sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72 };
  73
  74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  75 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
  76 static struct hammer2_mntlist hammer2_mntlist;
  77 static struct hammer2_pfslist hammer2_pfslist;
  78 static struct lock hammer2_mntlk;
  79
  80 int hammer2_debug;
  81 int hammer2_cluster_read = 4;           /* physical read-ahead */
  82 int hammer2_cluster_write = 0;          /* bdwrite() so later inval works */
  83 int hammer2_dedup_enable = 1;
  84 int hammer2_inval_enable = 0;
  85 int hammer2_flush_pipe = 100;
  86 int hammer2_synchronous_flush = 1;
  87 int hammer2_dio_count;
  88 long hammer2_chain_allocs;
  89 long hammer2_chain_frees;
  90 long hammer2_limit_dirty_chains;
  91 long hammer2_count_modified_chains;
  92 long hammer2_iod_invals;
  93 long hammer2_iod_file_read;
  94 long hammer2_iod_meta_read;
  95 long hammer2_iod_indr_read;
  96 long hammer2_iod_fmap_read;
  97 long hammer2_iod_volu_read;
  98 long hammer2_iod_file_write;
  99 long hammer2_iod_file_wembed;
 100 long hammer2_iod_file_wzero;
 101 long hammer2_iod_file_wdedup;
 102 long hammer2_iod_meta_write;
 103 long hammer2_iod_indr_write;
 104 long hammer2_iod_fmap_write;
 105 long hammer2_iod_volu_write;
 106
 107 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 108 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 109                 "Buffer used for compression.");
 110
 111 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 112 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 113                 "Buffer used for decompression.");
 114
 115 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 116
 117 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 118            &hammer2_debug, 0, "");
 119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW,
 120            &hammer2_cluster_read, 0, "");
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 122            &hammer2_cluster_write, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 124            &hammer2_dedup_enable, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 126            &hammer2_inval_enable, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 128            &hammer2_flush_pipe, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 130            &hammer2_synchronous_flush, 0, "");
 131 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 132            &hammer2_chain_allocs, 0, "");
 133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 134            &hammer2_chain_frees, 0, "");
 135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 136            &hammer2_limit_dirty_chains, 0, "");
 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 138            &hammer2_count_modified_chains, 0, "");
 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 140            &hammer2_dio_count, 0, "");
 141
 142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 143            &hammer2_iod_invals, 0, "");
 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 145            &hammer2_iod_file_read, 0, "");
 146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 147            &hammer2_iod_meta_read, 0, "");
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 149            &hammer2_iod_indr_read, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 151            &hammer2_iod_fmap_read, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 153            &hammer2_iod_volu_read, 0, "");
 154
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 156            &hammer2_iod_file_write, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 158            &hammer2_iod_file_wembed, 0, "");
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 160            &hammer2_iod_file_wzero, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 162            &hammer2_iod_file_wdedup, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 164            &hammer2_iod_meta_write, 0, "");
 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 166            &hammer2_iod_indr_write, 0, "");
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 168            &hammer2_iod_fmap_write, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 170            &hammer2_iod_volu_write, 0, "");
 171
 172 long hammer2_check_icrc32;
 173 long hammer2_check_xxhash64;
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 175            &hammer2_check_icrc32, 0, "");
 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 177            &hammer2_check_xxhash64, 0, "");
 178
 179 static int hammer2_vfs_init(struct vfsconf *conf);
 180 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 181 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 182                                 struct ucred *cred);
 183 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 184                                 struct vnode *, struct ucred *);
 185 static int hammer2_recovery(hammer2_dev_t *hmp);
 186 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 187 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 188 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 189                                 struct ucred *cred);
 190 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 191                                 struct ucred *cred);
 192 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 193                                 struct fid *fhp, struct vnode **vpp);
 194 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 195 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 196                                 int *exflagsp, struct ucred **credanonp);
 197
 198 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 199 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 200
 201 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 202
 203 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 204 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 205                                 hammer2_dev_t *hmp);
 206
 207 /*
 208  * HAMMER2 vfs operations.
 209  */
 210 static struct vfsops hammer2_vfsops = {
 211         .vfs_init       = hammer2_vfs_init,
 212         .vfs_uninit     = hammer2_vfs_uninit,
 213         .vfs_sync       = hammer2_vfs_sync,
 214         .vfs_mount      = hammer2_vfs_mount,
 215         .vfs_unmount    = hammer2_vfs_unmount,
 216         .vfs_root       = hammer2_vfs_root,
 217         .vfs_statfs     = hammer2_vfs_statfs,
 218         .vfs_statvfs    = hammer2_vfs_statvfs,
 219         .vfs_vget       = hammer2_vfs_vget,
 220         .vfs_vptofh     = hammer2_vfs_vptofh,
 221         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 222         .vfs_checkexp   = hammer2_vfs_checkexp
 223 };
 224
 225 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 226
 227 VFS_SET(hammer2_vfsops, hammer2, 0);
 228 MODULE_VERSION(hammer2, 1);
 229
 230 static
 231 int
 232 hammer2_vfs_init(struct vfsconf *conf)
 233 {
 234         static struct objcache_malloc_args margs_read;
 235         static struct objcache_malloc_args margs_write;
 236         static struct objcache_malloc_args margs_vop;
 237
 238         int error;
 239
 240         error = 0;
 241
 242         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 243                 error = EINVAL;
 244         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 245                 error = EINVAL;
 246         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 247                 error = EINVAL;
 248
 249         if (error)
 250                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 251
 252         margs_read.objsize = 65536;
 253         margs_read.mtype = M_HAMMER2_DEBUFFER;
 254
 255         margs_write.objsize = 32768;
 256         margs_write.mtype = M_HAMMER2_CBUFFER;
 257
 258         margs_vop.objsize = sizeof(hammer2_xop_t);
 259         margs_vop.mtype = M_HAMMER2;
 260
 261         /*
 262          * Note thaht for the XOPS cache we want backing store allocations
 263          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 264          * confusion), so use the backing store function that does it.  This
 265          * means that initial XOPS objects are zerod but REUSED objects are
 266          * not.  So we are responsible for cleaning the object up sufficiently
 267          * for our needs before objcache_put()ing it back (typically just the
 268          * FIFO indices).
 269          */
 270         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 271                                 0, 1, NULL, NULL, NULL,
 272                                 objcache_malloc_alloc,
 273                                 objcache_malloc_free,
 274                                 &margs_read);
 275         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 276                                 0, 1, NULL, NULL, NULL,
 277                                 objcache_malloc_alloc,
 278                                 objcache_malloc_free,
 279                                 &margs_write);
 280         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 281                                 0, 1, NULL, NULL, NULL,
 282                                 objcache_malloc_alloc_zero,
 283                                 objcache_malloc_free,
 284                                 &margs_vop);
 285
 286
 287         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 288         TAILQ_INIT(&hammer2_mntlist);
 289         TAILQ_INIT(&hammer2_pfslist);
 290
 291         hammer2_limit_dirty_chains = maxvnodes / 10;
 292         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 293                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 294
 295         return (error);
 296 }
 297
 298 static
 299 int
 300 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 301 {
 302         objcache_destroy(cache_buffer_read);
 303         objcache_destroy(cache_buffer_write);
 304         objcache_destroy(cache_xops);
 305         return 0;
 306 }
 307
 308 /*
 309  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
 310  * mounts and the spmp structure for media (hmp) structures.
 311  *
 312  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 313  * transactions.  Note that synchronization does not use this field.
 314  * (typically frontend operations and synchronization cannot run on the
 315  * same PFS node at the same time).
 316  *
 317  * XXX check locking
 318  */
 319 hammer2_pfs_t *
 320 hammer2_pfsalloc(hammer2_chain_t *chain, const hammer2_inode_data_t *ripdata,
 321                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 322 {
 323         hammer2_inode_t *iroot;
 324         hammer2_pfs_t *pmp;
 325         int count;
 326         int i;
 327         int j;
 328
 329         /*
 330          * Locate or create the PFS based on the cluster id.  If ripdata
 331          * is NULL this is a spmp which is unique and is always allocated.
 332          *
 333          * If the device is mounted in local mode all PFSs are considered
 334          * independent and not part of any cluster (for debugging only).
 335          */
 336         if (ripdata) {
 337                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 338                         if (force_local != pmp->force_local)
 339                                 continue;
 340                         if (force_local == NULL &&
 341                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 342                                  sizeof(pmp->pfs_clid)) == 0) {
 343                                         break;
 344                         } else if (force_local && pmp->pfs_names[0] &&
 345                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 346                                         break;
 347                         }
 348                 }
 349         } else {
 350                 pmp = NULL;
 351         }
 352
 353         if (pmp == NULL) {
 354                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 355                 pmp->force_local = force_local;
 356                 hammer2_trans_manage_init(pmp);
 357                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 358                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 359                 lockinit(&pmp->lock, "pfslk", 0, 0);
 360                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 361                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 362                 spin_init(&pmp->xop_spin, "h2xop");
 363                 spin_init(&pmp->lru_spin, "h2lru");
 364                 RB_INIT(&pmp->inum_tree);
 365                 TAILQ_INIT(&pmp->sideq);
 366                 TAILQ_INIT(&pmp->lru_list);
 367                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 368
 369                 /*
 370                  * Distribute backend operations to threads
 371                  */
 372                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 373                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 374
 375                 /*
 376                  * Save the last media transaction id for the flusher.  Set
 377                  * initial
 378                  */
 379                 if (ripdata)
 380                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 381                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 382
 383                 /*
 384                  * The synchronization thread may start too early, make
 385                  * sure it stays frozen until we are ready to let it go.
