sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2018 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         hammer2_spin_ex(&pmp->lru_spin);
 672         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 673                 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
 674                 atomic_add_int(&pmp->lru_count, -1);
 675                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
 676                 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
 677                 hammer2_chain_ref(chain);
 678                 hammer2_spin_unex(&pmp->lru_spin);
 679                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 680                 hammer2_chain_drop(chain);
 681                 hammer2_spin_ex(&pmp->lru_spin);
 682         }
 683         hammer2_spin_unex(&pmp->lru_spin);
 684
 685         /*
 686          * Free remaining pmp resources
 687          */
 688         kmalloc_destroy(&pmp->mmsg);
 689         kmalloc_destroy(&pmp->minode);
 690
 691         kfree(pmp, M_HAMMER2);
 692 }
 693
 694 /*
 695  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 696  * empty is terminated and freed.
 697  *
 698  * XXX inefficient.
 699  */
 700 static void
 701 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 702 {
 703         hammer2_pfs_t *pmp;
 704         hammer2_inode_t *iroot;
 705         hammer2_chain_t *rchain;
 706         int didfreeze;
 707         int i;
 708         int j;
 709
 710 again:
 711         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 712                 if ((iroot = pmp->iroot) == NULL)
 713                         continue;
 714                 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
 715                 hammer2_inode_run_sideq(pmp, 1);
 716                 hammer2_bioq_sync(pmp);
 717                 hammer2_trans_done(pmp);
 718                 if (hmp->spmp == pmp) {
 719                         hmp->spmp = NULL;
 720                         hmp->vchain.pmp = NULL;
 721                         hmp->fchain.pmp = NULL;
 722                 }
 723
 724                 /*
 725                  * Determine if this PFS is affected.  If it is we must
 726                  * freeze all management threads and lock its iroot.
 727                  *
 728                  * Freezing a management thread forces it idle, operations
 729                  * in-progress will be aborted and it will have to start
 730                  * over again when unfrozen, or exit if told to exit.
 731                  */
 732                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 733                         if (pmp->pfs_hmps[i] == hmp)
 734                                 break;
 735                 }
 736                 if (i != HAMMER2_MAXCLUSTER) {
 737                         /*
 738                          * Make sure all synchronization threads are locked
 739                          * down.
 740                          */
 741                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 742                                 if (pmp->pfs_hmps[i] == NULL)
 743                                         continue;
 744                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 745                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 746                                         hammer2_thr_freeze_async(
 747                                                 &pmp->xop_groups[j].thrs[i]);
 748                                 }
 749                         }
 750                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 751                                 if (pmp->pfs_hmps[i] == NULL)
 752                                         continue;
 753                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 754                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 755                                         hammer2_thr_freeze(
 756                                                 &pmp->xop_groups[j].thrs[i]);
 757                                 }
 758                         }
 759
 760                         /*
 761                          * Lock the inode and clean out matching chains.
 762                          * Note that we cannot use hammer2_inode_lock_*()
 763                          * here because that would attempt to validate the
 764                          * cluster that we are in the middle of ripping
 765                          * apart.
 766                          *
 767                          * WARNING! We are working directly on the inodes
 768                          *          embedded cluster.
 769                          */
 770                         hammer2_mtx_ex(&iroot->lock);
 771
 772                         /*
 773                          * Remove the chain from matching elements of the PFS.
 774                          */
 775                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 776                                 if (pmp->pfs_hmps[i] != hmp)
 777                                         continue;
 778                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 779                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 780                                         hammer2_thr_delete(
 781                                                 &pmp->xop_groups[j].thrs[i]);
 782                                 }
 783                                 rchain = iroot->cluster.array[i].chain;
 784                                 iroot->cluster.array[i].chain = NULL;
 785                                 pmp->pfs_types[i] = 0;
 786                                 if (pmp->pfs_names[i]) {
 787                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 788                                         pmp->pfs_names[i] = NULL;
 789                                 }
 790                                 if (rchain) {
 791                                         hammer2_chain_drop(rchain);
 792                                         /* focus hint */
 793                                         if (iroot->cluster.focus == rchain)
 794                                                 iroot->cluster.focus = NULL;
 795                                 }
 796                                 pmp->pfs_hmps[i] = NULL;
 797                         }
 798                         hammer2_mtx_unlock(&iroot->lock);
 799                         didfreeze = 1;  /* remaster, unfreeze down below */
 800                 } else {
 801                         didfreeze = 0;
 802                 }
 803
 804                 /*
 805                  * Cleanup trailing chains.  Gaps may remain.
 806                  */
 807                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 808                         if (pmp->pfs_hmps[i])
 809                                 break;
 810                 }
 811                 iroot->cluster.nchains = i + 1;
 812
 813                 /*
 814                  * If the PMP has no elements remaining we can destroy it.
 815                  * (this will transition management threads from frozen->exit).
 816                  */
 817                 if (iroot->cluster.nchains == 0) {
 818                         hammer2_pfsfree(pmp);
 819                         goto again;
 820                 }
 821
 822                 /*
 823                  * If elements still remain we need to set the REMASTER
 824                  * flag and unfreeze it.
 825                  */
 826                 if (didfreeze) {
 827                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 828                                 if (pmp->pfs_hmps[i] == NULL)
 829                                         continue;
 830                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 831                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 832                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 833                                         hammer2_thr_remaster(
 834                                                 &pmp->xop_groups[j].thrs[i]);
 835                                         hammer2_thr_unfreeze(
 836                                                 &pmp->xop_groups[j].thrs[i]);
 837                                 }
 838                         }
 839                 }
 840         }
 841 }
 842
 843 /*
 844  * Mount or remount HAMMER2 fileystem from physical media
 845  *
 846  *      mountroot
 847  *              mp              mount point structure
 848  *              path            NULL
 849  *              data            <unused>
 850  *              cred            <unused>
 851  *
 852  *      mount
 853  *              mp              mount point structure
 854  *              path            path to mount point
 855  *              data            pointer to argument structure in user space
 856  *                      volume  volume path (device@LABEL form)
 857  *                      hflags  user mount flags
 858  *              cred            user credentials
 859  *
 860  * RETURNS:     0       Success
 861  *              !0      error number
 862  */
 863 static
 864 int
 865 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 866                   struct ucred *cred)
 867 {
 868         struct hammer2_mount_info info;
 869         hammer2_pfs_t *pmp;
 870         hammer2_pfs_t *spmp;
 871         hammer2_dev_t *hmp;
 872         hammer2_dev_t *force_local;
 873         hammer2_key_t key_next;
 874         hammer2_key_t key_dummy;
 875         hammer2_key_t lhc;
 876         struct vnode *devvp;
 877         struct nlookupdata nd;
 878         hammer2_chain_t *parent;
 879         hammer2_chain_t *chain;
 880         hammer2_cluster_t *cluster;
 881         const hammer2_inode_data_t *ripdata;
 882         hammer2_blockref_t bref;
 883         struct file *fp;
 884         char devstr[MNAMELEN];
 885         size_t size;
 886         size_t done;
 887         char *dev;
 888         char *label;
 889         int ronly = 1;
 890         int error;
 891         int i;
 892
 893         hmp = NULL;
 894         pmp = NULL;
 895         dev = NULL;
 896         label = NULL;
 897         devvp = NULL;
 898
 899         kprintf("hammer2_mount\n");
 900
 901         if (path == NULL) {
 902                 /*
 903                  * Root mount
 904                  */
 905                 bzero(&info, sizeof(info));
 906                 info.cluster_fd = -1;
 907                 ksnprintf(devstr, sizeof(devstr), "%s",
 908                           mp->mnt_stat.f_mntfromname);
 909                 kprintf("hammer2_mount: root '%s'\n", devstr);
 910         } else {
 911                 /*
 912                  * Non-root mount or updating a mount
 913                  */
 914                 error = copyin(data, &info, sizeof(info));
 915                 if (error)
 916                         return (error);
 917
 918                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 919                 if (error)
 920                         return (error);
 921         }
 922
 923         /*
 924          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 925          * if no label specified, based on the partition id.  Error out if no
 926          * label or device (with partition id) is specified.  This is strictly
 927          * a convenience to match the default label created by newfs_hammer2,
 928          * our preference is that a label always be specified.
 929          *
 930          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
 931          *       that does not specify a device, as long as some H2 label
 932          *       has already been mounted from that device.  This makes
 933          *       mounting snapshots a lot easier.
