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_enable = 1;
  82 int hammer2_hardlink_enable = 1;
  83 int hammer2_flush_pipe = 100;
  84 int hammer2_synchronous_flush = 1;
  85 int hammer2_dio_count;
  86 long hammer2_limit_dirty_chains;
  87 long hammer2_iod_file_read;
  88 long hammer2_iod_meta_read;
  89 long hammer2_iod_indr_read;
  90 long hammer2_iod_fmap_read;
  91 long hammer2_iod_volu_read;
  92 long hammer2_iod_file_write;
  93 long hammer2_iod_meta_write;
  94 long hammer2_iod_indr_write;
  95 long hammer2_iod_fmap_write;
  96 long hammer2_iod_volu_write;
  97 long hammer2_ioa_file_read;
  98 long hammer2_ioa_meta_read;
  99 long hammer2_ioa_indr_read;
 100 long hammer2_ioa_fmap_read;
 101 long hammer2_ioa_volu_read;
 102 long hammer2_ioa_fmap_write;
 103 long hammer2_ioa_file_write;
 104 long hammer2_ioa_meta_write;
 105 long hammer2_ioa_indr_write;
 106 long hammer2_ioa_volu_write;
 107
 108 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 109 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 110                 "Buffer used for compression.");
 111
 112 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 113 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 114                 "Buffer used for decompression.");
 115
 116 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 117
 118 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 119            &hammer2_debug, 0, "");
 120 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW,
 121            &hammer2_cluster_enable, 0, "");
 122 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
 123            &hammer2_hardlink_enable, 0, "");
 124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 125            &hammer2_flush_pipe, 0, "");
 126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 127            &hammer2_synchronous_flush, 0, "");
 128 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 129            &hammer2_limit_dirty_chains, 0, "");
 130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 131            &hammer2_dio_count, 0, "");
 132
 133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 134            &hammer2_iod_file_read, 0, "");
 135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 136            &hammer2_iod_meta_read, 0, "");
 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 138            &hammer2_iod_indr_read, 0, "");
 139 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 140            &hammer2_iod_fmap_read, 0, "");
 141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 142            &hammer2_iod_volu_read, 0, "");
 143
 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 145            &hammer2_iod_file_write, 0, "");
 146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 147            &hammer2_iod_meta_write, 0, "");
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 149            &hammer2_iod_indr_write, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 151            &hammer2_iod_fmap_write, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 153            &hammer2_iod_volu_write, 0, "");
 154
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
 156            &hammer2_ioa_file_read, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
 158            &hammer2_ioa_meta_read, 0, "");
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
 160            &hammer2_ioa_indr_read, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
 162            &hammer2_ioa_fmap_read, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
 164            &hammer2_ioa_volu_read, 0, "");
 165
 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
 167            &hammer2_ioa_file_write, 0, "");
 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
 169            &hammer2_ioa_meta_write, 0, "");
 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
 171            &hammer2_ioa_indr_write, 0, "");
 172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
 173            &hammer2_ioa_fmap_write, 0, "");
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
 175            &hammer2_ioa_volu_write, 0, "");
 176
 177 static int hammer2_vfs_init(struct vfsconf *conf);
 178 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 179 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 180                                 struct ucred *cred);
 181 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 182                                 struct vnode *, struct ucred *);
 183 static int hammer2_recovery(hammer2_dev_t *hmp);
 184 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 185 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 186 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 187                                 struct ucred *cred);
 188 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 189                                 struct ucred *cred);
 190 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
 191                                 ino_t ino, struct vnode **vpp);
 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 = desiredvnodes / 10;
 292
 293         return (error);
 294 }
 295
 296 static
 297 int
 298 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 299 {
 300         objcache_destroy(cache_buffer_read);
 301         objcache_destroy(cache_buffer_write);
 302         objcache_destroy(cache_xops);
 303         return 0;
 304 }
 305
 306 /*
 307  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
 308  * mounts and the spmp structure for media (hmp) structures.
 309  *
 310  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 311  * transactions.  Note that synchronization does not use this field.
 312  * (typically frontend operations and synchronization cannot run on the
 313  * same PFS node at the same time).
 314  *
 315  * XXX check locking
 316  */
 317 hammer2_pfs_t *
 318 hammer2_pfsalloc(hammer2_cluster_t *cluster,
 319                  const hammer2_inode_data_t *ripdata,
 320                  hammer2_tid_t modify_tid)
 321 {
 322         hammer2_chain_t *rchain;
 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 (ripdata) {
 334                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 335                         if (bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 336                                  sizeof(pmp->pfs_clid)) == 0) {
 337                                         break;
 338                         }
 339                 }
 340         } else {
 341                 pmp = NULL;
 342         }
 343
 344         if (pmp == NULL) {
 345                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 346                 hammer2_trans_manage_init(pmp);
 347                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 348                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 349                 lockinit(&pmp->lock, "pfslk", 0, 0);
 350                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 351                 RB_INIT(&pmp->inum_tree);
 352                 TAILQ_INIT(&pmp->unlinkq);
 353                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 354
 355                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 356                         hammer2_xop_group_init(pmp, &pmp->xop_groups[j]);
 357
 358                 /*
 359                  * Save the last media transaction id for the flusher.  Set
 360                  * initial
 361                  */
 362                 if (ripdata)
 363                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 364                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 365
 366                 /*
 367                  * The synchronization thread may start too early, make
 368                  * sure it stays frozen until we are ready to let it go.
 369                  * XXX
 370                  */
 371                 /*
 372                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 373                                          HAMMER2_THREAD_REMASTER;
 374                 */
 375         }
 376
 377         /*
 378          * Create the PFS's root inode.
 379          */
 380         if ((iroot = pmp->iroot) == NULL) {
 381                 iroot = hammer2_inode_get(pmp, NULL, NULL);
 382                 pmp->iroot = iroot;
 383                 hammer2_inode_ref(iroot);
 384                 hammer2_inode_unlock(iroot, NULL);
 385         }
 386
 387         /*
 388          * Stop here if no cluster is passed in.
 389          */
 390         if (cluster == NULL)
 391                 goto done;
 392
 393         /*
 394          * When a cluster is passed in we must add the cluster's chains
 395          * to the PFS's root inode, update pmp->pfs_types[], and update
 396          * the syncronization threads.
 397          *
 398          * At the moment empty spots can develop due to removals or failures.