 386                  * XXX
 387                  */
 388                 /*
 389                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 390                                          HAMMER2_THREAD_REMASTER;
 391                 */
 392         }
 393
 394         /*
 395          * Create the PFS's root inode.
 396          */
 397         if ((iroot = pmp->iroot) == NULL) {
 398                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 399                 pmp->iroot = iroot;
 400                 hammer2_inode_ref(iroot);
 401                 hammer2_inode_unlock(iroot);
 402         }
 403
 404         /*
 405          * Stop here if no chain is passed in.
 406          */
 407         if (chain == NULL)
 408                 goto done;
 409
 410         /*
 411          * When a chain is passed in we must add it to the PFS's root
 412          * inode, update pmp->pfs_types[], and update the syncronization
 413          * threads.
 414          *
 415          * When forcing local mode, mark the PFS as a MASTER regardless.
 416          *
 417          * At the moment empty spots can develop due to removals or failures.
 418          * Ultimately we want to re-fill these spots but doing so might
 419          * confused running code. XXX
 420          */
 421         hammer2_inode_ref(iroot);
 422         hammer2_mtx_ex(&iroot->lock);
 423         j = iroot->cluster.nchains;
 424
 425         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 426
 427         if (j == HAMMER2_MAXCLUSTER) {
 428                 kprintf("hammer2_mount: cluster full!\n");
 429                 /* XXX fatal error? */
 430         } else {
 431                 KKASSERT(chain->pmp == NULL);
 432                 chain->pmp = pmp;
 433                 hammer2_chain_ref(chain);
 434                 iroot->cluster.array[j].chain = chain;
 435                 if (force_local)
 436                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 437                 else
 438                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 439                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 440                 pmp->pfs_hmps[j] = chain->hmp;
 441
 442                 /*
 443                  * If the PFS is already mounted we must account
 444                  * for the mount_count here.
 445                  */
 446                 if (pmp->mp)
 447                         ++chain->hmp->mount_count;
 448
 449                 /*
 450                  * May have to fixup dirty chain tracking.  Previous
 451                  * pmp was NULL so nothing to undo.
 452                  */
 453                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 454                         hammer2_pfs_memory_inc(pmp);
 455                 ++j;
 456         }
 457         iroot->cluster.nchains = j;
 458
 459         /*
 460          * Update nmasters from any PFS inode which is part of the cluster.
 461          * It is possible that this will result in a value which is too
 462          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 463          * override this value later on.
 464          *
 465          * (This informs us of masters that might not currently be
 466          *  discoverable by this mount).
 467          */
 468         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 469                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 470         }
 471
 472         /*
 473          * Count visible masters.  Masters are usually added with
 474          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 475          * are more (XXX and must update the master inodes).
 476          */
 477         count = 0;
 478         for (i = 0; i < iroot->cluster.nchains; ++i) {
 479                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 480                         ++count;
 481         }
 482         if (pmp->pfs_nmasters < count)
 483                 pmp->pfs_nmasters = count;
 484
 485         /*
 486          * Create missing synchronization and support threads.
 487          *
 488          * Single-node masters (including snapshots) have nothing to
 489          * synchronize and do not require this thread.
 490          *
 491          * Multi-node masters or any number of soft masters, slaves, copy,
 492          * or other PFS types need the thread.
 493          *
 494          * Each thread is responsible for its particular cluster index.
 495          * We use independent threads so stalls or mismatches related to
 496          * any given target do not affect other targets.
 497          */
 498         for (i = 0; i < iroot->cluster.nchains; ++i) {
 499                 /*
 500                  * Single-node masters (including snapshots) have nothing
 501                  * to synchronize and will make direct xops support calls,
 502                  * thus they do not require this thread.
 503                  *
 504                  * Note that there can be thousands of snapshots.  We do not
 505                  * want to create thousands of threads.
 506                  */
 507                 if (pmp->pfs_nmasters <= 1 &&
 508                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 509                         continue;
 510                 }
 511
 512                 /*
 513                  * Sync support thread
 514                  */
 515                 if (pmp->sync_thrs[i].td == NULL) {
 516                         hammer2_thr_create(&pmp->sync_thrs[i], pmp,
 517                                            "h2nod", i, -1,
 518                                            hammer2_primary_sync_thread);
 519                 }
 520         }
 521
 522         /*
 523          * Create missing Xop threads
 524          */
 525         if (pmp->mp)
 526                 hammer2_xop_helper_create(pmp);
 527
 528         hammer2_mtx_unlock(&iroot->lock);
 529         hammer2_inode_drop(iroot);
 530 done:
 531         return pmp;
 532 }
 533
 534 /*
 535  * Destroy a PFS, typically only occurs after the last mount on a device
 536  * has gone away.
 537  */
 538 static void
 539 hammer2_pfsfree(hammer2_pfs_t *pmp)
 540 {
 541         hammer2_inode_t *iroot;
 542         hammer2_chain_t *chain;
 543         int i;
 544         int j;
 545
 546         /*
 547          * Cleanup our reference on iroot.  iroot is (should) not be needed
 548          * by the flush code.
 549          */
 550         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 551
 552         iroot = pmp->iroot;
 553         if (iroot) {
 554                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 555                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 556                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 557                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 558                 }
 559 #if REPORT_REFS_ERRORS
 560                 if (pmp->iroot->refs != 1)
 561                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 562                                 pmp->iroot, pmp->iroot->refs);
 563 #else
 564                 KKASSERT(pmp->iroot->refs == 1);
 565 #endif
 566                 /* ref for pmp->iroot */
 567                 hammer2_inode_drop(pmp->iroot);
 568                 pmp->iroot = NULL;
 569         }
 570
 571         /*
 572          * Cleanup chains remaining on LRU list.
 573          */
 574         kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
 575         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 576                 hammer2_chain_ref(chain);
 577                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 578                 hammer2_chain_drop(chain);
 579         }
 580
 581         /*
 582          * Free remaining pmp resources
 583          */
 584         kmalloc_destroy(&pmp->mmsg);
 585         kmalloc_destroy(&pmp->minode);
 586
 587         kfree(pmp, M_HAMMER2);
 588 }
 589
 590 /*
 591  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 592  * empty is terminated and freed.
 593  *
 594  * XXX inefficient.
 595  */
 596 static void
 597 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 598 {
 599         hammer2_pfs_t *pmp;
 600         hammer2_inode_t *iroot;
 601         hammer2_chain_t *rchain;
 602         int didfreeze;
 603         int i;
 604         int j;
 605
 606 again:
 607         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 608                 if ((iroot = pmp->iroot) == NULL)
 609                         continue;
 610                 if (hmp->spmp == pmp) {
 611                         kprintf("unmount hmp %p remove spmp %p\n",
 612                                 hmp, pmp);
 613                         hmp->spmp = NULL;
 614                 }
 615
 616                 /*
 617                  * Determine if this PFS is affected.  If it is we must
 618                  * freeze all management threads and lock its iroot.
 619                  *
 620                  * Freezing a management thread forces it idle, operations
 621                  * in-progress will be aborted and it will have to start
 622                  * over again when unfrozen, or exit if told to exit.
 623                  */
 624                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 625                         if (pmp->pfs_hmps[i] == hmp)
 626                                 break;
 627                 }
 628                 if (i != HAMMER2_MAXCLUSTER) {
 629                         /*
 630                          * Make sure all synchronization threads are locked
 631                          * down.
 632                          */
 633                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 634                                 if (pmp->pfs_hmps[i] == NULL)
 635                                         continue;
 636                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 637                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 638                                         hammer2_thr_freeze_async(
 639                                                 &pmp->xop_groups[j].thrs[i]);
 640                                 }
 641                         }
 642                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 643                                 if (pmp->pfs_hmps[i] == NULL)
 644                                         continue;
 645                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 646                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 647                                         hammer2_thr_freeze(
 648                                                 &pmp->xop_groups[j].thrs[i]);
 649                                 }
 650                         }
 651
 652                         /*
 653                          * Lock the inode and clean out matching chains.
 654                          * Note that we cannot use hammer2_inode_lock_*()
 655                          * here because that would attempt to validate the
 656                          * cluster that we are in the middle of ripping
 657                          * apart.
 658                          *
 659                          * WARNING! We are working directly on the inodes
 660                          *          embedded cluster.
 661                          */
 662                         hammer2_mtx_ex(&iroot->lock);
 663
 664                         /*
 665                          * Remove the chain from matching elements of the PFS.