 934          */
 935         dev = devstr;
 936         label = strchr(devstr, '@');
 937         if (label && ((label + 1) - dev) > done)
 938                 return (EINVAL);
 939         if (label == NULL || label[1] == 0) {
 940                 char slice;
 941
 942                 if (label == NULL)
 943                         label = devstr + strlen(devstr);
 944                 slice = label[-1];
 945                 switch(slice) {
 946                 case 'a':
 947                         label = "BOOT";
 948                         break;
 949                 case 'd':
 950                         label = "ROOT";
 951                         break;
 952                 default:
 953                         label = "DATA";
 954                         break;
 955                 }
 956         } else {
 957                 *label = '\0';
 958                 label++;
 959         }
 960
 961         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
 962                 dev, label, (mp->mnt_flag & MNT_RDONLY));
 963
 964         if (mp->mnt_flag & MNT_UPDATE) {
 965                 /*
 966                  * Update mount.  Note that pmp->iroot->cluster is
 967                  * an inode-embedded cluster and thus cannot be
 968                  * directly locked.
 969                  *
 970                  * XXX HAMMER2 needs to implement NFS export via
 971                  *     mountctl.
 972                  */
 973                 pmp = MPTOPMP(mp);
 974                 pmp->hflags = info.hflags;
 975                 cluster = &pmp->iroot->cluster;
 976                 for (i = 0; i < cluster->nchains; ++i) {
 977                         if (cluster->array[i].chain == NULL)
 978                                 continue;
 979                         hmp = cluster->array[i].chain->hmp;
 980                         devvp = hmp->devvp;
 981                         error = hammer2_remount(hmp, mp, path,
 982                                                 devvp, cred);
 983                         if (error)
 984                                 break;
 985                 }
 986
 987                 return error;
 988         }
 989
 990         /*
 991          * HMP device mount
 992          *
 993          * If a path is specified and dev is not an empty string, lookup the
 994          * name and verify that it referes to a block device.
 995          *
 996          * If a path is specified and dev is an empty string we fall through
 997          * and locate the label in the hmp search.
 998          */
 999         if (path && *dev != 0) {
1000                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1001                 if (error == 0)
1002                         error = nlookup(&nd);
1003                 if (error == 0)
1004                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1005                 nlookup_done(&nd);
1006         } else if (path == NULL) {
1007                 /* root mount */
1008                 cdev_t cdev = kgetdiskbyname(dev);
1009                 error = bdevvp(cdev, &devvp);
1010                 if (error)
1011                         kprintf("hammer2: cannot find '%s'\n", dev);
1012         } else {
1013                 /*
1014                  * We will locate the hmp using the label in the hmp loop.
1015                  */
1016                 error = 0;
1017         }
1018
1019         /*
1020          * Make sure its a block device.  Do not check to see if it is
1021          * already mounted until we determine that its a fresh H2 device.
1022          */
1023         if (error == 0 && devvp) {
1024                 vn_isdisk(devvp, &error);
1025         }
1026
1027         /*
1028          * Determine if the device has already been mounted.  After this
1029          * check hmp will be non-NULL if we are doing the second or more
1030          * hammer2 mounts from the same device.
1031          */
1032         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1033         if (devvp) {
1034                 /*
1035                  * Match the device.  Due to the way devfs works,
1036                  * we may not be able to directly match the vnode pointer,
1037                  * so also check to see if the underlying device matches.
1038                  */
1039                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1040                         if (hmp->devvp == devvp)
1041                                 break;
1042                         if (devvp->v_rdev &&
1043                             hmp->devvp->v_rdev == devvp->v_rdev) {
1044                                 break;
1045                         }
1046                 }
1047
1048                 /*
1049                  * If no match this may be a fresh H2 mount, make sure
1050                  * the device is not mounted on anything else.
1051                  */
1052                 if (hmp == NULL)
1053                         error = vfs_mountedon(devvp);
1054         } else if (error == 0) {
1055                 /*
1056                  * Match the label to a pmp already probed.
1057                  */
1058                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1059                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1060                                 if (pmp->pfs_names[i] &&
1061                                     strcmp(pmp->pfs_names[i], label) == 0) {
1062                                         hmp = pmp->pfs_hmps[i];
1063                                         break;
1064                                 }
1065                         }
1066                         if (hmp)
1067                                 break;
1068                 }
1069                 if (hmp == NULL)
1070                         error = ENOENT;
1071         }
1072
1073         /*
1074          * Open the device if this isn't a secondary mount and construct
1075          * the H2 device mount (hmp).
1076          */
1077         if (hmp == NULL) {
1078                 hammer2_chain_t *schain;
1079                 hammer2_xid_t xid;
1080
1081                 if (error == 0 && vcount(devvp) > 0) {
1082                         kprintf("Primary device already has references\n");
1083                         error = EBUSY;
1084                 }
1085
1086                 /*
1087                  * Now open the device
1088                  */
1089                 if (error == 0) {
1090                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1091                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1092                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1093                         if (error == 0) {
1094                                 error = VOP_OPEN(devvp,
1095                                              (ronly ? FREAD : FREAD | FWRITE),
1096                                              FSCRED, NULL);
1097                         }
1098                         vn_unlock(devvp);
1099                 }
1100                 if (error && devvp) {
1101                         vrele(devvp);
1102                         devvp = NULL;
1103                 }
1104                 if (error) {
1105                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1106                         return error;
1107                 }
1108                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1109                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1110                 hmp->ronly = ronly;
1111                 hmp->devvp = devvp;
1112                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1113                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1114                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1115                 RB_INIT(&hmp->iotree);
1116                 spin_init(&hmp->io_spin, "hm2mount_io");
1117                 spin_init(&hmp->list_spin, "hm2mount_list");
1118                 TAILQ_INIT(&hmp->flushq);
1119
1120                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1121                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1122                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1123
1124                 /*
1125                  * vchain setup. vchain.data is embedded.
1126                  * vchain.refs is initialized and will never drop to 0.
1127                  *
1128                  * NOTE! voldata is not yet loaded.
1129                  */
1130                 hmp->vchain.hmp = hmp;
1131                 hmp->vchain.refs = 1;
1132                 hmp->vchain.data = (void *)&hmp->voldata;
1133                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1134                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1135                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1136
1137                 hammer2_chain_core_init(&hmp->vchain);
1138                 /* hmp->vchain.u.xxx is left NULL */
1139
1140                 /*
1141                  * fchain setup.  fchain.data is embedded.
1142                  * fchain.refs is initialized and will never drop to 0.
1143                  *
1144                  * The data is not used but needs to be initialized to
1145                  * pass assertion muster.  We use this chain primarily
1146                  * as a placeholder for the freemap's top-level RBTREE
1147                  * so it does not interfere with the volume's topology
1148                  * RBTREE.
1149                  */
1150                 hmp->fchain.hmp = hmp;
1151                 hmp->fchain.refs = 1;
1152                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1153                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1154                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1155                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1156                 hmp->fchain.bref.methods =
1157                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1158                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1159
1160                 hammer2_chain_core_init(&hmp->fchain);
1161                 /* hmp->fchain.u.xxx is left NULL */
1162
1163                 /*
1164                  * Install the volume header and initialize fields from
1165                  * voldata.
1166                  */
1167                 error = hammer2_install_volume_header(hmp);
1168                 if (error) {
1169                         hammer2_unmount_helper(mp, NULL, hmp);
1170                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1171                         hammer2_vfs_unmount(mp, MNT_FORCE);
1172                         return error;
1173                 }
1174
1175                 /*
1176                  * Really important to get these right or flush will get
1177                  * confused.
1178                  */
1179                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1180                 kprintf("alloc spmp %p tid %016jx\n",
1181                         hmp->spmp, hmp->voldata.mirror_tid);
1182                 spmp = hmp->spmp;
1183
1184                 /*
1185                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1186                  * is inherited from the volume header.
1187                  */
1188                 xid = 0;
1189                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1190                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1191                 hmp->vchain.pmp = spmp;
1192                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1193                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1194                 hmp->fchain.pmp = spmp;
1195
1196                 /*
1197                  * First locate the super-root inode, which is key 0
1198                  * relative to the volume header's blockset.
1199                  *
1200                  * Then locate the root inode by scanning the directory keyspace
1201                  * represented by the label.