 399          * Ultimately we want to re-fill these spots but doing so might
 400          * confused running code. XXX
 401          */
 402         hammer2_inode_ref(iroot);
 403         hammer2_mtx_ex(&iroot->lock);
 404         j = iroot->cluster.nchains;
 405
 406         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 407
 408         for (i = 0; i < cluster->nchains; ++i) {
 409                 if (j == HAMMER2_MAXCLUSTER)
 410                         break;
 411                 rchain = cluster->array[i].chain;
 412                 KKASSERT(rchain->pmp == NULL);
 413                 rchain->pmp = pmp;
 414                 hammer2_chain_ref(rchain);
 415                 iroot->cluster.array[j].chain = rchain;
 416                 pmp->pfs_types[j] = ripdata->meta.pfs_type;
 417                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 418
 419                 /*
 420                  * If the PFS is already mounted we must account
 421                  * for the mount_count here.
 422                  */
 423                 if (pmp->mp)
 424                         ++rchain->hmp->mount_count;
 425
 426                 /*
 427                  * May have to fixup dirty chain tracking.  Previous
 428                  * pmp was NULL so nothing to undo.
 429                  */
 430                 if (rchain->flags & HAMMER2_CHAIN_MODIFIED)
 431                         hammer2_pfs_memory_inc(pmp);
 432                 ++j;
 433         }
 434         iroot->cluster.nchains = j;
 435
 436         if (i != cluster->nchains) {
 437                 kprintf("hammer2_mount: cluster full!\n");
 438                 /* XXX fatal error? */
 439         }
 440
 441         /*
 442          * Update nmasters from any PFS inode which is part of the cluster.
 443          * It is possible that this will result in a value which is too
 444          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 445          * override this value later on.
 446          *
 447          * (This informs us of masters that might not currently be
 448          *  discoverable by this mount).
 449          */
 450         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 451                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 452         }
 453
 454         /*
 455          * Count visible masters.  Masters are usually added with
 456          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 457          * are more (XXX and must update the master inodes).
 458          */
 459         count = 0;
 460         for (i = 0; i < iroot->cluster.nchains; ++i) {
 461                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 462                         ++count;
 463         }
 464         if (pmp->pfs_nmasters < count)
 465                 pmp->pfs_nmasters = count;
 466
 467         /*
 468          * Create missing synchronization and support threads.
 469          *
 470          * Single-node masters (including snapshots) have nothing to
 471          * synchronize and do not require this thread.
 472          *
 473          * Multi-node masters or any number of soft masters, slaves, copy,
 474          * or other PFS types need the thread.
 475          *
 476          * Each thread is responsible for its particular cluster index.
 477          * We use independent threads so stalls or mismatches related to
 478          * any given target do not affect other targets.
 479          */
 480         for (i = 0; i < iroot->cluster.nchains; ++i) {
 481                 /*
 482                  * Single-node masters (including snapshots) have nothing
 483                  * to synchronize and will make direct xops support calls,
 484                  * thus they do not require this thread.
 485                  *
 486                  * Note that there can be thousands of snapshots.  We do not
 487                  * want to create thousands of threads.
 488                  */
 489                 if (pmp->pfs_nmasters <= 1 &&
 490                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 491                         continue;
 492                 }
 493
 494                 /*
 495                  * Sync support thread
 496                  */
 497                 if (pmp->sync_thrs[i].td == NULL) {
 498                         hammer2_thr_create(&pmp->sync_thrs[i], pmp,
 499                                            "h2nod", i, -1,
 500                                            hammer2_primary_sync_thread);
 501                 }
 502         }
 503
 504         /*
 505          * Create missing Xop threads
 506          */
 507         if (pmp->mp)
 508                 hammer2_xop_helper_create(pmp);
 509
 510         hammer2_mtx_unlock(&iroot->lock);
 511         hammer2_inode_drop(iroot);
 512 done:
 513         return pmp;
 514 }
 515
 516 /*
 517  * Destroy a PFS, typically only occurs after the last mount on a device
 518  * has gone away.
 519  */
 520 static void
 521 hammer2_pfsfree(hammer2_pfs_t *pmp)
 522 {
 523         hammer2_inode_t *iroot;
 524         int i;
 525         int j;
 526
 527         /*
 528          * Cleanup our reference on iroot.  iroot is (should) not be needed
 529          * by the flush code.
 530          */
 531         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 532
 533         iroot = pmp->iroot;
 534         if (iroot) {
 535                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 536                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 537                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 538                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 539                 }
 540 #if REPORT_REFS_ERRORS
 541                 if (pmp->iroot->refs != 1)
 542                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 543                                 pmp->iroot, pmp->iroot->refs);
 544 #else
 545                 KKASSERT(pmp->iroot->refs == 1);
 546 #endif
 547                 /* ref for pmp->iroot */
 548                 hammer2_inode_drop(pmp->iroot);
 549                 pmp->iroot = NULL;
 550         }
 551
 552         kmalloc_destroy(&pmp->mmsg);
 553         kmalloc_destroy(&pmp->minode);
 554
 555         kfree(pmp, M_HAMMER2);
 556 }
 557
 558 /*
 559  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 560  * empty is terminated and freed.
 561  *
 562  * XXX inefficient.
 563  */
 564 static void
 565 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 566 {
 567         hammer2_pfs_t *pmp;
 568         hammer2_inode_t *iroot;
 569         hammer2_cluster_t *cluster;
 570         hammer2_chain_t *rchain;
 571         int didfreeze;
 572         int i;
 573         int j;
 574
 575 again:
 576         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 577                 if ((iroot = pmp->iroot) == NULL)
 578                         continue;
 579                 if (hmp->spmp == pmp) {
 580                         kprintf("unmount hmp %p remove spmp %p\n",
 581                                 hmp, pmp);
 582                         hmp->spmp = NULL;
 583                 }
 584
 585                 /*
 586                  * Determine if this PFS is affected.  If it is we must
 587                  * freeze all management threads and lock its iroot.
 588                  *
 589                  * Freezing a management thread forces it idle, operations
 590                  * in-progress will be aborted and it will have to start
 591                  * over again when unfrozen, or exit if told to exit.
 592                  */
 593                 cluster = &iroot->cluster;
 594                 for (i = 0; i < cluster->nchains; ++i) {
 595                         rchain = cluster->array[i].chain;
 596                         if (rchain == NULL || rchain->hmp != hmp)
 597                                 continue;
 598                         break;
 599                 }
 600                 if (i != cluster->nchains) {
 601                         /*
 602                          * Make sure all synchronization threads are locked
 603                          * down.
 604                          */
 605                         for (i = 0; i < iroot->cluster.nchains; ++i) {
 606                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 607                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 608                                         hammer2_thr_freeze_async(
 609                                                 &pmp->xop_groups[j].thrs[i]);
 610                                 }
 611                         }
 612                         for (i = 0; i < iroot->cluster.nchains; ++i) {
 613                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 614                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 615                                         hammer2_thr_freeze(
 616                                                 &pmp->xop_groups[j].thrs[i]);
 617                                 }
 618                         }
 619
 620                         /*
 621                          * Lock the inode and clean out matching chains.