 666                          */
 667                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 668                                 if (pmp->pfs_hmps[i] != hmp)
 669                                         continue;
 670                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 671                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 672                                         hammer2_thr_delete(
 673                                                 &pmp->xop_groups[j].thrs[i]);
 674                                 }
 675                                 rchain = iroot->cluster.array[i].chain;
 676                                 iroot->cluster.array[i].chain = NULL;
 677                                 pmp->pfs_types[i] = 0;
 678                                 if (pmp->pfs_names[i]) {
 679                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 680                                         pmp->pfs_names[i] = NULL;
 681                                 }
 682                                 if (rchain) {
 683                                         hammer2_chain_drop(rchain);
 684                                         /* focus hint */
 685                                         if (iroot->cluster.focus == rchain)
 686                                                 iroot->cluster.focus = NULL;
 687                                 }
 688                                 pmp->pfs_hmps[i] = NULL;
 689                         }
 690                         hammer2_mtx_unlock(&iroot->lock);
 691                         didfreeze = 1;  /* remaster, unfreeze down below */
 692                 } else {
 693                         didfreeze = 0;
 694                 }
 695
 696                 /*
 697                  * Cleanup trailing chains.  Gaps may remain.
 698                  */
 699                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 700                         if (pmp->pfs_hmps[i])
 701                                 break;
 702                 }
 703                 iroot->cluster.nchains = i + 1;
 704
 705                 /*
 706                  * If the PMP has no elements remaining we can destroy it.
 707                  * (this will transition management threads from frozen->exit).
 708                  */
 709                 if (iroot->cluster.nchains == 0) {
 710                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 711                                 hmp, pmp);
 712                         hammer2_pfsfree(pmp);
 713                         goto again;
 714                 }
 715
 716                 /*
 717                  * If elements still remain we need to set the REMASTER
 718                  * flag and unfreeze it.
 719                  */
 720                 if (didfreeze) {
 721                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 722                                 if (pmp->pfs_hmps[i] == NULL)
 723                                         continue;
 724                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 725                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 726                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 727                                         hammer2_thr_remaster(
 728                                                 &pmp->xop_groups[j].thrs[i]);
 729                                         hammer2_thr_unfreeze(
 730                                                 &pmp->xop_groups[j].thrs[i]);
 731                                 }
 732                         }
 733                 }
 734         }
 735 }
 736
 737 /*
 738  * Mount or remount HAMMER2 fileystem from physical media
 739  *
 740  *      mountroot
 741  *              mp              mount point structure
 742  *              path            NULL
 743  *              data            <unused>
 744  *              cred            <unused>
 745  *
 746  *      mount
 747  *              mp              mount point structure
 748  *              path            path to mount point
 749  *              data            pointer to argument structure in user space
 750  *                      volume  volume path (device@LABEL form)
 751  *                      hflags  user mount flags
 752  *              cred            user credentials
 753  *
 754  * RETURNS:     0       Success
 755  *              !0      error number
 756  */
 757 static
 758 int
 759 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 760                   struct ucred *cred)
 761 {
 762         struct hammer2_mount_info info;
 763         hammer2_pfs_t *pmp;
 764         hammer2_pfs_t *spmp;
 765         hammer2_dev_t *hmp;
 766         hammer2_dev_t *force_local;
 767         hammer2_key_t key_next;
 768         hammer2_key_t key_dummy;
 769         hammer2_key_t lhc;
 770         struct vnode *devvp;
 771         struct nlookupdata nd;
 772         hammer2_chain_t *parent;
 773         hammer2_chain_t *chain;
 774         hammer2_cluster_t *cluster;
 775         const hammer2_inode_data_t *ripdata;
 776         hammer2_blockref_t bref;
 777         struct file *fp;
 778         char devstr[MNAMELEN];
 779         size_t size;
 780         size_t done;
 781         char *dev;
 782         char *label;
 783         int ronly = 1;
 784         int error;
 785         int cache_index;
 786         int i;
 787
 788         hmp = NULL;
 789         pmp = NULL;
 790         dev = NULL;
 791         label = NULL;
 792         devvp = NULL;
 793         cache_index = -1;
 794
 795         kprintf("hammer2_mount\n");
 796
 797         if (path == NULL) {
 798                 /*
 799                  * Root mount
 800                  */
 801                 bzero(&info, sizeof(info));
 802                 info.cluster_fd = -1;
 803                 ksnprintf(devstr, sizeof(devstr), "%s",
 804                           mp->mnt_stat.f_mntfromname);
 805                 kprintf("hammer2_mount: root '%s'\n", devstr);
 806         } else {
 807                 /*
 808                  * Non-root mount or updating a mount
 809                  */
 810                 error = copyin(data, &info, sizeof(info));
 811                 if (error)
 812                         return (error);
 813
 814                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 815                 if (error)
 816                         return (error);
 817         }
 818
 819         /* Extract device and label */
 820         dev = devstr;
 821         label = strchr(devstr, '@');
 822         if (label == NULL ||
 823             ((label + 1) - dev) > done) {
 824                 return (EINVAL);
 825         }
 826         *label = '\0';
 827         label++;
 828         if (*label == '\0')
 829                 return (EINVAL);
 830
 831         if (mp->mnt_flag & MNT_UPDATE) {
 832                 /*
 833                  * Update mount.  Note that pmp->iroot->cluster is
 834                  * an inode-embedded cluster and thus cannot be
 835                  * directly locked.
 836                  *
 837                  * XXX HAMMER2 needs to implement NFS export via
 838                  *     mountctl.
 839                  */
 840                 pmp = MPTOPMP(mp);
 841                 pmp->hflags = info.hflags;
 842                 cluster = &pmp->iroot->cluster;
 843                 for (i = 0; i < cluster->nchains; ++i) {
 844                         if (cluster->array[i].chain == NULL)
 845                                 continue;
 846                         hmp = cluster->array[i].chain->hmp;
 847                         devvp = hmp->devvp;
 848                         error = hammer2_remount(hmp, mp, path,
 849                                                 devvp, cred);
 850                         if (error)
 851                                 break;
 852                 }
 853
 854                 return error;
 855         }
 856
 857         /*
 858          * HMP device mount
 859          *
 860          * Lookup name and verify it refers to a block device.
 861          */
 862         if (path) {
 863                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 864                 if (error == 0)
 865                         error = nlookup(&nd);
 866                 if (error == 0)
 867                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 868                 nlookup_done(&nd);
 869         } else {
 870                 /* root mount */
 871                 cdev_t cdev = kgetdiskbyname(dev);
 872                 error = bdevvp(cdev, &devvp);
 873                 if (error)
 874                         kprintf("hammer2: cannot find '%s'\n", dev);
 875         }
 876
 877         if (error == 0) {
 878                 if (vn_isdisk(devvp, &error))
 879                         error = vfs_mountedon(devvp);
 880         }
 881
 882         /*
 883          * Determine if the device has already been mounted.  After this
 884          * check hmp will be non-NULL if we are doing the second or more
 885          * hammer2 mounts from the same device.
 886          */
 887         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 888         TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 889                 if (hmp->devvp == devvp)
 890                         break;
 891         }
 892
 893         /*
 894          * Open the device if this isn't a secondary mount and construct
 895          * the H2 device mount (hmp).
 896          */
 897         if (hmp == NULL) {
 898                 hammer2_chain_t *schain;
 899                 hammer2_xid_t xid;
 900
 901                 if (error == 0 && vcount(devvp) > 0)
 902                         error = EBUSY;
 903
 904                 /*
 905                  * Now open the device
 906                  */
 907                 if (error == 0) {
 908                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
 909                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
 910                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
 911                         if (error == 0) {
 912                                 error = VOP_OPEN(devvp,
 913                                                  ronly ? FREAD : FREAD | FWRITE,
 914                                                  FSCRED, NULL);
 915                         }
 916                         vn_unlock(devvp);
 917                 }
 918                 if (error && devvp) {
 919                         vrele(devvp);
 920                         devvp = NULL;
 921                 }
 922                 if (error) {
 923                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 924                         return error;
 925                 }
 926                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
 927                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
 928                 hmp->ronly = ronly;
 929                 hmp->devvp = devvp;
 930                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
 931                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
 932                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
 933                 RB_INIT(&hmp->iotree);
 934                 spin_init(&hmp->io_spin, "hm2mount_io");
 935                 spin_init(&hmp->list_spin, "hm2mount_list");
 936                 TAILQ_INIT(&hmp->flushq);
 937
 938                 lockinit(&hmp->vollk, "h2vol", 0, 0);
 939                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
 940
 941                 /*
 942                  * vchain setup. vchain.data is embedded.
 943                  * vchain.refs is initialized and will never drop to 0.
 944                  *
 945                  * NOTE! voldata is not yet loaded.
 946                  */
 947                 hmp->vchain.hmp = hmp;
 948                 hmp->vchain.refs = 1;
 949                 hmp->vchain.data = (void *)&hmp->voldata;
 950                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
 951                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 952                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 953
 954                 hammer2_chain_core_init(&hmp->vchain);
 955                 /* hmp->vchain.u.xxx is left NULL */
 956
 957                 /*
 958                  * fchain setup.  fchain.data is embedded.
 959                  * fchain.refs is initialized and will never drop to 0.
 960                  *
 961                  * The data is not used but needs to be initialized to
 962                  * pass assertion muster.  We use this chain primarily
 963                  * as a placeholder for the freemap's top-level RBTREE
 964                  * so it does not interfere with the volume's topology
 965                  * RBTREE.
 966                  */
 967                 hmp->fchain.hmp = hmp;
 968                 hmp->fchain.refs = 1;
 969                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
 970                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
 971                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 972                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 973                 hmp->fchain.bref.methods =
 974                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
 975                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
 976
 977                 hammer2_chain_core_init(&hmp->fchain);
 978                 /* hmp->fchain.u.xxx is left NULL */
 979
 980                 /*
 981                  * Install the volume header and initialize fields from
 982                  * voldata.