1202                  */
1203                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1204                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1205                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1206                                       &error, 0);
1207                 hammer2_chain_lookup_done(parent);
1208                 if (schain == NULL) {
1209                         kprintf("hammer2_mount: invalid super-root\n");
1210                         hammer2_unmount_helper(mp, NULL, hmp);
1211                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1212                         hammer2_vfs_unmount(mp, MNT_FORCE);
1213                         return EINVAL;
1214                 }
1215                 if (schain->error) {
1216                         kprintf("hammer2_mount: error %s reading super-root\n",
1217                                 hammer2_error_str(schain->error));
1218                         hammer2_chain_unlock(schain);
1219                         hammer2_chain_drop(schain);
1220                         schain = NULL;
1221                         hammer2_unmount_helper(mp, NULL, hmp);
1222                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1223                         hammer2_vfs_unmount(mp, MNT_FORCE);
1224                         return EINVAL;
1225                 }
1226
1227                 /*
1228                  * The super-root always uses an inode_tid of 1 when
1229                  * creating PFSs.
1230                  */
1231                 spmp->inode_tid = 1;
1232                 spmp->modify_tid = schain->bref.modify_tid + 1;
1233
1234                 /*
1235                  * Sanity-check schain's pmp and finish initialization.
1236                  * Any chain belonging to the super-root topology should
1237                  * have a NULL pmp (not even set to spmp).
1238                  */
1239                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1240                 KKASSERT(schain->pmp == NULL);
1241                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1242
1243                 /*
1244                  * Replace the dummy spmp->iroot with a real one.  It's
1245                  * easier to just do a wholesale replacement than to try
1246                  * to update the chain and fixup the iroot fields.
1247                  *
1248                  * The returned inode is locked with the supplied cluster.
1249                  */
1250                 cluster = hammer2_cluster_from_chain(schain);
1251                 hammer2_inode_drop(spmp->iroot);
1252                 spmp->iroot = NULL;
1253                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1254                 spmp->spmp_hmp = hmp;
1255                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1256                 spmp->pfs_hmps[0] = hmp;
1257                 hammer2_inode_ref(spmp->iroot);
1258                 hammer2_inode_unlock(spmp->iroot);
1259                 hammer2_cluster_unlock(cluster);
1260                 hammer2_cluster_drop(cluster);
1261                 schain = NULL;
1262                 /* leave spmp->iroot with one ref */
1263
1264                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1265                         error = hammer2_recovery(hmp);
1266                         /* XXX do something with error */
1267                 }
1268                 hammer2_update_pmps(hmp);
1269                 hammer2_iocom_init(hmp);
1270                 hammer2_bulkfree_init(hmp);
1271
1272                 /*
1273                  * Ref the cluster management messaging descriptor.  The mount
1274                  * program deals with the other end of the communications pipe.
1275                  *
1276                  * Root mounts typically do not supply one.
1277                  */
1278                 if (info.cluster_fd >= 0) {
1279                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1280                         if (fp) {
1281                                 hammer2_cluster_reconnect(hmp, fp);
1282                         } else {
1283                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1284                         }
1285                 }
1286         } else {
1287                 spmp = hmp->spmp;
1288                 if (info.hflags & HMNT2_DEVFLAGS) {
1289                         kprintf("hammer2: Warning: mount flags pertaining "
1290                                 "to the whole device may only be specified "
1291                                 "on the first mount of the device: %08x\n",
1292                                 info.hflags & HMNT2_DEVFLAGS);
1293                 }
1294         }
1295
1296         /*
1297          * Force local mount (disassociate all PFSs from their clusters).
1298          * Used primarily for debugging.
1299          */
1300         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1301
1302         /*
1303          * Lookup the mount point under the media-localized super-root.
1304          * Scanning hammer2_pfslist doesn't help us because it represents
1305          * PFS cluster ids which can aggregate several named PFSs together.
1306          *
1307          * cluster->pmp will incorrectly point to spmp and must be fixed
1308          * up later on.
1309          */
1310         hammer2_inode_lock(spmp->iroot, 0);
1311         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1312         lhc = hammer2_dirhash(label, strlen(label));
1313         chain = hammer2_chain_lookup(&parent, &key_next,
1314                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1315                                      &error, 0);
1316         while (chain) {
1317                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1318                     strcmp(label, chain->data->ipdata.filename) == 0) {
1319                         break;
1320                 }
1321                 chain = hammer2_chain_next(&parent, chain, &key_next,
1322                                             key_next,
1323                                             lhc + HAMMER2_DIRHASH_LOMASK,
1324                                             &error, 0);
1325         }
1326         if (parent) {
1327                 hammer2_chain_unlock(parent);
1328                 hammer2_chain_drop(parent);
1329         }
1330         hammer2_inode_unlock(spmp->iroot);
1331
1332         /*
1333          * PFS could not be found?
1334          */
1335         if (chain == NULL) {
1336                 if (error)
1337                         kprintf("hammer2_mount: PFS label I/O error\n");
1338                 else
1339                         kprintf("hammer2_mount: PFS label not found\n");
1340                 hammer2_unmount_helper(mp, NULL, hmp);
1341                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1342                 hammer2_vfs_unmount(mp, MNT_FORCE);
1343
1344                 return EINVAL;
1345         }
1346
1347         /*
1348          * Acquire the pmp structure (it should have already been allocated
1349          * via hammer2_update_pmps() so do not pass cluster in to add to
1350          * available chains).
1351          *
1352          * Check if the cluster has already been mounted.  A cluster can
1353          * only be mounted once, use null mounts to mount additional copies.
1354          */
1355         if (chain->error) {
1356                 kprintf("hammer2_mount: PFS label I/O error\n");
1357         } else {
1358                 ripdata = &chain->data->ipdata;
1359                 bref = chain->bref;
1360                 pmp = hammer2_pfsalloc(NULL, ripdata,
1361                                        bref.modify_tid, force_local);
1362         }
1363         hammer2_chain_unlock(chain);
1364         hammer2_chain_drop(chain);
1365
1366         /*
1367          * Finish the mount
1368          */
1369         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1370
1371         if (pmp->mp) {
1372                 kprintf("hammer2_mount: PFS already mounted!\n");
1373                 hammer2_unmount_helper(mp, NULL, hmp);
1374                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1375                 hammer2_vfs_unmount(mp, MNT_FORCE);
1376
1377                 return EBUSY;
1378         }
1379
1380         pmp->hflags = info.hflags;
1381         mp->mnt_flag |= MNT_LOCAL;
1382         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1383         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1384
1385         /*
1386          * required mount structure initializations
1387          */
1388         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1389         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1390
1391         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1392         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1393
1394         /*
1395          * Optional fields
1396          */
1397         mp->mnt_iosize_max = MAXPHYS;
1398
1399         /*
1400          * Connect up mount pointers.
1401          */
1402         hammer2_mount_helper(mp, pmp);
1403
1404         lockmgr(&hammer2_mntlk, LK_RELEASE);
1405
1406         /*
1407          * Finish setup
1408          */
1409         vfs_getnewfsid(mp);
1410         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1411         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1412         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1413
1414         if (path) {
1415                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1416                           MNAMELEN - 1, &size);
1417                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1418         } /* else root mount, already in there */
1419
1420         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1421         if (path) {
1422                 copyinstr(path, mp->mnt_stat.f_mntonname,
1423                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1424                           &size);
1425         } else {
1426                 /* root mount */
1427                 mp->mnt_stat.f_mntonname[0] = '/';
1428         }
1429
1430         /*
1431          * Initial statfs to prime mnt_stat.
1432          */
1433         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1434
1435         return 0;
1436 }
1437
1438 /*
1439  * Scan PFSs under the super-root and create hammer2_pfs structures.
1440  */
1441 static
1442 void
1443 hammer2_update_pmps(hammer2_dev_t *hmp)
1444 {
1445         const hammer2_inode_data_t *ripdata;
1446         hammer2_chain_t *parent;
1447         hammer2_chain_t *chain;
1448         hammer2_blockref_t bref;
1449         hammer2_dev_t *force_local;
1450         hammer2_pfs_t *spmp;
1451         hammer2_pfs_t *pmp;
1452         hammer2_key_t key_next;
1453         int error;
1454
1455         /*
1456          * Force local mount (disassociate all PFSs from their clusters).
1457          * Used primarily for debugging.
1458          */
1459         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1460
1461         /*
1462          * Lookup mount point under the media-localized super-root.
1463          *
1464          * cluster->pmp will incorrectly point to spmp and must be fixed
1465          * up later on.
