sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2017 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72         int pass;
  73 };
  74
  75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  76 static struct hammer2_mntlist hammer2_mntlist;
  77
  78 struct hammer2_pfslist hammer2_pfslist;
  79 struct lock hammer2_mntlk;
  80
  81 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
  82 int hammer2_debug;
  83 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
  84 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
  85 int hammer2_dedup_enable = 1;
  86 int hammer2_always_compress = 0;        /* always try to compress */
  87 int hammer2_inval_enable = 0;
  88 int hammer2_flush_pipe = 100;
  89 int hammer2_dio_count;
  90 int hammer2_dio_limit = 256;
  91 int hammer2_bulkfree_tps = 5000;
  92 long hammer2_chain_allocs;
  93 long hammer2_chain_frees;
  94 long hammer2_limit_dirty_chains;
  95 long hammer2_count_modified_chains;
  96 long hammer2_iod_invals;
  97 long hammer2_iod_file_read;
  98 long hammer2_iod_meta_read;
  99 long hammer2_iod_indr_read;
 100 long hammer2_iod_fmap_read;
 101 long hammer2_iod_volu_read;
 102 long hammer2_iod_file_write;
 103 long hammer2_iod_file_wembed;
 104 long hammer2_iod_file_wzero;
 105 long hammer2_iod_file_wdedup;
 106 long hammer2_iod_meta_write;
 107 long hammer2_iod_indr_write;
 108 long hammer2_iod_fmap_write;
 109 long hammer2_iod_volu_write;
 110
 111 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 112 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 113                 "Buffer used for compression.");
 114
 115 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 116 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 117                 "Buffer used for decompression.");
 118
 119 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 120
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
 122            &hammer2_supported_version, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 124            &hammer2_debug, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 126            &hammer2_cluster_meta_read, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 128            &hammer2_cluster_data_read, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 130            &hammer2_dedup_enable, 0, "");
 131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 132            &hammer2_always_compress, 0, "");
 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 134            &hammer2_inval_enable, 0, "");
 135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 136            &hammer2_flush_pipe, 0, "");
 137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
 138            &hammer2_bulkfree_tps, 0, "");
 139 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 140            &hammer2_chain_allocs, 0, "");
 141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 142            &hammer2_chain_frees, 0, "");
 143 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 144            &hammer2_limit_dirty_chains, 0, "");
 145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 146            &hammer2_count_modified_chains, 0, "");
 147 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 148            &hammer2_dio_count, 0, "");
 149 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
 150            &hammer2_dio_limit, 0, "");
 151
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 153            &hammer2_iod_invals, 0, "");
 154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 155            &hammer2_iod_file_read, 0, "");
 156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 157            &hammer2_iod_meta_read, 0, "");
 158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 159            &hammer2_iod_indr_read, 0, "");
 160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 161            &hammer2_iod_fmap_read, 0, "");
 162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 163            &hammer2_iod_volu_read, 0, "");
 164
 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 166            &hammer2_iod_file_write, 0, "");
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 168            &hammer2_iod_file_wembed, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 170            &hammer2_iod_file_wzero, 0, "");
 171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 172            &hammer2_iod_file_wdedup, 0, "");
 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 174            &hammer2_iod_meta_write, 0, "");
 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 176            &hammer2_iod_indr_write, 0, "");
 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 178            &hammer2_iod_fmap_write, 0, "");
 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 180            &hammer2_iod_volu_write, 0, "");
 181
 182 long hammer2_process_icrc32;
 183 long hammer2_process_xxhash64;
 184 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
 185            &hammer2_process_icrc32, 0, "");
 186 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
 187            &hammer2_process_xxhash64, 0, "");
 188
 189 static int hammer2_vfs_init(struct vfsconf *conf);
 190 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 191 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 192                                 struct ucred *cred);
 193 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 194                                 struct vnode *, struct ucred *);
 195 static int hammer2_recovery(hammer2_dev_t *hmp);
 196 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 197 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 198 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 199                                 struct ucred *cred);
 200 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 201                                 struct ucred *cred);
 202 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 203                                 struct fid *fhp, struct vnode **vpp);
 204 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 205 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 206                                 int *exflagsp, struct ucred **credanonp);
 207
 208 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 209 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 210
 211 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 212
 213 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 214 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 215                                 hammer2_dev_t *hmp);
 216
 217 /*
 218  * HAMMER2 vfs operations.
 219  */
 220 static struct vfsops hammer2_vfsops = {
 221         .vfs_init       = hammer2_vfs_init,
 222         .vfs_uninit     = hammer2_vfs_uninit,
 223         .vfs_sync       = hammer2_vfs_sync,
 224         .vfs_mount      = hammer2_vfs_mount,
 225         .vfs_unmount    = hammer2_vfs_unmount,
 226         .vfs_root       = hammer2_vfs_root,
 227         .vfs_statfs     = hammer2_vfs_statfs,
 228         .vfs_statvfs    = hammer2_vfs_statvfs,
 229         .vfs_vget       = hammer2_vfs_vget,
 230         .vfs_vptofh     = hammer2_vfs_vptofh,
 231         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 232         .vfs_checkexp   = hammer2_vfs_checkexp
 233 };
 234
 235 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 236
 237 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 238 MODULE_VERSION(hammer2, 1);
 239
 240 static
 241 int
 242 hammer2_vfs_init(struct vfsconf *conf)
 243 {
 244         static struct objcache_malloc_args margs_read;
 245         static struct objcache_malloc_args margs_write;
 246         static struct objcache_malloc_args margs_vop;
 247
 248         int error;
 249
 250         error = 0;
 251
 252         /*
 253          * A large DIO cache is needed to retain dedup enablement masks.
 254          * The bulkfree code clears related masks as part of the disk block
 255          * recycling algorithm, preventing it from being used for a later
 256          * dedup.
 257          *
 258          * NOTE: A large buffer cache can actually interfere with dedup
 259          *       operation because we dedup based on media physical buffers
 260          *       and not logical buffers.  Try to make the DIO case large
 261          *       enough to avoid this problem, but also cap it.
 262          */
 263         hammer2_dio_limit = nbuf * 2;
 264         if (hammer2_dio_limit > 100000)
 265                 hammer2_dio_limit = 100000;
 266
 267         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 268                 error = EINVAL;
 269         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 270                 error = EINVAL;
 271         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 272                 error = EINVAL;
 273
 274         if (error)
 275                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 276
 277         margs_read.objsize = 65536;
 278         margs_read.mtype = M_HAMMER2_DEBUFFER;
 279
 280         margs_write.objsize = 32768;
 281         margs_write.mtype = M_HAMMER2_CBUFFER;
 282
 283         margs_vop.objsize = sizeof(hammer2_xop_t);
 284         margs_vop.mtype = M_HAMMER2;
 285
 286         /*
 287          * Note thaht for the XOPS cache we want backing store allocations
 288          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 289          * confusion), so use the backing store function that does it.  This
 290          * means that initial XOPS objects are zerod but REUSED objects are
 291          * not.  So we are responsible for cleaning the object up sufficiently
 292          * for our needs before objcache_put()ing it back (typically just the
 293          * FIFO indices).
 294          */
 295         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 296                                 0, 1, NULL, NULL, NULL,
 297                                 objcache_malloc_alloc,
 298                                 objcache_malloc_free,
 299                                 &margs_read);
 300         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 301                                 0, 1, NULL, NULL, NULL,
 302                                 objcache_malloc_alloc,
 303                                 objcache_malloc_free,
 304                                 &margs_write);
 305         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 306                                 0, 1, NULL, NULL, NULL,
 307                                 objcache_malloc_alloc_zero,
 308                                 objcache_malloc_free,
 309                                 &margs_vop);
 310
 311
 312         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 313         TAILQ_INIT(&hammer2_mntlist);
 314         TAILQ_INIT(&hammer2_pfslist);
 315
 316         hammer2_limit_dirty_chains = maxvnodes / 10;
 317         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 318                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 319
 320         return (error);
 321 }
 322
 323 static
 324 int
 325 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 326 {
 327         objcache_destroy(cache_buffer_read);
 328         objcache_destroy(cache_buffer_write);
 329         objcache_destroy(cache_xops);
 330         return 0;
 331 }
 332
 333 /*
 334  * Core PFS allocator.  Used to allocate or reference the pmp structure
 335  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 336  * The pmp can be passed in or loaded by this function using the chain and
 337  * inode data.
 338  *
 339  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 340  * transactions.  Note that synchronization does not use this field.
 341  * (typically frontend operations and synchronization cannot run on the
 342  * same PFS node at the same time).
 343  *
 344  * XXX check locking
 345  */
 346 hammer2_pfs_t *
 347 hammer2_pfsalloc(hammer2_chain_t *chain,
 348                  const hammer2_inode_data_t *ripdata,
 349                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 350 {
 351         hammer2_pfs_t *pmp;
 352         hammer2_inode_t *iroot;
 353         int count;
 354         int i;
 355         int j;
 356
 357         pmp = NULL;
 358
 359         /*
 360          * Locate or create the PFS based on the cluster id.  If ripdata
 361          * is NULL this is a spmp which is unique and is always allocated.
 362          *
 363          * If the device is mounted in local mode all PFSs are considered
 364          * independent and not part of any cluster (for debugging only).
 365          */
 366         if (ripdata) {
 367                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 368                         if (force_local != pmp->force_local)
 369                                 continue;
 370                         if (force_local == NULL &&
 371                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 372                                  sizeof(pmp->pfs_clid)) == 0) {
 373                                         break;
 374                         } else if (force_local && pmp->pfs_names[0] &&
 375                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 376                                         break;
 377                         }
 378                 }
 379         }
 380
 381         if (pmp == NULL) {
 382                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 383                 pmp->force_local = force_local;
 384                 hammer2_trans_manage_init(pmp);
 385                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 386                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 387                 lockinit(&pmp->lock, "pfslk", 0, 0);
 388                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 389                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 390                 spin_init(&pmp->xop_spin, "h2xop");
 391                 spin_init(&pmp->lru_spin, "h2lru");
 392                 RB_INIT(&pmp->inum_tree);
 393                 TAILQ_INIT(&pmp->sideq);
 394                 TAILQ_INIT(&pmp->lru_list);
 395                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 396
 397                 /*
 398                  * Distribute backend operations to threads
 399                  */
 400                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 401                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 402
 403                 /*
 404                  * Save the last media transaction id for the flusher.  Set
 405                  * initial
 406                  */
 407                 if (ripdata)
 408                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 409                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 410
 411                 /*
 412                  * The synchronization thread may start too early, make
 413                  * sure it stays frozen until we are ready to let it go.
