_file_wzero, 0, "");
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 168            &hammer2_iod_file_wdedup, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 170            &hammer2_iod_meta_write, 0, "");
 171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 172            &hammer2_iod_indr_write, 0, "");
 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 174            &hammer2_iod_fmap_write, 0, "");
 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 176            &hammer2_iod_volu_write, 0, "");
 177
 178 long hammer2_check_icrc32;
 179 long hammer2_check_xxhash64;
 180 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 181            &hammer2_check_icrc32, 0, "");
 182 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 183            &hammer2_check_xxhash64, 0, "");
 184
 185 static int hammer2_vfs_init(struct vfsconf *conf);
 186 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 187 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 188                                 struct ucred *cred);
 189 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 190                                 struct vnode *, struct ucred *);
 191 static int hammer2_recovery(hammer2_dev_t *hmp);
 192 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 193 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 194 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 195                                 struct ucred *cred);
 196 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 197                                 struct ucred *cred);
 198 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 199                                 struct fid *fhp, struct vnode **vpp);
 200 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 201 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 202                                 int *exflagsp, struct ucred **credanonp);
 203
 204 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 205 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 206
 207 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 208
 209 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 210 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 211                                 hammer2_dev_t *hmp);
 212
 213 /*
 214  * HAMMER2 vfs operations.
 215  */
 216 static struct vfsops hammer2_vfsops = {
 217         .vfs_init       = hammer2_vfs_init,
 218         .vfs_uninit     = hammer2_vfs_uninit,
 219         .vfs_sync       = hammer2_vfs_sync,
 220         .vfs_mount      = hammer2_vfs_mount,
 221         .vfs_unmount    = hammer2_vfs_unmount,
 222         .vfs_root       = hammer2_vfs_root,
 223         .vfs_statfs     = hammer2_vfs_statfs,
 224         .vfs_statvfs    = hammer2_vfs_statvfs,
 225         .vfs_vget       = hammer2_vfs_vget,
 226         .vfs_vptofh     = hammer2_vfs_vptofh,
 227         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 228         .vfs_checkexp   = hammer2_vfs_checkexp
 229 };
 230
 231 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 232
 233 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 234 MODULE_VERSION(hammer2, 1);
 235
 236 static
 237 int
 238 hammer2_vfs_init(struct vfsconf *conf)
 239 {
 240         static struct objcache_malloc_args margs_read;
 241         static struct objcache_malloc_args margs_write;
 242         static struct objcache_malloc_args margs_vop;
 243
 244         int error;
 245
 246         error = 0;
 247
 248         /*
 249          * A large DIO cache is needed to retain dedup enablement masks.
 250          * The bulkfree code clears related masks as part of the disk block
 251          * recycling algorithm, preventing it from being used for a later
 252          * dedup.
 253          *
 254          * NOTE: A large buffer cache can actually interfere with dedup
 255          *       operation because we dedup based on media physical buffers
 256          *       and not logical buffers.  Try to make the DIO chace large
 257          *       enough to avoid this problem, but also cap it.
 258          */
 259         hammer2_limit_dio = nbuf * 2;
 260         if (hammer2_limit_dio > 100000)
 261                 hammer2_limit_dio = 100000;
 262
 263         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 264                 error = EINVAL;
 265         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 266                 error = EINVAL;
 267         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 268                 error = EINVAL;
 269
 270         if (error)
 271                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 272
 273         margs_read.objsize = 65536;
 274         margs_read.mtype = M_HAMMER2_DEBUFFER;
 275
 276         margs_write.objsize = 32768;
 277         margs_write.mtype = M_HAMMER2_CBUFFER;
 278
 279         margs_vop.objsize = sizeof(hammer2_xop_t);
 280         margs_vop.mtype = M_HAMMER2;
 281
 282         /*
 283          * Note thaht for the XOPS cache we want backing store allocations
 284          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 285          * confusion), so use the backing store function that does it.  This
 286          * means that initial XOPS objects are zerod but REUSED objects are
 287          * not.  So we are responsible for cleaning the object up sufficiently
 288          * for our needs before objcache_put()ing it back (typically just the
 289          * FIFO indices).
 290          */
 291         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 292                                 0, 1, NULL, NULL, NULL,
 293                                 objcache_malloc_alloc,
 294                                 objcache_malloc_free,
 295                                 &margs_read);
 296         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 297                                 0, 1, NULL, NULL, NULL,
 298                                 objcache_malloc_alloc,
 299                                 objcache_malloc_free,
 300                                 &margs_write);
 301         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 302                                 0, 1, NULL, NULL, NULL,
 303                                 objcache_malloc_alloc_zero,
 304                                 objcache_malloc_free,
 305                                 &margs_vop);
 306
 307
 308         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 309         TAILQ_INIT(&hammer2_mntlist);
 310         TAILQ_INIT(&hammer2_pfslist);
 311
 312         hammer2_limit_dirty_chains = maxvnodes / 10;
 313         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 314                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 315
 316         return (error);
 317 }
 318
 319 static
 320 int
 321 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 322 {
 323         objcache_destroy(cache_buffer_read);
 324         objcache_destroy(cache_buffer_write);
 325         objcache_destroy(cache_xops);
 326         return 0;
 327 }
 328
 329 /*
 330  * Core PFS allocator.  Used to allocate or reference the pmp structure
 331  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 332  * The pmp can be passed in or loaded by this function using the chain and
 333  * inode data.
 334  *
 335  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 336  * transactions.  Note that synchronization does not use this field.
 337  * (typically frontend operations and synchronization cannot run on the
 338  * same PFS node at the same time).
 339  *
 340  * XXX check locking
 341  */
 342 hammer2_pfs_t *
 343 hammer2_pfsalloc(hammer2_chain_t *chain,
 344                  const hammer2_inode_data_t *ripdata,
 345                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 346 {
 347         hammer2_pfs_t *pmp;
 348         hammer2_inode_t *iroot;
 349         int count;
 350         int i;
 351         int j;
 352
 353         pmp = NULL;
 354
 355         /*
 356          * Locate or create the PFS based on the cluster id.  If ripdata
 357          * is NULL this is a spmp which is unique and is always allocated.
 358          *
 359          * If the device is mounted in local mode all PFSs are considered
 360          * independent and not part of any cluster (for debugging only).
 361          */
 362         if (ripdata) {
 363                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 364                         if (force_local != pmp->force_local)
 365                                 continue;
 366                         if (force_local == NULL &&
 367                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 368                                  sizeof(pmp->pfs_clid)) == 0) {
 369                                         break;
 370                         } else if (force_local && pmp->pfs_names[0] &&
 371                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 372                                         break;
 373                         }
 374                 }
 375         }
 376
 377         if (pmp == NULL) {
 378                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 379                 pmp->force_local = force_local;
 380                 hammer2_trans_manage_init(pmp);
 381                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 382                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 383                 lockinit(&pmp->lock, "pfslk", 0, 0);
 384                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 385                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 386                 spin_init(&pmp->xop_spin, "h2xop");
 387                 spin_init(&pmp->lru_spin, "h2lru");
 388                 RB_INIT(&pmp->inum_tree);
 389                 TAILQ_INIT(&pmp->sideq);
 390                 TAILQ_INIT(&pmp->lru_list);
 391                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 392
 393                 /*
 394                  * Distribute backend operations to threads
 395                  */
 396                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 397                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 398
 399                 /*
 400                  * Save the last media transaction id for the flusher.  Set
 401                  * initial
 402                  */
 403                 if (ripdata)
 404                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 405                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 406
 407                 /*
 408                  * The synchronization thread may start too early, make
 409                  * sure it stays frozen until we are ready to let it go.
 410                  * XXX
 411                  */
 412                 /*
 413                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 414                                          HAMMER2_THREAD_REMASTER;
 415                 */
 416         }
 417
 418         /*
 419          * Create the PFS's root inode and any missing XOP helper threads.
 420          */
 421         if ((iroot = pmp->iroot) == NULL) {
 422                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 423                 if (ripdata)
 424                         iroot->meta = ripdata->meta;
 425                 pmp->iroot = iroot;
 426                 hammer2_inode_ref(iroot);
 427                 hammer2_inode_unlock(iroot);
 428         }
 429
 430         /*
 431          * Stop here if no chain is passed in.
 432          */
 433         if (chain == NULL)
 434                 goto done;
 435
 436         /*
 437          * When a chain is passed in we must add it to the PFS's root
 438          * inode, update pmp->pfs_types[], and update the syncronization
 439          * threads.
 440          *
 441          * When forcing local mode, mark the PFS as a MASTER regardless.
 442          *
 443          * At the moment empty spots can develop due to removals or failures.
 444          * Ultimately we want to re-fill these spots but doing so might
 445          * confused running code. XXX
 446          */
 447         hammer2_inode_ref(iroot);
 448         hammer2_mtx_ex(&iroot->lock);
 449         j = iroot->cluster.nchains;
 450
 451         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 452
 453         if (j == HAMMER2_MAXCLUSTER) {
 454                 kprintf("hammer2_mount: cluster full!\n");
 455                 /* XXX fatal error? */
 456         } else {
 457                 KKASSERT(chain->pmp == NULL);
 458                 chain->pmp = pmp;
 459                 hammer2_chain_ref(chain);
 460                 iroot->cluster.array[j].chain = chain;
 461                 if (force_local)
 462                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 463                 else
 464                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 465                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 466                 pmp->pfs_hmps[j] = chain->hmp;
 467
 468                 /*
 469                  * If the PFS is already mounted we must account
 470                  * for the mount_count here.