1466          */
1467         spmp = hmp->spmp;
1468         hammer2_inode_lock(spmp->iroot, 0);
1469         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1470         chain = hammer2_chain_lookup(&parent, &key_next,
1471                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1472                                          &error, 0);
1473         while (chain) {
1474                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1475                         continue;
1476                 if (chain->error) {
1477                         kprintf("I/O error scanning PFS labels\n");
1478                 } else {
1479                         ripdata = &chain->data->ipdata;
1480                         bref = chain->bref;
1481
1482                         pmp = hammer2_pfsalloc(chain, ripdata,
1483                                                bref.modify_tid, force_local);
1484                 }
1485                 chain = hammer2_chain_next(&parent, chain, &key_next,
1486                                            key_next, HAMMER2_KEY_MAX,
1487                                            &error, 0);
1488         }
1489         if (parent) {
1490                 hammer2_chain_unlock(parent);
1491                 hammer2_chain_drop(parent);
1492         }
1493         hammer2_inode_unlock(spmp->iroot);
1494 }
1495
1496 static
1497 int
1498 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1499                 struct vnode *devvp, struct ucred *cred)
1500 {
1501         int error;
1502
1503         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1504                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1505                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1506                 vn_unlock(devvp);
1507                 error = hammer2_recovery(hmp);
1508                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1509                 if (error == 0) {
1510                         VOP_CLOSE(devvp, FREAD, NULL);
1511                         hmp->ronly = 0;
1512                 } else {
1513                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1514                 }
1515                 vn_unlock(devvp);
1516         } else {
1517                 error = 0;
1518         }
1519         return error;
1520 }
1521
1522 static
1523 int
1524 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1525 {
1526         hammer2_pfs_t *pmp;
1527         int flags;
1528         int error = 0;
1529
1530         pmp = MPTOPMP(mp);
1531
1532         if (pmp == NULL)
1533                 return(0);
1534
1535         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1536
1537         /*
1538          * If mount initialization proceeded far enough we must flush
1539          * its vnodes and sync the underlying mount points.  Three syncs
1540          * are required to fully flush the filesystem (freemap updates lag
1541          * by one flush, and one extra for safety).
1542          */
1543         if (mntflags & MNT_FORCE)
1544                 flags = FORCECLOSE;
1545         else
1546                 flags = 0;
1547         if (pmp->iroot) {
1548                 error = vflush(mp, 0, flags);
1549                 if (error)
1550                         goto failed;
1551                 hammer2_vfs_sync(mp, MNT_WAIT);
1552                 hammer2_vfs_sync(mp, MNT_WAIT);
1553                 hammer2_vfs_sync(mp, MNT_WAIT);
1554         }
1555
1556         /*
1557          * Cleanup the frontend support XOPS threads
1558          */
1559         hammer2_xop_helper_cleanup(pmp);
1560
1561         if (pmp->mp)
1562                 hammer2_unmount_helper(mp, pmp, NULL);
1563
1564         error = 0;
1565 failed:
1566         lockmgr(&hammer2_mntlk, LK_RELEASE);
1567
1568         return (error);
1569 }
1570
1571 /*
1572  * Mount helper, hook the system mount into our PFS.
1573  * The mount lock is held.
1574  *
1575  * We must bump the mount_count on related devices for any
1576  * mounted PFSs.
1577  */
1578 static
1579 void
1580 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1581 {
1582         hammer2_cluster_t *cluster;
1583         hammer2_chain_t *rchain;
1584         int i;
1585
1586         mp->mnt_data = (qaddr_t)pmp;
1587         pmp->mp = mp;
1588
1589         /*
1590          * After pmp->mp is set we have to adjust hmp->mount_count.
1591          */
1592         cluster = &pmp->iroot->cluster;
1593         for (i = 0; i < cluster->nchains; ++i) {
1594                 rchain = cluster->array[i].chain;
1595                 if (rchain == NULL)
1596                         continue;
1597                 ++rchain->hmp->mount_count;
1598                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1599                         rchain->hmp, rchain->hmp->mount_count);
1600         }
1601
1602         /*
1603          * Create missing Xop threads
1604          */
1605         hammer2_xop_helper_create(pmp);
1606 }
1607
1608 /*
1609  * Mount helper, unhook the system mount from our PFS.
1610  * The mount lock is held.
1611  *
1612  * If hmp is supplied a mount responsible for being the first to open
1613  * the block device failed and the block device and all PFSs using the
1614  * block device must be cleaned up.
1615  *
1616  * If pmp is supplied multiple devices might be backing the PFS and each
1617  * must be disconnected.  This might not be the last PFS using some of the
1618  * underlying devices.  Also, we have to adjust our hmp->mount_count
1619  * accounting for the devices backing the pmp which is now undergoing an
1620  * unmount.
1621  */
1622 static
1623 void
1624 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1625 {
1626         hammer2_cluster_t *cluster;
1627         hammer2_chain_t *rchain;
1628         struct vnode *devvp;
1629         int dumpcnt;
1630         int ronly;
1631         int i;
1632
1633         /*
1634          * If no device supplied this is a high-level unmount and we have to
1635          * to disconnect the mount, adjust mount_count, and locate devices
1636          * that might now have no mounts.
1637          */
1638         if (pmp) {
1639                 KKASSERT(hmp == NULL);
1640                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1641                 pmp->mp = NULL;
1642                 mp->mnt_data = NULL;
1643
1644                 /*
1645                  * After pmp->mp is cleared we have to account for
1646                  * mount_count.
1647                  */
1648                 cluster = &pmp->iroot->cluster;
1649                 for (i = 0; i < cluster->nchains; ++i) {
1650                         rchain = cluster->array[i].chain;
1651                         if (rchain == NULL)
1652                                 continue;
1653                         --rchain->hmp->mount_count;
1654                         /* scrapping hmp now may invalidate the pmp */
1655                 }
1656 again:
1657                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1658                         if (hmp->mount_count == 0) {
1659                                 hammer2_unmount_helper(NULL, NULL, hmp);
1660                                 goto again;
1661                         }
1662                 }
1663                 return;
1664         }
1665
1666         /*
1667          * Try to terminate the block device.  We can't terminate it if
1668          * there are still PFSs referencing it.
1669          */
1670         if (hmp->mount_count)
1671                 return;
1672
1673         /*
1674          * Decomission the network before we start messing with the
1675          * device and PFS.
1676          */
1677         hammer2_iocom_uninit(hmp);
1678
1679         hammer2_bulkfree_uninit(hmp);
1680         hammer2_pfsfree_scan(hmp);
1681         hammer2_dev_exlock(hmp);        /* XXX order */
1682
1683         /*
1684          * Cycle the volume data lock as a safety (probably not needed any
1685          * more).  To ensure everything is out we need to flush at least
1686          * three times.  (1) The running of the sideq can dirty the
1687          * filesystem, (2) A normal flush can dirty the freemap, and
1688          * (3) ensure that the freemap is fully synchronized.
1689          *
1690          * The next mount's recovery scan can clean everything up but we want
1691          * to leave the filesystem in a 100% clean state on a normal unmount.
1692          */
1693 #if 0
1694         hammer2_voldata_lock(hmp);
1695         hammer2_voldata_unlock(hmp);
1696 #endif
1697
1698         /*
1699          * Flush whatever is left.  Unmounted but modified PFS's might still
1700          * have some dirty chains on them.
1701          */
1702         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1703         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1704         hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP | HAMMER2_FLUSH_ALL);
1705         hammer2_chain_unlock(&hmp->fchain);
1706         hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP | HAMMER2_FLUSH_ALL);
1707         hammer2_chain_unlock(&hmp->vchain);
1708
1709         if ((hmp->vchain.flags | hmp->fchain.flags) &
1710             HAMMER2_CHAIN_FLUSH_MASK) {
1711                 kprintf("hammer2_unmount: chains left over "
1712                         "after final sync\n");
1713                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1714                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1715
1716                 if (hammer2_debug & 0x0010)
1717                         Debugger("entered debugger");
1718         }
1719
1720         KKASSERT(hmp->spmp == NULL);
1721
1722         /*
1723          * Finish up with the device vnode
1724          */
1725         if ((devvp = hmp->devvp) != NULL) {
1726                 ronly = hmp->ronly;
1727                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1728                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1729                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1730                         ronly);
1731                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1732                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1733                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1734                 hmp->devvp = NULL;
1735                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1736                 vn_unlock(devvp);
1737                 vrele(devvp);
1738                 devvp = NULL;
1739         }
1740
1741         /*
1742          * Clear vchain/fchain flags that might prevent final cleanup
1743          * of these chains.