 414                  * XXX
 415                  */
 416                 /*
 417                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 418                                          HAMMER2_THREAD_REMASTER;
 419                 */
 420         }
 421
 422         /*
 423          * Create the PFS's root inode and any missing XOP helper threads.
 424          */
 425         if ((iroot = pmp->iroot) == NULL) {
 426                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 427                 if (ripdata)
 428                         iroot->meta = ripdata->meta;
 429                 pmp->iroot = iroot;
 430                 hammer2_inode_ref(iroot);
 431                 hammer2_inode_unlock(iroot);
 432         }
 433
 434         /*
 435          * Stop here if no chain is passed in.
 436          */
 437         if (chain == NULL)
 438                 goto done;
 439
 440         /*
 441          * When a chain is passed in we must add it to the PFS's root
 442          * inode, update pmp->pfs_types[], and update the syncronization
 443          * threads.
 444          *
 445          * When forcing local mode, mark the PFS as a MASTER regardless.
 446          *
 447          * At the moment empty spots can develop due to removals or failures.
 448          * Ultimately we want to re-fill these spots but doing so might
 449          * confused running code. XXX
 450          */
 451         hammer2_inode_ref(iroot);
 452         hammer2_mtx_ex(&iroot->lock);
 453         j = iroot->cluster.nchains;
 454
 455         if (j == HAMMER2_MAXCLUSTER) {
 456                 kprintf("hammer2_mount: cluster full!\n");
 457                 /* XXX fatal error? */
 458         } else {
 459                 KKASSERT(chain->pmp == NULL);
 460                 chain->pmp = pmp;
 461                 hammer2_chain_ref(chain);
 462                 iroot->cluster.array[j].chain = chain;
 463                 if (force_local)
 464                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 465                 else
 466                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 467                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 468                 pmp->pfs_hmps[j] = chain->hmp;
 469
 470                 /*
 471                  * If the PFS is already mounted we must account
 472                  * for the mount_count here.
 473                  */
 474                 if (pmp->mp)
 475                         ++chain->hmp->mount_count;
 476
 477                 /*
 478                  * May have to fixup dirty chain tracking.  Previous
 479                  * pmp was NULL so nothing to undo.
 480                  */
 481                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 482                         hammer2_pfs_memory_inc(pmp);
 483                 ++j;
 484         }
 485         iroot->cluster.nchains = j;
 486
 487         /*
 488          * Update nmasters from any PFS inode which is part of the cluster.
 489          * It is possible that this will result in a value which is too
 490          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 491          * override this value later on.
 492          *
 493          * (This informs us of masters that might not currently be
 494          *  discoverable by this mount).
 495          */
 496         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 497                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 498         }
 499
 500         /*
 501          * Count visible masters.  Masters are usually added with
 502          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 503          * are more (XXX and must update the master inodes).
 504          */
 505         count = 0;
 506         for (i = 0; i < iroot->cluster.nchains; ++i) {
 507                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 508                         ++count;
 509         }
 510         if (pmp->pfs_nmasters < count)
 511                 pmp->pfs_nmasters = count;
 512
 513         /*
 514          * Create missing synchronization and support threads.
 515          *
 516          * Single-node masters (including snapshots) have nothing to
 517          * synchronize and do not require this thread.
 518          *
 519          * Multi-node masters or any number of soft masters, slaves, copy,
 520          * or other PFS types need the thread.
 521          *
 522          * Each thread is responsible for its particular cluster index.
 523          * We use independent threads so stalls or mismatches related to
 524          * any given target do not affect other targets.
 525          */
 526         for (i = 0; i < iroot->cluster.nchains; ++i) {
 527                 /*
 528                  * Single-node masters (including snapshots) have nothing
 529                  * to synchronize and will make direct xops support calls,
 530                  * thus they do not require this thread.
 531                  *
 532                  * Note that there can be thousands of snapshots.  We do not
 533                  * want to create thousands of threads.
 534                  */
 535                 if (pmp->pfs_nmasters <= 1 &&
 536                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 537                         continue;
 538                 }
 539
 540                 /*
 541                  * Sync support thread
 542                  */
 543                 if (pmp->sync_thrs[i].td == NULL) {
 544                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 545                                            "h2nod", i, -1,
 546                                            hammer2_primary_sync_thread);
 547                 }
 548         }
 549
 550         /*
 551          * Create missing Xop threads
 552          *
 553          * NOTE: We create helper threads for all mounted PFSs or any
 554          *       PFSs with 2+ nodes (so the sync thread can update them,
 555          *       even if not mounted).
 556          */
 557         if (pmp->mp || iroot->cluster.nchains >= 2)
 558                 hammer2_xop_helper_create(pmp);
 559
 560         hammer2_mtx_unlock(&iroot->lock);
 561         hammer2_inode_drop(iroot);
 562 done:
 563         return pmp;
 564 }
 565
 566 /*
 567  * Deallocate an element of a probed PFS.  If destroying and this is a
 568  * MASTER, adjust nmasters.
 569  *
 570  * This function does not physically destroy the PFS element in its device
 571  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 572  */
 573 void
 574 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 575 {
 576         hammer2_inode_t *iroot;
 577         hammer2_chain_t *chain;
 578         int j;
 579
 580         /*
 581          * Cleanup our reference on iroot.  iroot is (should) not be needed
 582          * by the flush code.
 583          */
 584         iroot = pmp->iroot;
 585         if (iroot) {
 586                 /*
 587                  * Stop synchronizing
 588                  *
 589                  * XXX flush after acquiring the iroot lock.
 590                  * XXX clean out the cluster index from all inode structures.
 591                  */
 592                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 593
 594                 /*
 595                  * Remove the cluster index from the group.  If destroying
 596                  * the PFS and this is a master, adjust pfs_nmasters.
 597                  */
 598                 hammer2_mtx_ex(&iroot->lock);
 599                 chain = iroot->cluster.array[clindex].chain;
 600                 iroot->cluster.array[clindex].chain = NULL;
 601
 602                 switch(pmp->pfs_types[clindex]) {
 603                 case HAMMER2_PFSTYPE_MASTER:
 604                         if (destroying && pmp->pfs_nmasters > 0)
 605                                 --pmp->pfs_nmasters;
 606                         /* XXX adjust ripdata->meta.pfs_nmasters */
 607                         break;
 608                 default:
 609                         break;
 610                 }
 611                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 612
 613                 hammer2_mtx_unlock(&iroot->lock);
 614
 615                 /*
 616                  * Release the chain.
 617                  */
 618                 if (chain) {
 619                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 620                         hammer2_chain_drop(chain);
 621                 }
 622
 623                 /*
 624                  * Terminate all XOP threads for the cluster index.
 625                  */
 626                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 627                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 628         }
 629 }
 630
 631 /*
 632  * Destroy a PFS, typically only occurs after the last mount on a device
 633  * has gone away.
 634  */
 635 static void
 636 hammer2_pfsfree(hammer2_pfs_t *pmp)
 637 {
 638         hammer2_inode_t *iroot;
 639         hammer2_chain_t *chain;
 640         int i;
 641         int j;
 642
 643         /*
 644          * Cleanup our reference on iroot.  iroot is (should) not be needed
 645          * by the flush code.
 646          */
 647         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 648
 649         iroot = pmp->iroot;
 650         if (iroot) {
 651                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 652                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 653                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 654                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 655                 }
 656 #if REPORT_REFS_ERRORS
 657                 if (pmp->iroot->refs != 1)
 658                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 659                                 pmp->iroot, pmp->iroot->refs);
 660 #else
 661                 KKASSERT(pmp->iroot->refs == 1);
 662 #endif
 663                 /* ref for pmp->iroot */
 664                 hammer2_inode_drop(pmp->iroot);
 665                 pmp->iroot = NULL;
 666         }
 667
 668         /*
 669          * Cleanup chains remaining on LRU list.
 670          */
 671         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 672                 hammer2_chain_ref(chain);
 673                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 674                 hammer2_chain_drop(chain);
 675         }
 676
 677         /*
 678          * Free remaining pmp resources
 679          */
 680         kmalloc_destroy(&pmp->mmsg);
 681         kmalloc_destroy(&pmp->minode);
 682
 683         kfree(pmp, M_HAMMER2);
 684 }
 685
 686 /*
 687  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 688  * empty is terminated and freed.
 689  *
 690  * XXX inefficient.
 691  */
 692 static void
 693 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 694 {
 695         hammer2_pfs_t *pmp;
 696         hammer2_inode_t *iroot;
 697         hammer2_chain_t *rchain;
 698         int didfreeze;
 699         int i;
 700         int j;
 701
 702 again:
 703         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 704                 if ((iroot = pmp->iroot) == NULL)
 705                         continue;
 706                 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
 707                 hammer2_inode_run_sideq(pmp, 1);
 708                 hammer2_bioq_sync(pmp);
 709                 hammer2_trans_done(pmp);
 710                 if (hmp->spmp == pmp) {
 711                         hmp->spmp = NULL;
 712                         hmp->vchain.pmp = NULL;
 713                         hmp->fchain.pmp = NULL;
 714                 }
 715
 716                 /*
 717                  * Determine if this PFS is affected.  If it is we must
 718                  * freeze all management threads and lock its iroot.
 719                  *
 720                  * Freezing a management thread forces it idle, operations
 721                  * in-progress will be aborted and it will have to start
 722                  * over again when unfrozen, or exit if told to exit.
 723                  */
 724                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 725                         if (pmp->pfs_hmps[i] == hmp)
 726                                 break;
 727                 }
 728                 if (i != HAMMER2_MAXCLUSTER) {
 729                         /*
 730                          * Make sure all synchronization threads are locked
 731                          * down.
 732                          */
 733                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 734                                 if (pmp->pfs_hmps[i] == NULL)
 735                                         continue;
 736                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 737                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 738                                         hammer2_thr_freeze_async(
 739                                                 &pmp->xop_groups[j].thrs[i]);
 740                                 }
 741                         }
 742                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 743                                 if (pmp->pfs_hmps[i] == NULL)
 744                                         continue;
 745                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 746                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 747                                         hammer2_thr_freeze(
 748                                                 &pmp->xop_groups[j].thrs[i]);
 749                                 }
 750                         }
 751
 752                         /*
 753                          * Lock the inode and clean out matching chains.
 754                          * Note that we cannot use hammer2_inode_lock_*()
 755                          * here because that would attempt to validate the
 756                          * cluster that we are in the middle of ripping
 757                          * apart.
 758                          *
 759                          * WARNING! We are working directly on the inodes
 760                          *          embedded cluster.
 761                          */
 762                         hammer2_mtx_ex(&iroot->lock);
 763
 764                         /*
 765                          * Remove the chain from matching elements of the PFS.