 471                  */
 472                 if (pmp->mp)
 473                         ++chain->hmp->mount_count;
 474
 475                 /*
 476                  * May have to fixup dirty chain tracking.  Previous
 477                  * pmp was NULL so nothing to undo.
 478                  */
 479                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 480                         hammer2_pfs_memory_inc(pmp);
 481                 ++j;
 482         }
 483         iroot->cluster.nchains = j;
 484
 485         /*
 486          * Update nmasters from any PFS inode which is part of the cluster.
 487          * It is possible that this will result in a value which is too
 488          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 489          * override this value later on.
 490          *
 491          * (This informs us of masters that might not currently be
 492          *  discoverable by this mount).
 493          */
 494         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 495                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 496         }
 497
 498         /*
 499          * Count visible masters.  Masters are usually added with
 500          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 501          * are more (XXX and must update the master inodes).
 502          */
 503         count = 0;
 504         for (i = 0; i < iroot->cluster.nchains; ++i) {
 505                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 506                         ++count;
 507         }
 508         if (pmp->pfs_nmasters < count)
 509                 pmp->pfs_nmasters = count;
 510
 511         /*
 512          * Create missing synchronization and support threads.
 513          *
 514          * Single-node masters (including snapshots) have nothing to
 515          * synchronize and do not require this thread.
 516          *
 517          * Multi-node masters or any number of soft masters, slaves, copy,
 518          * or other PFS types need the thread.
 519          *
 520          * Each thread is responsible for its particular cluster index.
 521          * We use independent threads so stalls or mismatches related to
 522          * any given target do not affect other targets.
 523          */
 524         for (i = 0; i < iroot->cluster.nchains; ++i) {
 525                 /*
 526                  * Single-node masters (including snapshots) have nothing
 527                  * to synchronize and will make direct xops support calls,
 528                  * thus they do not require this thread.
 529                  *
 530                  * Note that there can be thousands of snapshots.  We do not
 531                  * want to create thousands of threads.
 532                  */
 533                 if (pmp->pfs_nmasters <= 1 &&
 534                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 535                         continue;
 536                 }
 537
 538                 /*
 539                  * Sync support thread
 540                  */
 541                 if (pmp->sync_thrs[i].td == NULL) {
 542                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 543                                            "h2nod", i, -1,
 544                                            hammer2_primary_sync_thread);
 545                 }
 546         }
 547
 548         /*
 549          * Create missing Xop threads
 550          *
 551          * NOTE: We create helper threads for all mounted PFSs or any
 552          *       PFSs with 2+ nodes (so the sync thread can update them,
 553          *       even if not mounted).
 554          */
 555         if (pmp->mp || iroot->cluster.nchains >= 2)
 556                 hammer2_xop_helper_create(pmp);
 557
 558         hammer2_mtx_unlock(&iroot->lock);
 559         hammer2_inode_drop(iroot);
 560 done:
 561         return pmp;
 562 }
 563
 564 /*
 565  * Deallocate an element of a probed PFS.  If destroying and this is a
 566  * MASTER, adjust nmasters.
 567  *
 568  * This function does not physically destroy the PFS element in its device
 569  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 570  */
 571 void
 572 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 573 {
 574         hammer2_inode_t *iroot;
 575         hammer2_chain_t *chain;
 576         int j;
 577
 578         /*
 579          * Cleanup our reference on iroot.  iroot is (should) not be needed
 580          * by the flush code.
 581          */
 582         iroot = pmp->iroot;
 583         if (iroot) {
 584                 /*
 585                  * Stop synchronizing
 586                  *
 587                  * XXX flush after acquiring the iroot lock.
 588                  * XXX clean out the cluster index from all inode structures.
 589                  */
 590                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 591
 592                 /*
 593                  * Remove the cluster index from the group.  If destroying
 594                  * the PFS and this is a master, adjust pfs_nmasters.
 595                  */
 596                 hammer2_mtx_ex(&iroot->lock);
 597                 chain = iroot->cluster.array[clindex].chain;
 598                 iroot->cluster.array[clindex].chain = NULL;
 599
 600                 switch(pmp->pfs_types[clindex]) {
 601                 case HAMMER2_PFSTYPE_MASTER:
 602                         if (destroying && pmp->pfs_nmasters > 0)
 603                                 --pmp->pfs_nmasters;
 604                         /* XXX adjust ripdata->meta.pfs_nmasters */
 605                         break;
 606                 default:
 607                         break;
 608                 }
 609                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 610
 611                 hammer2_mtx_unlock(&iroot->lock);
 612
 613                 /*
 614                  * Release the chain.
 615                  */
 616                 if (chain) {
 617                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 618                         hammer2_chain_drop(chain);
 619                 }
 620
 621                 /*
 622                  * Terminate all XOP threads for the cluster index.
 623                  */
 624                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 625                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 626         }
 627 }
 628
 629 /*
 630  * Destroy a PFS, typically only occurs after the last mount on a device
 631  * has gone away.
 632  */
 633 static void
 634 hammer2_pfsfree(hammer2_pfs_t *pmp)
 635 {
 636         hammer2_inode_t *iroot;
 637         hammer2_chain_t *chain;
 638         int i;
 639         int j;
 640
 641         /*
 642          * Cleanup our reference on iroot.  iroot is (should) not be needed
 643          * by the flush code.
 644          */
 645         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 646
 647         iroot = pmp->iroot;
 648         if (iroot) {
 649                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 650                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 651                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 652                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 653                 }
 654 #if REPORT_REFS_ERRORS
 655                 if (pmp->iroot->refs != 1)
 656                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 657                                 pmp->iroot, pmp->iroot->refs);
 658 #else
 659                 KKASSERT(pmp->iroot->refs == 1);
 660 #endif
 661                 /* ref for pmp->iroot */
 662                 hammer2_inode_drop(pmp->iroot);
 663                 pmp->iroot = NULL;
 664         }
 665
 666         /*
 667          * Cleanup chains remaining on LRU list.
 668          */
 669         kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
 670         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 671                 hammer2_chain_ref(chain);
 672                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 673                 hammer2_chain_drop(chain);
 674         }
 675
 676         /*
 677          * Free remaining pmp resources
 678          */
 679         kmalloc_destroy(&pmp->mmsg);
 680         kmalloc_destroy(&pmp->minode);
 681
 682         kfree(pmp, M_HAMMER2);
 683 }
 684
 685 /*
 686  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 687  * empty is terminated and freed.
 688  *
 689  * XXX inefficient.
 690  */
 691 static void
 692 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 693 {
 694         hammer2_pfs_t *pmp;
 695         hammer2_inode_t *iroot;
 696         hammer2_chain_t *rchain;
 697         int didfreeze;
 698         int i;
 699         int j;
 700
 701 again:
 702         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 703                 if ((iroot = pmp->iroot) == NULL)
 704                         continue;
 705                 if (hmp->spmp == pmp) {
 706                         kprintf("unmount hmp %p remove spmp %p\n",
 707                                 hmp, pmp);
 708                         hmp->spmp = NULL;
 709                 }
 710
 711                 /*
 712                  * Determine if this PFS is affected.  If it is we must
 713                  * freeze all management threads and lock its iroot.
 714                  *
 715                  * Freezing a management thread forces it idle, operations
 716                  * in-progress will be aborted and it will have to start
 717                  * over again when unfrozen, or exit if told to exit.
 718                  */
 719                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 720                         if (pmp->pfs_hmps[i] == hmp)
 721                                 break;
 722                 }
 723                 if (i != HAMMER2_MAXCLUSTER) {
 724                         /*
 725                          * Make sure all synchronization threads are locked
 726                          * down.
 727                          */
 728                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 729                                 if (pmp->pfs_hmps[i] == NULL)
 730                                         continue;
 731                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 732                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 733                                         hammer2_thr_freeze_async(
 734                                                 &pmp->xop_groups[j].thrs[i]);
 735                                 }
 736                         }
 737                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 738                                 if (pmp->pfs_hmps[i] == NULL)
 739                                         continue;
 740                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 741                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 742                                         hammer2_thr_freeze(
 743                                                 &pmp->xop_groups[j].thrs[i]);
 744                                 }
 745                         }
 746
 747                         /*
 748                          * Lock the inode and clean out matching chains.
 749                          * Note that we cannot use hammer2_inode_lock_*()
 750                          * here because that would attempt to validate the
 751                          * cluster that we are in the middle of ripping
 752                          * apart.
 753                          *
 754                          * WARNING! We are working directly on the inodes
 755                          *          embedded cluster.
 756                          */
 757                         hammer2_mtx_ex(&iroot->lock);
 758
 759                         /*
 760                          * Remove the chain from matching elements of the PFS.
 761                          */
 762                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 763                                 if (pmp->pfs_hmps[i] != hmp)
 764                                         continue;
 765                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 766                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 767                                         hammer2_thr_delete(
 768                                                 &pmp->xop_groups[j].thrs[i]);
 769                                 }
 770                                 rchain = iroot->cluster.array[i].chain;
 771                                 iroot->cluster.array[i].chain = NULL;
 772                                 pmp->pfs_types[i] = 0;
 773                                 if (pmp->pfs_names[i]) {
 774                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 775                                         pmp->pfs_names[i] = NULL;
 776                                 }
 777                                 if (rchain) {
 778                                         hammer2_chain_drop(rchain);
 779                                         /* focus hint */
 780                                         if (iroot->cluster.focus == rchain)
 781                                                 iroot->cluster.focus = NULL;
 782                                 }
 783                                 pmp->pfs_hmps[i] = NULL;
 784                         }
 785                         hammer2_mtx_unlock(&iroot->lock);
 786                         didfreeze = 1;  /* remaster, unfreeze down below */
 787                 } else {
 788                         didfreeze = 0;
 789                 }
 790
 791                 /*
 792                  * Cleanup trailing chains.  Gaps may remain.