 983                  */
 984                 error = hammer2_install_volume_header(hmp);
 985                 if (error) {
 986                         hammer2_unmount_helper(mp, NULL, hmp);
 987                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 988                         hammer2_vfs_unmount(mp, MNT_FORCE);
 989                         return error;
 990                 }
 991
 992                 /*
 993                  * Really important to get these right or flush will get
 994                  * confused.
 995                  */
 996                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
 997                 kprintf("alloc spmp %p tid %016jx\n",
 998                         hmp->spmp, hmp->voldata.mirror_tid);
 999                 spmp = hmp->spmp;
1000
1001                 /*
1002                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1003                  * is inherited from the volume header.
1004                  */
1005                 xid = 0;
1006                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1007                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1008                 hmp->vchain.pmp = spmp;
1009                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1010                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1011                 hmp->fchain.pmp = spmp;
1012
1013                 /*
1014                  * First locate the super-root inode, which is key 0
1015                  * relative to the volume header's blockset.
1016                  *
1017                  * Then locate the root inode by scanning the directory keyspace
1018                  * represented by the label.
1019                  */
1020                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1021                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1022                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1023                                       &cache_index, 0);
1024                 hammer2_chain_lookup_done(parent);
1025                 if (schain == NULL) {
1026                         kprintf("hammer2_mount: invalid super-root\n");
1027                         hammer2_unmount_helper(mp, NULL, hmp);
1028                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1029                         hammer2_vfs_unmount(mp, MNT_FORCE);
1030                         return EINVAL;
1031                 }
1032                 if (schain->error) {
1033                         kprintf("hammer2_mount: error %s reading super-root\n",
1034                                 hammer2_error_str(schain->error));
1035                         hammer2_chain_unlock(schain);
1036                         hammer2_chain_drop(schain);
1037                         schain = NULL;
1038                         hammer2_unmount_helper(mp, NULL, hmp);
1039                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1040                         hammer2_vfs_unmount(mp, MNT_FORCE);
1041                         return EINVAL;
1042                 }
1043
1044                 /*
1045                  * The super-root always uses an inode_tid of 1 when
1046                  * creating PFSs.
1047                  */
1048                 spmp->inode_tid = 1;
1049                 spmp->modify_tid = schain->bref.modify_tid + 1;
1050
1051                 /*
1052                  * Sanity-check schain's pmp and finish initialization.
1053                  * Any chain belonging to the super-root topology should
1054                  * have a NULL pmp (not even set to spmp).
1055                  */
1056                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1057                 KKASSERT(schain->pmp == NULL);
1058                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1059
1060                 /*
1061                  * Replace the dummy spmp->iroot with a real one.  It's
1062                  * easier to just do a wholesale replacement than to try
1063                  * to update the chain and fixup the iroot fields.
1064                  *
1065                  * The returned inode is locked with the supplied cluster.
1066                  */
1067                 cluster = hammer2_cluster_from_chain(schain);
1068                 hammer2_inode_drop(spmp->iroot);
1069                 spmp->iroot = NULL;
1070                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1071                 spmp->spmp_hmp = hmp;
1072                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1073                 spmp->pfs_hmps[0] = hmp;
1074                 hammer2_inode_ref(spmp->iroot);
1075                 hammer2_inode_unlock(spmp->iroot);
1076                 hammer2_cluster_unlock(cluster);
1077                 hammer2_cluster_drop(cluster);
1078                 schain = NULL;
1079                 /* leave spmp->iroot with one ref */
1080
1081                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1082                         error = hammer2_recovery(hmp);
1083                         /* XXX do something with error */
1084                 }
1085                 hammer2_update_pmps(hmp);
1086                 hammer2_iocom_init(hmp);
1087
1088                 /*
1089                  * Ref the cluster management messaging descriptor.  The mount
1090                  * program deals with the other end of the communications pipe.
1091                  *
1092                  * Root mounts typically do not supply one.
1093                  */
1094                 if (info.cluster_fd >= 0) {
1095                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1096                         if (fp) {
1097                                 hammer2_cluster_reconnect(hmp, fp);
1098                         } else {
1099                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1100                         }
1101                 }
1102         } else {
1103                 spmp = hmp->spmp;
1104                 if (info.hflags & HMNT2_DEVFLAGS) {
1105                         kprintf("hammer2: Warning: mount flags pertaining "
1106                                 "to the whole device may only be specified "
1107                                 "on the first mount of the device: %08x\n",
1108                                 info.hflags & HMNT2_DEVFLAGS);
1109                 }
1110         }
1111
1112         /*
1113          * Force local mount (disassociate all PFSs from their clusters).
1114          * Used primarily for debugging.
1115          */
1116         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1117
1118         /*
1119          * Lookup the mount point under the media-localized super-root.
1120          * Scanning hammer2_pfslist doesn't help us because it represents
1121          * PFS cluster ids which can aggregate several named PFSs together.
1122          *
1123          * cluster->pmp will incorrectly point to spmp and must be fixed
1124          * up later on.
1125          */
1126         hammer2_inode_lock(spmp->iroot, 0);
1127         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1128         lhc = hammer2_dirhash(label, strlen(label));
1129         chain = hammer2_chain_lookup(&parent, &key_next,
1130                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1131                                      &cache_index, 0);
1132         while (chain) {
1133                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1134                     strcmp(label, chain->data->ipdata.filename) == 0) {
1135                         break;
1136                 }
1137                 chain = hammer2_chain_next(&parent, chain, &key_next,
1138                                             key_next,
1139                                             lhc + HAMMER2_DIRHASH_LOMASK,
1140                                             &cache_index, 0);
1141         }
1142         if (parent) {
1143                 hammer2_chain_unlock(parent);
1144                 hammer2_chain_drop(parent);
1145         }
1146         hammer2_inode_unlock(spmp->iroot);
1147
1148         /*
1149          * PFS could not be found?
1150          */
1151         if (chain == NULL) {
1152                 kprintf("hammer2_mount: PFS label not found\n");
1153                 hammer2_unmount_helper(mp, NULL, hmp);
1154                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1155                 hammer2_vfs_unmount(mp, MNT_FORCE);
1156
1157                 return EINVAL;
1158         }
1159
1160         /*
1161          * Acquire the pmp structure (it should have already been allocated
1162          * via hammer2_update_pmps() so do not pass cluster in to add to
1163          * available chains).
1164          *
1165          * Check if the cluster has already been mounted.  A cluster can
1166          * only be mounted once, use null mounts to mount additional copies.
1167          */
1168         ripdata = &chain->data->ipdata;
1169         bref = chain->bref;
1170         pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid, force_local);
1171         hammer2_chain_unlock(chain);
1172         hammer2_chain_drop(chain);
1173
1174         if (pmp->mp) {
1175                 kprintf("hammer2_mount: PFS already mounted!\n");
1176                 hammer2_unmount_helper(mp, NULL, hmp);
1177                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1178                 hammer2_vfs_unmount(mp, MNT_FORCE);
1179
1180                 return EBUSY;
1181         }
1182
1183         /*
1184          * Finish the mount
1185          */
1186         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1187
1188         pmp->hflags = info.hflags;
1189         mp->mnt_flag = MNT_LOCAL;
1190         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1191         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1192
1193         /*
1194          * required mount structure initializations
1195          */
1196         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1197         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1198
1199         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1200         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1201
1202         /*
1203          * Optional fields
1204          */
1205         mp->mnt_iosize_max = MAXPHYS;
1206
1207         /*
1208          * Connect up mount pointers.
1209          */
1210         hammer2_mount_helper(mp, pmp);
1211
1212         lockmgr(&hammer2_mntlk, LK_RELEASE);
1213
1214         /*
1215          * Finish setup
1216          */
1217         vfs_getnewfsid(mp);
1218         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1219         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1220         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1221
1222         if (path) {
1223                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1224                           MNAMELEN - 1, &size);
1225                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1226         } /* else root mount, already in there */
1227
1228         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1229         if (path) {
1230                 copyinstr(path, mp->mnt_stat.f_mntonname,
1231                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1232                           &size);
1233         } else {
1234                 /* root mount */
1235                 mp->mnt_stat.f_mntonname[0] = '/';
1236         }
1237
1238         /*
1239          * Initial statfs to prime mnt_stat.
1240          */
1241         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1242
1243         return 0;
1244 }
1245
1246 /*
1247  * Scan PFSs under the super-root and create hammer2_pfs structures.
1248  */
1249 static
1250 void
1251 hammer2_update_pmps(hammer2_dev_t *hmp)
1252 {
1253         const hammer2_inode_data_t *ripdata;
1254         hammer2_chain_t *parent;
1255         hammer2_chain_t *chain;
1256         hammer2_blockref_t bref;
1257         hammer2_dev_t *force_local;
1258         hammer2_pfs_t *spmp;
1259         hammer2_pfs_t *pmp;
1260         hammer2_key_t key_next;
1261         int cache_index = -1;
1262
1263         /*
1264          * Force local mount (disassociate all PFSs from their clusters).
1265          * Used primarily for debugging.
1266          */
1267         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1268
1269         /*
1270          * Lookup mount point under the media-localized super-root.
1271          *
1272          * cluster->pmp will incorrectly point to spmp and must be fixed
1273          * up later on.