 622                          * Note that we cannot use hammer2_inode_lock_*()
 623                          * here because that would attempt to validate the
 624                          * cluster that we are in the middle of ripping
 625                          * apart.
 626                          *
 627                          * WARNING! We are working directly on the inodes
 628                          *          embedded cluster.
 629                          */
 630                         hammer2_mtx_ex(&iroot->lock);
 631
 632                         /*
 633                          * Remove the chain from matching elements of the PFS.
 634                          */
 635                         for (i = 0; i < cluster->nchains; ++i) {
 636                                 rchain = cluster->array[i].chain;
 637                                 if (rchain == NULL || rchain->hmp != hmp)
 638                                         continue;
 639                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 640                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 641                                         hammer2_thr_delete(
 642                                                 &pmp->xop_groups[j].thrs[i]);
 643                                 }
 644                                 rchain = cluster->array[i].chain;
 645                                 cluster->array[i].chain = NULL;
 646                                 pmp->pfs_types[i] = 0;
 647                                 if (pmp->pfs_names[i]) {
 648                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 649                                         pmp->pfs_names[i] = NULL;
 650                                 }
 651                                 hammer2_chain_drop(rchain);
 652
 653                                 /* focus hint */
 654                                 if (cluster->focus == rchain)
 655                                         cluster->focus = NULL;
 656                         }
 657                         hammer2_mtx_unlock(&iroot->lock);
 658                         didfreeze = 1;  /* remaster, unfreeze down below */
 659                 } else {
 660                         didfreeze = 0;
 661                 }
 662
 663                 /*
 664                  * Cleanup trailing chains.  Do not reorder chains (for now).
 665                  * XXX might remove more than we intended.
 666                  */
 667                 while (i > 0) {
 668                         if (cluster->array[i - 1].chain)
 669                                 break;
 670                         --i;
 671                 }
 672                 cluster->nchains = i;
 673
 674                 /*
 675                  * If the PMP has no elements remaining we can destroy it.
 676                  * (this will transition management threads from frozen->exit).
 677                  */
 678                 if (cluster->nchains == 0) {
 679                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 680                                 hmp, pmp);
 681                         hammer2_pfsfree(pmp);
 682                         goto again;
 683                 }
 684
 685                 /*
 686                  * If elements still remain we need to set the REMASTER
 687                  * flag and unfreeze it.
 688                  */
 689                 if (didfreeze) {
 690                         for (i = 0; i < iroot->cluster.nchains; ++i) {
 691                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 692                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 693                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 694                                         hammer2_thr_remaster(
 695                                                 &pmp->xop_groups[j].thrs[i]);
 696                                         hammer2_thr_unfreeze(
 697                                                 &pmp->xop_groups[j].thrs[i]);
 698                                 }
 699                         }
 700                 }
 701         }
 702 }
 703
 704 /*
 705  * Mount or remount HAMMER2 fileystem from physical media
 706  *
 707  *      mountroot
 708  *              mp              mount point structure
 709  *              path            NULL
 710  *              data            <unused>
 711  *              cred            <unused>
 712  *
 713  *      mount
 714  *              mp              mount point structure
 715  *              path            path to mount point
 716  *              data            pointer to argument structure in user space
 717  *                      volume  volume path (device@LABEL form)
 718  *                      hflags  user mount flags
 719  *              cred            user credentials
 720  *
 721  * RETURNS:     0       Success
 722  *              !0      error number
 723  */
 724 static
 725 int
 726 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 727                   struct ucred *cred)
 728 {
 729         struct hammer2_mount_info info;
 730         hammer2_pfs_t *pmp;
 731         hammer2_pfs_t *spmp;
 732         hammer2_dev_t *hmp;
 733         hammer2_key_t key_next;
 734         hammer2_key_t key_dummy;
 735         hammer2_key_t lhc;
 736         struct vnode *devvp;
 737         struct nlookupdata nd;
 738         hammer2_chain_t *parent;
 739         hammer2_cluster_t *cluster;
 740         hammer2_cluster_t *cparent;
 741         const hammer2_inode_data_t *ripdata;
 742         hammer2_blockref_t bref;
 743         struct file *fp;
 744         char devstr[MNAMELEN];
 745         size_t size;
 746         size_t done;
 747         char *dev;
 748         char *label;
 749         int ronly = 1;
 750         int error;
 751         int cache_index;
 752         int i;
 753
 754         hmp = NULL;
 755         pmp = NULL;
 756         dev = NULL;
 757         label = NULL;
 758         devvp = NULL;
 759         cache_index = -1;
 760
 761         kprintf("hammer2_mount\n");
 762
 763         if (path == NULL) {
 764                 /*
 765                  * Root mount
 766                  */
 767                 bzero(&info, sizeof(info));
 768                 info.cluster_fd = -1;
 769                 return (EOPNOTSUPP);
 770         } else {
 771                 /*
 772                  * Non-root mount or updating a mount
 773                  */
 774                 error = copyin(data, &info, sizeof(info));
 775                 if (error)
 776                         return (error);
 777
 778                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 779                 if (error)
 780                         return (error);
 781
 782                 /* Extract device and label */
 783                 dev = devstr;
 784                 label = strchr(devstr, '@');
 785                 if (label == NULL ||
 786                     ((label + 1) - dev) > done) {
 787                         return (EINVAL);
 788                 }
 789                 *label = '\0';
 790                 label++;
 791                 if (*label == '\0')
 792                         return (EINVAL);
 793
 794                 if (mp->mnt_flag & MNT_UPDATE) {
 795                         /*
 796                          * Update mount.  Note that pmp->iroot->cluster is
 797                          * an inode-embedded cluster and thus cannot be
 798                          * directly locked.
 799                          *
 800                          * XXX HAMMER2 needs to implement NFS export via
 801                          *     mountctl.
 802                          */
 803                         pmp = MPTOPMP(mp);
 804                         cluster = &pmp->iroot->cluster;
 805                         for (i = 0; i < cluster->nchains; ++i) {
 806                                 if (cluster->array[i].chain == NULL)
 807                                         continue;
 808                                 hmp = cluster->array[i].chain->hmp;
 809                                 devvp = hmp->devvp;
 810                                 error = hammer2_remount(hmp, mp, path,
 811                                                         devvp, cred);
 812                                 if (error)
 813                                         break;
 814                         }
 815                         /*hammer2_inode_install_hidden(pmp);*/
 816
 817                         return error;
 818                 }
 819         }
 820
 821         /*
 822          * HMP device mount
 823          *
 824          * Lookup name and verify it refers to a block device.