 793                  */
 794                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 795                         if (pmp->pfs_hmps[i])
 796                                 break;
 797                 }
 798                 iroot->cluster.nchains = i + 1;
 799
 800                 /*
 801                  * If the PMP has no elements remaining we can destroy it.
 802                  * (this will transition management threads from frozen->exit).
 803                  */
 804                 if (iroot->cluster.nchains == 0) {
 805                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 806                                 hmp, pmp);
 807                         hammer2_pfsfree(pmp);
 808                         goto again;
 809                 }
 810
 811                 /*
 812                  * If elements still remain we need to set the REMASTER
 813                  * flag and unfreeze it.
 814                  */
 815                 if (didfreeze) {
 816                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 817                                 if (pmp->pfs_hmps[i] == NULL)
 818                                         continue;
 819                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 820                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 821                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 822                                         hammer2_thr_remaster(
 823                                                 &pmp->xop_groups[j].thrs[i]);
 824                                         hammer2_thr_unfreeze(
 825                                                 &pmp->xop_groups[j].thrs[i]);
 826                                 }
 827                         }
 828                 }
 829         }
 830 }
 831
 832 /*
 833  * Mount or remount HAMMER2 fileystem from physical media
 834  *
 835  *      mountroot
 836  *              mp              mount point structure
 837  *              path            NULL
 838  *              data            <unused>
 839  *              cred            <unused>
 840  *
 841  *      mount
 842  *              mp              mount point structure
 843  *              path            path to mount point
 844  *              data            pointer to argument structure in user space
 845  *                      volume  volume path (device@LABEL form)
 846  *                      hflags  user mount flags
 847  *              cred            user credentials
 848  *
 849  * RETURNS:     0       Success
 850  *              !0      error number
 851  */
 852 static
 853 int
 854 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 855                   struct ucred *cred)
 856 {
 857         struct hammer2_mount_info info;
 858         hammer2_pfs_t *pmp;
 859         hammer2_pfs_t *spmp;
 860         hammer2_dev_t *hmp;
 861         hammer2_dev_t *force_local;
 862         hammer2_key_t key_next;
 863         hammer2_key_t key_dummy;
 864         hammer2_key_t lhc;
 865         struct vnode *devvp;
 866         struct nlookupdata nd;
 867         hammer2_chain_t *parent;
 868         hammer2_chain_t *chain;
 869         hammer2_cluster_t *cluster;
 870         const hammer2_inode_data_t *ripdata;
 871         hammer2_blockref_t bref;
 872         struct file *fp;
 873         char devstr[MNAMELEN];
 874         size_t size;
 875         size_t done;
 876         char *dev;
 877         char *label;
 878         int ronly = 1;
 879         int error;
 880         int cache_index;
 881         int i;
 882
 883         hmp = NULL;
 884         pmp = NULL;
 885         dev = NULL;
 886         label = NULL;
 887         devvp = NULL;
 888         cache_index = -1;
 889
 890         kprintf("hammer2_mount\n");
 891
 892         if (path == NULL) {
 893                 /*
 894                  * Root mount
 895                  */
 896                 bzero(&info, sizeof(info));
 897                 info.cluster_fd = -1;
 898                 ksnprintf(devstr, sizeof(devstr), "%s",
 899                           mp->mnt_stat.f_mntfromname);
 900                 kprintf("hammer2_mount: root '%s'\n", devstr);
 901         } else {
 902                 /*
 903                  * Non-root mount or updating a mount
 904                  */
 905                 error = copyin(data, &info, sizeof(info));
 906                 if (error)
 907                         return (error);
 908
 909                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 910                 if (error)
 911                         return (error);
 912         }
 913
 914         /*
 915          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 916          * if no label specified, based on the partition id.  Error out if no
 917          * partition id.  This is strictly a convenience to match the
 918          * default label created by newfs_hammer2, our preference is
 919          * that a label always be specified.
 920          */
 921         dev = devstr;
 922         label = strchr(devstr, '@');
 923         if (label && ((label + 1) - dev) > done)
 924                 return (EINVAL);
 925         if (label && label == devstr)
 926                 return (EINVAL);
 927         if (label == NULL || label[1] == 0) {
 928                 char slice;
 929
 930                 if (label == NULL)
 931                         label = devstr + strlen(devstr);
 932                 slice = label[-1];
 933                 switch(slice) {
 934                 case 'a':
 935                         label = "BOOT";
 936                         break;
 937                 case 'd':
 938                         label = "ROOT";
 939                         break;
 940                 default:
 941                         label = "DATA";
 942                         break;
 943                 }
 944         } else {
 945                 *label = '\0';
 946                 label++;
 947         }
 948
 949         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
 950                 dev, label, (mp->mnt_flag & MNT_RDONLY));
 951
 952         if (mp->mnt_flag & MNT_UPDATE) {
 953                 /*
 954                  * Update mount.  Note that pmp->iroot->cluster is
 955                  * an inode-embedded cluster and thus cannot be
 956                  * directly locked.
 957                  *
 958                  * XXX HAMMER2 needs to implement NFS export via
 959                  *     mountctl.
 960                  */
 961                 pmp = MPTOPMP(mp);
 962                 pmp->hflags = info.hflags;
 963                 cluster = &pmp->iroot->cluster;
 964                 for (i = 0; i < cluster->nchains; ++i) {
 965                         if (cluster->array[i].chain == NULL)
 966                                 continue;
 967                         hmp = cluster->array[i].chain->hmp;
 968                         devvp = hmp->devvp;
 969                         error = hammer2_remount(hmp, mp, path,
 970                                                 devvp, cred);
 971                         if (error)
 972                                 break;
 973                 }
 974
 975                 return error;
 976         }
 977
 978         /*
 979          * HMP device mount
 980          *
 981          * If a path is specified and dev is not an empty string, lookup the
 982          * name and verify that it referes to a block device.
 983          *
 984          * If a path is specified and dev is an empty string we fall through
 985          * and locate the label in the hmp search.
 986          */
 987         if (path && *dev != 0) {
 988                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 989                 if (error == 0)
 990                         error = nlookup(&nd);
 991                 if (error == 0)
 992                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 993                 nlookup_done(&nd);
 994         } else if (path == NULL) {
 995                 /* root mount */
 996                 cdev_t cdev = kgetdiskbyname(dev);
 997                 error = bdevvp(cdev, &devvp);
 998                 if (error)
 999                         kprintf("hammer2: cannot find '%s'\n", dev);
1000         } else {
1001                 /*
1002                  * We will locate the hmp using the label in the hmp loop.
1003                  */
1004                 error = 0;
1005         }
1006
1007         /*
1008          * Make sure its a block device.  Do not check to see if it is
1009          * already mounted until we determine that its a fresh H2 device.
1010          */
1011         if (error == 0 && devvp) {
1012                 vn_isdisk(devvp, &error);
1013         }
1014
1015         /*
1016          * Determine if the device has already been mounted.  After this
1017          * check hmp will be non-NULL if we are doing the second or more
1018          * hammer2 mounts from the same device.
1019          */
1020         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1021         if (devvp) {
1022                 /*
1023                  * Match the device.  Due to the way devfs works,
1024                  * we may not be able to directly match the vnode pointer,
1025                  * so also check to see if the underlying device matches.
1026                  */
1027                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1028                         if (hmp->devvp == devvp)
1029                                 break;
1030                         if (devvp->v_rdev &&
1031                             hmp->devvp->v_rdev == devvp->v_rdev) {
1032                                 break;
1033                         }
1034                 }
1035
1036                 /*
1037                  * If no match this may be a fresh H2 mount, make sure
1038                  * the device is not mounted on anything else.
1039                  */
1040                 if (hmp == NULL)
1041                         error = vfs_mountedon(devvp);
1042         } else if (error == 0) {
1043                 /*
1044                  * Match the label to a pmp already probed.
1045                  */
1046                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1047                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1048                                 if (pmp->pfs_names[i] &&
1049                                     strcmp(pmp->pfs_names[i], label) == 0) {
1050                                         hmp = pmp->pfs_hmps[i];
1051                                         break;
1052                                 }
1053                         }
1054                         if (hmp)
1055                                 break;
1056                 }
1057                 if (hmp == NULL)
1058                         error = ENOENT;
1059         }
1060
1061         /*
1062          * Open the device if this isn't a secondary mount and construct
1063          * the H2 device mount (hmp).
1064          */
1065         if (hmp == NULL) {
1066                 hammer2_chain_t *schain;
1067                 hammer2_xid_t xid;
1068
1069                 if (error == 0 && vcount(devvp) > 0) {
1070                         kprintf("Primary device already has references\n");
1071                         error = EBUSY;
1072                 }
1073
1074                 /*
1075                  * Now open the device
1076                  */
1077                 if (error == 0) {
1078                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1079                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1080                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1081                         if (error == 0) {
1082                                 error = VOP_OPEN(devvp,
1083                                              (ronly ? FREAD : FREAD | FWRITE),
1084                                              FSCRED, NULL);
1085                         }
1086                         vn_unlock(devvp);
1087                 }
1088                 if (error && devvp) {
1089                         vrele(devvp);
1090                         devvp = NULL;
1091                 }
1092                 if (error) {
1093                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1094                         return error;
1095                 }
1096                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1097                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1098                 hmp->ronly = ronly;
1099                 hmp->devvp = devvp;
1100                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1101                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1102                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1103                 RB_INIT(&hmp->iotree);
1104                 spin_init(&hmp->io_spin, "hm2mount_io");
1105                 spin_init(&hmp->list_spin, "hm2mount_list");
1106                 TAILQ_INIT(&hmp->flushq);
1107
1108                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1109                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1110                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1111
1112                 /*
1113                  * vchain setup. vchain.data is embedded.