1274          */
1275         spmp = hmp->spmp;
1276         hammer2_inode_lock(spmp->iroot, 0);
1277         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1278         chain = hammer2_chain_lookup(&parent, &key_next,
1279                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1280                                          &cache_index, 0);
1281         while (chain) {
1282                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1283                         continue;
1284                 ripdata = &chain->data->ipdata;
1285                 bref = chain->bref;
1286                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1287
1288                 pmp = hammer2_pfsalloc(chain, ripdata,
1289                                        bref.modify_tid, force_local);
1290                 chain = hammer2_chain_next(&parent, chain, &key_next,
1291                                            key_next, HAMMER2_KEY_MAX,
1292                                            &cache_index, 0);
1293         }
1294         if (parent) {
1295                 hammer2_chain_unlock(parent);
1296                 hammer2_chain_drop(parent);
1297         }
1298         hammer2_inode_unlock(spmp->iroot);
1299 }
1300
1301 static
1302 int
1303 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1304                 struct vnode *devvp, struct ucred *cred)
1305 {
1306         int error;
1307
1308         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1309                 error = hammer2_recovery(hmp);
1310         } else {
1311                 error = 0;
1312         }
1313         return error;
1314 }
1315
1316 static
1317 int
1318 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1319 {
1320         hammer2_pfs_t *pmp;
1321         int flags;
1322         int error = 0;
1323
1324         pmp = MPTOPMP(mp);
1325
1326         if (pmp == NULL)
1327                 return(0);
1328
1329         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1330
1331         /*
1332          * If mount initialization proceeded far enough we must flush
1333          * its vnodes and sync the underlying mount points.  Three syncs
1334          * are required to fully flush the filesystem (freemap updates lag
1335          * by one flush, and one extra for safety).
1336          */
1337         if (mntflags & MNT_FORCE)
1338                 flags = FORCECLOSE;
1339         else
1340                 flags = 0;
1341         if (pmp->iroot) {
1342                 error = vflush(mp, 0, flags);
1343                 if (error)
1344                         goto failed;
1345                 hammer2_vfs_sync(mp, MNT_WAIT);
1346                 hammer2_vfs_sync(mp, MNT_WAIT);
1347                 hammer2_vfs_sync(mp, MNT_WAIT);
1348         }
1349
1350         /*
1351          * Cleanup the frontend support XOPS threads
1352          */
1353         hammer2_xop_helper_cleanup(pmp);
1354
1355         if (pmp->mp)
1356                 hammer2_unmount_helper(mp, pmp, NULL);
1357
1358         error = 0;
1359 failed:
1360         lockmgr(&hammer2_mntlk, LK_RELEASE);
1361
1362         return (error);
1363 }
1364
1365 /*
1366  * Mount helper, hook the system mount into our PFS.
1367  * The mount lock is held.
1368  *
1369  * We must bump the mount_count on related devices for any
1370  * mounted PFSs.
1371  */
1372 static
1373 void
1374 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1375 {
1376         hammer2_cluster_t *cluster;
1377         hammer2_chain_t *rchain;
1378         int i;
1379
1380         mp->mnt_data = (qaddr_t)pmp;
1381         pmp->mp = mp;
1382
1383         /*
1384          * After pmp->mp is set we have to adjust hmp->mount_count.
1385          */
1386         cluster = &pmp->iroot->cluster;
1387         for (i = 0; i < cluster->nchains; ++i) {
1388                 rchain = cluster->array[i].chain;
1389                 if (rchain == NULL)
1390                         continue;
1391                 ++rchain->hmp->mount_count;
1392                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1393                         rchain->hmp, rchain->hmp->mount_count);
1394         }
1395
1396         /*
1397          * Create missing Xop threads
1398          */
1399         hammer2_xop_helper_create(pmp);
1400 }
1401
1402 /*
1403  * Mount helper, unhook the system mount from our PFS.
1404  * The mount lock is held.
1405  *
1406  * If hmp is supplied a mount responsible for being the first to open
1407  * the block device failed and the block device and all PFSs using the
1408  * block device must be cleaned up.
1409  *
1410  * If pmp is supplied multiple devices might be backing the PFS and each
1411  * must be disconnected.  This might not be the last PFS using some of the
1412  * underlying devices.  Also, we have to adjust our hmp->mount_count
1413  * accounting for the devices backing the pmp which is now undergoing an
1414  * unmount.
1415  */
1416 static
1417 void
1418 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1419 {
1420         hammer2_cluster_t *cluster;
1421         hammer2_chain_t *rchain;
1422         struct vnode *devvp;
1423         int dumpcnt;
1424         int ronly = 0;
1425         int i;
1426
1427         /*
1428          * If no device supplied this is a high-level unmount and we have to
1429          * to disconnect the mount, adjust mount_count, and locate devices
1430          * that might now have no mounts.
1431          */
1432         if (pmp) {
1433                 KKASSERT(hmp == NULL);
1434                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1435                 pmp->mp = NULL;
1436                 mp->mnt_data = NULL;
1437
1438                 /*
1439                  * After pmp->mp is cleared we have to account for
1440                  * mount_count.
1441                  */
1442                 cluster = &pmp->iroot->cluster;
1443                 for (i = 0; i < cluster->nchains; ++i) {
1444                         rchain = cluster->array[i].chain;
1445                         if (rchain == NULL)
1446                                 continue;
1447                         --rchain->hmp->mount_count;
1448                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1449                                 rchain->hmp, rchain->hmp->mount_count);
1450                         /* scrapping hmp now may invalidate the pmp */
1451                 }
1452 again:
1453                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1454                         if (hmp->mount_count == 0) {
1455                                 hammer2_unmount_helper(NULL, NULL, hmp);
1456                                 goto again;
1457                         }
1458                 }
1459                 return;
1460         }
1461
1462         /*
1463          * Try to terminate the block device.  We can't terminate it if
1464          * there are still PFSs referencing it.
1465          */
1466         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1467                 hmp, hmp->mount_count);
1468         if (hmp->mount_count)
1469                 return;
1470
1471         hammer2_pfsfree_scan(hmp);
1472         hammer2_dev_exlock(hmp);        /* XXX order */
1473
1474         /*
1475          * Cycle the volume data lock as a safety (probably not needed any
1476          * more).  To ensure everything is out we need to flush at least
1477          * three times.  (1) The running of the sideq can dirty the
1478          * filesystem, (2) A normal flush can dirty the freemap, and
1479          * (3) ensure that the freemap is fully synchronized.
1480          *
1481          * The next mount's recovery scan can clean everything up but we want
1482          * to leave the filesystem in a 100% clean state on a normal unmount.
1483          */
1484 #if 0
1485         hammer2_voldata_lock(hmp);
1486         hammer2_voldata_unlock(hmp);
1487 #endif
1488         hammer2_iocom_uninit(hmp);
1489
1490         if ((hmp->vchain.flags | hmp->fchain.flags) &
1491             HAMMER2_CHAIN_FLUSH_MASK) {
1492                 kprintf("hammer2_unmount: chains left over "
1493                         "after final sync\n");
1494                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1495                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1496
1497                 if (hammer2_debug & 0x0010)
1498                         Debugger("entered debugger");
1499         }
1500
1501         KKASSERT(hmp->spmp == NULL);
1502
1503         /*
1504          * Finish up with the device vnode
1505          */
1506         if ((devvp = hmp->devvp) != NULL) {
1507                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1508                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1509                 hmp->devvp = NULL;
1510                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1511                 vn_unlock(devvp);
1512                 vrele(devvp);
1513                 devvp = NULL;
1514         }
1515
1516         /*
1517          * Clear vchain/fchain flags that might prevent final cleanup
1518          * of these chains.
1519          */
1520         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1521                 atomic_add_long(&hammer2_count_modified_chains, -1);
1522                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1523                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1524         }
1525         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1526                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1527         }
1528
1529         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1530                 atomic_add_long(&hammer2_count_modified_chains, -1);
1531                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1532                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1533         }
1534         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1535                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1536         }
1537
1538         /*
1539          * Final drop of embedded freemap root chain to
1540          * clean up fchain.core (fchain structure is not
1541          * flagged ALLOCATED so it is cleaned out and then
1542          * left to rot).
1543          */
1544         hammer2_chain_drop(&hmp->fchain);
1545
1546         /*
1547          * Final drop of embedded volume root chain to clean
1548          * up vchain.core (vchain structure is not flagged
1549          * ALLOCATED so it is cleaned out and then left to
1550          * rot).