 766                          */
 767                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 768                                 if (pmp->pfs_hmps[i] != hmp)
 769                                         continue;
 770                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 771                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 772                                         hammer2_thr_delete(
 773                                                 &pmp->xop_groups[j].thrs[i]);
 774                                 }
 775                                 rchain = iroot->cluster.array[i].chain;
 776                                 iroot->cluster.array[i].chain = NULL;
 777                                 pmp->pfs_types[i] = 0;
 778                                 if (pmp->pfs_names[i]) {
 779                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 780                                         pmp->pfs_names[i] = NULL;
 781                                 }
 782                                 if (rchain) {
 783                                         hammer2_chain_drop(rchain);
 784                                         /* focus hint */
 785                                         if (iroot->cluster.focus == rchain)
 786                                                 iroot->cluster.focus = NULL;
 787                                 }
 788                                 pmp->pfs_hmps[i] = NULL;
 789                         }
 790                         hammer2_mtx_unlock(&iroot->lock);
 791                         didfreeze = 1;  /* remaster, unfreeze down below */
 792                 } else {
 793                         didfreeze = 0;
 794                 }
 795
 796                 /*
 797                  * Cleanup trailing chains.  Gaps may remain.
 798                  */
 799                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 800                         if (pmp->pfs_hmps[i])
 801                                 break;
 802                 }
 803                 iroot->cluster.nchains = i + 1;
 804
 805                 /*
 806                  * If the PMP has no elements remaining we can destroy it.
 807                  * (this will transition management threads from frozen->exit).
 808                  */
 809                 if (iroot->cluster.nchains == 0) {
 810                         hammer2_pfsfree(pmp);
 811                         goto again;
 812                 }
 813
 814                 /*
 815                  * If elements still remain we need to set the REMASTER
 816                  * flag and unfreeze it.
 817                  */
 818                 if (didfreeze) {
 819                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 820                                 if (pmp->pfs_hmps[i] == NULL)
 821                                         continue;
 822                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 823                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 824                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 825                                         hammer2_thr_remaster(
 826                                                 &pmp->xop_groups[j].thrs[i]);
 827                                         hammer2_thr_unfreeze(
 828                                                 &pmp->xop_groups[j].thrs[i]);
 829                                 }
 830                         }
 831                 }
 832         }
 833 }
 834
 835 /*
 836  * Mount or remount HAMMER2 fileystem from physical media
 837  *
 838  *      mountroot
 839  *              mp              mount point structure
 840  *              path            NULL
 841  *              data            <unused>
 842  *              cred            <unused>
 843  *
 844  *      mount
 845  *              mp              mount point structure
 846  *              path            path to mount point
 847  *              data            pointer to argument structure in user space
 848  *                      volume  volume path (device@LABEL form)
 849  *                      hflags  user mount flags
 850  *              cred            user credentials
 851  *
 852  * RETURNS:     0       Success
 853  *              !0      error number
 854  */
 855 static
 856 int
 857 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 858                   struct ucred *cred)
 859 {
 860         struct hammer2_mount_info info;
 861         hammer2_pfs_t *pmp;
 862         hammer2_pfs_t *spmp;
 863         hammer2_dev_t *hmp;
 864         hammer2_dev_t *force_local;
 865         hammer2_key_t key_next;
 866         hammer2_key_t key_dummy;
 867         hammer2_key_t lhc;
 868         struct vnode *devvp;
 869         struct nlookupdata nd;
 870         hammer2_chain_t *parent;
 871         hammer2_chain_t *chain;
 872         hammer2_cluster_t *cluster;
 873         const hammer2_inode_data_t *ripdata;
 874         hammer2_blockref_t bref;
 875         struct file *fp;
 876         char devstr[MNAMELEN];
 877         size_t size;
 878         size_t done;
 879         char *dev;
 880         char *label;
 881         int ronly = 1;
 882         int error;
 883         int i;
 884
 885         hmp = NULL;
 886         pmp = NULL;
 887         dev = NULL;
 888         label = NULL;
 889         devvp = NULL;
 890
 891         kprintf("hammer2_mount\n");
 892
 893         if (path == NULL) {
 894                 /*
 895                  * Root mount
 896                  */
 897                 bzero(&info, sizeof(info));
 898                 info.cluster_fd = -1;
 899                 ksnprintf(devstr, sizeof(devstr), "%s",
 900                           mp->mnt_stat.f_mntfromname);
 901                 kprintf("hammer2_mount: root '%s'\n", devstr);
 902         } else {
 903                 /*
 904                  * Non-root mount or updating a mount
 905                  */
 906                 error = copyin(data, &info, sizeof(info));
 907                 if (error)
 908                         return (error);
 909
 910                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 911                 if (error)
 912                         return (error);
 913         }
 914
 915         /*
 916          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 917          * if no label specified, based on the partition id.  Error out if no
 918          * label or device (with partition id) is specified.  This is strictly
 919          * a convenience to match the default label created by newfs_hammer2,
 920          * our preference is that a label always be specified.
 921          *
 922          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
 923          *       that does not specify a device, as long as some H2 label
 924          *       has already been mounted from that device.  This makes
 925          *       mounting snapshots a lot easier.
 926          */
 927         dev = devstr;
 928         label = strchr(devstr, '@');
 929         if (label && ((label + 1) - dev) > done)
 930                 return (EINVAL);
 931         if (label == NULL || label[1] == 0) {
 932                 char slice;
 933
 934                 if (label == NULL)
 935                         label = devstr + strlen(devstr);
 936                 slice = label[-1];
 937                 switch(slice) {
 938                 case 'a':
 939                         label = "BOOT";
 940                         break;
 941                 case 'd':
 942                         label = "ROOT";
 943                         break;
 944                 default:
 945                         label = "DATA";
 946                         break;
 947                 }
 948         } else {
 949                 *label = '\0';
 950                 label++;
 951         }
 952
 953         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
 954                 dev, label, (mp->mnt_flag & MNT_RDONLY));
 955
 956         if (mp->mnt_flag & MNT_UPDATE) {
 957                 /*
 958                  * Update mount.  Note that pmp->iroot->cluster is
 959                  * an inode-embedded cluster and thus cannot be
 960                  * directly locked.
 961                  *
 962                  * XXX HAMMER2 needs to implement NFS export via
 963                  *     mountctl.
 964                  */
 965                 pmp = MPTOPMP(mp);
 966                 pmp->hflags = info.hflags;
 967                 cluster = &pmp->iroot->cluster;
 968                 for (i = 0; i < cluster->nchains; ++i) {
 969                         if (cluster->array[i].chain == NULL)
 970                                 continue;
 971                         hmp = cluster->array[i].chain->hmp;
 972                         devvp = hmp->devvp;
 973                         error = hammer2_remount(hmp, mp, path,
 974                                                 devvp, cred);
 975                         if (error)
 976                                 break;
 977                 }
 978
 979                 return error;
 980         }
 981
 982         /*
 983          * HMP device mount
 984          *
 985          * If a path is specified and dev is not an empty string, lookup the
 986          * name and verify that it referes to a block device.
 987          *
 988          * If a path is specified and dev is an empty string we fall through
 989          * and locate the label in the hmp search.
 990          */
 991         if (path && *dev != 0) {
 992                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 993                 if (error == 0)
 994                         error = nlookup(&nd);
 995                 if (error == 0)
 996                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 997                 nlookup_done(&nd);
 998         } else if (path == NULL) {
 999                 /* root mount */
1000                 cdev_t cdev = kgetdiskbyname(dev);
1001                 error = bdevvp(cdev, &devvp);
1002                 if (error)
1003                         kprintf("hammer2: cannot find '%s'\n", dev);
1004         } else {
1005                 /*
1006                  * We will locate the hmp using the label in the hmp loop.
1007                  */
1008                 error = 0;
1009         }
1010
1011         /*
1012          * Make sure its a block device.  Do not check to see if it is
1013          * already mounted until we determine that its a fresh H2 device.
1014          */
1015         if (error == 0 && devvp) {
1016                 vn_isdisk(devvp, &error);
1017         }
1018
1019         /*
1020          * Determine if the device has already been mounted.  After this
1021          * check hmp will be non-NULL if we are doing the second or more
1022          * hammer2 mounts from the same device.
1023          */
1024         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1025         if (devvp) {
1026                 /*
1027                  * Match the device.  Due to the way devfs works,
1028                  * we may not be able to directly match the vnode pointer,
1029                  * so also check to see if the underlying device matches.
1030                  */
1031                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1032                         if (hmp->devvp == devvp)
1033                                 break;
1034                         if (devvp->v_rdev &&
1035                             hmp->devvp->v_rdev == devvp->v_rdev) {
1036                                 break;
1037                         }
1038                 }
1039
1040                 /*
1041                  * If no match this may be a fresh H2 mount, make sure
1042                  * the device is not mounted on anything else.
1043                  */
1044                 if (hmp == NULL)
1045                         error = vfs_mountedon(devvp);
1046         } else if (error == 0) {
1047                 /*
1048                  * Match the label to a pmp already probed.
1049                  */
1050                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1051                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1052                                 if (pmp->pfs_names[i] &&
1053                                     strcmp(pmp->pfs_names[i], label) == 0) {
1054                                         hmp = pmp->pfs_hmps[i];
1055                                         break;
1056                                 }
1057                         }
1058                         if (hmp)
1059                                 break;
1060                 }
1061                 if (hmp == NULL)
1062                         error = ENOENT;
1063         }
1064
1065         /*
1066          * Open the device if this isn't a secondary mount and construct
1067          * the H2 device mount (hmp).
1068          */
1069         if (hmp == NULL) {
1070                 hammer2_chain_t *schain;
1071                 hammer2_xid_t xid;
1072
1073                 if (error == 0 && vcount(devvp) > 0) {
1074                         kprintf("Primary device already has references\n");
1075                         error = EBUSY;
1076                 }
1077
1078                 /*
1079                  * Now open the device
1080                  */
1081                 if (error == 0) {
1082                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1083                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1084                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1085                         if (error == 0) {
1086                                 error = VOP_OPEN(devvp,
1087                                              (ronly ? FREAD : FREAD | FWRITE),
1088                                              FSCRED, NULL);
1089                         }
1090                         vn_unlock(devvp);
1091                 }
1092                 if (error && devvp) {
1093                         vrele(devvp);
1094                         devvp = NULL;
1095                 }
1096                 if (error) {
1097                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1098                         return error;
1099                 }
1100                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1101                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1102                 hmp->ronly = ronly;
1103                 hmp->devvp = devvp;
1104                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1105                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1106                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1107                 RB_INIT(&hmp->iotree);
1108                 spin_init(&hmp->io_spin, "hm2mount_io");
1109                 spin_init(&hmp->list_spin, "hm2mount_list");
1110                 TAILQ_INIT(&hmp->flushq);
1111
1112                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1113                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1114                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1115
1116                 /*
1117                  * vchain setup. vchain.data is embedded.