1744          */
1745         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1746                 atomic_add_long(&hammer2_count_modified_chains, -1);
1747                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1748                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1749         }
1750         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1751                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1752         }
1753
1754         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1755                 atomic_add_long(&hammer2_count_modified_chains, -1);
1756                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1757                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1758         }
1759         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1760                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1761         }
1762
1763         /*
1764          * Final drop of embedded freemap root chain to
1765          * clean up fchain.core (fchain structure is not
1766          * flagged ALLOCATED so it is cleaned out and then
1767          * left to rot).
1768          */
1769         hammer2_chain_drop(&hmp->fchain);
1770
1771         /*
1772          * Final drop of embedded volume root chain to clean
1773          * up vchain.core (vchain structure is not flagged
1774          * ALLOCATED so it is cleaned out and then left to
1775          * rot).
1776          */
1777         dumpcnt = 50;
1778         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1779         dumpcnt = 50;
1780         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1781         hammer2_dev_unlock(hmp);
1782         hammer2_chain_drop(&hmp->vchain);
1783
1784         hammer2_io_cleanup(hmp, &hmp->iotree);
1785         if (hmp->iofree_count) {
1786                 kprintf("io_cleanup: %d I/O's left hanging\n",
1787                         hmp->iofree_count);
1788         }
1789
1790         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1791         kmalloc_destroy(&hmp->mchain);
1792         kfree(hmp, M_HAMMER2);
1793 }
1794
1795 int
1796 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1797                  ino_t ino, struct vnode **vpp)
1798 {
1799         hammer2_xop_lookup_t *xop;
1800         hammer2_pfs_t *pmp;
1801         hammer2_inode_t *ip;
1802         hammer2_tid_t inum;
1803         int error;
1804
1805         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1806
1807         error = 0;
1808         pmp = MPTOPMP(mp);
1809
1810         /*
1811          * Easy if we already have it cached
1812          */
1813         ip = hammer2_inode_lookup(pmp, inum);
1814         if (ip) {
1815                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1816                 *vpp = hammer2_igetv(ip, &error);
1817                 hammer2_inode_unlock(ip);
1818                 hammer2_inode_drop(ip);         /* from lookup */
1819
1820                 return error;
1821         }
1822
1823         /*
1824          * Otherwise we have to find the inode
1825          */
1826         xop = hammer2_xop_alloc(pmp->iroot, 0);
1827         xop->lhc = inum;
1828         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1829         error = hammer2_xop_collect(&xop->head, 0);
1830
1831         if (error == 0) {
1832                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1833                         kprintf("vget: no collect error but also no rdata\n");
1834                         kprintf("xop %p\n", xop);
1835                         while ((hammer2_debug & 0x80000) == 0) {
1836                                 tsleep(xop, PCATCH, "wait", hz * 10);
1837                         }
1838                         ip = NULL;
1839                 } else {
1840                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1841                 }
1842         }
1843         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1844
1845         if (ip) {
1846                 *vpp = hammer2_igetv(ip, &error);
1847                 hammer2_inode_unlock(ip);
1848         } else {
1849                 *vpp = NULL;
1850                 error = ENOENT;
1851         }
1852         return (error);
1853 }
1854
1855 static
1856 int
1857 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1858 {
1859         hammer2_pfs_t *pmp;
1860         struct vnode *vp;
1861         int error;
1862
1863         pmp = MPTOPMP(mp);
1864         if (pmp->iroot == NULL) {
1865                 *vpp = NULL;
1866                 return EINVAL;
1867         }
1868
1869         error = 0;
1870         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1871
1872         while (pmp->inode_tid == 0) {
1873                 hammer2_xop_ipcluster_t *xop;
1874                 hammer2_inode_meta_t *meta;
1875
1876                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1877                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1878                 error = hammer2_xop_collect(&xop->head, 0);
1879
1880                 if (error == 0) {
1881                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1882                         pmp->iroot->meta = *meta;
1883                         pmp->inode_tid = meta->pfs_inum + 1;
1884                         if (pmp->inode_tid < HAMMER2_INODE_START)
1885                                 pmp->inode_tid = HAMMER2_INODE_START;
1886                         pmp->modify_tid =
1887                                 xop->head.cluster.focus->bref.modify_tid + 1;
1888                         kprintf("PFS: Starting inode %jd\n",
1889                                 (intmax_t)pmp->inode_tid);
1890                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1891                                 pmp->inode_tid, pmp->modify_tid);
1892                         wakeup(&pmp->iroot);
1893
1894                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1895
1896                         /*
1897                          * Prime the mount info.
1898                          */
1899                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1900                         break;
1901                 }
1902
1903                 /*
1904                  * Loop, try again
1905                  */
1906                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1907                 hammer2_inode_unlock(pmp->iroot);
1908                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1909                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1910                 if (error == EINTR)
1911                         break;
1912         }
1913
1914         if (error) {
1915                 hammer2_inode_unlock(pmp->iroot);
1916                 *vpp = NULL;
1917         } else {
1918                 vp = hammer2_igetv(pmp->iroot, &error);
1919                 hammer2_inode_unlock(pmp->iroot);
1920                 *vpp = vp;
1921         }
1922
1923         return (error);
1924 }
1925
1926 /*
1927  * Filesystem status
1928  *
1929  * XXX incorporate ipdata->meta.inode_quota and data_quota
1930  */
1931 static
1932 int
1933 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1934 {
1935         hammer2_pfs_t *pmp;
1936         hammer2_dev_t *hmp;
1937         hammer2_blockref_t bref;
1938         struct statfs tmp;
1939         int i;
1940
1941         /*
1942          * NOTE: iroot might not have validated the cluster yet.
1943          */
1944         pmp = MPTOPMP(mp);
1945
1946         bzero(&tmp, sizeof(tmp));
1947
1948         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1949                 hmp = pmp->pfs_hmps[i];
1950                 if (hmp == NULL)
1951                         continue;
1952                 if (pmp->iroot->cluster.array[i].chain)
1953                         bref = pmp->iroot->cluster.array[i].chain->bref;
1954                 else
1955                         bzero(&bref, sizeof(bref));
1956
1957                 tmp.f_files = bref.embed.stats.inode_count;
1958                 tmp.f_ffree = 0;
1959                 tmp.f_blocks = hmp->voldata.allocator_size /
1960                                mp->mnt_vstat.f_bsize;
1961                 tmp.f_bfree = hmp->voldata.allocator_free /
1962                               mp->mnt_vstat.f_bsize;
1963                 tmp.f_bavail = tmp.f_bfree;
1964
1965                 if (cred && cred->cr_uid != 0) {
1966                         uint64_t adj;
1967
1968                         /* 5% */
1969                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
1970                         tmp.f_blocks -= adj;
1971                         tmp.f_bfree -= adj;
1972                         tmp.f_bavail -= adj;
1973                 }
1974
1975                 mp->mnt_stat.f_blocks = tmp.f_blocks;
1976                 mp->mnt_stat.f_bfree = tmp.f_bfree;
1977                 mp->mnt_stat.f_bavail = tmp.f_bavail;
1978                 mp->mnt_stat.f_files = tmp.f_files;
1979                 mp->mnt_stat.f_ffree = tmp.f_ffree;
1980
1981                 *sbp = mp->mnt_stat;
1982         }
1983         return (0);
1984 }
1985
1986 static
1987 int
1988 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1989 {
1990         hammer2_pfs_t *pmp;
1991         hammer2_dev_t *hmp;
1992         hammer2_blockref_t bref;
1993         struct statvfs tmp;
1994         int i;
1995
1996         /*
1997          * NOTE: iroot might not have validated the cluster yet.
1998          */
1999         pmp = MPTOPMP(mp);
2000         bzero(&tmp, sizeof(tmp));
2001
2002         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2003                 hmp = pmp->pfs_hmps[i];
2004                 if (hmp == NULL)
2005                         continue;
2006                 if (pmp->iroot->cluster.array[i].chain)
2007                         bref = pmp->iroot->cluster.array[i].chain->bref;
2008                 else
2009                         bzero(&bref, sizeof(bref));
2010
2011                 tmp.f_files = bref.embed.stats.inode_count;
2012                 tmp.f_ffree = 0;
2013                 tmp.f_blocks = hmp->voldata.allocator_size /
2014                                mp->mnt_vstat.f_bsize;
2015                 tmp.f_bfree = hmp->voldata.allocator_free /
2016                               mp->mnt_vstat.f_bsize;
2017                 tmp.f_bavail = tmp.f_bfree;
2018
2019                 if (cred && cred->cr_uid != 0) {
2020                         uint64_t adj;
2021
2022                         /* 5% */
2023                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2024                         tmp.f_blocks -= adj;
2025                         tmp.f_bfree -= adj;
2026                         tmp.f_bavail -= adj;
2027                 }
2028
2029                 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2030                 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2031                 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2032                 mp->mnt_vstat.f_files = tmp.f_files;
2033                 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2034
2035                 *sbp = mp->mnt_vstat;
2036         }
2037         return (0);
2038 }
2039
2040 /*
2041  * Mount-time recovery (RW mounts)
2042  *
2043  * Updates to the free block table are allowed to lag flushes by one
2044  * transaction.  In case of a crash, then on a fresh mount we must do an
2045  * incremental scan of the last committed transaction id and make sure that
2046  * all related blocks have been marked allocated.