1114                  * vchain.refs is initialized and will never drop to 0.
1115                  *
1116                  * NOTE! voldata is not yet loaded.
1117                  */
1118                 hmp->vchain.hmp = hmp;
1119                 hmp->vchain.refs = 1;
1120                 hmp->vchain.data = (void *)&hmp->voldata;
1121                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1122                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1123                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1124
1125                 hammer2_chain_core_init(&hmp->vchain);
1126                 /* hmp->vchain.u.xxx is left NULL */
1127
1128                 /*
1129                  * fchain setup.  fchain.data is embedded.
1130                  * fchain.refs is initialized and will never drop to 0.
1131                  *
1132                  * The data is not used but needs to be initialized to
1133                  * pass assertion muster.  We use this chain primarily
1134                  * as a placeholder for the freemap's top-level RBTREE
1135                  * so it does not interfere with the volume's topology
1136                  * RBTREE.
1137                  */
1138                 hmp->fchain.hmp = hmp;
1139                 hmp->fchain.refs = 1;
1140                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1141                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1142                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1143                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1144                 hmp->fchain.bref.methods =
1145                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1146                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1147
1148                 hammer2_chain_core_init(&hmp->fchain);
1149                 /* hmp->fchain.u.xxx is left NULL */
1150
1151                 /*
1152                  * Install the volume header and initialize fields from
1153                  * voldata.
1154                  */
1155                 error = hammer2_install_volume_header(hmp);
1156                 if (error) {
1157                         hammer2_unmount_helper(mp, NULL, hmp);
1158                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1159                         hammer2_vfs_unmount(mp, MNT_FORCE);
1160                         return error;
1161                 }
1162
1163                 /*
1164                  * Really important to get these right or flush will get
1165                  * confused.
1166                  */
1167                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1168                 kprintf("alloc spmp %p tid %016jx\n",
1169                         hmp->spmp, hmp->voldata.mirror_tid);
1170                 spmp = hmp->spmp;
1171
1172                 /*
1173                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1174                  * is inherited from the volume header.
1175                  */
1176                 xid = 0;
1177                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1178                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1179                 hmp->vchain.pmp = spmp;
1180                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1181                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1182                 hmp->fchain.pmp = spmp;
1183
1184                 /*
1185                  * First locate the super-root inode, which is key 0
1186                  * relative to the volume header's blockset.
1187                  *
1188                  * Then locate the root inode by scanning the directory keyspace
1189                  * represented by the label.
1190                  */
1191                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1192                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1193                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1194                                       &cache_index, 0);
1195                 hammer2_chain_lookup_done(parent);
1196                 if (schain == NULL) {
1197                         kprintf("hammer2_mount: invalid super-root\n");
1198                         hammer2_unmount_helper(mp, NULL, hmp);
1199                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1200                         hammer2_vfs_unmount(mp, MNT_FORCE);
1201                         return EINVAL;
1202                 }
1203                 if (schain->error) {
1204                         kprintf("hammer2_mount: error %s reading super-root\n",
1205                                 hammer2_error_str(schain->error));
1206                         hammer2_chain_unlock(schain);
1207                         hammer2_chain_drop(schain);
1208                         schain = NULL;
1209                         hammer2_unmount_helper(mp, NULL, hmp);
1210                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1211                         hammer2_vfs_unmount(mp, MNT_FORCE);
1212                         return EINVAL;
1213                 }
1214
1215                 /*
1216                  * The super-root always uses an inode_tid of 1 when
1217                  * creating PFSs.
1218                  */
1219                 spmp->inode_tid = 1;
1220                 spmp->modify_tid = schain->bref.modify_tid + 1;
1221
1222                 /*
1223                  * Sanity-check schain's pmp and finish initialization.
1224                  * Any chain belonging to the super-root topology should
1225                  * have a NULL pmp (not even set to spmp).
1226                  */
1227                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1228                 KKASSERT(schain->pmp == NULL);
1229                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1230
1231                 /*
1232                  * Replace the dummy spmp->iroot with a real one.  It's
1233                  * easier to just do a wholesale replacement than to try
1234                  * to update the chain and fixup the iroot fields.
1235                  *
1236                  * The returned inode is locked with the supplied cluster.
1237                  */
1238                 cluster = hammer2_cluster_from_chain(schain);
1239                 hammer2_inode_drop(spmp->iroot);
1240                 spmp->iroot = NULL;
1241                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1242                 spmp->spmp_hmp = hmp;
1243                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1244                 spmp->pfs_hmps[0] = hmp;
1245                 hammer2_inode_ref(spmp->iroot);
1246                 hammer2_inode_unlock(spmp->iroot);
1247                 hammer2_cluster_unlock(cluster);
1248                 hammer2_cluster_drop(cluster);
1249                 schain = NULL;
1250                 /* leave spmp->iroot with one ref */
1251
1252                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1253                         error = hammer2_recovery(hmp);
1254                         /* XXX do something with error */
1255                 }
1256                 hammer2_update_pmps(hmp);
1257                 hammer2_iocom_init(hmp);
1258                 hammer2_bulkfree_init(hmp);
1259
1260                 /*
1261                  * Ref the cluster management messaging descriptor.  The mount
1262                  * program deals with the other end of the communications pipe.
1263                  *
1264                  * Root mounts typically do not supply one.
1265                  */
1266                 if (info.cluster_fd >= 0) {
1267                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1268                         if (fp) {
1269                                 hammer2_cluster_reconnect(hmp, fp);
1270                         } else {
1271                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1272                         }
1273                 }
1274         } else {
1275                 spmp = hmp->spmp;
1276                 if (info.hflags & HMNT2_DEVFLAGS) {
1277                         kprintf("hammer2: Warning: mount flags pertaining "
1278                                 "to the whole device may only be specified "
1279                                 "on the first mount of the device: %08x\n",
1280                                 info.hflags & HMNT2_DEVFLAGS);
1281                 }
1282         }
1283
1284         /*
1285          * Force local mount (disassociate all PFSs from their clusters).
1286          * Used primarily for debugging.
1287          */
1288         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1289
1290         /*
1291          * Lookup the mount point under the media-localized super-root.
1292          * Scanning hammer2_pfslist doesn't help us because it represents
1293          * PFS cluster ids which can aggregate several named PFSs together.
1294          *
1295          * cluster->pmp will incorrectly point to spmp and must be fixed
1296          * up later on.
1297          */
1298         hammer2_inode_lock(spmp->iroot, 0);
1299         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1300         lhc = hammer2_dirhash(label, strlen(label));
1301         chain = hammer2_chain_lookup(&parent, &key_next,
1302                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1303                                      &cache_index, 0);
1304         while (chain) {
1305                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1306                     strcmp(label, chain->data->ipdata.filename) == 0) {
1307                         break;
1308                 }
1309                 chain = hammer2_chain_next(&parent, chain, &key_next,
1310                                             key_next,
1311                                             lhc + HAMMER2_DIRHASH_LOMASK,
1312                                             &cache_index, 0);
1313         }
1314         if (parent) {
1315                 hammer2_chain_unlock(parent);
1316                 hammer2_chain_drop(parent);
1317         }
1318         hammer2_inode_unlock(spmp->iroot);
1319
1320         /*
1321          * PFS could not be found?
1322          */
1323         if (chain == NULL) {
1324                 kprintf("hammer2_mount: PFS label not found\n");
1325                 hammer2_unmount_helper(mp, NULL, hmp);
1326                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1327                 hammer2_vfs_unmount(mp, MNT_FORCE);
1328
1329                 return EINVAL;
1330         }
1331
1332         /*
1333          * Acquire the pmp structure (it should have already been allocated
1334          * via hammer2_update_pmps() so do not pass cluster in to add to
1335          * available chains).
1336          *
1337          * Check if the cluster has already been mounted.  A cluster can
1338          * only be mounted once, use null mounts to mount additional copies.
1339          */
1340         ripdata = &chain->data->ipdata;
1341         bref = chain->bref;
1342         pmp = hammer2_pfsalloc(NULL, ripdata,
1343                                bref.modify_tid, force_local);
1344         hammer2_chain_unlock(chain);
1345         hammer2_chain_drop(chain);
1346
1347         /*
1348          * Finish the mount
1349          */
1350         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1351
1352         if (pmp->mp) {
1353                 kprintf("hammer2_mount: PFS already mounted!\n");
1354                 hammer2_unmount_helper(mp, NULL, hmp);
1355                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1356                 hammer2_vfs_unmount(mp, MNT_FORCE);
1357
1358                 return EBUSY;
1359         }
1360
1361         pmp->hflags = info.hflags;
1362         mp->mnt_flag |= MNT_LOCAL;
1363         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1364         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1365
1366         /*
1367          * required mount structure initializations
1368          */
1369         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1370         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1371
1372         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1373         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1374
1375         /*
1376          * Optional fields
1377          */
1378         mp->mnt_iosize_max = MAXPHYS;
1379
1380         /*
1381          * Connect up mount pointers.