 825          */
 826         error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 827         if (error == 0)
 828                 error = nlookup(&nd);
 829         if (error == 0)
 830                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 831         nlookup_done(&nd);
 832
 833         if (error == 0) {
 834                 if (vn_isdisk(devvp, &error))
 835                         error = vfs_mountedon(devvp);
 836         }
 837
 838         /*
 839          * Determine if the device has already been mounted.  After this
 840          * check hmp will be non-NULL if we are doing the second or more
 841          * hammer2 mounts from the same device.
 842          */
 843         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 844         TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 845                 if (hmp->devvp == devvp)
 846                         break;
 847         }
 848
 849         /*
 850          * Open the device if this isn't a secondary mount and construct
 851          * the H2 device mount (hmp).
 852          */
 853         if (hmp == NULL) {
 854                 hammer2_chain_t *schain;
 855                 hammer2_xid_t xid;
 856
 857                 if (error == 0 && vcount(devvp) > 0)
 858                         error = EBUSY;
 859
 860                 /*
 861                  * Now open the device
 862                  */
 863                 if (error == 0) {
 864                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
 865                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
 866                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
 867                         if (error == 0) {
 868                                 error = VOP_OPEN(devvp,
 869                                                  ronly ? FREAD : FREAD | FWRITE,
 870                                                  FSCRED, NULL);
 871                         }
 872                         vn_unlock(devvp);
 873                 }
 874                 if (error && devvp) {
 875                         vrele(devvp);
 876                         devvp = NULL;
 877                 }
 878                 if (error) {
 879                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 880                         return error;
 881                 }
 882                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
 883                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
 884                 hmp->ronly = ronly;
 885                 hmp->devvp = devvp;
 886                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
 887                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
 888                 RB_INIT(&hmp->iotree);
 889                 spin_init(&hmp->io_spin, "hm2mount_io");
 890                 spin_init(&hmp->list_spin, "hm2mount_list");
 891                 TAILQ_INIT(&hmp->flushq);
 892
 893                 lockinit(&hmp->vollk, "h2vol", 0, 0);
 894
 895                 /*
 896                  * vchain setup. vchain.data is embedded.
 897                  * vchain.refs is initialized and will never drop to 0.
 898                  *
 899                  * NOTE! voldata is not yet loaded.
 900                  */
 901                 hmp->vchain.hmp = hmp;
 902                 hmp->vchain.refs = 1;
 903                 hmp->vchain.data = (void *)&hmp->voldata;
 904                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
 905                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 906                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 907
 908                 hammer2_chain_core_init(&hmp->vchain);
 909                 /* hmp->vchain.u.xxx is left NULL */
 910
 911                 /*
 912                  * fchain setup.  fchain.data is embedded.
 913                  * fchain.refs is initialized and will never drop to 0.
 914                  *
 915                  * The data is not used but needs to be initialized to
 916                  * pass assertion muster.  We use this chain primarily
 917                  * as a placeholder for the freemap's top-level RBTREE
 918                  * so it does not interfere with the volume's topology
 919                  * RBTREE.
 920                  */
 921                 hmp->fchain.hmp = hmp;
 922                 hmp->fchain.refs = 1;
 923                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
 924                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
 925                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 926                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 927                 hmp->fchain.bref.methods =
 928                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
 929                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
 930
 931                 hammer2_chain_core_init(&hmp->fchain);
 932                 /* hmp->fchain.u.xxx is left NULL */
 933
 934                 /*
 935                  * Install the volume header and initialize fields from
 936                  * voldata.
 937                  */
 938                 error = hammer2_install_volume_header(hmp);
 939                 if (error) {
 940                         hammer2_unmount_helper(mp, NULL, hmp);
 941                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 942                         hammer2_vfs_unmount(mp, MNT_FORCE);
 943                         return error;
 944                 }
 945
 946                 /*
 947                  * Really important to get these right or flush will get
 948                  * confused.
 949                  */
 950                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0);
 951                 kprintf("alloc spmp %p tid %016jx\n",
 952                         hmp->spmp, hmp->voldata.mirror_tid);
 953                 spmp = hmp->spmp;
 954
 955                 /*
 956                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
 957                  * is inherited from the volume header.
 958                  */
 959                 xid = 0;
 960                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 961                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
 962                 hmp->vchain.pmp = spmp;
 963                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 964                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
 965                 hmp->fchain.pmp = spmp;
 966
 967                 /*
 968                  * First locate the super-root inode, which is key 0
 969                  * relative to the volume header's blockset.
 970                  *
 971                  * Then locate the root inode by scanning the directory keyspace
 972                  * represented by the label.
 973                  */
 974                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
 975                 schain = hammer2_chain_lookup(&parent, &key_dummy,
 976                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
 977                                       &cache_index, 0);
 978                 hammer2_chain_lookup_done(parent);
 979                 if (schain == NULL) {
 980                         kprintf("hammer2_mount: invalid super-root\n");
 981                         hammer2_unmount_helper(mp, NULL, hmp);
 982                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 983                         hammer2_vfs_unmount(mp, MNT_FORCE);
 984                         return EINVAL;
 985                 }
 986                 if (schain->error) {
 987                         kprintf("hammer2_mount: error %s reading super-root\n",
 988                                 hammer2_error_str(schain->error));
 989                         hammer2_chain_unlock(schain);
 990                         hammer2_chain_drop(schain);
 991                         schain = NULL;
 992                         hammer2_unmount_helper(mp, NULL, hmp);
 993                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 994                         hammer2_vfs_unmount(mp, MNT_FORCE);
 995                         return EINVAL;
 996                 }
 997
 998                 /*
 999                  * The super-root always uses an inode_tid of 1 when
1000                  * creating PFSs.
1001                  */
1002                 spmp->inode_tid = 1;
1003                 spmp->modify_tid = schain->bref.modify_tid + 1;
1004
1005                 /*
1006                  * Sanity-check schain's pmp and finish initialization.
1007                  * Any chain belonging to the super-root topology should
1008                  * have a NULL pmp (not even set to spmp).
1009                  */
1010                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1011                 KKASSERT(schain->pmp == NULL);
1012                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1013
1014                 /*
1015                  * Replace the dummy spmp->iroot with a real one.  It's
1016                  * easier to just do a wholesale replacement than to try
1017                  * to update the chain and fixup the iroot fields.
1018                  *
1019                  * The returned inode is locked with the supplied cluster.