2047  *
2048  * The super-root topology and each PFS has its own transaction id domain,
2049  * so we must track PFS boundary transitions.
2050  */
2051 struct hammer2_recovery_elm {
2052         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2053         hammer2_chain_t *chain;
2054         hammer2_tid_t sync_tid;
2055 };
2056
2057 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2058
2059 struct hammer2_recovery_info {
2060         struct hammer2_recovery_list list;
2061         hammer2_tid_t   mtid;
2062         int     depth;
2063 };
2064
2065 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2066                         hammer2_chain_t *parent,
2067                         struct hammer2_recovery_info *info,
2068                         hammer2_tid_t sync_tid);
2069
2070 #define HAMMER2_RECOVERY_MAXDEPTH       10
2071
2072 static
2073 int
2074 hammer2_recovery(hammer2_dev_t *hmp)
2075 {
2076         struct hammer2_recovery_info info;
2077         struct hammer2_recovery_elm *elm;
2078         hammer2_chain_t *parent;
2079         hammer2_tid_t sync_tid;
2080         hammer2_tid_t mirror_tid;
2081         int error;
2082
2083         hammer2_trans_init(hmp->spmp, 0);
2084
2085         sync_tid = hmp->voldata.freemap_tid;
2086         mirror_tid = hmp->voldata.mirror_tid;
2087
2088         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2089         if (sync_tid >= mirror_tid) {
2090                 kprintf(" no recovery needed\n");
2091         } else {
2092                 kprintf(" freemap recovery %016jx-%016jx\n",
2093                         sync_tid + 1, mirror_tid);
2094         }
2095
2096         TAILQ_INIT(&info.list);
2097         info.depth = 0;
2098         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2099         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2100         hammer2_chain_lookup_done(parent);
2101
2102         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2103                 TAILQ_REMOVE(&info.list, elm, entry);
2104                 parent = elm->chain;
2105                 sync_tid = elm->sync_tid;
2106                 kfree(elm, M_HAMMER2);
2107
2108                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2109                 error |= hammer2_recovery_scan(hmp, parent, &info,
2110                                               hmp->voldata.freemap_tid);
2111                 hammer2_chain_unlock(parent);
2112                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2113         }
2114         hammer2_trans_done(hmp->spmp);
2115
2116         return error;
2117 }
2118
2119 static
2120 int
2121 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2122                       struct hammer2_recovery_info *info,
2123                       hammer2_tid_t sync_tid)
2124 {
2125         const hammer2_inode_data_t *ripdata;
2126         hammer2_chain_t *chain;
2127         hammer2_blockref_t bref;
2128         int tmp_error;
2129         int rup_error;
2130         int error;
2131         int first;
2132
2133         /*
2134          * Adjust freemap to ensure that the block(s) are marked allocated.
2135          */
2136         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2137                 hammer2_freemap_adjust(hmp, &parent->bref,
2138                                        HAMMER2_FREEMAP_DORECOVER);
2139         }
2140
2141         /*
2142          * Check type for recursive scan
2143          */
2144         switch(parent->bref.type) {
2145         case HAMMER2_BREF_TYPE_VOLUME:
2146                 /* data already instantiated */
2147                 break;
2148         case HAMMER2_BREF_TYPE_INODE:
2149                 /*
2150                  * Must instantiate data for DIRECTDATA test and also
2151                  * for recursion.
2152                  */
2153                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2154                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2155                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2156                         /* not applicable to recovery scan */
2157                         hammer2_chain_unlock(parent);
2158                         return 0;
2159                 }
2160                 hammer2_chain_unlock(parent);
2161                 break;
2162         case HAMMER2_BREF_TYPE_INDIRECT:
2163                 /*
2164                  * Must instantiate data for recursion
2165                  */
2166                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2167                 hammer2_chain_unlock(parent);
2168                 break;
2169         case HAMMER2_BREF_TYPE_DIRENT:
2170         case HAMMER2_BREF_TYPE_DATA:
2171         case HAMMER2_BREF_TYPE_FREEMAP:
2172         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2173         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2174                 /* not applicable to recovery scan */
2175                 return 0;
2176                 break;
2177         default:
2178                 return HAMMER2_ERROR_BADBREF;
2179         }
2180
2181         /*
2182          * Defer operation if depth limit reached or if we are crossing a
2183          * PFS boundary.
2184          */
2185         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2186                 struct hammer2_recovery_elm *elm;
2187
2188                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2189                 elm->chain = parent;
2190                 elm->sync_tid = sync_tid;
2191                 hammer2_chain_ref(parent);
2192                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2193                 /* unlocked by caller */
2194
2195                 return(0);
2196         }
2197
2198
2199         /*
2200          * Recursive scan of the last flushed transaction only.  We are
2201          * doing this without pmp assignments so don't leave the chains
2202          * hanging around after we are done with them.
2203          *
2204          * error        Cumulative error this level only
2205          * rup_error    Cumulative error for recursion
2206          * tmp_error    Specific non-cumulative recursion error
2207          */
2208         chain = NULL;
2209         first = 1;
2210         rup_error = 0;
2211         error = 0;
2212
2213         for (;;) {
2214                 error |= hammer2_chain_scan(parent, &chain, &bref,
2215                                             &first,
2216                                             HAMMER2_LOOKUP_NODATA);
2217
2218                 /*
2219                  * Problem during scan or EOF
2220                  */
2221                 if (error)
2222                         break;
2223
2224                 /*
2225                  * If this is a leaf
2226                  */
2227                 if (chain == NULL) {
2228                         if (bref.mirror_tid > sync_tid) {
2229                                 hammer2_freemap_adjust(hmp, &bref,
2230                                                      HAMMER2_FREEMAP_DORECOVER);
2231                         }
2232                         continue;
2233                 }
2234
2235                 /*
2236                  * This may or may not be a recursive node.
2237                  */
2238                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2239                 if (bref.mirror_tid > sync_tid) {
2240                         ++info->depth;
2241                         tmp_error = hammer2_recovery_scan(hmp, chain,
2242                                                            info, sync_tid);
2243                         --info->depth;
2244                 } else {
2245                         tmp_error = 0;
2246                 }
2247
2248                 /*
2249                  * Flush the recovery at the PFS boundary to stage it for
2250                  * the final flush of the super-root topology.
2251                  */
2252                 if (tmp_error == 0 &&
2253                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2254                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2255                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2256                                              HAMMER2_FLUSH_ALL);
2257                 }
2258                 rup_error |= tmp_error;
2259         }
2260         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2261 }
2262
2263 /*
2264  * Sync a mount point; this is called periodically on a per-mount basis from
2265  * the filesystem syncer, and whenever a user issues a sync.
2266  */
2267 int
2268 hammer2_vfs_sync(struct mount *mp, int waitfor)
2269 {
2270         hammer2_xop_flush_t *xop;
2271         struct hammer2_sync_info info;
2272         hammer2_inode_t *iroot;
2273         hammer2_pfs_t *pmp;
2274         int flags;
2275         int error;
2276
2277         pmp = MPTOPMP(mp);
2278         iroot = pmp->iroot;
2279         KKASSERT(iroot);
2280         KKASSERT(iroot->pmp == pmp);
2281
2282         /*
2283          * We can't acquire locks on existing vnodes while in a transaction
2284          * without risking a deadlock.  This assumes that vfsync() can be
2285          * called without the vnode locked (which it can in DragonFly).
2286          * Otherwise we'd have to implement a multi-pass or flag the lock
2287          * failures and retry.
2288          *
2289          * The reclamation code interlocks with the sync list's token
2290          * (by removing the vnode from the scan list) before unlocking
2291          * the inode, giving us time to ref the inode.
2292          */
2293         /*flags = VMSC_GETVP;*/
2294         flags = 0;
2295         if (waitfor & MNT_LAZY)
2296                 flags |= VMSC_ONEPASS;
2297
2298         /*
2299          * Flush vnodes individually using a normal transaction to avoid
2300          * stalling any concurrent operations.  This will flush the related
2301          * buffer cache buffers and inodes to the media.