1382          */
1383         hammer2_mount_helper(mp, pmp);
1384
1385         lockmgr(&hammer2_mntlk, LK_RELEASE);
1386
1387         /*
1388          * Finish setup
1389          */
1390         vfs_getnewfsid(mp);
1391         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1392         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1393         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1394
1395         if (path) {
1396                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1397                           MNAMELEN - 1, &size);
1398                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1399         } /* else root mount, already in there */
1400
1401         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1402         if (path) {
1403                 copyinstr(path, mp->mnt_stat.f_mntonname,
1404                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1405                           &size);
1406         } else {
1407                 /* root mount */
1408                 mp->mnt_stat.f_mntonname[0] = '/';
1409         }
1410
1411         /*
1412          * Initial statfs to prime mnt_stat.
1413          */
1414         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1415
1416         return 0;
1417 }
1418
1419 /*
1420  * Scan PFSs under the super-root and create hammer2_pfs structures.
1421  */
1422 static
1423 void
1424 hammer2_update_pmps(hammer2_dev_t *hmp)
1425 {
1426         const hammer2_inode_data_t *ripdata;
1427         hammer2_chain_t *parent;
1428         hammer2_chain_t *chain;
1429         hammer2_blockref_t bref;
1430         hammer2_dev_t *force_local;
1431         hammer2_pfs_t *spmp;
1432         hammer2_pfs_t *pmp;
1433         hammer2_key_t key_next;
1434         int cache_index = -1;
1435
1436         /*
1437          * Force local mount (disassociate all PFSs from their clusters).
1438          * Used primarily for debugging.
1439          */
1440         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1441
1442         /*
1443          * Lookup mount point under the media-localized super-root.
1444          *
1445          * cluster->pmp will incorrectly point to spmp and must be fixed
1446          * up later on.
1447          */
1448         spmp = hmp->spmp;
1449         hammer2_inode_lock(spmp->iroot, 0);
1450         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1451         chain = hammer2_chain_lookup(&parent, &key_next,
1452                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1453                                          &cache_index, 0);
1454         while (chain) {
1455                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1456                         continue;
1457                 ripdata = &chain->data->ipdata;
1458                 bref = chain->bref;
1459                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1460
1461                 pmp = hammer2_pfsalloc(chain, ripdata,
1462                                        bref.modify_tid, force_local);
1463                 chain = hammer2_chain_next(&parent, chain, &key_next,
1464                                            key_next, HAMMER2_KEY_MAX,
1465                                            &cache_index, 0);
1466         }
1467         if (parent) {
1468                 hammer2_chain_unlock(parent);
1469                 hammer2_chain_drop(parent);
1470         }
1471         hammer2_inode_unlock(spmp->iroot);
1472 }
1473
1474 static
1475 int
1476 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1477                 struct vnode *devvp, struct ucred *cred)
1478 {
1479         int error;
1480
1481         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1482                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1483                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1484                 vn_unlock(devvp);
1485                 error = hammer2_recovery(hmp);
1486                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1487                 if (error == 0) {
1488                         VOP_CLOSE(devvp, FREAD, NULL);
1489                         hmp->ronly = 0;
1490                 } else {
1491                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1492                 }
1493                 vn_unlock(devvp);
1494         } else {
1495                 error = 0;
1496         }
1497         return error;
1498 }
1499
1500 static
1501 int
1502 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1503 {
1504         hammer2_pfs_t *pmp;
1505         int flags;
1506         int error = 0;
1507
1508         pmp = MPTOPMP(mp);
1509
1510         if (pmp == NULL)
1511                 return(0);
1512
1513         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1514
1515         /*
1516          * If mount initialization proceeded far enough we must flush
1517          * its vnodes and sync the underlying mount points.  Three syncs
1518          * are required to fully flush the filesystem (freemap updates lag
1519          * by one flush, and one extra for safety).
1520          */
1521         if (mntflags & MNT_FORCE)
1522                 flags = FORCECLOSE;
1523         else
1524                 flags = 0;
1525         if (pmp->iroot) {
1526                 error = vflush(mp, 0, flags);
1527                 if (error)
1528                         goto failed;
1529                 hammer2_vfs_sync(mp, MNT_WAIT);
1530                 hammer2_vfs_sync(mp, MNT_WAIT);
1531                 hammer2_vfs_sync(mp, MNT_WAIT);
1532         }
1533
1534         /*
1535          * Cleanup the frontend support XOPS threads
1536          */
1537         hammer2_xop_helper_cleanup(pmp);
1538
1539         if (pmp->mp)
1540                 hammer2_unmount_helper(mp, pmp, NULL);
1541
1542         error = 0;
1543 failed:
1544         lockmgr(&hammer2_mntlk, LK_RELEASE);
1545
1546         return (error);
1547 }
1548
1549 /*
1550  * Mount helper, hook the system mount into our PFS.
1551  * The mount lock is held.
1552  *
1553  * We must bump the mount_count on related devices for any
1554  * mounted PFSs.
1555  */
1556 static
1557 void
1558 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1559 {
1560         hammer2_cluster_t *cluster;
1561         hammer2_chain_t *rchain;
1562         int i;
1563
1564         mp->mnt_data = (qaddr_t)pmp;
1565         pmp->mp = mp;
1566
1567         /*
1568          * After pmp->mp is set we have to adjust hmp->mount_count.
1569          */
1570         cluster = &pmp->iroot->cluster;
1571         for (i = 0; i < cluster->nchains; ++i) {
1572                 rchain = cluster->array[i].chain;
1573                 if (rchain == NULL)
1574                         continue;
1575                 ++rchain->hmp->mount_count;
1576                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1577                         rchain->hmp, rchain->hmp->mount_count);
1578         }
1579
1580         /*
1581          * Create missing Xop threads
1582          */
1583         hammer2_xop_helper_create(pmp);
1584 }
1585
1586 /*
1587  * Mount helper, unhook the system mount from our PFS.
1588  * The mount lock is held.
1589  *
1590  * If hmp is supplied a mount responsible for being the first to open
1591  * the block device failed and the block device and all PFSs using the
1592  * block device must be cleaned up.
1593  *
1594  * If pmp is supplied multiple devices might be backing the PFS and each
1595  * must be disconnected.  This might not be the last PFS using some of the
1596  * underlying devices.  Also, we have to adjust our hmp->mount_count
1597  * accounting for the devices backing the pmp which is now undergoing an
1598  * unmount.
1599  */
1600 static
1601 void
1602 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1603 {
1604         hammer2_cluster_t *cluster;
1605         hammer2_chain_t *rchain;
1606         struct vnode *devvp;
1607         int dumpcnt;
1608         int ronly;
1609         int i;
1610
1611         /*
1612          * If no device supplied this is a high-level unmount and we have to
1613          * to disconnect the mount, adjust mount_count, and locate devices
1614          * that might now have no mounts.
1615          */
1616         if (pmp) {
1617                 KKASSERT(hmp == NULL);
1618                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1619                 pmp->mp = NULL;
1620                 mp->mnt_data = NULL;
1621
1622                 /*
1623                  * After pmp->mp is cleared we have to account for
1624                  * mount_count.
1625                  */
1626                 cluster = &pmp->iroot->cluster;
1627                 for (i = 0; i < cluster->nchains; ++i) {
1628                         rchain = cluster->array[i].chain;
1629                         if (rchain == NULL)
1630                                 continue;
1631                         --rchain->hmp->mount_count;
1632                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1633                                 rchain->hmp, rchain->hmp->mount_count);
1634                         /* scrapping hmp now may invalidate the pmp */
1635                 }
1636 again:
1637                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1638                         if (hmp->mount_count == 0) {
1639                                 hammer2_unmount_helper(NULL, NULL, hmp);
1640                                 goto again;
1641                         }
1642                 }
1643                 return;
1644         }
1645
1646         /*
1647          * Try to terminate the block device.  We can't terminate it if
1648          * there are still PFSs referencing it.
1649          */
1650         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1651                 hmp, hmp->mount_count);
1652         if (hmp->mount_count)
1653                 return;
1654
1655         hammer2_bulkfree_uninit(hmp);
1656         hammer2_pfsfree_scan(hmp);
1657         hammer2_dev_exlock(hmp);        /* XXX order */
1658
1659         /*
1660          * Cycle the volume data lock as a safety (probably not needed any
1661          * more).  To ensure everything is out we need to flush at least
1662          * three times.  (1) The running of the sideq can dirty the
1663          * filesystem, (2) A normal flush can dirty the freemap, and
1664          * (3) ensure that the freemap is fully synchronized.
1665          *
1666          * The next mount's recovery scan can clean everything up but we want
1667          * to leave the filesystem in a 100% clean state on a normal unmount.
1668          */
1669 #if 0
1670         hammer2_voldata_lock(hmp);
1671         hammer2_voldata_unlock(hmp);
1672 #endif
1673         hammer2_iocom_uninit(hmp);
1674
1675         if ((hmp->vchain.flags | hmp->fchain.flags) &
1676             HAMMER2_CHAIN_FLUSH_MASK) {
1677                 kprintf("hammer2_unmount: chains left over "
1678                         "after final sync\n");
1679                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1680                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1681
1682                 if (hammer2_debug & 0x0010)
1683                         Debugger("entered debugger");
1684         }
1685
1686         KKASSERT(hmp->spmp == NULL);
1687
1688         /*
1689          * Finish up with the device vnode
1690          */
1691         if ((devvp = hmp->devvp) != NULL) {
1692                 ronly = hmp->ronly;
1693                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1694                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1695                 hmp->devvp = NULL;
1696                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1697                 vn_unlock(devvp);
1698                 vrele(devvp);
1699                 devvp = NULL;
1700         }
1701
1702         /*
1703          * Clear vchain/fchain flags that might prevent final cleanup
1704          * of these chains.