1020                  */
1021                 cluster = hammer2_cluster_from_chain(schain);
1022                 hammer2_inode_drop(spmp->iroot);
1023                 spmp->iroot = NULL;
1024                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster);
1025                 spmp->spmp_hmp = hmp;
1026                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1027                 hammer2_inode_ref(spmp->iroot);
1028                 hammer2_inode_unlock(spmp->iroot, cluster);
1029                 schain = NULL;
1030                 /* leave spmp->iroot with one ref */
1031
1032                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1033                         error = hammer2_recovery(hmp);
1034                         /* XXX do something with error */
1035                 }
1036                 hammer2_update_pmps(hmp);
1037                 hammer2_iocom_init(hmp);
1038
1039                 /*
1040                  * Ref the cluster management messaging descriptor.  The mount
1041                  * program deals with the other end of the communications pipe.
1042                  */
1043                 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1044                 if (fp) {
1045                         hammer2_cluster_reconnect(hmp, fp);
1046                 } else {
1047                         kprintf("hammer2_mount: bad cluster_fd!\n");
1048                 }
1049         } else {
1050                 spmp = hmp->spmp;
1051         }
1052
1053         /*
1054          * Lookup the mount point under the media-localized super-root.
1055          * Scanning hammer2_pfslist doesn't help us because it represents
1056          * PFS cluster ids which can aggregate several named PFSs together.
1057          *
1058          * cluster->pmp will incorrectly point to spmp and must be fixed
1059          * up later on.
1060          */
1061         hammer2_inode_lock(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1062         cparent = hammer2_inode_cluster(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1063         lhc = hammer2_dirhash(label, strlen(label));
1064         cluster = hammer2_cluster_lookup(cparent, &key_next,
1065                                       lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1066                                       0);
1067         while (cluster) {
1068                 if (hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE &&
1069                     strcmp(label,
1070                        hammer2_cluster_rdata(cluster)->ipdata.filename) == 0) {
1071                         break;
1072                 }
1073                 cluster = hammer2_cluster_next(cparent, cluster, &key_next,
1074                                             key_next,
1075                                             lhc + HAMMER2_DIRHASH_LOMASK, 0);
1076         }
1077         hammer2_inode_unlock(spmp->iroot, cparent);
1078
1079         /*
1080          * PFS could not be found?
1081          */
1082         if (cluster == NULL) {
1083                 kprintf("hammer2_mount: PFS label not found\n");
1084                 hammer2_unmount_helper(mp, NULL, hmp);
1085                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1086                 hammer2_vfs_unmount(mp, MNT_FORCE);
1087
1088                 return EINVAL;
1089         }
1090
1091         /*
1092          * Acquire the pmp structure (it should have already been allocated
1093          * via hammer2_update_pmps() so do not pass cluster in to add to
1094          * available chains).
1095          *
1096          * Check if the cluster has already been mounted.  A cluster can
1097          * only be mounted once, use null mounts to mount additional copies.
1098          */
1099         ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1100         hammer2_cluster_bref(cluster, &bref);
1101         pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid);
1102         hammer2_cluster_unlock(cluster);
1103         hammer2_cluster_drop(cluster);
1104
1105         if (pmp->mp) {
1106                 kprintf("hammer2_mount: PFS already mounted!\n");
1107                 hammer2_unmount_helper(mp, NULL, hmp);
1108                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1109                 hammer2_vfs_unmount(mp, MNT_FORCE);
1110
1111                 return EBUSY;
1112         }
1113
1114         /*
1115          * Finish the mount
1116          */
1117         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1118
1119         mp->mnt_flag = MNT_LOCAL;
1120         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1121         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1122
1123         /*
1124          * required mount structure initializations
1125          */
1126         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1127         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1128
1129         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1130         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1131
1132         /*
1133          * Optional fields
1134          */
1135         mp->mnt_iosize_max = MAXPHYS;
1136
1137         /*
1138          * Connect up mount pointers.
1139          */
1140         hammer2_mount_helper(mp, pmp);
1141
1142         lockmgr(&hammer2_mntlk, LK_RELEASE);
1143
1144         /*
1145          * With the cluster operational install ihidden.
1146          * (only applicable to pfs mounts, not applicable to spmp)
1147          */
1148         hammer2_inode_install_hidden(pmp);
1149
1150         /*
1151          * Finish setup
1152          */
1153         vfs_getnewfsid(mp);
1154         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1155         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1156         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1157
1158         copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
1159         bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1160         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1161         copyinstr(path, mp->mnt_stat.f_mntonname,
1162                   sizeof(mp->mnt_stat.f_mntonname) - 1,
1163                   &size);
1164
1165         /*
1166          * Initial statfs to prime mnt_stat.
1167          */
1168         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1169
1170         return 0;
1171 }
1172
1173 /*
1174  * Scan PFSs under the super-root and create hammer2_pfs structures.
1175  */
1176 static
1177 void
1178 hammer2_update_pmps(hammer2_dev_t *hmp)
1179 {
1180         const hammer2_inode_data_t *ripdata;
1181         hammer2_cluster_t *cparent;
1182         hammer2_cluster_t *cluster;
1183         hammer2_blockref_t bref;
1184         hammer2_pfs_t *spmp;
1185         hammer2_pfs_t *pmp;
1186         hammer2_key_t key_next;
1187
1188         /*
1189          * Lookup mount point under the media-localized super-root.
1190          *
1191          * cluster->pmp will incorrectly point to spmp and must be fixed
1192          * up later on.
1193          */
1194         spmp = hmp->spmp;
1195         hammer2_inode_lock(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1196         cparent = hammer2_inode_cluster(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1197         cluster = hammer2_cluster_lookup(cparent, &key_next,
1198                                          HAMMER2_KEY_MIN,
1199                                          HAMMER2_KEY_MAX,
1200                                          0);
1201         while (cluster) {
1202                 if (hammer2_cluster_type(cluster) != HAMMER2_BREF_TYPE_INODE)
1203                         continue;
1204                 ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1205                 hammer2_cluster_bref(cluster, &bref);
1206                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1207
1208                 pmp = hammer2_pfsalloc(cluster, ripdata, bref.modify_tid);
1209                 cluster = hammer2_cluster_next(cparent, cluster,
1210                                                &key_next,
1211                                                key_next,
1212                                                HAMMER2_KEY_MAX,
1213                                                0);
1214         }
1215         hammer2_inode_unlock(spmp->iroot, cparent);
1216 }
1217
1218 static
1219 int
1220 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path,
1221                 struct vnode *devvp, struct ucred *cred)
1222 {
1223         int error;
1224
1225         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1226                 error = hammer2_recovery(hmp);
1227         } else {
1228                 error = 0;
1229         }
1230         return error;
1231 }
1232
1233 static
1234 int
1235 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1236 {
1237         hammer2_pfs_t *pmp;
1238         int flags;
1239         int error = 0;
1240
1241         pmp = MPTOPMP(mp);
1242
1243         if (pmp == NULL)
1244                 return(0);
1245
1246         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1247
1248         /*
1249          * If mount initialization proceeded far enough we must flush
1250          * its vnodes and sync the underlying mount points.  Three syncs
1251          * are required to fully flush the filesystem (freemap updates lag
1252          * by one flush, and one extra for safety).