1551          */
1552         dumpcnt = 50;
1553         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1554         dumpcnt = 50;
1555         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1556         hammer2_dev_unlock(hmp);
1557         hammer2_chain_drop(&hmp->vchain);
1558
1559         hammer2_io_cleanup(hmp, &hmp->iotree);
1560         if (hmp->iofree_count) {
1561                 kprintf("io_cleanup: %d I/O's left hanging\n",
1562                         hmp->iofree_count);
1563         }
1564
1565         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1566         kmalloc_destroy(&hmp->mchain);
1567         kfree(hmp, M_HAMMER2);
1568 }
1569
1570 int
1571 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1572                  ino_t ino, struct vnode **vpp)
1573 {
1574         hammer2_xop_lookup_t *xop;
1575         hammer2_pfs_t *pmp;
1576         hammer2_inode_t *ip;
1577         hammer2_tid_t inum;
1578         int error;
1579
1580         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1581
1582         error = 0;
1583         pmp = MPTOPMP(mp);
1584
1585         /*
1586          * Easy if we already have it cached
1587          */
1588         ip = hammer2_inode_lookup(pmp, inum);
1589         if (ip) {
1590                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1591                 *vpp = hammer2_igetv(ip, &error);
1592                 hammer2_inode_unlock(ip);
1593                 hammer2_inode_drop(ip);         /* from lookup */
1594
1595                 return error;
1596         }
1597
1598         /*
1599          * Otherwise we have to find the inode
1600          */
1601         xop = hammer2_xop_alloc(pmp->iroot, 0);
1602         xop->lhc = inum;
1603         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1604         error = hammer2_xop_collect(&xop->head, 0);
1605
1606         if (error == 0) {
1607                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1608                         kprintf("vget: no collect error but also no rdata\n");
1609                         kprintf("xop %p\n", xop);
1610                         while ((hammer2_debug & 0x80000) == 0) {
1611                                 tsleep(xop, PCATCH, "wait", hz * 10);
1612                         }
1613                         ip = NULL;
1614                 } else {
1615                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1616                 }
1617         }
1618         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1619
1620         if (ip) {
1621                 *vpp = hammer2_igetv(ip, &error);
1622                 hammer2_inode_unlock(ip);
1623         } else {
1624                 *vpp = NULL;
1625                 error = ENOENT;
1626         }
1627         return (error);
1628 }
1629
1630 static
1631 int
1632 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1633 {
1634         hammer2_pfs_t *pmp;
1635         struct vnode *vp;
1636         int error;
1637
1638         pmp = MPTOPMP(mp);
1639         if (pmp->iroot == NULL) {
1640                 *vpp = NULL;
1641                 return EINVAL;
1642         }
1643
1644         error = 0;
1645         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1646
1647         while (pmp->inode_tid == 0) {
1648                 hammer2_xop_ipcluster_t *xop;
1649                 hammer2_inode_meta_t *meta;
1650
1651                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1652                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1653                 error = hammer2_xop_collect(&xop->head, 0);
1654
1655                 if (error == 0) {
1656                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1657                         pmp->iroot->meta = *meta;
1658                         pmp->inode_tid = meta->pfs_inum + 1;
1659                         if (pmp->inode_tid < HAMMER2_INODE_START)
1660                                 pmp->inode_tid = HAMMER2_INODE_START;
1661                         pmp->modify_tid =
1662                                 xop->head.cluster.focus->bref.modify_tid + 1;
1663                         kprintf("PFS: Starting inode %jd\n",
1664                                 (intmax_t)pmp->inode_tid);
1665                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1666                                 pmp->inode_tid, pmp->modify_tid);
1667                         wakeup(&pmp->iroot);
1668
1669                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1670
1671                         /*
1672                          * Prime the mount info.
1673                          */
1674                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1675                         break;
1676                 }
1677
1678                 /*
1679                  * Loop, try again
1680                  */
1681                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1682                 hammer2_inode_unlock(pmp->iroot);
1683                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1684                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1685                 if (error == EINTR)
1686                         break;
1687         }
1688
1689         if (error) {
1690                 hammer2_inode_unlock(pmp->iroot);
1691                 *vpp = NULL;
1692         } else {
1693                 vp = hammer2_igetv(pmp->iroot, &error);
1694                 hammer2_inode_unlock(pmp->iroot);
1695                 *vpp = vp;
1696         }
1697
1698         return (error);
1699 }
1700
1701 /*
1702  * Filesystem status
1703  *
1704  * XXX incorporate ipdata->meta.inode_quota and data_quota
1705  */
1706 static
1707 int
1708 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1709 {
1710         hammer2_pfs_t *pmp;
1711         hammer2_dev_t *hmp;
1712         hammer2_blockref_t bref;
1713         int i;
1714
1715         /*
1716          * NOTE: iroot might not have validated the cluster yet.
1717          */
1718         pmp = MPTOPMP(mp);
1719
1720         mp->mnt_stat.f_files = 0;
1721         mp->mnt_stat.f_ffree = 0;
1722         mp->mnt_stat.f_blocks = 0;
1723         mp->mnt_stat.f_bfree = 0;
1724         mp->mnt_stat.f_bavail = 0;
1725
1726         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1727                 hmp = pmp->pfs_hmps[i];
1728                 if (hmp == NULL)
1729                         continue;
1730                 if (pmp->iroot->cluster.array[i].chain)
1731                         bref = pmp->iroot->cluster.array[i].chain->bref;
1732                 else
1733                         bzero(&bref, sizeof(bref));
1734
1735                 mp->mnt_stat.f_files = bref.inode_count;
1736                 mp->mnt_stat.f_ffree = 0;
1737 #if 0
1738                 mp->mnt_stat.f_blocks = (bref.data_count +
1739                                          hmp->voldata.allocator_free) /
1740                                         mp->mnt_vstat.f_bsize;
1741                 mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free /
1742                                         mp->mnt_vstat.f_bsize;
1743 #endif
1744                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1745                                         mp->mnt_vstat.f_bsize;
1746                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1747                                         mp->mnt_vstat.f_bsize;
1748                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1749
1750                 *sbp = mp->mnt_stat;
1751         }
1752         return (0);
1753 }
1754
1755 static
1756 int
1757 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1758 {
1759         hammer2_pfs_t *pmp;
1760         hammer2_dev_t *hmp;
1761         hammer2_blockref_t bref;
1762         int i;
1763
1764         /*
1765          * NOTE: iroot might not have validated the cluster yet.
1766          */
1767         pmp = MPTOPMP(mp);
1768
1769         mp->mnt_vstat.f_bsize = 0;
1770         mp->mnt_vstat.f_files = 0;
1771         mp->mnt_vstat.f_ffree = 0;
1772         mp->mnt_vstat.f_blocks = 0;
1773         mp->mnt_vstat.f_bfree = 0;
1774         mp->mnt_vstat.f_bavail = 0;
1775
1776         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1777                 hmp = pmp->pfs_hmps[i];
1778                 if (hmp == NULL)
1779                         continue;
1780                 if (pmp->iroot->cluster.array[i].chain)
1781                         bref = pmp->iroot->cluster.array[i].chain->bref;
1782                 else
1783                         bzero(&bref, sizeof(bref));
1784
1785                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1786                 mp->mnt_vstat.f_files = bref.inode_count;
1787                 mp->mnt_vstat.f_ffree = 0;
1788 #if 0
1789                 mp->mnt_vstat.f_blocks = (bref.data_count +
1790                                          hmp->voldata.allocator_free) /
1791                                         mp->mnt_vstat.f_bsize;
1792                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1793                                         mp->mnt_vstat.f_bsize;
1794 #endif
1795                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1796                                         mp->mnt_vstat.f_bsize;
1797                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1798                                         mp->mnt_vstat.f_bsize;
1799                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1800
1801                 *sbp = mp->mnt_vstat;
1802         }
1803         return (0);
1804 }
1805
1806 /*
1807  * Mount-time recovery (RW mounts)
1808  *
1809  * Updates to the free block table are allowed to lag flushes by one
1810  * transaction.  In case of a crash, then on a fresh mount we must do an
1811  * incremental scan of the last committed transaction id and make sure that
1812  * all related blocks have been marked allocated.
1813  *
1814  * The super-root topology and each PFS has its own transaction id domain,
1815  * so we must track PFS boundary transitions.
1816  */
1817 struct hammer2_recovery_elm {
1818         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1819         hammer2_chain_t *chain;
1820         hammer2_tid_t sync_tid;
1821 };
1822
1823 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1824
1825 struct hammer2_recovery_info {
1826         struct hammer2_recovery_list list;
1827         hammer2_tid_t   mtid;
1828         int     depth;
1829 };
1830
1831 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1832                         hammer2_chain_t *parent,
1833                         struct hammer2_recovery_info *info,
1834                         hammer2_tid_t sync_tid);
1835
1836 #define HAMMER2_RECOVERY_MAXDEPTH       10
1837
1838 static
1839 int
1840 hammer2_recovery(hammer2_dev_t *hmp)
1841 {
1842         struct hammer2_recovery_info info;
1843         struct hammer2_recovery_elm *elm;
1844         hammer2_chain_t *parent;
1845         hammer2_tid_t sync_tid;
1846         hammer2_tid_t mirror_tid;
1847         int error;
1848         int cumulative_error = 0;
1849
1850         hammer2_trans_init(hmp->spmp, 0);
1851
1852         sync_tid = hmp->voldata.freemap_tid;
1853         mirror_tid = hmp->voldata.mirror_tid;
1854
1855         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1856         if (sync_tid >= mirror_tid) {
1857                 kprintf(" no recovery needed\n");
1858         } else {
1859                 kprintf(" freemap recovery %016jx-%016jx\n",
1860                         sync_tid + 1, mirror_tid);
1861         }
1862
1863         TAILQ_INIT(&info.list);
1864         info.depth = 0;
1865         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1866         cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
1867         hammer2_chain_lookup_done(parent);
1868
1869         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1870                 TAILQ_REMOVE(&info.list, elm, entry);
1871                 parent = elm->chain;
1872                 sync_tid = elm->sync_tid;
1873                 kfree(elm, M_HAMMER2);
1874
1875                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1876                 error = hammer2_recovery_scan(hmp, parent, &info,
1877                                               hmp->voldata.freemap_tid);
1878                 hammer2_chain_unlock(parent);
1879                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
1880                 if (error)
1881                         cumulative_error = error;
1882         }
1883         hammer2_trans_done(hmp->spmp);
1884
1885         return cumulative_error;
1886 }
1887
1888 static
1889 int
1890 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1891                       struct hammer2_recovery_info *info,
1892                       hammer2_tid_t sync_tid)
1893 {
1894         const hammer2_inode_data_t *ripdata;
1895         hammer2_chain_t *chain;
1896         hammer2_blockref_t bref;
1897         int cache_index;
1898         int cumulative_error = 0;
1899         int error;
1900         int first;
1901
1902         /*
1903          * Adjust freemap to ensure that the block(s) are marked allocated.