1705          */
1706         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1707                 atomic_add_long(&hammer2_count_modified_chains, -1);
1708                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1709                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1710         }
1711         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1712                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1713         }
1714
1715         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1716                 atomic_add_long(&hammer2_count_modified_chains, -1);
1717                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1718                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1719         }
1720         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1721                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1722         }
1723
1724         /*
1725          * Final drop of embedded freemap root chain to
1726          * clean up fchain.core (fchain structure is not
1727          * flagged ALLOCATED so it is cleaned out and then
1728          * left to rot).
1729          */
1730         hammer2_chain_drop(&hmp->fchain);
1731
1732         /*
1733          * Final drop of embedded volume root chain to clean
1734          * up vchain.core (vchain structure is not flagged
1735          * ALLOCATED so it is cleaned out and then left to
1736          * rot).
1737          */
1738         dumpcnt = 50;
1739         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1740         dumpcnt = 50;
1741         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1742         hammer2_dev_unlock(hmp);
1743         hammer2_chain_drop(&hmp->vchain);
1744
1745         hammer2_io_cleanup(hmp, &hmp->iotree);
1746         if (hmp->iofree_count) {
1747                 kprintf("io_cleanup: %d I/O's left hanging\n",
1748                         hmp->iofree_count);
1749         }
1750
1751         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1752         kmalloc_destroy(&hmp->mchain);
1753         kfree(hmp, M_HAMMER2);
1754 }
1755
1756 int
1757 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1758                  ino_t ino, struct vnode **vpp)
1759 {
1760         hammer2_xop_lookup_t *xop;
1761         hammer2_pfs_t *pmp;
1762         hammer2_inode_t *ip;
1763         hammer2_tid_t inum;
1764         int error;
1765
1766         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1767
1768         error = 0;
1769         pmp = MPTOPMP(mp);
1770
1771         /*
1772          * Easy if we already have it cached
1773          */
1774         ip = hammer2_inode_lookup(pmp, inum);
1775         if (ip) {
1776                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1777                 *vpp = hammer2_igetv(ip, &error);
1778                 hammer2_inode_unlock(ip);
1779                 hammer2_inode_drop(ip);         /* from lookup */
1780
1781                 return error;
1782         }
1783
1784         /*
1785          * Otherwise we have to find the inode
1786          */
1787         xop = hammer2_xop_alloc(pmp->iroot, 0);
1788         xop->lhc = inum;
1789         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1790         error = hammer2_xop_collect(&xop->head, 0);
1791
1792         if (error == 0) {
1793                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1794                         kprintf("vget: no collect error but also no rdata\n");
1795                         kprintf("xop %p\n", xop);
1796                         while ((hammer2_debug & 0x80000) == 0) {
1797                                 tsleep(xop, PCATCH, "wait", hz * 10);
1798                         }
1799                         ip = NULL;
1800                 } else {
1801                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1802                 }
1803         }
1804         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1805
1806         if (ip) {
1807                 *vpp = hammer2_igetv(ip, &error);
1808                 hammer2_inode_unlock(ip);
1809         } else {
1810                 *vpp = NULL;
1811                 error = ENOENT;
1812         }
1813         return (error);
1814 }
1815
1816 static
1817 int
1818 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1819 {
1820         hammer2_pfs_t *pmp;
1821         struct vnode *vp;
1822         int error;
1823
1824         pmp = MPTOPMP(mp);
1825         if (pmp->iroot == NULL) {
1826                 *vpp = NULL;
1827                 return EINVAL;
1828         }
1829
1830         error = 0;
1831         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1832
1833         while (pmp->inode_tid == 0) {
1834                 hammer2_xop_ipcluster_t *xop;
1835                 hammer2_inode_meta_t *meta;
1836
1837                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1838                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1839                 error = hammer2_xop_collect(&xop->head, 0);
1840
1841                 if (error == 0) {
1842                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1843                         pmp->iroot->meta = *meta;
1844                         pmp->inode_tid = meta->pfs_inum + 1;
1845                         if (pmp->inode_tid < HAMMER2_INODE_START)
1846                                 pmp->inode_tid = HAMMER2_INODE_START;
1847                         pmp->modify_tid =
1848                                 xop->head.cluster.focus->bref.modify_tid + 1;
1849                         kprintf("PFS: Starting inode %jd\n",
1850                                 (intmax_t)pmp->inode_tid);
1851                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1852                                 pmp->inode_tid, pmp->modify_tid);
1853                         wakeup(&pmp->iroot);
1854
1855                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1856
1857                         /*
1858                          * Prime the mount info.
1859                          */
1860                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1861                         break;
1862                 }
1863
1864                 /*
1865                  * Loop, try again
1866                  */
1867                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1868                 hammer2_inode_unlock(pmp->iroot);
1869                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1870                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1871                 if (error == EINTR)
1872                         break;
1873         }
1874
1875         if (error) {
1876                 hammer2_inode_unlock(pmp->iroot);
1877                 *vpp = NULL;
1878         } else {
1879                 vp = hammer2_igetv(pmp->iroot, &error);
1880                 hammer2_inode_unlock(pmp->iroot);
1881                 *vpp = vp;
1882         }
1883
1884         return (error);
1885 }
1886
1887 /*
1888  * Filesystem status
1889  *
1890  * XXX incorporate ipdata->meta.inode_quota and data_quota
1891  */
1892 static
1893 int
1894 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1895 {
1896         hammer2_pfs_t *pmp;
1897         hammer2_dev_t *hmp;
1898         hammer2_blockref_t bref;
1899         int i;
1900
1901         /*
1902          * NOTE: iroot might not have validated the cluster yet.
1903          */
1904         pmp = MPTOPMP(mp);
1905
1906         mp->mnt_stat.f_files = 0;
1907         mp->mnt_stat.f_ffree = 0;
1908         mp->mnt_stat.f_blocks = 0;
1909         mp->mnt_stat.f_bfree = 0;
1910         mp->mnt_stat.f_bavail = 0;
1911
1912         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1913                 hmp = pmp->pfs_hmps[i];
1914                 if (hmp == NULL)
1915                         continue;
1916                 if (pmp->iroot->cluster.array[i].chain)
1917                         bref = pmp->iroot->cluster.array[i].chain->bref;
1918                 else
1919                         bzero(&bref, sizeof(bref));
1920
1921                 mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1922                 mp->mnt_stat.f_ffree = 0;
1923                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1924                                         mp->mnt_vstat.f_bsize;
1925                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1926                                         mp->mnt_vstat.f_bsize;
1927                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1928
1929                 *sbp = mp->mnt_stat;
1930         }
1931         return (0);
1932 }
1933
1934 static
1935 int
1936 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1937 {
1938         hammer2_pfs_t *pmp;
1939         hammer2_dev_t *hmp;
1940         hammer2_blockref_t bref;
1941         int i;
1942
1943         /*
1944          * NOTE: iroot might not have validated the cluster yet.
1945          */
1946         pmp = MPTOPMP(mp);
1947
1948         mp->mnt_vstat.f_bsize = 0;
1949         mp->mnt_vstat.f_files = 0;
1950         mp->mnt_vstat.f_ffree = 0;
1951         mp->mnt_vstat.f_blocks = 0;
1952         mp->mnt_vstat.f_bfree = 0;
1953         mp->mnt_vstat.f_bavail = 0;
1954
1955         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1956                 hmp = pmp->pfs_hmps[i];
1957                 if (hmp == NULL)
1958                         continue;
1959                 if (pmp->iroot->cluster.array[i].chain)
1960                         bref = pmp->iroot->cluster.array[i].chain->bref;
1961                 else
1962                         bzero(&bref, sizeof(bref));
1963
1964                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1965                 mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
1966                 mp->mnt_vstat.f_ffree = 0;
1967                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1968                                         mp->mnt_vstat.f_bsize;
1969                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1970                                         mp->mnt_vstat.f_bsize;
1971                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1972
1973                 *sbp = mp->mnt_vstat;
1974         }
1975         return (0);
1976 }
1977
1978 /*
1979  * Mount-time recovery (RW mounts)
1980  *
1981  * Updates to the free block table are allowed to lag flushes by one
1982  * transaction.  In case of a crash, then on a fresh mount we must do an
1983  * incremental scan of the last committed transaction id and make sure that
1984  * all related blocks have been marked allocated.
1985  *
1986  * The super-root topology and each PFS has its own transaction id domain,
1987  * so we must track PFS boundary transitions.
1988  */
1989 struct hammer2_recovery_elm {
1990         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1991         hammer2_chain_t *chain;
1992         hammer2_tid_t sync_tid;
1993 };
1994
1995 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1996
1997 struct hammer2_recovery_info {
1998         struct hammer2_recovery_list list;
1999         hammer2_tid_t   mtid;
2000         int     depth;
2001 };
2002
2003 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2004                         hammer2_chain_t *parent,
2005                         struct hammer2_recovery_info *info,
2006                         hammer2_tid_t sync_tid);
2007
2008 #define HAMMER2_RECOVERY_MAXDEPTH       10
2009
2010 static
2011 int
2012 hammer2_recovery(hammer2_dev_t *hmp)
2013 {
2014         struct hammer2_recovery_info info;
2015         struct hammer2_recovery_elm *elm;
2016         hammer2_chain_t *parent;
2017         hammer2_tid_t sync_tid;
2018         hammer2_tid_t mirror_tid;
2019         int error;
2020         int cumulative_error = 0;
2021
2022         hammer2_trans_init(hmp->spmp, 0);
2023
2024         sync_tid = hmp->voldata.freemap_tid;
2025         mirror_tid = hmp->voldata.mirror_tid;
2026
2027         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2028         if (sync_tid >= mirror_tid) {
2029                 kprintf(" no recovery needed\n");
2030         } else {
2031                 kprintf(" freemap recovery %016jx-%016jx\n",
2032                         sync_tid + 1, mirror_tid);
2033         }
2034
2035         TAILQ_INIT(&info.list);
2036         info.depth = 0;
2037         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2038         cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2039         hammer2_chain_lookup_done(parent);
2040
2041         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2042                 TAILQ_REMOVE(&info.list, elm, entry);
2043                 parent = elm->chain;
2044                 sync_tid = elm->sync_tid;
2045                 kfree(elm, M_HAMMER2);
2046
2047                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2048                 error = hammer2_recovery_scan(hmp, parent, &info,
2049                                               hmp->voldata.freemap_tid);
2050                 hammer2_chain_unlock(parent);
2051                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2052                 if (error)
2053                         cumulative_error = error;
2054         }
2055         hammer2_trans_done(hmp->spmp);
2056
2057         return cumulative_error;
2058 }
2059
2060 static
2061 int
2062 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2063                       struct hammer2_recovery_info *info,
2064                       hammer2_tid_t sync_tid)
2065 {
2066         const hammer2_inode_data_t *ripdata;
2067         hammer2_chain_t *chain;
2068         hammer2_blockref_t bref;
2069         int cache_index;
2070         int cumulative_error = 0;
2071         int error;
2072         int first;
2073
2074         /*
2075          * Adjust freemap to ensure that the block(s) are marked allocated.