1253          */
1254         if (mntflags & MNT_FORCE)
1255                 flags = FORCECLOSE;
1256         else
1257                 flags = 0;
1258         if (pmp->iroot) {
1259                 error = vflush(mp, 0, flags);
1260                 if (error)
1261                         goto failed;
1262                 hammer2_vfs_sync(mp, MNT_WAIT);
1263                 hammer2_vfs_sync(mp, MNT_WAIT);
1264                 hammer2_vfs_sync(mp, MNT_WAIT);
1265         }
1266
1267         /*
1268          * Cleanup the frontend support XOPS threads
1269          */
1270         hammer2_xop_helper_cleanup(pmp);
1271
1272         /*
1273          * Cleanup our reference on ihidden.
1274          */
1275         if (pmp->ihidden) {
1276                 hammer2_inode_drop(pmp->ihidden);
1277                 pmp->ihidden = NULL;
1278         }
1279         if (pmp->mp)
1280                 hammer2_unmount_helper(mp, pmp, NULL);
1281
1282         error = 0;
1283 failed:
1284         lockmgr(&hammer2_mntlk, LK_RELEASE);
1285
1286         return (error);
1287 }
1288
1289 /*
1290  * Mount helper, hook the system mount into our PFS.
1291  * The mount lock is held.
1292  *
1293  * We must bump the mount_count on related devices for any
1294  * mounted PFSs.
1295  */
1296 static
1297 void
1298 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1299 {
1300         hammer2_cluster_t *cluster;
1301         hammer2_chain_t *rchain;
1302         int i;
1303
1304         mp->mnt_data = (qaddr_t)pmp;
1305         pmp->mp = mp;
1306
1307         /*
1308          * After pmp->mp is set we have to adjust hmp->mount_count.
1309          */
1310         cluster = &pmp->iroot->cluster;
1311         for (i = 0; i < cluster->nchains; ++i) {
1312                 rchain = cluster->array[i].chain;
1313                 if (rchain == NULL)
1314                         continue;
1315                 ++rchain->hmp->mount_count;
1316                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1317                         rchain->hmp, rchain->hmp->mount_count);
1318         }
1319
1320         /*
1321          * Create missing Xop threads
1322          */
1323         hammer2_xop_helper_create(pmp);
1324 }
1325
1326 /*
1327  * Mount helper, unhook the system mount from our PFS.
1328  * The mount lock is held.
1329  *
1330  * If hmp is supplied a mount responsible for being the first to open
1331  * the block device failed and the block device and all PFSs using the
1332  * block device must be cleaned up.
1333  *
1334  * If pmp is supplied multiple devices might be backing the PFS and each
1335  * must be disconnect.  This might not be the last PFS using some of the
1336  * underlying devices.  Also, we have to adjust our hmp->mount_count
1337  * accounting for the devices backing the pmp which is now undergoing an
1338  * unmount.
1339  */
1340 static
1341 void
1342 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1343 {
1344         hammer2_cluster_t *cluster;
1345         hammer2_chain_t *rchain;
1346         struct vnode *devvp;
1347         int dumpcnt;
1348         int ronly = 0;
1349         int i;
1350
1351         /*
1352          * If no device supplied this is a high-level unmount and we have to
1353          * to disconnect the mount, adjust mount_count, and locate devices
1354          * that might now have no mounts.
1355          */
1356         if (pmp) {
1357                 KKASSERT(hmp == NULL);
1358                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1359                 pmp->mp = NULL;
1360                 mp->mnt_data = NULL;
1361
1362                 /*
1363                  * After pmp->mp is cleared we have to account for
1364                  * mount_count.
1365                  */
1366                 cluster = &pmp->iroot->cluster;
1367                 for (i = 0; i < cluster->nchains; ++i) {
1368                         rchain = cluster->array[i].chain;
1369                         if (rchain == NULL)
1370                                 continue;
1371                         --rchain->hmp->mount_count;
1372                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1373                                 rchain->hmp, rchain->hmp->mount_count);
1374                         /* scrapping hmp now may invalidate the pmp */
1375                 }
1376 again:
1377                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1378                         if (hmp->mount_count == 0) {
1379                                 hammer2_unmount_helper(NULL, NULL, hmp);
1380                                 goto again;
1381                         }
1382                 }
1383                 return;
1384         }
1385
1386         /*
1387          * Try to terminate the block device.  We can't terminate it if
1388          * there are still PFSs referencing it.
1389          */
1390         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1391                 hmp, hmp->mount_count);
1392         if (hmp->mount_count)
1393                 return;
1394
1395         hammer2_pfsfree_scan(hmp);
1396         hammer2_dev_exlock(hmp);        /* XXX order */
1397
1398         /*
1399          * Cycle the volume data lock as a safety (probably not needed any
1400          * more).  To ensure everything is out we need to flush at least
1401          * three times.  (1) The running of the unlinkq can dirty the
1402          * filesystem, (2) A normal flush can dirty the freemap, and
1403          * (3) ensure that the freemap is fully synchronized.
1404          *
1405          * The next mount's recovery scan can clean everything up but we want
1406          * to leave the filesystem in a 100% clean state on a normal unmount.
1407          */
1408 #if 0
1409         hammer2_voldata_lock(hmp);
1410         hammer2_voldata_unlock(hmp);
1411 #endif
1412         hammer2_iocom_uninit(hmp);
1413
1414         if ((hmp->vchain.flags | hmp->fchain.flags) &
1415             HAMMER2_CHAIN_FLUSH_MASK) {
1416                 kprintf("hammer2_unmount: chains left over "
1417                         "after final sync\n");
1418                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1419                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1420
1421                 if (hammer2_debug & 0x0010)
1422                         Debugger("entered debugger");
1423         }
1424
1425         KKASSERT(hmp->spmp == NULL);
1426
1427         /*
1428          * Finish up with the device vnode
1429          */
1430         if ((devvp = hmp->devvp) != NULL) {
1431                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1432                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1433                 hmp->devvp = NULL;
1434                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1435                 vn_unlock(devvp);
1436                 vrele(devvp);
1437                 devvp = NULL;
1438         }
1439
1440         /*
1441          * Clear vchain/fchain flags that might prevent final cleanup
1442          * of these chains.