2302          *
2303          * For efficiency do an async pass before making sure with a
2304          * synchronous pass on all related buffer cache buffers.
2305          *
2306          * Do a single synchronous pass to avoid double-flushing vnodes,
2307          * which can waste copy-on-write blocks.  XXX do not do two passes.
2308          */
2309         hammer2_trans_init(pmp, 0);
2310         info.error = 0;
2311         info.waitfor = MNT_NOWAIT;
2312         info.pass = 1;
2313         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2314         info.waitfor = MNT_WAIT;
2315         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2316
2317         info.pass = 2;
2318         info.waitfor = MNT_WAIT;
2319         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2320
2321         /*
2322          * We must also run the sideq to handle any disconnected inodes
2323          * as the vnode scan will not see these.
2324          */
2325         hammer2_inode_run_sideq(pmp, 1);
2326         hammer2_trans_done(pmp);
2327
2328         /*
2329          * Start our flush transaction and flush the root topology down to
2330          * the inodes, but not the inodes themselves (which we already flushed
2331          * above).  Any concurrent activity effecting inode contents will not
2332          * be part of this flush cycle.
2333          *
2334          * (1) vfsync() all dirty vnodes via vfsyncscan().
2335          *
2336          * (2) Flush any remaining dirty inodes (the sideq), including any
2337          *     which may have been created during or raced against the
2338          *     vfsync().  To catch all cases this must be done after the
2339          *     vfsync().
2340          *
2341          * (3) Wait for any pending BIO I/O to complete (hammer2_bioq_sync()).
2342          *
2343          * NOTE! It is still possible for the paging code to push pages
2344          *       out via a UIO_NOCOPY hammer2_vop_write() during the main
2345          *       flush.
2346          */
2347         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2348
2349         /*
2350          * Use the XOP interface to concurrently flush all nodes to
2351          * synchronize the PFSROOT subtopology to the media.  A standard
2352          * end-of-scan ENOENT error indicates cluster sufficiency.
2353          *
2354          * Note that this flush will not be visible on crash recovery until
2355          * we flush the super-root topology in the next loop.
2356          *
2357          * XXX For now wait for all flushes to complete.
2358          */
2359         if (iroot) {
2360                 /*
2361                  * If unmounting try to flush everything including any
2362                  * sub-trees under inodes, just in case there is dangling
2363                  * modified data, as a safety.  Otherwise just flush up to
2364                  * the inodes in this stage.
2365                  */
2366                 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2367                         xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2368                 } else {
2369                         xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2370                                                        HAMMER2_XOP_INODE_STOP);
2371                 }
2372                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2373                 error = hammer2_xop_collect(&xop->head,
2374                                             HAMMER2_XOP_COLLECT_WAITALL);
2375                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2376                 if (error == HAMMER2_ERROR_ENOENT)
2377                         error = 0;
2378                 else
2379                         error = hammer2_error_to_errno(error);
2380         } else {
2381                 error = 0;
2382         }
2383         hammer2_trans_done(pmp);
2384
2385         return (error);
2386 }
2387
2388 /*
2389  * Sync passes.
2390  *
2391  * Note that we ignore the tranasction mtid we got above.  Instead,
2392  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2393  * transactions.
2394  *
2395  * WARNING! The frontend might be waiting on chnmem (limit_dirty_chains)
2396  * while holding a vnode locked.  When this situation occurs we cannot
2397  * safely test whether it is ok to clear the dirty bit on the vnode.
2398  * However, we can still flush the inode's topology.
2399  */
2400 static int
2401 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2402 {
2403         struct hammer2_sync_info *info = data;
2404         hammer2_inode_t *ip;
2405         int error;
2406
2407         /*
2408          * Degenerate cases.  Note that ip == NULL typically means the
2409          * syncer vnode itself and we don't want to vclrisdirty() in that
2410          * situation.
2411          */
2412         ip = VTOI(vp);
2413         if (ip == NULL) {
2414                 return(0);
2415         }
2416         if (vp->v_type == VNON || vp->v_type == VBAD) {
2417                 vclrisdirty(vp);
2418                 return(0);
2419         }
2420
2421         /*
2422          * Synchronize the buffer cche and inode meta-data to the backing
2423          * chain topology.
2424          *
2425          * vfsync is not necessarily synchronous, so it is best NOT to try
2426          * to flush the backing topology to media at this point.
2427          */
2428         hammer2_inode_ref(ip);
2429         if ((ip->flags & (HAMMER2_INODE_RESIZED|HAMMER2_INODE_MODIFIED)) ||
2430             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2431                 if (info->pass == 1)
2432                         vfsync(vp, info->waitfor, 1, NULL, NULL);
2433                 else
2434                         bio_track_wait(&vp->v_track_write, 0, 0);
2435         }
2436         if (info->pass == 2 && (vp->v_flag & VISDIRTY)) {
2437                 /*
2438                  * v_token is needed to interlock v_rbdirty_tree.
2439                  */
2440                 lwkt_gettoken(&vp->v_token);
2441                 hammer2_inode_lock(ip, 0);
2442                 hammer2_inode_chain_sync(ip);
2443                 hammer2_inode_chain_flush(ip);
2444                 if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2445                                   HAMMER2_INODE_RESIZED |
2446                                   HAMMER2_INODE_DIRTYDATA)) == 0 &&
2447                     RB_EMPTY(&vp->v_rbdirty_tree) &&
2448                     !bio_track_active(&vp->v_track_write)) {
2449                         vclrisdirty(vp);
2450                 }
2451                 hammer2_inode_unlock(ip);
2452                 lwkt_reltoken(&vp->v_token);
2453         }
2454         hammer2_inode_drop(ip);
2455 #if 1
2456         error = 0;
2457         if (error)
2458                 info->error = error;
2459 #endif
2460         return(0);
2461 }
2462
2463 static
2464 int
2465 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2466 {
2467         hammer2_inode_t *ip;
2468
2469         KKASSERT(MAXFIDSZ >= 16);
2470         ip = VTOI(vp);
2471         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2472         fhp->fid_ext = 0;
2473         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2474         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2475
2476         return 0;
2477 }
2478
2479 static
2480 int
2481 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2482                struct fid *fhp, struct vnode **vpp)
2483 {
2484         hammer2_pfs_t *pmp;
2485         hammer2_tid_t inum;
2486         int error;
2487
2488         pmp = MPTOPMP(mp);
2489         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2490         if (vpp) {
2491                 if (inum == 1)
2492                         error = hammer2_vfs_root(mp, vpp);
2493                 else
2494                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2495         } else {
2496                 error = 0;
2497         }
2498         if (error)
2499                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2500         return error;
2501 }
2502
2503 static
2504 int
2505 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2506                  int *exflagsp, struct ucred **credanonp)
2507 {
2508         hammer2_pfs_t *pmp;
2509         struct netcred *np;
2510         int error;
2511
2512         pmp = MPTOPMP(mp);
2513         np = vfs_export_lookup(mp, &pmp->export, nam);
2514         if (np) {
2515                 *exflagsp = np->netc_exflags;
2516                 *credanonp = &np->netc_anon;
2517                 error = 0;
2518         } else {
2519                 error = EACCES;
2520         }
2521         return error;
2522 }
2523
2524 /*
2525  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2526  * header into the HMP
2527  *
2528  * XXX read four volhdrs and use the one with the highest TID whos CRC
2529  *     matches.
2530  *
2531  * XXX check iCRCs.
2532  *
2533  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2534  *     nonexistant locations.
2535  *
2536  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2537  */
2538 static
2539 int
2540 hammer2_install_volume_header(hammer2_dev_t *hmp)
2541 {
2542         hammer2_volume_data_t *vd;
2543         struct buf *bp;
2544         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2545         int error_reported;
2546         int error;
2547         int valid;
2548         int i;
2549
2550         error_reported = 0;
2551         error = 0;
2552         valid = 0;
2553         bp = NULL;
2554
2555         /*
2556          * There are up to 4 copies of the volume header (syncs iterate
2557          * between them so there is no single master).  We don't trust the
2558          * volu_size field so we don't know precisely how large the filesystem
2559          * is, so depend on the OS to return an error if we go beyond the
2560          * block device's EOF.