1118                  * vchain.refs is initialized and will never drop to 0.
1119                  *
1120                  * NOTE! voldata is not yet loaded.
1121                  */
1122                 hmp->vchain.hmp = hmp;
1123                 hmp->vchain.refs = 1;
1124                 hmp->vchain.data = (void *)&hmp->voldata;
1125                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1126                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1127                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1128
1129                 hammer2_chain_core_init(&hmp->vchain);
1130                 /* hmp->vchain.u.xxx is left NULL */
1131
1132                 /*
1133                  * fchain setup.  fchain.data is embedded.
1134                  * fchain.refs is initialized and will never drop to 0.
1135                  *
1136                  * The data is not used but needs to be initialized to
1137                  * pass assertion muster.  We use this chain primarily
1138                  * as a placeholder for the freemap's top-level RBTREE
1139                  * so it does not interfere with the volume's topology
1140                  * RBTREE.
1141                  */
1142                 hmp->fchain.hmp = hmp;
1143                 hmp->fchain.refs = 1;
1144                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1145                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1146                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1147                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1148                 hmp->fchain.bref.methods =
1149                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1150                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1151
1152                 hammer2_chain_core_init(&hmp->fchain);
1153                 /* hmp->fchain.u.xxx is left NULL */
1154
1155                 /*
1156                  * Install the volume header and initialize fields from
1157                  * voldata.
1158                  */
1159                 error = hammer2_install_volume_header(hmp);
1160                 if (error) {
1161                         hammer2_unmount_helper(mp, NULL, hmp);
1162                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1163                         hammer2_vfs_unmount(mp, MNT_FORCE);
1164                         return error;
1165                 }
1166
1167                 /*
1168                  * Really important to get these right or flush will get
1169                  * confused.
1170                  */
1171                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1172                 kprintf("alloc spmp %p tid %016jx\n",
1173                         hmp->spmp, hmp->voldata.mirror_tid);
1174                 spmp = hmp->spmp;
1175
1176                 /*
1177                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1178                  * is inherited from the volume header.
1179                  */
1180                 xid = 0;
1181                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1182                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1183                 hmp->vchain.pmp = spmp;
1184                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1185                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1186                 hmp->fchain.pmp = spmp;
1187
1188                 /*
1189                  * First locate the super-root inode, which is key 0
1190                  * relative to the volume header's blockset.
1191                  *
1192                  * Then locate the root inode by scanning the directory keyspace
1193                  * represented by the label.
1194                  */
1195                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1196                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1197                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1198                                       &error, 0);
1199                 hammer2_chain_lookup_done(parent);
1200                 if (schain == NULL) {
1201                         kprintf("hammer2_mount: invalid super-root\n");
1202                         hammer2_unmount_helper(mp, NULL, hmp);
1203                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1204                         hammer2_vfs_unmount(mp, MNT_FORCE);
1205                         return EINVAL;
1206                 }
1207                 if (schain->error) {
1208                         kprintf("hammer2_mount: error %s reading super-root\n",
1209                                 hammer2_error_str(schain->error));
1210                         hammer2_chain_unlock(schain);
1211                         hammer2_chain_drop(schain);
1212                         schain = NULL;
1213                         hammer2_unmount_helper(mp, NULL, hmp);
1214                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1215                         hammer2_vfs_unmount(mp, MNT_FORCE);
1216                         return EINVAL;
1217                 }
1218
1219                 /*
1220                  * The super-root always uses an inode_tid of 1 when
1221                  * creating PFSs.
1222                  */
1223                 spmp->inode_tid = 1;
1224                 spmp->modify_tid = schain->bref.modify_tid + 1;
1225
1226                 /*
1227                  * Sanity-check schain's pmp and finish initialization.
1228                  * Any chain belonging to the super-root topology should
1229                  * have a NULL pmp (not even set to spmp).
1230                  */
1231                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1232                 KKASSERT(schain->pmp == NULL);
1233                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1234
1235                 /*
1236                  * Replace the dummy spmp->iroot with a real one.  It's
1237                  * easier to just do a wholesale replacement than to try
1238                  * to update the chain and fixup the iroot fields.
1239                  *
1240                  * The returned inode is locked with the supplied cluster.
1241                  */
1242                 cluster = hammer2_cluster_from_chain(schain);
1243                 hammer2_inode_drop(spmp->iroot);
1244                 spmp->iroot = NULL;
1245                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1246                 spmp->spmp_hmp = hmp;
1247                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1248                 spmp->pfs_hmps[0] = hmp;
1249                 hammer2_inode_ref(spmp->iroot);
1250                 hammer2_inode_unlock(spmp->iroot);
1251                 hammer2_cluster_unlock(cluster);
1252                 hammer2_cluster_drop(cluster);
1253                 schain = NULL;
1254                 /* leave spmp->iroot with one ref */
1255
1256                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1257                         error = hammer2_recovery(hmp);
1258                         /* XXX do something with error */
1259                 }
1260                 hammer2_update_pmps(hmp);
1261                 hammer2_iocom_init(hmp);
1262                 hammer2_bulkfree_init(hmp);
1263
1264                 /*
1265                  * Ref the cluster management messaging descriptor.  The mount
1266                  * program deals with the other end of the communications pipe.
1267                  *
1268                  * Root mounts typically do not supply one.
1269                  */
1270                 if (info.cluster_fd >= 0) {
1271                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1272                         if (fp) {
1273                                 hammer2_cluster_reconnect(hmp, fp);
1274                         } else {
1275                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1276                         }
1277                 }
1278         } else {
1279                 spmp = hmp->spmp;
1280                 if (info.hflags & HMNT2_DEVFLAGS) {
1281                         kprintf("hammer2: Warning: mount flags pertaining "
1282                                 "to the whole device may only be specified "
1283                                 "on the first mount of the device: %08x\n",
1284                                 info.hflags & HMNT2_DEVFLAGS);
1285                 }
1286         }
1287
1288         /*
1289          * Force local mount (disassociate all PFSs from their clusters).
1290          * Used primarily for debugging.
1291          */
1292         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1293
1294         /*
1295          * Lookup the mount point under the media-localized super-root.
1296          * Scanning hammer2_pfslist doesn't help us because it represents
1297          * PFS cluster ids which can aggregate several named PFSs together.
1298          *
1299          * cluster->pmp will incorrectly point to spmp and must be fixed
1300          * up later on.
1301          */
1302         hammer2_inode_lock(spmp->iroot, 0);
1303         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1304         lhc = hammer2_dirhash(label, strlen(label));
1305         chain = hammer2_chain_lookup(&parent, &key_next,
1306                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1307                                      &error, 0);
1308         while (chain) {
1309                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1310                     strcmp(label, chain->data->ipdata.filename) == 0) {
1311                         break;
1312                 }
1313                 chain = hammer2_chain_next(&parent, chain, &key_next,
1314                                             key_next,
1315                                             lhc + HAMMER2_DIRHASH_LOMASK,
1316                                             &error, 0);
1317         }
1318         if (parent) {
1319                 hammer2_chain_unlock(parent);
1320                 hammer2_chain_drop(parent);
1321         }
1322         hammer2_inode_unlock(spmp->iroot);
1323
1324         /*
1325          * PFS could not be found?
1326          */
1327         if (chain == NULL) {
1328                 if (error)
1329                         kprintf("hammer2_mount: PFS label I/O error\n");
1330                 else
1331                         kprintf("hammer2_mount: PFS label not found\n");
1332                 hammer2_unmount_helper(mp, NULL, hmp);
1333                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1334                 hammer2_vfs_unmount(mp, MNT_FORCE);
1335
1336                 return EINVAL;
1337         }
1338
1339         /*
1340          * Acquire the pmp structure (it should have already been allocated
1341          * via hammer2_update_pmps() so do not pass cluster in to add to
1342          * available chains).
1343          *
1344          * Check if the cluster has already been mounted.  A cluster can
1345          * only be mounted once, use null mounts to mount additional copies.
1346          */
1347         if (chain->error) {
1348                 kprintf("hammer2_mount: PFS label I/O error\n");
1349         } else {
1350                 ripdata = &chain->data->ipdata;
1351                 bref = chain->bref;
1352                 pmp = hammer2_pfsalloc(NULL, ripdata,
1353                                        bref.modify_tid, force_local);
1354         }
1355         hammer2_chain_unlock(chain);
1356         hammer2_chain_drop(chain);
1357
1358         /*
1359          * Finish the mount
1360          */
1361         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1362
1363         if (pmp->mp) {
1364                 kprintf("hammer2_mount: PFS already mounted!\n");
1365                 hammer2_unmount_helper(mp, NULL, hmp);
1366                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1367                 hammer2_vfs_unmount(mp, MNT_FORCE);
1368
1369                 return EBUSY;
1370         }
1371
1372         pmp->hflags = info.hflags;
1373         mp->mnt_flag |= MNT_LOCAL;
1374         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1375         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1376
1377         /*
1378          * required mount structure initializations
1379          */
1380         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1381         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1382
1383         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1384         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1385
1386         /*
1387          * Optional fields
1388          */
1389         mp->mnt_iosize_max = MAXPHYS;
1390
1391         /*
1392          * Connect up mount pointers.
1393          */
1394         hammer2_mount_helper(mp, pmp);
1395
1396         lockmgr(&hammer2_mntlk, LK_RELEASE);
1397
1398         /*
1399          * Finish setup
1400          */
1401         vfs_getnewfsid(mp);
1402         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1403         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1404         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1405
1406         if (path) {
1407                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1408                           MNAMELEN - 1, &size);
1409                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1410         } /* else root mount, already in there */
1411
1412         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1413         if (path) {
1414                 copyinstr(path, mp->mnt_stat.f_mntonname,
1415                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1416                           &size);
1417         } else {
1418                 /* root mount */
1419                 mp->mnt_stat.f_mntonname[0] = '/';
1420         }
1421
1422         /*
1423          * Initial statfs to prime mnt_stat.
1424          */
1425         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1426
1427         return 0;
1428 }
1429
1430 /*
1431  * Scan PFSs under the super-root and create hammer2_pfs structures.
1432  */
1433 static
1434 void
1435 hammer2_update_pmps(hammer2_dev_t *hmp)
1436 {
1437         const hammer2_inode_data_t *ripdata;
1438         hammer2_chain_t *parent;
1439         hammer2_chain_t *chain;
1440         hammer2_blockref_t bref;
1441         hammer2_dev_t *force_local;
1442         hammer2_pfs_t *spmp;
1443         hammer2_pfs_t *pmp;
1444         hammer2_key_t key_next;
1445         int error;
1446
1447         /*
1448          * Force local mount (disassociate all PFSs from their clusters).
1449          * Used primarily for debugging.