2076          */
2077         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2078                 hammer2_freemap_adjust(hmp, &parent->bref,
2079                                        HAMMER2_FREEMAP_DORECOVER);
2080         }
2081
2082         /*
2083          * Check type for recursive scan
2084          */
2085         switch(parent->bref.type) {
2086         case HAMMER2_BREF_TYPE_VOLUME:
2087                 /* data already instantiated */
2088                 break;
2089         case HAMMER2_BREF_TYPE_INODE:
2090                 /*
2091                  * Must instantiate data for DIRECTDATA test and also
2092                  * for recursion.
2093                  */
2094                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2095                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2096                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2097                         /* not applicable to recovery scan */
2098                         hammer2_chain_unlock(parent);
2099                         return 0;
2100                 }
2101                 hammer2_chain_unlock(parent);
2102                 break;
2103         case HAMMER2_BREF_TYPE_INDIRECT:
2104                 /*
2105                  * Must instantiate data for recursion
2106                  */
2107                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2108                 hammer2_chain_unlock(parent);
2109                 break;
2110         case HAMMER2_BREF_TYPE_DIRENT:
2111         case HAMMER2_BREF_TYPE_DATA:
2112         case HAMMER2_BREF_TYPE_FREEMAP:
2113         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2114         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2115                 /* not applicable to recovery scan */
2116                 return 0;
2117                 break;
2118         default:
2119                 return EDOM;
2120         }
2121
2122         /*
2123          * Defer operation if depth limit reached or if we are crossing a
2124          * PFS boundary.
2125          */
2126         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2127                 struct hammer2_recovery_elm *elm;
2128
2129                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2130                 elm->chain = parent;
2131                 elm->sync_tid = sync_tid;
2132                 hammer2_chain_ref(parent);
2133                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2134                 /* unlocked by caller */
2135
2136                 return(0);
2137         }
2138
2139
2140         /*
2141          * Recursive scan of the last flushed transaction only.  We are
2142          * doing this without pmp assignments so don't leave the chains
2143          * hanging around after we are done with them.
2144          */
2145         cache_index = 0;
2146         chain = NULL;
2147         first = 1;
2148
2149         while (hammer2_chain_scan(parent, &chain, &bref,
2150                                   &first, &cache_index,
2151                                   HAMMER2_LOOKUP_NODATA) != NULL) {
2152                 /*
2153                  * If this is a leaf
2154                  */
2155                 if (chain == NULL) {
2156                         if (bref.mirror_tid > sync_tid) {
2157                                 hammer2_freemap_adjust(hmp, &bref,
2158                                                      HAMMER2_FREEMAP_DORECOVER);
2159                         }
2160                         continue;
2161                 }
2162
2163                 /*
2164                  * This may or may not be a recursive node.
2165                  */
2166                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2167                 if (bref.mirror_tid > sync_tid) {
2168                         ++info->depth;
2169                         error = hammer2_recovery_scan(hmp, chain,
2170                                                       info, sync_tid);
2171                         --info->depth;
2172                         if (error)
2173                                 cumulative_error = error;
2174                 }
2175
2176                 /*
2177                  * Flush the recovery at the PFS boundary to stage it for
2178                  * the final flush of the super-root topology.
2179                  */
2180                 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2181                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2182                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2183                 }
2184         }
2185
2186         return cumulative_error;
2187 }
2188
2189 /*
2190  * Sync a mount point; this is called on a per-mount basis from the
2191  * filesystem syncer process periodically and whenever a user issues
2192  * a sync.
2193  */
2194 int
2195 hammer2_vfs_sync(struct mount *mp, int waitfor)
2196 {
2197         hammer2_xop_flush_t *xop;
2198         struct hammer2_sync_info info;
2199         hammer2_inode_t *iroot;
2200         hammer2_pfs_t *pmp;
2201         int flags;
2202         int error;
2203
2204         pmp = MPTOPMP(mp);
2205         iroot = pmp->iroot;
2206         KKASSERT(iroot);
2207         KKASSERT(iroot->pmp == pmp);
2208
2209         /*
2210          * We can't acquire locks on existing vnodes while in a transaction
2211          * without risking a deadlock.  This assumes that vfsync() can be
2212          * called without the vnode locked (which it can in DragonFly).
2213          * Otherwise we'd have to implement a multi-pass or flag the lock
2214          * failures and retry.
2215          *
2216          * The reclamation code interlocks with the sync list's token
2217          * (by removing the vnode from the scan list) before unlocking
2218          * the inode, giving us time to ref the inode.
2219          */
2220         /*flags = VMSC_GETVP;*/
2221         flags = 0;
2222         if (waitfor & MNT_LAZY)
2223                 flags |= VMSC_ONEPASS;
2224
2225         /*
2226          * Preflush the vnodes using a normal transaction before interlocking
2227          * with a flush transaction.  We do this to try to run as much of
2228          * the compression as possible outside the flush transaction.
2229          *
2230          * For efficiency do an async pass before making sure with a
2231          * synchronous pass on all related buffer cache buffers.
2232          */
2233         hammer2_trans_init(pmp, 0);
2234         info.error = 0;
2235         info.waitfor = MNT_NOWAIT;
2236         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2237         info.waitfor = MNT_WAIT;
2238         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2239         hammer2_trans_done(pmp);
2240
2241         /*
2242          * Start our flush transaction.  This does not return until all
2243          * concurrent transactions have completed and will prevent any
2244          * new transactions from running concurrently, except for the
2245          * buffer cache transactions.
2246          *
2247          * NOTE!  It is still possible for the paging code to push pages
2248          *        out via a UIO_NOCOPY hammer2_vop_write() during the main
2249          *        flush.
2250          */
2251         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2252         hammer2_inode_run_sideq(pmp);
2253
2254         info.error = 0;
2255         info.waitfor = MNT_NOWAIT;
2256         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2257         info.waitfor = MNT_WAIT;
2258         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2259         hammer2_bioq_sync(pmp);
2260
2261         /*
2262          * Use the XOP interface to concurrently flush all nodes to
2263          * synchronize the PFSROOT subtopology to the media.  A standard
2264          * end-of-scan ENOENT error indicates cluster sufficiency.
2265          *
2266          * Note that this flush will not be visible on crash recovery until
2267          * we flush the super-root topology in the next loop.
2268          *
2269          * XXX For now wait for all flushes to complete.
2270          */
2271         if (iroot) {
2272                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2273                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2274                 error = hammer2_xop_collect(&xop->head,
2275                                             HAMMER2_XOP_COLLECT_WAITALL);
2276                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2277                 if (error == ENOENT)
2278                         error = 0;
2279         } else {
2280                 error = 0;
2281         }
2282         hammer2_trans_done(pmp);
2283
2284         return (error);
2285 }
2286
2287 /*
2288  * Sync passes.
2289  *
2290  * Note that we ignore the tranasction mtid we got above.  Instead,
2291  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2292  * transactions.
2293  */
2294 static int
2295 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2296 {
2297         struct hammer2_sync_info *info = data;
2298         hammer2_inode_t *ip;
2299         int error;
2300
2301         /*
2302          * Degenerate cases.  Note that ip == NULL typically means the
2303          * syncer vnode itself and we don't want to vclrisdirty() in that
2304          * situation.
2305          */
2306         ip = VTOI(vp);
2307         if (ip == NULL) {
2308                 return(0);
2309         }
2310         if (vp->v_type == VNON || vp->v_type == VBAD) {
2311                 vclrisdirty(vp);
2312                 return(0);
2313         }
2314
2315         /*
2316          * VOP_FSYNC will start a new transaction so replicate some code
2317          * here to do it inline (see hammer2_vop_fsync()).
2318          *
2319          * WARNING: The vfsync interacts with the buffer cache and might
2320          *          block, we can't hold the inode lock at that time.
2321          *          However, we MUST ref ip before blocking to ensure that
2322          *          it isn't ripped out from under us (since we do not
2323          *          hold a lock on the vnode).