1443          */
1444         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1445                 atomic_clear_int(&hmp->vchain.flags,
1446                                  HAMMER2_CHAIN_MODIFIED);
1447                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1448                 hammer2_chain_drop(&hmp->vchain);
1449         }
1450         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1451                 atomic_clear_int(&hmp->vchain.flags,
1452                                  HAMMER2_CHAIN_UPDATE);
1453                 hammer2_chain_drop(&hmp->vchain);
1454         }
1455
1456         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1457                 atomic_clear_int(&hmp->fchain.flags,
1458                                  HAMMER2_CHAIN_MODIFIED);
1459                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1460                 hammer2_chain_drop(&hmp->fchain);
1461         }
1462         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1463                 atomic_clear_int(&hmp->fchain.flags,
1464                                  HAMMER2_CHAIN_UPDATE);
1465                 hammer2_chain_drop(&hmp->fchain);
1466         }
1467
1468         /*
1469          * Final drop of embedded freemap root chain to
1470          * clean up fchain.core (fchain structure is not
1471          * flagged ALLOCATED so it is cleaned out and then
1472          * left to rot).
1473          */
1474         hammer2_chain_drop(&hmp->fchain);
1475
1476         /*
1477          * Final drop of embedded volume root chain to clean
1478          * up vchain.core (vchain structure is not flagged
1479          * ALLOCATED so it is cleaned out and then left to
1480          * rot).
1481          */
1482         dumpcnt = 50;
1483         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1484         dumpcnt = 50;
1485         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1486         hammer2_dev_unlock(hmp);
1487         hammer2_chain_drop(&hmp->vchain);
1488
1489         hammer2_io_cleanup(hmp, &hmp->iotree);
1490         if (hmp->iofree_count) {
1491                 kprintf("io_cleanup: %d I/O's left hanging\n",
1492                         hmp->iofree_count);
1493         }
1494
1495         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1496         kmalloc_destroy(&hmp->mchain);
1497         kfree(hmp, M_HAMMER2);
1498 }
1499
1500 static
1501 int
1502 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1503              ino_t ino, struct vnode **vpp)
1504 {
1505         kprintf("hammer2_vget\n");
1506         return (EOPNOTSUPP);
1507 }
1508
1509 static
1510 int
1511 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1512 {
1513         hammer2_pfs_t *pmp;
1514         hammer2_cluster_t *cparent;
1515         int error;
1516         struct vnode *vp;
1517
1518         pmp = MPTOPMP(mp);
1519         if (pmp->iroot == NULL) {
1520                 *vpp = NULL;
1521                 error = EINVAL;
1522         } else {
1523                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_ALWAYS |
1524                                                HAMMER2_RESOLVE_SHARED);
1525                 cparent = hammer2_inode_cluster(pmp->iroot,
1526                                                 HAMMER2_RESOLVE_ALWAYS |
1527                                                 HAMMER2_RESOLVE_SHARED);
1528
1529                 /*
1530                  * Initialize pmp->inode_tid and pmp->modify_tid on first access
1531                  * to the root of mount that resolves good.
1532                  * XXX probably not the best place for this.
1533                  */
1534                 if (pmp->inode_tid == 0 &&
1535                     cparent->error == 0 && cparent->focus) {
1536                         const hammer2_inode_data_t *ripdata;
1537                         hammer2_blockref_t bref;
1538
1539                         ripdata = &hammer2_cluster_rdata(cparent)->ipdata;
1540                         hammer2_cluster_bref(cparent, &bref);
1541                         pmp->inode_tid = ripdata->meta.pfs_inum + 1;
1542                         if (pmp->inode_tid < HAMMER2_INODE_START)
1543                                 pmp->inode_tid = HAMMER2_INODE_START;
1544                         pmp->modify_tid = bref.modify_tid + 1;
1545                         pmp->iroot->meta = ripdata->meta;
1546                         hammer2_cluster_bref(cparent, &pmp->iroot->bref);
1547                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1548                                 pmp->inode_tid, pmp->modify_tid);
1549                 }
1550
1551                 vp = hammer2_igetv(pmp->iroot, &error);
1552                 hammer2_inode_unlock(pmp->iroot, cparent);
1553                 *vpp = vp;
1554                 if (vp == NULL)
1555                         kprintf("vnodefail\n");
1556         }
1557
1558         return (error);
1559 }
1560
1561 /*
1562  * Filesystem status
1563  *
1564  * XXX incorporate ipdata->meta.inode_quota and data_quota
1565  */
1566 static
1567 int
1568 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1569 {
1570         hammer2_pfs_t *pmp;
1571         hammer2_dev_t *hmp;
1572         hammer2_blockref_t bref;
1573
1574         pmp = MPTOPMP(mp);
1575         KKASSERT(pmp->iroot->cluster.nchains >= 1);
1576         hmp = pmp->iroot->cluster.focus->hmp;   /* iroot retains focus */
1577         bref = pmp->iroot->cluster.focus->bref; /* no lock */
1578
1579         mp->mnt_stat.f_files = bref.inode_count;
1580         mp->mnt_stat.f_ffree = 0;
1581         mp->mnt_stat.f_blocks = (bref.data_count +
1582                                  hmp->voldata.allocator_free) /
1583                                 mp->mnt_vstat.f_bsize;
1584         mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free /
1585                                 mp->mnt_vstat.f_bsize;
1586         mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1587
1588         *sbp = mp->mnt_stat;
1589         return (0);
1590 }
1591
1592 static
1593 int
1594 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1595 {
1596         hammer2_pfs_t *pmp;
1597         hammer2_dev_t *hmp;
1598         hammer2_blockref_t bref;
1599
1600         pmp = MPTOPMP(mp);
1601         KKASSERT(pmp->iroot->cluster.nchains >= 1);
1602         hmp = pmp->iroot->cluster.focus->hmp;   /* iroot retains focus */
1603         bref = pmp->iroot->cluster.focus->bref; /* no lock */
1604
1605         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1606         mp->mnt_vstat.f_files = bref.inode_count;
1607         mp->mnt_vstat.f_ffree = 0;
1608         mp->mnt_vstat.f_blocks = (bref.data_count +
1609                                  hmp->voldata.allocator_free) /
1610                                 mp->mnt_vstat.f_bsize;
1611         mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1612                                 mp->mnt_vstat.f_bsize;
1613         mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1614
1615         *sbp = mp->mnt_vstat;
1616         return (0);
1617 }
1618
1619 /*
1620  * Mount-time recovery (RW mounts)
1621  *
1622  * Updates to the free block table are allowed to lag flushes by one
1623  * transaction.  In case of a crash, then on a fresh mount we must do an
1624  * incremental scan of the last committed transaction id and make sure that
1625  * all related blocks have been marked allocated.