2561          */
2562         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2563                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2564                               HAMMER2_VOLUME_BYTES, &bp);
2565                 if (error) {
2566                         brelse(bp);
2567                         bp = NULL;
2568                         continue;
2569                 }
2570
2571                 vd = (struct hammer2_volume_data *) bp->b_data;
2572                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2573                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2574                         brelse(bp);
2575                         bp = NULL;
2576                         continue;
2577                 }
2578
2579                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2580                         /* XXX: Reversed-endianness filesystem */
2581                         kprintf("hammer2: reverse-endian filesystem detected");
2582                         brelse(bp);
2583                         bp = NULL;
2584                         continue;
2585                 }
2586
2587                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2588                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2589                                       HAMMER2_VOLUME_ICRC0_SIZE);
2590                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2591                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2592                                        HAMMER2_VOLUME_ICRC1_SIZE);
2593                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2594                         kprintf("hammer2 volume header crc "
2595                                 "mismatch copy #%d %08x/%08x\n",
2596                                 i, crc0, crc);
2597                         error_reported = 1;
2598                         brelse(bp);
2599                         bp = NULL;
2600                         continue;
2601                 }
2602                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2603                         valid = 1;
2604                         hmp->voldata = *vd;
2605                         hmp->volhdrno = i;
2606                 }
2607                 brelse(bp);
2608                 bp = NULL;
2609         }
2610         if (valid) {
2611                 hmp->volsync = hmp->voldata;
2612                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2613                 error = 0;
2614                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2615                         kprintf("hammer2: using volume header #%d\n",
2616                                 hmp->volhdrno);
2617                 }
2618         } else {
2619                 error = EINVAL;
2620                 kprintf("hammer2: no valid volume headers found!\n");
2621         }
2622         return (error);
2623 }
2624
2625 /*
2626  * This handles hysteresis on regular file flushes.  Because the BIOs are
2627  * routed to a thread it is possible for an excessive number to build up
2628  * and cause long front-end stalls long before the runningbuffspace limit
2629  * is hit, so we implement hammer2_flush_pipe to control the
2630  * hysteresis.
2631  *
2632  * This is a particular problem when compression is used.
2633  */
2634 void
2635 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2636 {
2637         atomic_add_int(&pmp->count_lwinprog, 1);
2638 }
2639
2640 void
2641 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2642 {
2643         int lwinprog;
2644
2645         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2646         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2647             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2648                 atomic_clear_int(&pmp->count_lwinprog,
2649                                  HAMMER2_LWINPROG_WAITING);
2650                 wakeup(&pmp->count_lwinprog);
2651         }
2652         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2653             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2654                 atomic_clear_int(&pmp->count_lwinprog,
2655                                  HAMMER2_LWINPROG_WAITING0);
2656                 wakeup(&pmp->count_lwinprog);
2657         }
2658 }
2659
2660 void
2661 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2662 {
2663         int lwinprog;
2664         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2665                                     HAMMER2_LWINPROG_WAITING0;
2666
2667         for (;;) {
2668                 lwinprog = pmp->count_lwinprog;
2669                 cpu_ccfence();
2670                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2671                         break;
2672                 tsleep_interlock(&pmp->count_lwinprog, 0);
2673                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2674                 lwinprog = pmp->count_lwinprog;
2675                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2676                         break;
2677                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2678         }
2679 }
2680
2681 /*
2682  * Manage excessive memory resource use for chain and related
2683  * structures.
2684  */
2685 void
2686 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2687 {
2688         uint32_t waiting;
2689         uint32_t count;
2690         uint32_t limit;
2691 #if 0
2692         static int zzticks;
2693 #endif
2694
2695         /*
2696          * Atomic check condition and wait.  Also do an early speedup of
2697          * the syncer to try to avoid hitting the wait.
2698          */
2699         for (;;) {
2700                 waiting = pmp->inmem_dirty_chains;
2701                 cpu_ccfence();
2702                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2703
2704                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2705                 if (limit < hammer2_limit_dirty_chains)
2706                         limit = hammer2_limit_dirty_chains;
2707                 if (limit < 1000)
2708                         limit = 1000;
2709
2710 #if 0
2711                 if ((int)(ticks - zzticks) > hz) {
2712                         zzticks = ticks;
2713                         kprintf("count %ld %ld\n", count, limit);
2714                 }
2715 #endif
2716
2717                 /*
2718                  * Block if there are too many dirty chains present, wait
2719                  * for the flush to clean some out.
2720                  */
2721                 if (count > limit) {
2722                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2723                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2724                                                waiting,
2725                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2726                                 speedup_syncer(pmp->mp);
2727                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2728                                        "chnmem", hz);
2729                         }
2730                         continue;       /* loop on success or fail */
2731                 }
2732
2733                 /*
2734                  * Try to start an early flush before we are forced to block.
2735                  */
2736                 if (count > limit * 5 / 10)
2737                         speedup_syncer(pmp->mp);
2738                 break;
2739         }
2740 }
2741
2742 void
2743 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2744 {
2745         if (pmp) {
2746                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2747         }
2748 }
2749
2750 void
2751 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2752 {
2753         uint32_t waiting;
2754
2755         if (pmp) {
2756                 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
2757                 /* don't need --waiting to test flag */
2758                 if (waiting & HAMMER2_DIRTYCHAIN_WAITING) {
2759                         atomic_clear_int(&pmp->inmem_dirty_chains,
2760                                          HAMMER2_DIRTYCHAIN_WAITING);
2761                         wakeup(&pmp->inmem_dirty_chains);
2762                 }
2763         }
2764 }
2765
2766 /*
2767  * Returns 0 if the filesystem has tons of free space
2768  * Returns 1 if the filesystem has less than 10% remaining
2769  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2770  */
2771 int
2772 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2773 {
2774         hammer2_pfs_t *pmp;
2775         hammer2_dev_t *hmp;
2776         hammer2_off_t free_reserved;
2777         hammer2_off_t free_nominal;
2778         int i;
2779
2780         pmp = ip->pmp;
2781
2782         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2783                 free_reserved = HAMMER2_SEGSIZE;
2784                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2785                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2786                         hmp = pmp->pfs_hmps[i];
2787                         if (hmp == NULL)
2788                                 continue;
2789                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2790                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2791                                 continue;
2792
2793                         if (free_nominal > hmp->voldata.allocator_free)
2794                                 free_nominal = hmp->voldata.allocator_free;
2795                         if (free_reserved < hmp->free_reserved)
2796                                 free_reserved = hmp->free_reserved;
2797                 }
2798
2799                 /*
2800                  * SMP races ok
2801                  */
2802                 pmp->free_reserved = free_reserved;
2803                 pmp->free_nominal = free_nominal;
2804                 pmp->free_ticks = ticks;
2805         } else {
2806                 free_reserved = pmp->free_reserved;
2807                 free_nominal = pmp->free_nominal;
2808         }
2809         if (cred && cred->cr_uid != 0) {
2810                 if ((int64_t)(free_nominal - bytes) <
2811                     (int64_t)free_reserved) {
2812                         return 2;
2813                 }
2814         } else {
2815                 if ((int64_t)(free_nominal - bytes) <
2816                     (int64_t)free_reserved / 2) {
2817                         return 2;
2818                 }
2819         }
2820         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2821                 return 1;
2822         return 0;
2823 }
2824
2825 /*
2826  * Debugging
2827  */
2828 void
2829 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
2830                    u_int flags)
2831 {
2832         hammer2_chain_t *scan;
2833         hammer2_chain_t *parent;
2834
2835         --*countp;
2836         if (*countp == 0) {
2837                 kprintf("%*.*s...\n", tab, tab, "");
2838                 return;
2839         }
2840         if (*countp < 0)
2841                 return;
2842         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2843                 tab, tab, "", pfx,
2844                 chain, chain->bref.type,
2845                 chain->bref.key, chain->bref.keybits,
2846                 chain->bref.mirror_tid);
2847
2848         kprintf("%*.*s      [%08x] (%s) refs=%d",
2849                 tab, tab, "",
2850                 chain->flags,
2851                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2852                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2853                 chain->refs);
2854
2855         parent = chain->parent;
2856         if (parent)
2857                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2858                         tab, tab, "",
2859                         parent, parent->flags, parent->refs);
2860         if (RB_EMPTY(&chain->core.rbtree)) {
2861                 kprintf("\n");
2862         } else {
2863                 kprintf(" {\n");
2864                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
2865                         if ((scan->flags & flags) || flags == (u_int)-1) {
2866                                 hammer2_dump_chain(scan, tab + 4, countp, 'a',
2867                                                    flags);
2868                         }
2869                 }
2870                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2871                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2872                                 chain->data->ipdata.filename);
2873                 else
2874                         kprintf("%*.*s}\n", tab, tab, "");
2875         }
2876 }