2324          */
2325         hammer2_inode_ref(ip);
2326         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2327             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2328                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2329                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2330                                  HAMMER2_INODE_MODIFIED)) {
2331                         hammer2_inode_lock(ip, 0);
2332                         if (ip->flags & (HAMMER2_INODE_RESIZED |
2333                                          HAMMER2_INODE_MODIFIED)) {
2334                                 hammer2_inode_chain_sync(ip);
2335                         }
2336                         hammer2_inode_unlock(ip);
2337                 }
2338         }
2339         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2340             RB_EMPTY(&vp->v_rbdirty_tree)) {
2341                 vclrisdirty(vp);
2342         }
2343
2344         hammer2_inode_drop(ip);
2345 #if 1
2346         error = 0;
2347         if (error)
2348                 info->error = error;
2349 #endif
2350         return(0);
2351 }
2352
2353 static
2354 int
2355 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2356 {
2357         hammer2_inode_t *ip;
2358
2359         KKASSERT(MAXFIDSZ >= 16);
2360         ip = VTOI(vp);
2361         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2362         fhp->fid_ext = 0;
2363         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2364         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2365
2366         return 0;
2367 }
2368
2369 static
2370 int
2371 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2372                struct fid *fhp, struct vnode **vpp)
2373 {
2374         hammer2_pfs_t *pmp;
2375         hammer2_tid_t inum;
2376         int error;
2377
2378         pmp = MPTOPMP(mp);
2379         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2380         if (vpp) {
2381                 if (inum == 1)
2382                         error = hammer2_vfs_root(mp, vpp);
2383                 else
2384                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2385         } else {
2386                 error = 0;
2387         }
2388         if (error)
2389                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2390         return error;
2391 }
2392
2393 static
2394 int
2395 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2396                  int *exflagsp, struct ucred **credanonp)
2397 {
2398         hammer2_pfs_t *pmp;
2399         struct netcred *np;
2400         int error;
2401
2402         pmp = MPTOPMP(mp);
2403         np = vfs_export_lookup(mp, &pmp->export, nam);
2404         if (np) {
2405                 *exflagsp = np->netc_exflags;
2406                 *credanonp = &np->netc_anon;
2407                 error = 0;
2408         } else {
2409                 error = EACCES;
2410         }
2411         return error;
2412 }
2413
2414 /*
2415  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2416  * header into the HMP
2417  *
2418  * XXX read four volhdrs and use the one with the highest TID whos CRC
2419  *     matches.
2420  *
2421  * XXX check iCRCs.
2422  *
2423  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2424  *     nonexistant locations.
2425  *
2426  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2427  */
2428 static
2429 int
2430 hammer2_install_volume_header(hammer2_dev_t *hmp)
2431 {
2432         hammer2_volume_data_t *vd;
2433         struct buf *bp;
2434         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2435         int error_reported;
2436         int error;
2437         int valid;
2438         int i;
2439
2440         error_reported = 0;
2441         error = 0;
2442         valid = 0;
2443         bp = NULL;
2444
2445         /*
2446          * There are up to 4 copies of the volume header (syncs iterate
2447          * between them so there is no single master).  We don't trust the
2448          * volu_size field so we don't know precisely how large the filesystem
2449          * is, so depend on the OS to return an error if we go beyond the
2450          * block device's EOF.
2451          */
2452         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2453                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2454                               HAMMER2_VOLUME_BYTES, &bp);
2455                 if (error) {
2456                         brelse(bp);
2457                         bp = NULL;
2458                         continue;
2459                 }
2460
2461                 vd = (struct hammer2_volume_data *) bp->b_data;
2462                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2463                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2464                         brelse(bp);
2465                         bp = NULL;
2466                         continue;
2467                 }
2468
2469                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2470                         /* XXX: Reversed-endianness filesystem */
2471                         kprintf("hammer2: reverse-endian filesystem detected");
2472                         brelse(bp);
2473                         bp = NULL;
2474                         continue;
2475                 }
2476
2477                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2478                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2479                                       HAMMER2_VOLUME_ICRC0_SIZE);
2480                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2481                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2482                                        HAMMER2_VOLUME_ICRC1_SIZE);
2483                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2484                         kprintf("hammer2 volume header crc "
2485                                 "mismatch copy #%d %08x/%08x\n",
2486                                 i, crc0, crc);
2487                         error_reported = 1;
2488                         brelse(bp);
2489                         bp = NULL;
2490                         continue;
2491                 }
2492                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2493                         valid = 1;
2494                         hmp->voldata = *vd;
2495                         hmp->volhdrno = i;
2496                 }
2497                 brelse(bp);
2498                 bp = NULL;
2499         }
2500         if (valid) {
2501                 hmp->volsync = hmp->voldata;
2502                 error = 0;
2503                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2504                         kprintf("hammer2: using volume header #%d\n",
2505                                 hmp->volhdrno);
2506                 }
2507         } else {
2508                 error = EINVAL;
2509                 kprintf("hammer2: no valid volume headers found!\n");
2510         }
2511         return (error);
2512 }
2513
2514 /*
2515  * This handles hysteresis on regular file flushes.  Because the BIOs are
2516  * routed to a thread it is possible for an excessive number to build up
2517  * and cause long front-end stalls long before the runningbuffspace limit
2518  * is hit, so we implement hammer2_flush_pipe to control the
2519  * hysteresis.
2520  *
2521  * This is a particular problem when compression is used.
2522  */
2523 void
2524 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2525 {
2526         atomic_add_int(&pmp->count_lwinprog, 1);
2527 }
2528
2529 void
2530 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2531 {
2532         int lwinprog;
2533
2534         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2535         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2536             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2537                 atomic_clear_int(&pmp->count_lwinprog,
2538                                  HAMMER2_LWINPROG_WAITING);
2539                 wakeup(&pmp->count_lwinprog);
2540         }
2541         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2542             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2543                 atomic_clear_int(&pmp->count_lwinprog,
2544                                  HAMMER2_LWINPROG_WAITING0);
2545                 wakeup(&pmp->count_lwinprog);
2546         }
2547 }
2548
2549 void
2550 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2551 {
2552         int lwinprog;
2553         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2554                                     HAMMER2_LWINPROG_WAITING0;
2555
2556         for (;;) {
2557                 lwinprog = pmp->count_lwinprog;
2558                 cpu_ccfence();
2559                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2560                         break;
2561                 tsleep_interlock(&pmp->count_lwinprog, 0);
2562                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2563                 lwinprog = pmp->count_lwinprog;
2564                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2565                         break;
2566                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2567         }
2568 }
2569
2570 /*
2571  * Manage excessive memory resource use for chain and related
2572  * structures.
2573  */
2574 void
2575 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2576 {
2577         uint32_t waiting;
2578         uint32_t count;
2579         uint32_t limit;
2580 #if 0
2581         static int zzticks;
2582 #endif
2583
2584         /*
2585          * Atomic check condition and wait.  Also do an early speedup of
2586          * the syncer to try to avoid hitting the wait.
2587          */
2588         for (;;) {
2589                 waiting = pmp->inmem_dirty_chains;
2590                 cpu_ccfence();
2591                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2592
2593                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2594                 if (limit < hammer2_limit_dirty_chains)
2595                         limit = hammer2_limit_dirty_chains;
2596                 if (limit < 1000)
2597                         limit = 1000;
2598
2599 #if 0
2600                 if ((int)(ticks - zzticks) > hz) {
2601                         zzticks = ticks;
2602                         kprintf("count %ld %ld\n", count, limit);
2603                 }
2604 #endif
2605
2606                 /*
2607                  * Block if there are too many dirty chains present, wait
2608                  * for the flush to clean some out.
2609                  */
2610                 if (count > limit) {
2611                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2612                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2613                                                waiting,
2614                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2615                                 speedup_syncer(pmp->mp);
2616                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2617                                        "chnmem", hz);
2618                         }
2619                         continue;       /* loop on success or fail */
2620                 }
2621
2622                 /*
2623                  * Try to start an early flush before we are forced to block.
2624                  */
2625                 if (count > limit * 7 / 10)
2626                         speedup_syncer(pmp->mp);
2627                 break;
2628         }
2629 }
2630
2631 void
2632 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2633 {
2634         if (pmp) {
2635                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2636         }
2637 }
2638
2639 void
2640 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2641 {
2642         uint32_t waiting;
2643
2644         if (pmp == NULL)
2645                 return;
2646
2647         for (;;) {
2648                 waiting = pmp->inmem_dirty_chains;
2649                 cpu_ccfence();
2650                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2651                                        waiting,
2652                                        (waiting - 1) &
2653                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2654                         break;
2655                 }
2656         }
2657
2658         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2659                 wakeup(&pmp->inmem_dirty_chains);
2660 }
2661
2662 /*
2663  * Debugging
2664  */
2665 void
2666 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2667 {
2668         hammer2_chain_t *scan;
2669         hammer2_chain_t *parent;
2670
2671         --*countp;
2672         if (*countp == 0) {
2673                 kprintf("%*.*s...\n", tab, tab, "");
2674                 return;
2675         }
2676         if (*countp < 0)
2677                 return;
2678         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2679                 tab, tab, "", pfx,
2680                 chain, chain->bref.type,
2681                 chain->bref.key, chain->bref.keybits,
2682                 chain->bref.mirror_tid);
2683
2684         kprintf("%*.*s      [%08x] (%s) refs=%d",
2685                 tab, tab, "",
2686                 chain->flags,
2687                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2688                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2689                 chain->refs);
2690
2691         parent = chain->parent;
2692         if (parent)
2693                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2694                         tab, tab, "",
2695                         parent, parent->flags, parent->refs);
2696         if (RB_EMPTY(&chain->core.rbtree)) {
2697                 kprintf("\n");
2698         } else {
2699                 kprintf(" {\n");
2700                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2701                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2702                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2703                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2704                                 chain->data->ipdata.filename);
2705                 else
2706                         kprintf("%*.*s}\n", tab, tab, "");
2707         }
2708 }