1450          */
1451         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1452
1453         /*
1454          * Lookup mount point under the media-localized super-root.
1455          *
1456          * cluster->pmp will incorrectly point to spmp and must be fixed
1457          * up later on.
1458          */
1459         spmp = hmp->spmp;
1460         hammer2_inode_lock(spmp->iroot, 0);
1461         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1462         chain = hammer2_chain_lookup(&parent, &key_next,
1463                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1464                                          &error, 0);
1465         while (chain) {
1466                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1467                         continue;
1468                 if (chain->error) {
1469                         kprintf("I/O error scanning PFS labels\n");
1470                 } else {
1471                         ripdata = &chain->data->ipdata;
1472                         bref = chain->bref;
1473
1474                         pmp = hammer2_pfsalloc(chain, ripdata,
1475                                                bref.modify_tid, force_local);
1476                 }
1477                 chain = hammer2_chain_next(&parent, chain, &key_next,
1478                                            key_next, HAMMER2_KEY_MAX,
1479                                            &error, 0);
1480         }
1481         if (parent) {
1482                 hammer2_chain_unlock(parent);
1483                 hammer2_chain_drop(parent);
1484         }
1485         hammer2_inode_unlock(spmp->iroot);
1486 }
1487
1488 static
1489 int
1490 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1491                 struct vnode *devvp, struct ucred *cred)
1492 {
1493         int error;
1494
1495         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1496                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1497                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1498                 vn_unlock(devvp);
1499                 error = hammer2_recovery(hmp);
1500                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1501                 if (error == 0) {
1502                         VOP_CLOSE(devvp, FREAD, NULL);
1503                         hmp->ronly = 0;
1504                 } else {
1505                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1506                 }
1507                 vn_unlock(devvp);
1508         } else {
1509                 error = 0;
1510         }
1511         return error;
1512 }
1513
1514 static
1515 int
1516 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1517 {
1518         hammer2_pfs_t *pmp;
1519         int flags;
1520         int error = 0;
1521
1522         pmp = MPTOPMP(mp);
1523
1524         if (pmp == NULL)
1525                 return(0);
1526
1527         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1528
1529         /*
1530          * If mount initialization proceeded far enough we must flush
1531          * its vnodes and sync the underlying mount points.  Three syncs
1532          * are required to fully flush the filesystem (freemap updates lag
1533          * by one flush, and one extra for safety).
1534          */
1535         if (mntflags & MNT_FORCE)
1536                 flags = FORCECLOSE;
1537         else
1538                 flags = 0;
1539         if (pmp->iroot) {
1540                 error = vflush(mp, 0, flags);
1541                 if (error)
1542                         goto failed;
1543                 hammer2_vfs_sync(mp, MNT_WAIT);
1544                 hammer2_vfs_sync(mp, MNT_WAIT);
1545                 hammer2_vfs_sync(mp, MNT_WAIT);
1546         }
1547
1548         /*
1549          * Cleanup the frontend support XOPS threads
1550          */
1551         hammer2_xop_helper_cleanup(pmp);
1552
1553         if (pmp->mp)
1554                 hammer2_unmount_helper(mp, pmp, NULL);
1555
1556         error = 0;
1557 failed:
1558         lockmgr(&hammer2_mntlk, LK_RELEASE);
1559
1560         return (error);
1561 }
1562
1563 /*
1564  * Mount helper, hook the system mount into our PFS.
1565  * The mount lock is held.
1566  *
1567  * We must bump the mount_count on related devices for any
1568  * mounted PFSs.
1569  */
1570 static
1571 void
1572 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1573 {
1574         hammer2_cluster_t *cluster;
1575         hammer2_chain_t *rchain;
1576         int i;
1577
1578         mp->mnt_data = (qaddr_t)pmp;
1579         pmp->mp = mp;
1580
1581         /*
1582          * After pmp->mp is set we have to adjust hmp->mount_count.
1583          */
1584         cluster = &pmp->iroot->cluster;
1585         for (i = 0; i < cluster->nchains; ++i) {
1586                 rchain = cluster->array[i].chain;
1587                 if (rchain == NULL)
1588                         continue;
1589                 ++rchain->hmp->mount_count;
1590                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1591                         rchain->hmp, rchain->hmp->mount_count);
1592         }
1593
1594         /*
1595          * Create missing Xop threads
1596          */
1597         hammer2_xop_helper_create(pmp);
1598 }
1599
1600 /*
1601  * Mount helper, unhook the system mount from our PFS.
1602  * The mount lock is held.
1603  *
1604  * If hmp is supplied a mount responsible for being the first to open
1605  * the block device failed and the block device and all PFSs using the
1606  * block device must be cleaned up.
1607  *
1608  * If pmp is supplied multiple devices might be backing the PFS and each
1609  * must be disconnected.  This might not be the last PFS using some of the
1610  * underlying devices.  Also, we have to adjust our hmp->mount_count
1611  * accounting for the devices backing the pmp which is now undergoing an
1612  * unmount.
1613  */
1614 static
1615 void
1616 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1617 {
1618         hammer2_cluster_t *cluster;
1619         hammer2_chain_t *rchain;
1620         struct vnode *devvp;
1621         int dumpcnt;
1622         int ronly;
1623         int i;
1624
1625         /*
1626          * If no device supplied this is a high-level unmount and we have to
1627          * to disconnect the mount, adjust mount_count, and locate devices
1628          * that might now have no mounts.
1629          */
1630         if (pmp) {
1631                 KKASSERT(hmp == NULL);
1632                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1633                 pmp->mp = NULL;
1634                 mp->mnt_data = NULL;
1635
1636                 /*
1637                  * After pmp->mp is cleared we have to account for
1638                  * mount_count.
1639                  */
1640                 cluster = &pmp->iroot->cluster;
1641                 for (i = 0; i < cluster->nchains; ++i) {
1642                         rchain = cluster->array[i].chain;
1643                         if (rchain == NULL)
1644                                 continue;
1645                         --rchain->hmp->mount_count;
1646                         /* scrapping hmp now may invalidate the pmp */
1647                 }
1648 again:
1649                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1650                         if (hmp->mount_count == 0) {
1651                                 hammer2_unmount_helper(NULL, NULL, hmp);
1652                                 goto again;
1653                         }
1654                 }
1655                 return;
1656         }
1657
1658         /*
1659          * Try to terminate the block device.  We can't terminate it if
1660          * there are still PFSs referencing it.
1661          */
1662         if (hmp->mount_count)
1663                 return;
1664
1665         /*
1666          * Decomission the network before we start messing with the
1667          * device and PFS.
1668          */
1669         hammer2_iocom_uninit(hmp);
1670
1671         hammer2_bulkfree_uninit(hmp);
1672         hammer2_pfsfree_scan(hmp);
1673         hammer2_dev_exlock(hmp);        /* XXX order */
1674
1675         /*
1676          * Cycle the volume data lock as a safety (probably not needed any
1677          * more).  To ensure everything is out we need to flush at least
1678          * three times.  (1) The running of the sideq can dirty the
1679          * filesystem, (2) A normal flush can dirty the freemap, and
1680          * (3) ensure that the freemap is fully synchronized.
1681          *
1682          * The next mount's recovery scan can clean everything up but we want
1683          * to leave the filesystem in a 100% clean state on a normal unmount.
1684          */
1685 #if 0
1686         hammer2_voldata_lock(hmp);
1687         hammer2_voldata_unlock(hmp);
1688 #endif
1689
1690         /*
1691          * Flush whatever is left.  Unmounted but modified PFS's might still
1692          * have some dirty chains on them.
1693          */
1694         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1695         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1696         hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP | HAMMER2_FLUSH_ALL);
1697         hammer2_chain_unlock(&hmp->fchain);
1698         hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP | HAMMER2_FLUSH_ALL);
1699         hammer2_chain_unlock(&hmp->vchain);
1700
1701         if ((hmp->vchain.flags | hmp->fchain.flags) &
1702             HAMMER2_CHAIN_FLUSH_MASK) {
1703                 kprintf("hammer2_unmount: chains left over "
1704                         "after final sync\n");
1705                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1706                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1707
1708                 if (hammer2_debug & 0x0010)
1709                         Debugger("entered debugger");
1710         }
1711
1712         KKASSERT(hmp->spmp == NULL);
1713
1714         /*
1715          * Finish up with the device vnode
1716          */
1717         if ((devvp = hmp->devvp) != NULL) {
1718                 ronly = hmp->ronly;
1719                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1720                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1721                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1722                         ronly);
1723                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1724                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1725                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1726                 hmp->devvp = NULL;
1727                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1728                 vn_unlock(devvp);
1729                 vrele(devvp);
1730                 devvp = NULL;
1731         }
1732
1733         /*
1734          * Clear vchain/fchain flags that might prevent final cleanup
1735          * of these chains.
1736          */
1737         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1738                 atomic_add_long(&hammer2_count_modified_chains, -1);
1739                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1740                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1741         }
1742         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1743                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1744         }
1745
1746         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1747                 atomic_add_long(&hammer2_count_modified_chains, -1);
1748                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1749                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1750         }
1751         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1752                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1753         }
1754
1755         /*
1756          * Final drop of embedded freemap root chain to
1757          * clean up fchain.core (fchain structure is not
1758          * flagged ALLOCATED so it is cleaned out and then
1759          * left to rot).
1760          */
1761         hammer2_chain_drop(&hmp->fchain);
1762
1763         /*
1764          * Final drop of embedded volume root chain to clean
1765          * up vchain.core (vchain structure is not flagged
1766          * ALLOCATED so it is cleaned out and then left to
1767          * rot).