1904          */
1905         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1906                 hammer2_freemap_adjust(hmp, &parent->bref,
1907                                        HAMMER2_FREEMAP_DORECOVER);
1908         }
1909
1910         /*
1911          * Check type for recursive scan
1912          */
1913         switch(parent->bref.type) {
1914         case HAMMER2_BREF_TYPE_VOLUME:
1915                 /* data already instantiated */
1916                 break;
1917         case HAMMER2_BREF_TYPE_INODE:
1918                 /*
1919                  * Must instantiate data for DIRECTDATA test and also
1920                  * for recursion.
1921                  */
1922                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1923                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
1924                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1925                         /* not applicable to recovery scan */
1926                         hammer2_chain_unlock(parent);
1927                         return 0;
1928                 }
1929                 hammer2_chain_unlock(parent);
1930                 break;
1931         case HAMMER2_BREF_TYPE_INDIRECT:
1932                 /*
1933                  * Must instantiate data for recursion
1934                  */
1935                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1936                 hammer2_chain_unlock(parent);
1937                 break;
1938         case HAMMER2_BREF_TYPE_DATA:
1939         case HAMMER2_BREF_TYPE_FREEMAP:
1940         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1941         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1942                 /* not applicable to recovery scan */
1943                 return 0;
1944                 break;
1945         default:
1946                 return EDOM;
1947         }
1948
1949         /*
1950          * Defer operation if depth limit reached or if we are crossing a
1951          * PFS boundary.
1952          */
1953         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1954                 struct hammer2_recovery_elm *elm;
1955
1956                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1957                 elm->chain = parent;
1958                 elm->sync_tid = sync_tid;
1959                 hammer2_chain_ref(parent);
1960                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
1961                 /* unlocked by caller */
1962
1963                 return(0);
1964         }
1965
1966
1967         /*
1968          * Recursive scan of the last flushed transaction only.  We are
1969          * doing this without pmp assignments so don't leave the chains
1970          * hanging around after we are done with them.
1971          */
1972         cache_index = 0;
1973         chain = NULL;
1974         first = 1;
1975
1976         while (hammer2_chain_scan(parent, &chain, &bref,
1977                                   &first, &cache_index,
1978                                   HAMMER2_LOOKUP_NODATA) != NULL) {
1979                 /*
1980                  * If this is a leaf
1981                  */
1982                 if (chain == NULL) {
1983                         if (bref.mirror_tid > sync_tid) {
1984                                 hammer2_freemap_adjust(hmp, &bref,
1985                                                      HAMMER2_FREEMAP_DORECOVER);
1986                         }
1987                         continue;
1988                 }
1989
1990                 /*
1991                  * This may or may not be a recursive node.
1992                  */
1993                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
1994                 if (bref.mirror_tid > sync_tid) {
1995                         ++info->depth;
1996                         error = hammer2_recovery_scan(hmp, chain,
1997                                                       info, sync_tid);
1998                         --info->depth;
1999                         if (error)
2000                                 cumulative_error = error;
2001                 }
2002
2003                 /*
2004                  * Flush the recovery at the PFS boundary to stage it for
2005                  * the final flush of the super-root topology.
2006                  */
2007                 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2008                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2009                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2010                 }
2011         }
2012
2013         return cumulative_error;
2014 }
2015
2016 /*
2017  * Sync a mount point; this is called on a per-mount basis from the
2018  * filesystem syncer process periodically and whenever a user issues
2019  * a sync.
2020  */
2021 int
2022 hammer2_vfs_sync(struct mount *mp, int waitfor)
2023 {
2024         hammer2_xop_flush_t *xop;
2025         struct hammer2_sync_info info;
2026         hammer2_inode_t *iroot;
2027         hammer2_pfs_t *pmp;
2028         int flags;
2029         int error;
2030
2031         pmp = MPTOPMP(mp);
2032         iroot = pmp->iroot;
2033         KKASSERT(iroot);
2034         KKASSERT(iroot->pmp == pmp);
2035
2036         /*
2037          * We can't acquire locks on existing vnodes while in a transaction
2038          * without risking a deadlock.  This assumes that vfsync() can be
2039          * called without the vnode locked (which it can in DragonFly).
2040          * Otherwise we'd have to implement a multi-pass or flag the lock
2041          * failures and retry.
2042          *
2043          * The reclamation code interlocks with the sync list's token
2044          * (by removing the vnode from the scan list) before unlocking
2045          * the inode, giving us time to ref the inode.
2046          */
2047         /*flags = VMSC_GETVP;*/
2048         flags = 0;
2049         if (waitfor & MNT_LAZY)
2050                 flags |= VMSC_ONEPASS;
2051
2052         /*
2053          * Preflush the vnodes using a normal transaction before interlocking
2054          * with a flush transaction.  We do this to try to run as much of
2055          * the compression as possible outside the flush transaction.
2056          *
2057          * For efficiency do an async pass before making sure with a
2058          * synchronous pass on all related buffer cache buffers.
2059          */
2060         hammer2_trans_init(pmp, 0);
2061         info.error = 0;
2062         info.waitfor = MNT_NOWAIT;
2063         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2064         info.waitfor = MNT_WAIT;
2065         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2066         hammer2_trans_done(pmp);
2067
2068         /*
2069          * Start our flush transaction.  This does not return until all
2070          * concurrent transactions have completed and will prevent any
2071          * new transactions from running concurrently, except for the
2072          * buffer cache transactions.
2073          *
2074          * NOTE!  It is still possible for the paging code to push pages
2075          *        out via a UIO_NOCOPY hammer2_vop_write() during the main
2076          *        flush.
2077          */
2078         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2079         hammer2_inode_run_sideq(pmp);
2080
2081         info.error = 0;
2082         info.waitfor = MNT_NOWAIT;
2083         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2084         info.waitfor = MNT_WAIT;
2085         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2086         hammer2_bioq_sync(pmp);
2087
2088         /*
2089          * Use the XOP interface to concurrently flush all nodes to
2090          * synchronize the PFSROOT subtopology to the media.  A standard
2091          * end-of-scan ENOENT error indicates cluster sufficiency.
2092          *
2093          * Note that this flush will not be visible on crash recovery until
2094          * we flush the super-root topology in the next loop.
2095          *
2096          * XXX For now wait for all flushes to complete.
2097          */
2098         if (iroot) {
2099                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2100                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2101                 error = hammer2_xop_collect(&xop->head,
2102                                             HAMMER2_XOP_COLLECT_WAITALL);
2103                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2104                 if (error == ENOENT)
2105                         error = 0;
2106         } else {
2107                 error = 0;
2108         }
2109         hammer2_trans_done(pmp);
2110
2111         return (error);
2112 }
2113
2114 /*
2115  * Sync passes.
2116  *
2117  * Note that we ignore the tranasction mtid we got above.  Instead,
2118  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2119  * transactions.
2120  */
2121 static int
2122 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2123 {
2124         struct hammer2_sync_info *info = data;
2125         hammer2_inode_t *ip;
2126         int error;
2127
2128         /*
2129          * Degenerate cases.  Note that ip == NULL typically means the
2130          * syncer vnode itself and we don't want to vclrisdirty() in that
2131          * situation.
2132          */
2133         ip = VTOI(vp);
2134         if (ip == NULL) {
2135                 return(0);
2136         }
2137         if (vp->v_type == VNON || vp->v_type == VBAD) {
2138                 vclrisdirty(vp);
2139                 return(0);
2140         }
2141
2142         /*
2143          * VOP_FSYNC will start a new transaction so replicate some code
2144          * here to do it inline (see hammer2_vop_fsync()).
2145          *
2146          * WARNING: The vfsync interacts with the buffer cache and might
2147          *          block, we can't hold the inode lock at that time.
2148          *          However, we MUST ref ip before blocking to ensure that
2149          *          it isn't ripped out from under us (since we do not
2150          *          hold a lock on the vnode).
2151          */
2152         hammer2_inode_ref(ip);
2153         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2154             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2155                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2156                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2157                                  HAMMER2_INODE_MODIFIED)) {
2158                         hammer2_inode_lock(ip, 0);
2159                         hammer2_inode_chain_sync(ip);
2160                         hammer2_inode_unlock(ip);
2161                 }
2162         }
2163         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2164             RB_EMPTY(&vp->v_rbdirty_tree)) {
2165                 vclrisdirty(vp);
2166         }
2167
2168         hammer2_inode_drop(ip);
2169 #if 1
2170         error = 0;
2171         if (error)
2172                 info->error = error;
2173 #endif
2174         return(0);
2175 }
2176
2177 static
2178 int
2179 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2180 {
2181         hammer2_inode_t *ip;
2182
2183         KKASSERT(MAXFIDSZ >= 16);
2184         ip = VTOI(vp);
2185         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2186         fhp->fid_ext = 0;
2187         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2188         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2189
2190         return 0;
2191 }
2192
2193 static
2194 int
2195 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2196                struct fid *fhp, struct vnode **vpp)
2197 {
2198         hammer2_pfs_t *pmp;
2199         hammer2_tid_t inum;
2200         int error;
2201
2202         pmp = MPTOPMP(mp);
2203         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2204         if (vpp) {
2205                 if (inum == 1)
2206                         error = hammer2_vfs_root(mp, vpp);
2207                 else
2208                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2209         } else {
2210                 error = 0;
2211         }
2212         if (error)
2213                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2214         return error;
2215 }
2216
2217 static
2218 int
2219 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2220                  int *exflagsp, struct ucred **credanonp)
2221 {
2222         hammer2_pfs_t *pmp;
2223         struct netcred *np;
2224         int error;
2225
2226         pmp = MPTOPMP(mp);
2227         np = vfs_export_lookup(mp, &pmp->export, nam);
2228         if (np) {
2229                 *exflagsp = np->netc_exflags;
2230                 *credanonp = &np->netc_anon;
2231                 error = 0;
2232         } else {
2233                 error = EACCES;
2234         }
2235         return error;
2236 }
2237
2238 /*
2239  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2240  * header into the HMP
2241  *
2242  * XXX read four volhdrs and use the one with the highest TID whos CRC
2243  *     matches.
2244  *
2245  * XXX check iCRCs.
2246  *
2247  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2248  *     nonexistant locations.
2249  *
2250  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2251  */
2252 static
2253 int
2254 hammer2_install_volume_header(hammer2_dev_t *hmp)
2255 {
2256         hammer2_volume_data_t *vd;
2257         struct buf *bp;
2258         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2259         int error_reported;
2260         int error;
2261         int valid;
2262         int i;
2263
2264         error_reported = 0;
2265         error = 0;
2266         valid = 0;
2267         bp = NULL;
2268
2269         /*
2270          * There are up to 4 copies of the volume header (syncs iterate
2271          * between them so there is no single master).  We don't trust the
2272          * volu_size field so we don't know precisely how large the filesystem
2273          * is, so depend on the OS to return an error if we go beyond the
2274          * block device's EOF.
2275          */
2276         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2277                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2278                               HAMMER2_VOLUME_BYTES, &bp);
2279                 if (error) {
2280                         brelse(bp);
2281                         bp = NULL;
2282                         continue;
2283                 }
2284
2285                 vd = (struct hammer2_volume_data *) bp->b_data;
2286                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2287                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2288                         brelse(bp);
2289                         bp = NULL;
2290                         continue;
2291                 }
2292
2293                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2294                         /* XXX: Reversed-endianness filesystem */
2295                         kprintf("hammer2: reverse-endian filesystem detected");
2296                         brelse(bp);
2297                         bp = NULL;
2298                         continue;
2299                 }
2300
2301                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2302                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2303                                       HAMMER2_VOLUME_ICRC0_SIZE);
2304                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2305                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2306                                        HAMMER2_VOLUME_ICRC1_SIZE);
2307                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2308                         kprintf("hammer2 volume header crc "
2309                                 "mismatch copy #%d %08x/%08x\n",
2310                                 i, crc0, crc);
2311                         error_reported = 1;
2312                         brelse(bp);
2313                         bp = NULL;
2314                         continue;
2315                 }
2316                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2317                         valid = 1;
2318                         hmp->voldata = *vd;
2319                         hmp->volhdrno = i;
2320                 }
2321                 brelse(bp);
2322                 bp = NULL;
2323         }
2324         if (valid) {
2325                 hmp->volsync = hmp->voldata;
2326                 error = 0;
2327                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2328                         kprintf("hammer2: using volume header #%d\n",
2329                                 hmp->volhdrno);
2330                 }
2331         } else {
2332                 error = EINVAL;
2333                 kprintf("hammer2: no valid volume headers found!\n");
2334         }
2335         return (error);
2336 }
2337
2338 /*
2339  * This handles hysteresis on regular file flushes.  Because the BIOs are
2340  * routed to a thread it is possible for an excessive number to build up
2341  * and cause long front-end stalls long before the runningbuffspace limit
2342  * is hit, so we implement hammer2_flush_pipe to control the
2343  * hysteresis.
2344  *
2345  * This is a particular problem when compression is used.
2346  */
2347 void
2348 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2349 {
2350         atomic_add_int(&pmp->count_lwinprog, 1);
2351 }
2352
2353 void
2354 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2355 {
2356         int lwinprog;
2357
2358         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2359         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2360             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2361                 atomic_clear_int(&pmp->count_lwinprog,
2362                                  HAMMER2_LWINPROG_WAITING);
2363                 wakeup(&pmp->count_lwinprog);
2364         }
2365         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2366             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2367                 atomic_clear_int(&pmp->count_lwinprog,
2368                                  HAMMER2_LWINPROG_WAITING0);
2369                 wakeup(&pmp->count_lwinprog);
2370         }
2371 }
2372
2373 void
2374 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2375 {
2376         int lwinprog;
2377         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2378                                     HAMMER2_LWINPROG_WAITING0;
2379
2380         for (;;) {
2381                 lwinprog = pmp->count_lwinprog;
2382                 cpu_ccfence();
2383                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2384                         break;
2385                 tsleep_interlock(&pmp->count_lwinprog, 0);
2386                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2387                 lwinprog = pmp->count_lwinprog;
2388                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2389                         break;
2390                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2391         }
2392 }
2393
2394 /*
2395  * Manage excessive memory resource use for chain and related
2396  * structures.
2397  */
2398 void
2399 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2400 {
2401         uint32_t waiting;
2402         uint32_t count;
2403         uint32_t limit;
2404 #if 0
2405         static int zzticks;
2406 #endif
2407
2408         /*
2409          * Atomic check condition and wait.  Also do an early speedup of
2410          * the syncer to try to avoid hitting the wait.
2411          */
2412         for (;;) {
2413                 waiting = pmp->inmem_dirty_chains;
2414                 cpu_ccfence();
2415                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2416
2417                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2418                 if (limit < hammer2_limit_dirty_chains)
2419                         limit = hammer2_limit_dirty_chains;
2420                 if (limit < 1000)
2421                         limit = 1000;
2422
2423 #if 0
2424                 if ((int)(ticks - zzticks) > hz) {
2425                         zzticks = ticks;
2426                         kprintf("count %ld %ld\n", count, limit);
2427                 }
2428 #endif
2429
2430                 /*
2431                  * Block if there are too many dirty chains present, wait
2432                  * for the flush to clean some out.
2433                  */
2434                 if (count > limit) {
2435                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2436                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2437                                                waiting,
2438                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2439                                 speedup_syncer(pmp->mp);
2440                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2441                                        "chnmem", hz);
2442                         }
2443                         continue;       /* loop on success or fail */
2444                 }
2445
2446                 /*
2447                  * Try to start an early flush before we are forced to block.
2448                  */
2449                 if (count > limit * 7 / 10)
2450                         speedup_syncer(pmp->mp);
2451                 break;
2452         }
2453 }
2454
2455 void
2456 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2457 {
2458         if (pmp) {
2459                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2460         }
2461 }
2462
2463 void
2464 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2465 {
2466         uint32_t waiting;
2467
2468         if (pmp == NULL)
2469                 return;
2470
2471         for (;;) {
2472                 waiting = pmp->inmem_dirty_chains;
2473                 cpu_ccfence();
2474                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2475                                        waiting,
2476                                        (waiting - 1) &
2477                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2478                         break;
2479                 }
2480         }
2481
2482         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2483                 wakeup(&pmp->inmem_dirty_chains);
2484 }
2485
2486 /*
2487  * Debugging
2488  */
2489 void
2490 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2491 {
2492         hammer2_chain_t *scan;
2493         hammer2_chain_t *parent;
2494
2495         --*countp;
2496         if (*countp == 0) {
2497                 kprintf("%*.*s...\n", tab, tab, "");
2498                 return;
2499         }
2500         if (*countp < 0)
2501                 return;
2502         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2503                 tab, tab, "", pfx,
2504                 chain, chain->bref.type,
2505                 chain->bref.key, chain->bref.keybits,
2506                 chain->bref.mirror_tid);
2507
2508         kprintf("%*.*s      [%08x] (%s) refs=%d",
2509                 tab, tab, "",
2510                 chain->flags,
2511                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2512                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2513                 chain->refs);
2514
2515         parent = chain->parent;
2516         if (parent)
2517                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2518                         tab, tab, "",
2519                         parent, parent->flags, parent->refs);
2520         if (RB_EMPTY(&chain->core.rbtree)) {
2521                 kprintf("\n");
2522         } else {
2523                 kprintf(" {\n");
2524                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2525                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2526                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2527                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2528                                 chain->data->ipdata.filename);
2529                 else
2530                         kprintf("%*.*s}\n", tab, tab, "");
2531         }
2532 }