1768          */
1769         dumpcnt = 50;
1770         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1771         dumpcnt = 50;
1772         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1773         hammer2_dev_unlock(hmp);
1774         hammer2_chain_drop(&hmp->vchain);
1775
1776         hammer2_io_cleanup(hmp, &hmp->iotree);
1777         if (hmp->iofree_count) {
1778                 kprintf("io_cleanup: %d I/O's left hanging\n",
1779                         hmp->iofree_count);
1780         }
1781
1782         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1783         kmalloc_destroy(&hmp->mchain);
1784         kfree(hmp, M_HAMMER2);
1785 }
1786
1787 int
1788 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1789                  ino_t ino, struct vnode **vpp)
1790 {
1791         hammer2_xop_lookup_t *xop;
1792         hammer2_pfs_t *pmp;
1793         hammer2_inode_t *ip;
1794         hammer2_tid_t inum;
1795         int error;
1796
1797         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1798
1799         error = 0;
1800         pmp = MPTOPMP(mp);
1801
1802         /*
1803          * Easy if we already have it cached
1804          */
1805         ip = hammer2_inode_lookup(pmp, inum);
1806         if (ip) {
1807                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1808                 *vpp = hammer2_igetv(ip, &error);
1809                 hammer2_inode_unlock(ip);
1810                 hammer2_inode_drop(ip);         /* from lookup */
1811
1812                 return error;
1813         }
1814
1815         /*
1816          * Otherwise we have to find the inode
1817          */
1818         xop = hammer2_xop_alloc(pmp->iroot, 0);
1819         xop->lhc = inum;
1820         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1821         error = hammer2_xop_collect(&xop->head, 0);
1822
1823         if (error == 0) {
1824                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1825                         kprintf("vget: no collect error but also no rdata\n");
1826                         kprintf("xop %p\n", xop);
1827                         while ((hammer2_debug & 0x80000) == 0) {
1828                                 tsleep(xop, PCATCH, "wait", hz * 10);
1829                         }
1830                         ip = NULL;
1831                 } else {
1832                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1833                 }
1834         }
1835         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1836
1837         if (ip) {
1838                 *vpp = hammer2_igetv(ip, &error);
1839                 hammer2_inode_unlock(ip);
1840         } else {
1841                 *vpp = NULL;
1842                 error = ENOENT;
1843         }
1844         return (error);
1845 }
1846
1847 static
1848 int
1849 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1850 {
1851         hammer2_pfs_t *pmp;
1852         struct vnode *vp;
1853         int error;
1854
1855         pmp = MPTOPMP(mp);
1856         if (pmp->iroot == NULL) {
1857                 *vpp = NULL;
1858                 return EINVAL;
1859         }
1860
1861         error = 0;
1862         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1863
1864         while (pmp->inode_tid == 0) {
1865                 hammer2_xop_ipcluster_t *xop;
1866                 hammer2_inode_meta_t *meta;
1867
1868                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1869                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1870                 error = hammer2_xop_collect(&xop->head, 0);
1871
1872                 if (error == 0) {
1873                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1874                         pmp->iroot->meta = *meta;
1875                         pmp->inode_tid = meta->pfs_inum + 1;
1876                         if (pmp->inode_tid < HAMMER2_INODE_START)
1877                                 pmp->inode_tid = HAMMER2_INODE_START;
1878                         pmp->modify_tid =
1879                                 xop->head.cluster.focus->bref.modify_tid + 1;
1880                         kprintf("PFS: Starting inode %jd\n",
1881                                 (intmax_t)pmp->inode_tid);
1882                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1883                                 pmp->inode_tid, pmp->modify_tid);
1884                         wakeup(&pmp->iroot);
1885
1886                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1887
1888                         /*
1889                          * Prime the mount info.
1890                          */
1891                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1892                         break;
1893                 }
1894
1895                 /*
1896                  * Loop, try again
1897                  */
1898                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1899                 hammer2_inode_unlock(pmp->iroot);
1900                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1901                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1902                 if (error == EINTR)
1903                         break;
1904         }
1905
1906         if (error) {
1907                 hammer2_inode_unlock(pmp->iroot);
1908                 *vpp = NULL;
1909         } else {
1910                 vp = hammer2_igetv(pmp->iroot, &error);
1911                 hammer2_inode_unlock(pmp->iroot);
1912                 *vpp = vp;
1913         }
1914
1915         return (error);
1916 }
1917
1918 /*
1919  * Filesystem status
1920  *
1921  * XXX incorporate ipdata->meta.inode_quota and data_quota
1922  */
1923 static
1924 int
1925 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1926 {
1927         hammer2_pfs_t *pmp;
1928         hammer2_dev_t *hmp;
1929         hammer2_blockref_t bref;
1930         struct statfs tmp;
1931         int i;
1932
1933         /*
1934          * NOTE: iroot might not have validated the cluster yet.
1935          */
1936         pmp = MPTOPMP(mp);
1937
1938         bzero(&tmp, sizeof(tmp));
1939
1940         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1941                 hmp = pmp->pfs_hmps[i];
1942                 if (hmp == NULL)
1943                         continue;
1944                 if (pmp->iroot->cluster.array[i].chain)
1945                         bref = pmp->iroot->cluster.array[i].chain->bref;
1946                 else
1947                         bzero(&bref, sizeof(bref));
1948
1949                 tmp.f_files = bref.embed.stats.inode_count;
1950                 tmp.f_ffree = 0;
1951                 tmp.f_blocks = hmp->voldata.allocator_size /
1952                                mp->mnt_vstat.f_bsize;
1953                 tmp.f_bfree = hmp->voldata.allocator_free /
1954                               mp->mnt_vstat.f_bsize;
1955                 tmp.f_bavail = tmp.f_bfree;
1956
1957                 if (cred && cred->cr_uid != 0) {
1958                         uint64_t adj;
1959
1960                         /* 5% */
1961                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
1962                         tmp.f_blocks -= adj;
1963                         tmp.f_bfree -= adj;
1964                         tmp.f_bavail -= adj;
1965                 }
1966
1967                 mp->mnt_stat.f_blocks = tmp.f_blocks;
1968                 mp->mnt_stat.f_bfree = tmp.f_bfree;
1969                 mp->mnt_stat.f_bavail = tmp.f_bavail;
1970                 mp->mnt_stat.f_files = tmp.f_files;
1971                 mp->mnt_stat.f_ffree = tmp.f_ffree;
1972
1973                 *sbp = mp->mnt_stat;
1974         }
1975         return (0);
1976 }
1977
1978 static
1979 int
1980 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1981 {
1982         hammer2_pfs_t *pmp;
1983         hammer2_dev_t *hmp;
1984         hammer2_blockref_t bref;
1985         struct statvfs tmp;
1986         int i;
1987
1988         /*
1989          * NOTE: iroot might not have validated the cluster yet.
1990          */
1991         pmp = MPTOPMP(mp);
1992         bzero(&tmp, sizeof(tmp));
1993
1994         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1995                 hmp = pmp->pfs_hmps[i];
1996                 if (hmp == NULL)
1997                         continue;
1998                 if (pmp->iroot->cluster.array[i].chain)
1999                         bref = pmp->iroot->cluster.array[i].chain->bref;
2000                 else
2001                         bzero(&bref, sizeof(bref));
2002
2003                 tmp.f_files = bref.embed.stats.inode_count;
2004                 tmp.f_ffree = 0;
2005                 tmp.f_blocks = hmp->voldata.allocator_size /
2006                                mp->mnt_vstat.f_bsize;
2007                 tmp.f_bfree = hmp->voldata.allocator_free /
2008                               mp->mnt_vstat.f_bsize;
2009                 tmp.f_bavail = tmp.f_bfree;
2010
2011                 if (cred && cred->cr_uid != 0) {
2012                         uint64_t adj;
2013
2014                         /* 5% */
2015                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2016                         tmp.f_blocks -= adj;
2017                         tmp.f_bfree -= adj;
2018                         tmp.f_bavail -= adj;
2019                 }
2020
2021                 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2022                 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2023                 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2024                 mp->mnt_vstat.f_files = tmp.f_files;
2025                 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2026
2027                 *sbp = mp->mnt_vstat;
2028         }
2029         return (0);
2030 }
2031
2032 /*
2033  * Mount-time recovery (RW mounts)
2034  *
2035  * Updates to the free block table are allowed to lag flushes by one
2036  * transaction.  In case of a crash, then on a fresh mount we must do an
2037  * incremental scan of the last committed transaction id and make sure that
2038  * all related blocks have been marked allocated.
2039  *
2040  * The super-root topology and each PFS has its own transaction id domain,
2041  * so we must track PFS boundary transitions.
2042  */
2043 struct hammer2_recovery_elm {
2044         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2045         hammer2_chain_t *chain;
2046         hammer2_tid_t sync_tid;
2047 };
2048
2049 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2050
2051 struct hammer2_recovery_info {
2052         struct hammer2_recovery_list list;
2053         hammer2_tid_t   mtid;
2054         int     depth;
2055 };
2056
2057 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2058                         hammer2_chain_t *parent,
2059                         struct hammer2_recovery_info *info,
2060                         hammer2_tid_t sync_tid);
2061
2062 #define HAMMER2_RECOVERY_MAXDEPTH       10
2063
2064 static
2065 int
2066 hammer2_recovery(hammer2_dev_t *hmp)
2067 {
2068         struct hammer2_recovery_info info;
2069         struct hammer2_recovery_elm *elm;
2070         hammer2_chain_t *parent;
2071         hammer2_tid_t sync_tid;
2072         hammer2_tid_t mirror_tid;
2073         int error;
2074
2075         hammer2_trans_init(hmp->spmp, 0);
2076
2077         sync_tid = hmp->voldata.freemap_tid;
2078         mirror_tid = hmp->voldata.mirror_tid;
2079
2080         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2081         if (sync_tid >= mirror_tid) {
2082                 kprintf(" no recovery needed\n");
2083         } else {
2084                 kprintf(" freemap recovery %016jx-%016jx\n",
2085                         sync_tid + 1, mirror_tid);
2086         }
2087
2088         TAILQ_INIT(&info.list);
2089         info.depth = 0;
2090         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2091         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2092         hammer2_chain_lookup_done(parent);
2093
2094         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2095                 TAILQ_REMOVE(&info.list, elm, entry);
2096                 parent = elm->chain;
2097                 sync_tid = elm->sync_tid;
2098                 kfree(elm, M_HAMMER2);
2099
2100                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2101                 error |= hammer2_recovery_scan(hmp, parent, &info,
2102                                               hmp->voldata.freemap_tid);
2103                 hammer2_chain_unlock(parent);
2104                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2105         }
2106         hammer2_trans_done(hmp->spmp);
2107
2108         return error;
2109 }
2110
2111 static
2112 int
2113 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2114                       struct hammer2_recovery_info *info,
2115                       hammer2_tid_t sync_tid)
2116 {
2117         const hammer2_inode_data_t *ripdata;
2118         hammer2_chain_t *chain;
2119         hammer2_blockref_t bref;
2120         int tmp_error;
2121         int rup_error;
2122         int error;
2123         int first;
2124
2125         /*
2126          * Adjust freemap to ensure that the block(s) are marked allocated.
2127          */
2128         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2129                 hammer2_freemap_adjust(hmp, &parent->bref,
2130                                        HAMMER2_FREEMAP_DORECOVER);
2131         }
2132
2133         /*
2134          * Check type for recursive scan
2135          */
2136         switch(parent->bref.type) {
2137         case HAMMER2_BREF_TYPE_VOLUME:
2138                 /* data already instantiated */
2139                 break;
2140         case HAMMER2_BREF_TYPE_INODE:
2141                 /*
2142                  * Must instantiate data for DIRECTDATA test and also
2143                  * for recursion.
2144                  */
2145                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2146                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2147                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2148                         /* not applicable to recovery scan */
2149                         hammer2_chain_unlock(parent);
2150                         return 0;
2151                 }
2152                 hammer2_chain_unlock(parent);
2153                 break;
2154         case HAMMER2_BREF_TYPE_INDIRECT:
2155                 /*
2156                  * Must instantiate data for recursion
2157                  */
2158                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2159                 hammer2_chain_unlock(parent);
2160                 break;
2161         case HAMMER2_BREF_TYPE_DIRENT:
2162         case HAMMER2_BREF_TYPE_DATA:
2163         case HAMMER2_BREF_TYPE_FREEMAP:
2164         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2165         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2166                 /* not applicable to recovery scan */
2167                 return 0;
2168                 break;
2169         default:
2170                 return HAMMER2_ERROR_BADBREF;
2171         }
2172
2173         /*
2174          * Defer operation if depth limit reached or if we are crossing a
2175          * PFS boundary.
2176          */
2177         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2178                 struct hammer2_recovery_elm *elm;
2179
2180                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2181                 elm->chain = parent;
2182                 elm->sync_tid = sync_tid;
2183                 hammer2_chain_ref(parent);
2184                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2185                 /* unlocked by caller */
2186
2187                 return(0);
2188         }
2189
2190
2191         /*
2192          * Recursive scan of the last flushed transaction only.  We are
2193          * doing this without pmp assignments so don't leave the chains
2194          * hanging around after we are done with them.
2195          *
2196          * error        Cumulative error this level only
2197          * rup_error    Cumulative error for recursion
2198          * tmp_error    Specific non-cumulative recursion error
2199          */
2200         chain = NULL;
2201         first = 1;
2202         rup_error = 0;
2203         error = 0;
2204
2205         for (;;) {
2206                 error |= hammer2_chain_scan(parent, &chain, &bref,
2207                                             &first,
2208                                             HAMMER2_LOOKUP_NODATA);
2209
2210                 /*
2211                  * Problem during scan or EOF
2212                  */
2213                 if (error)
2214                         break;
2215
2216                 /*
2217                  * If this is a leaf
2218                  */
2219                 if (chain == NULL) {
2220                         if (bref.mirror_tid > sync_tid) {
2221                                 hammer2_freemap_adjust(hmp, &bref,
2222                                                      HAMMER2_FREEMAP_DORECOVER);
2223                         }
2224                         continue;
2225                 }
2226
2227                 /*
2228                  * This may or may not be a recursive node.
2229                  */
2230                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2231                 if (bref.mirror_tid > sync_tid) {
2232                         ++info->depth;
2233                         tmp_error = hammer2_recovery_scan(hmp, chain,
2234                                                            info, sync_tid);
2235                         --info->depth;
2236                 } else {
2237                         tmp_error = 0;
2238                 }
2239
2240                 /*
2241                  * Flush the recovery at the PFS boundary to stage it for
2242                  * the final flush of the super-root topology.
2243                  */
2244                 if (tmp_error == 0 &&
2245                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2246                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2247                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2248                                              HAMMER2_FLUSH_ALL);
2249                 }
2250                 rup_error |= tmp_error;
2251         }
2252         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2253 }
2254
2255 /*
2256  * Sync a mount point; this is called periodically on a per-mount basis from
2257  * the filesystem syncer, and whenever a user issues a sync.
2258  */
2259 int
2260 hammer2_vfs_sync(struct mount *mp, int waitfor)
2261 {
2262         hammer2_xop_flush_t *xop;
2263         struct hammer2_sync_info info;
2264         hammer2_inode_t *iroot;
2265         hammer2_pfs_t *pmp;
2266         int flags;
2267         int error;
2268
2269         pmp = MPTOPMP(mp);
2270         iroot = pmp->iroot;
2271         KKASSERT(iroot);
2272         KKASSERT(iroot->pmp == pmp);
2273
2274         /*
2275          * We can't acquire locks on existing vnodes while in a transaction
2276          * without risking a deadlock.  This assumes that vfsync() can be
2277          * called without the vnode locked (which it can in DragonFly).
2278          * Otherwise we'd have to implement a multi-pass or flag the lock
2279          * failures and retry.
2280          *
2281          * The reclamation code interlocks with the sync list's token
2282          * (by removing the vnode from the scan list) before unlocking
2283          * the inode, giving us time to ref the inode.
2284          */
2285         /*flags = VMSC_GETVP;*/
2286         flags = 0;
2287         if (waitfor & MNT_LAZY)
2288                 flags |= VMSC_ONEPASS;
2289
2290         /*
2291          * Flush vnodes individually using a normal transaction to avoid
2292          * stalling any concurrent operations.  This will flush the related
2293          * buffer cache buffers and inodes to the media.
2294          *
2295          * For efficiency do an async pass before making sure with a
2296          * synchronous pass on all related buffer cache buffers.
2297          *
2298          * Do a single synchronous pass to avoid double-flushing vnodes,
2299          * which can waste copy-on-write blocks.  XXX do not do two passes.
2300          */
2301         hammer2_trans_init(pmp, 0);
2302         info.error = 0;
2303         info.waitfor = MNT_NOWAIT;
2304         info.pass = 1;
2305         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2306         info.waitfor = MNT_WAIT;
2307         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2308
2309         info.pass = 2;
2310         info.waitfor = MNT_WAIT;
2311         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2312
2313         /*
2314          * We must also run the sideq to handle any disconnected inodes
2315          * as the vnode scan will not see these.
2316          */
2317         hammer2_inode_run_sideq(pmp, 1);
2318         hammer2_trans_done(pmp);
2319
2320         /*
2321          * Start our flush transaction and flush the root topology down to
2322          * the inodes, but not the inodes themselves (which we already flushed
2323          * above).  Any concurrent activity effecting inode contents will not
2324          * be part of this flush cycle.
2325          *
2326          * (1) vfsync() all dirty vnodes via vfsyncscan().
2327          *
2328          * (2) Flush any remaining dirty inodes (the sideq), including any
2329          *     which may have been created during or raced against the
2330          *     vfsync().  To catch all cases this must be done after the
2331          *     vfsync().
2332          *
2333          * (3) Wait for any pending BIO I/O to complete (hammer2_bioq_sync()).
2334          *
2335          * NOTE! It is still possible for the paging code to push pages
2336          *       out via a UIO_NOCOPY hammer2_vop_write() during the main
2337          *       flush.
2338          */
2339         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2340
2341         /*
2342          * Use the XOP interface to concurrently flush all nodes to
2343          * synchronize the PFSROOT subtopology to the media.  A standard
2344          * end-of-scan ENOENT error indicates cluster sufficiency.
2345          *
2346          * Note that this flush will not be visible on crash recovery until
2347          * we flush the super-root topology in the next loop.
2348          *
2349          * XXX For now wait for all flushes to complete.
2350          */
2351         if (iroot) {
2352                 /*
2353                  * If unmounting try to flush everything including any
2354                  * sub-trees under inodes, just in case there is dangling
2355                  * modified data, as a safety.  Otherwise just flush up to
2356                  * the inodes in this stage.
2357                  */
2358                 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2359                         xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2360                 } else {
2361                         xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2362                                                        HAMMER2_XOP_INODE_STOP);
2363                 }
2364                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2365                 error = hammer2_xop_collect(&xop->head,
2366                                             HAMMER2_XOP_COLLECT_WAITALL);
2367                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2368                 if (error == HAMMER2_ERROR_ENOENT)
2369                         error = 0;
2370                 else
2371                         error = hammer2_error_to_errno(error);
2372         } else {
2373                 error = 0;
2374         }
2375         hammer2_trans_done(pmp);
2376
2377         return (error);
2378 }
2379
2380 /*
2381  * Sync passes.
2382  *
2383  * Note that we ignore the tranasction mtid we got above.  Instead,
2384  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2385  * transactions.
2386  *
2387  * WARNING! The frontend might be waiting on chnmem (limit_dirty_chains)
2388  * while holding a vnode locked.  When this situation occurs we cannot
2389  * safely test whether it is ok to clear the dirty bit on the vnode.
2390  * However, we can still flush the inode's topology.
2391  */
2392 static int
2393 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2394 {
2395         struct hammer2_sync_info *info = data;
2396         hammer2_inode_t *ip;
2397         int error;
2398
2399         /*
2400          * Degenerate cases.  Note that ip == NULL typically means the
2401          * syncer vnode itself and we don't want to vclrisdirty() in that
2402          * situation.
2403          */
2404         ip = VTOI(vp);
2405         if (ip == NULL) {
2406                 return(0);
2407         }
2408         if (vp->v_type == VNON || vp->v_type == VBAD) {
2409                 vclrisdirty(vp);
2410                 return(0);
2411         }
2412
2413         /*
2414          * Synchronize the buffer cche and inode meta-data to the backing
2415          * chain topology.
2416          *
2417          * vfsync is not necessarily synchronous, so it is best NOT to try
2418          * to flush the backing topology to media at this point.
2419          */
2420         hammer2_inode_ref(ip);
2421         if ((ip->flags & (HAMMER2_INODE_RESIZED|HAMMER2_INODE_MODIFIED)) ||
2422             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2423                 if (info->pass == 1)
2424                         vfsync(vp, info->waitfor, 1, NULL, NULL);
2425                 else
2426                         bio_track_wait(&vp->v_track_write, 0, 0);
2427         }
2428         if (info->pass == 2 && (vp->v_flag & VISDIRTY)) {
2429                 if (vx_get_nonblock(vp) == 0) {
2430                         hammer2_inode_lock(ip, 0);
2431                         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2432                             RB_EMPTY(&vp->v_rbdirty_tree) &&
2433                             !bio_track_active(&vp->v_track_write)) {
2434                                 vclrisdirty(vp);
2435                         }
2436                         hammer2_inode_chain_sync(ip);
2437                         hammer2_inode_chain_flush(ip);
2438                         hammer2_inode_unlock(ip);
2439                         vx_put(vp);
2440                 } else {
2441                         /* can't safely clear isdirty */
2442                         hammer2_inode_lock(ip, 0);
2443                         hammer2_inode_chain_flush(ip);
2444                         hammer2_inode_unlock(ip);
2445                 }
2446         }
2447         hammer2_inode_drop(ip);
2448 #if 1
2449         error = 0;
2450         if (error)
2451                 info->error = error;
2452 #endif
2453         return(0);
2454 }
2455
2456 static
2457 int
2458 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2459 {
2460         hammer2_inode_t *ip;
2461
2462         KKASSERT(MAXFIDSZ >= 16);
2463         ip = VTOI(vp);
2464         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2465         fhp->fid_ext = 0;
2466         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2467         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2468
2469         return 0;
2470 }
2471
2472 static
2473 int
2474 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2475                struct fid *fhp, struct vnode **vpp)
2476 {
2477         hammer2_pfs_t *pmp;
2478         hammer2_tid_t inum;
2479         int error;
2480
2481         pmp = MPTOPMP(mp);
2482         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2483         if (vpp) {
2484                 if (inum == 1)
2485                         error = hammer2_vfs_root(mp, vpp);
2486                 else
2487                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2488         } else {
2489                 error = 0;
2490         }
2491         if (error)
2492                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2493         return error;
2494 }
2495
2496 static
2497 int
2498 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2499                  int *exflagsp, struct ucred **credanonp)
2500 {
2501         hammer2_pfs_t *pmp;
2502         struct netcred *np;
2503         int error;
2504
2505         pmp = MPTOPMP(mp);
2506         np = vfs_export_lookup(mp, &pmp->export, nam);
2507         if (np) {
2508                 *exflagsp = np->netc_exflags;
2509                 *credanonp = &np->netc_anon;
2510                 error = 0;
2511         } else {
2512                 error = EACCES;
2513         }
2514         return error;
2515 }
2516
2517 /*
2518  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2519  * header into the HMP
2520  *
2521  * XXX read four volhdrs and use the one with the highest TID whos CRC
2522  *     matches.
2523  *
2524  * XXX check iCRCs.
2525  *
2526  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2527  *     nonexistant locations.
2528  *
2529  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2530  */
2531 static
2532 int
2533 hammer2_install_volume_header(hammer2_dev_t *hmp)
2534 {
2535         hammer2_volume_data_t *vd;
2536         struct buf *bp;
2537         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2538         int error_reported;
2539         int error;
2540         int valid;
2541         int i;
2542
2543         error_reported = 0;
2544         error = 0;
2545         valid = 0;
2546         bp = NULL;
2547
2548         /*
2549          * There are up to 4 copies of the volume header (syncs iterate
2550          * between them so there is no single master).  We don't trust the
2551          * volu_size field so we don't know precisely how large the filesystem
2552          * is, so depend on the OS to return an error if we go beyond the
2553          * block device's EOF.
2554          */
2555         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2556                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2557                               HAMMER2_VOLUME_BYTES, &bp);
2558                 if (error) {
2559                         brelse(bp);
2560                         bp = NULL;
2561                         continue;
2562                 }
2563
2564                 vd = (struct hammer2_volume_data *) bp->b_data;
2565                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2566                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2567                         brelse(bp);
2568                         bp = NULL;
2569                         continue;
2570                 }
2571
2572                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2573                         /* XXX: Reversed-endianness filesystem */
2574                         kprintf("hammer2: reverse-endian filesystem detected");
2575                         brelse(bp);
2576                         bp = NULL;
2577                         continue;
2578                 }
2579
2580                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2581                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2582                                       HAMMER2_VOLUME_ICRC0_SIZE);
2583                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2584                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2585                                        HAMMER2_VOLUME_ICRC1_SIZE);
2586                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2587                         kprintf("hammer2 volume header crc "
2588                                 "mismatch copy #%d %08x/%08x\n",
2589                                 i, crc0, crc);
2590                         error_reported = 1;
2591                         brelse(bp);
2592                         bp = NULL;
2593                         continue;
2594                 }
2595                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2596                         valid = 1;
2597                         hmp->voldata = *vd;
2598                         hmp->volhdrno = i;
2599                 }
2600                 brelse(bp);
2601                 bp = NULL;
2602         }
2603         if (valid) {
2604                 hmp->volsync = hmp->voldata;
2605                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2606                 error = 0;
2607                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2608                         kprintf("hammer2: using volume header #%d\n",
2609                                 hmp->volhdrno);
2610                 }
2611         } else {
2612                 error = EINVAL;
2613                 kprintf("hammer2: no valid volume headers found!\n");
2614         }
2615         return (error);
2616 }
2617
2618 /*
2619  * This handles hysteresis on regular file flushes.  Because the BIOs are
2620  * routed to a thread it is possible for an excessive number to build up
2621  * and cause long front-end stalls long before the runningbuffspace limit
2622  * is hit, so we implement hammer2_flush_pipe to control the
2623  * hysteresis.
2624  *
2625  * This is a particular problem when compression is used.
2626  */
2627 void
2628 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2629 {
2630         atomic_add_int(&pmp->count_lwinprog, 1);
2631 }
2632
2633 void
2634 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2635 {
2636         int lwinprog;
2637
2638         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2639         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2640             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2641                 atomic_clear_int(&pmp->count_lwinprog,
2642                                  HAMMER2_LWINPROG_WAITING);
2643                 wakeup(&pmp->count_lwinprog);
2644         }
2645         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2646             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2647                 atomic_clear_int(&pmp->count_lwinprog,
2648                                  HAMMER2_LWINPROG_WAITING0);
2649                 wakeup(&pmp->count_lwinprog);
2650         }
2651 }
2652
2653 void
2654 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2655 {
2656         int lwinprog;
2657         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2658                                     HAMMER2_LWINPROG_WAITING0;
2659
2660         for (;;) {
2661                 lwinprog = pmp->count_lwinprog;
2662                 cpu_ccfence();
2663                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2664                         break;
2665                 tsleep_interlock(&pmp->count_lwinprog, 0);
2666                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2667                 lwinprog = pmp->count_lwinprog;
2668                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2669                         break;
2670                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2671         }
2672 }
2673
2674 /*
2675  * Manage excessive memory resource use for chain and related
2676  * structures.
2677  */
2678 void
2679 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2680 {
2681         uint32_t waiting;
2682         uint32_t count;
2683         uint32_t limit;
2684 #if 0
2685         static int zzticks;
2686 #endif
2687
2688         /*
2689          * Atomic check condition and wait.  Also do an early speedup of
2690          * the syncer to try to avoid hitting the wait.
2691          */
2692         for (;;) {
2693                 waiting = pmp->inmem_dirty_chains;
2694                 cpu_ccfence();
2695                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2696
2697                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2698                 if (limit < hammer2_limit_dirty_chains)
2699                         limit = hammer2_limit_dirty_chains;
2700                 if (limit < 1000)
2701                         limit = 1000;
2702
2703 #if 0
2704                 if ((int)(ticks - zzticks) > hz) {
2705                         zzticks = ticks;
2706                         kprintf("count %ld %ld\n", count, limit);
2707                 }
2708 #endif
2709
2710                 /*
2711                  * Block if there are too many dirty chains present, wait
2712                  * for the flush to clean some out.
2713                  */
2714                 if (count > limit) {
2715                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2716                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2717                                                waiting,
2718                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2719                                 speedup_syncer(pmp->mp);
2720                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2721                                        "chnmem", hz);
2722                         }
2723                         continue;       /* loop on success or fail */
2724                 }
2725
2726                 /*
2727                  * Try to start an early flush before we are forced to block.
2728                  */
2729                 if (count > limit * 5 / 10)
2730                         speedup_syncer(pmp->mp);
2731                 break;
2732         }
2733 }
2734
2735 void
2736 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2737 {
2738         if (pmp) {
2739                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2740         }
2741 }
2742
2743 void
2744 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2745 {
2746         uint32_t waiting;
2747
2748         if (pmp) {
2749                 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
2750                 /* don't need --waiting to test flag */
2751                 if (waiting & HAMMER2_DIRTYCHAIN_WAITING) {
2752                         atomic_clear_int(&pmp->inmem_dirty_chains,
2753                                          HAMMER2_DIRTYCHAIN_WAITING);
2754                         wakeup(&pmp->inmem_dirty_chains);
2755                 }
2756         }
2757 }
2758
2759 /*
2760  * Returns 0 if the filesystem has tons of free space
2761  * Returns 1 if the filesystem has less than 10% remaining
2762  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2763  */
2764 int
2765 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2766 {
2767         hammer2_pfs_t *pmp;
2768         hammer2_dev_t *hmp;
2769         hammer2_off_t free_reserved;
2770         hammer2_off_t free_nominal;
2771         int i;
2772
2773         pmp = ip->pmp;
2774
2775         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2776                 free_reserved = HAMMER2_SEGSIZE;
2777                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2778                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2779                         hmp = pmp->pfs_hmps[i];
2780                         if (hmp == NULL)
2781                                 continue;
2782                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2783                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2784                                 continue;
2785
2786                         if (free_nominal > hmp->voldata.allocator_free)
2787                                 free_nominal = hmp->voldata.allocator_free;
2788                         if (free_reserved < hmp->free_reserved)
2789                                 free_reserved = hmp->free_reserved;
2790                 }
2791
2792                 /*
2793                  * SMP races ok
2794                  */
2795                 pmp->free_reserved = free_reserved;
2796                 pmp->free_nominal = free_nominal;
2797                 pmp->free_ticks = ticks;
2798         } else {
2799                 free_reserved = pmp->free_reserved;
2800                 free_nominal = pmp->free_nominal;
2801         }
2802         if (cred && cred->cr_uid != 0) {
2803                 if ((int64_t)(free_nominal - bytes) <
2804                     (int64_t)free_reserved) {
2805                         return 2;
2806                 }
2807         } else {
2808                 if ((int64_t)(free_nominal - bytes) <
2809                     (int64_t)free_reserved / 2) {
2810                         return 2;
2811                 }
2812         }
2813         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2814                 return 1;
2815         return 0;
2816 }
2817
2818 /*
2819  * Debugging
2820  */
2821 void
2822 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
2823                    u_int flags)
2824 {
2825         hammer2_chain_t *scan;
2826         hammer2_chain_t *parent;
2827
2828         --*countp;
2829         if (*countp == 0) {
2830                 kprintf("%*.*s...\n", tab, tab, "");
2831                 return;
2832         }
2833         if (*countp < 0)
2834                 return;
2835         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2836                 tab, tab, "", pfx,
2837                 chain, chain->bref.type,
2838                 chain->bref.key, chain->bref.keybits,
2839                 chain->bref.mirror_tid);
2840
2841         kprintf("%*.*s      [%08x] (%s) refs=%d",
2842                 tab, tab, "",
2843                 chain->flags,
2844                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2845                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2846                 chain->refs);
2847
2848         parent = chain->parent;
2849         if (parent)
2850                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2851                         tab, tab, "",
2852                         parent, parent->flags, parent->refs);
2853         if (RB_EMPTY(&chain->core.rbtree)) {
2854                 kprintf("\n");
2855         } else {
2856                 kprintf(" {\n");
2857                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
2858                         if ((scan->flags & flags) || flags == (u_int)-1) {
2859                                 hammer2_dump_chain(scan, tab + 4, countp, 'a',
2860                                                    flags);
2861                         }
2862                 }
2863                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2864                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2865                                 chain->data->ipdata.filename);
2866                 else
2867                         kprintf("%*.*s}\n", tab, tab, "");
2868         }
2869 }