1626  *
1627  * The super-root topology and each PFS has its own transaction id domain,
1628  * so we must track PFS boundary transitions.
1629  */
1630 struct hammer2_recovery_elm {
1631         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1632         hammer2_chain_t *chain;
1633         hammer2_tid_t sync_tid;
1634 };
1635
1636 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1637
1638 struct hammer2_recovery_info {
1639         struct hammer2_recovery_list list;
1640         int     depth;
1641 };
1642
1643 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1644                         hammer2_chain_t *parent,
1645                         struct hammer2_recovery_info *info,
1646                         hammer2_tid_t sync_tid);
1647
1648 #define HAMMER2_RECOVERY_MAXDEPTH       10
1649
1650 static
1651 int
1652 hammer2_recovery(hammer2_dev_t *hmp)
1653 {
1654         struct hammer2_recovery_info info;
1655         struct hammer2_recovery_elm *elm;
1656         hammer2_chain_t *parent;
1657         hammer2_tid_t sync_tid;
1658         hammer2_tid_t mirror_tid;
1659         int error;
1660         int cumulative_error = 0;
1661
1662         hammer2_trans_init(hmp->spmp, 0);
1663
1664         sync_tid = hmp->voldata.freemap_tid;
1665         mirror_tid = hmp->voldata.mirror_tid;
1666
1667         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1668         if (sync_tid >= mirror_tid) {
1669                 kprintf(" no recovery needed\n");
1670         } else {
1671                 kprintf(" freemap recovery %016jx-%016jx\n",
1672                         sync_tid + 1, mirror_tid);
1673         }
1674
1675         TAILQ_INIT(&info.list);
1676         info.depth = 0;
1677         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1678         cumulative_error = hammer2_recovery_scan(hmp, parent,
1679                                                  &info, sync_tid);
1680         hammer2_chain_lookup_done(parent);
1681
1682         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1683                 TAILQ_REMOVE(&info.list, elm, entry);
1684                 parent = elm->chain;
1685                 sync_tid = elm->sync_tid;
1686                 kfree(elm, M_HAMMER2);
1687
1688                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1689                 error = hammer2_recovery_scan(hmp, parent,
1690                                               &info, hmp->voldata.freemap_tid);
1691                 hammer2_chain_unlock(parent);
1692                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
1693                 if (error)
1694                         cumulative_error = error;
1695         }
1696         hammer2_trans_done(hmp->spmp);
1697
1698         return cumulative_error;
1699 }
1700
1701 static
1702 int
1703 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1704                       struct hammer2_recovery_info *info,
1705                       hammer2_tid_t sync_tid)
1706 {
1707         const hammer2_inode_data_t *ripdata;
1708         hammer2_chain_t *chain;
1709         int cache_index;
1710         int cumulative_error = 0;
1711         int error;
1712
1713         /*
1714          * Adjust freemap to ensure that the block(s) are marked allocated.
1715          */
1716         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1717                 hammer2_freemap_adjust(hmp, &parent->bref,
1718                                        HAMMER2_FREEMAP_DORECOVER);
1719         }
1720
1721         /*
1722          * Check type for recursive scan
1723          */
1724         switch(parent->bref.type) {
1725         case HAMMER2_BREF_TYPE_VOLUME:
1726                 /* data already instantiated */
1727                 break;
1728         case HAMMER2_BREF_TYPE_INODE:
1729                 /*
1730                  * Must instantiate data for DIRECTDATA test and also
1731                  * for recursion.
1732                  */
1733                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1734                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
1735                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1736                         /* not applicable to recovery scan */
1737                         hammer2_chain_unlock(parent);
1738                         return 0;
1739                 }
1740                 hammer2_chain_unlock(parent);
1741                 break;
1742         case HAMMER2_BREF_TYPE_INDIRECT:
1743                 /*
1744                  * Must instantiate data for recursion
1745                  */
1746                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1747                 hammer2_chain_unlock(parent);
1748                 break;
1749         case HAMMER2_BREF_TYPE_DATA:
1750         case HAMMER2_BREF_TYPE_FREEMAP:
1751         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1752         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1753                 /* not applicable to recovery scan */
1754                 return 0;
1755                 break;
1756         default:
1757                 return EDOM;
1758         }
1759
1760         /*
1761          * Defer operation if depth limit reached or if we are crossing a
1762          * PFS boundary.
1763          */
1764         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1765                 struct hammer2_recovery_elm *elm;
1766
1767                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1768                 elm->chain = parent;
1769                 elm->sync_tid = sync_tid;
1770                 hammer2_chain_ref(parent);
1771                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
1772                 /* unlocked by caller */
1773
1774                 return(0);
1775         }
1776
1777
1778         /*
1779          * Recursive scan of the last flushed transaction only.  We are
1780          * doing this without pmp assignments so don't leave the chains
1781          * hanging around after we are done with them.
1782          */
1783         cache_index = 0;
1784         chain = hammer2_chain_scan(parent, NULL, &cache_index,
1785                                    HAMMER2_LOOKUP_NODATA);
1786         while (chain) {
1787                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
1788                 if (chain->bref.mirror_tid > sync_tid) {
1789                         ++info->depth;
1790                         error = hammer2_recovery_scan(hmp, chain,
1791                                                       info, sync_tid);
1792                         --info->depth;
1793                         if (error)
1794                                 cumulative_error = error;
1795                 }
1796
1797                 /*
1798                  * Flush the recovery at the PFS boundary to stage it for
1799                  * the final flush of the super-root topology.
1800                  */
1801                 if ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
1802                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
1803                         hammer2_flush(chain, 1);
1804                 }
1805                 chain = hammer2_chain_scan(parent, chain, &cache_index,
1806                                            HAMMER2_LOOKUP_NODATA);
1807         }
1808
1809         return cumulative_error;
1810 }
1811
1812 /*
1813  * Sync a mount point; this is called on a per-mount basis from the
1814  * filesystem syncer process periodically and whenever a user issues
1815  * a sync.
1816  */
1817 int
1818 hammer2_vfs_sync(struct mount *mp, int waitfor)
1819 {
1820         hammer2_xop_flush_t *xop;
1821         struct hammer2_sync_info info;
1822         hammer2_inode_t *iroot;
1823         hammer2_pfs_t *pmp;
1824         int flags;
1825         int error;
1826
1827         pmp = MPTOPMP(mp);
1828         iroot = pmp->iroot;
1829         KKASSERT(iroot);
1830         KKASSERT(iroot->pmp == pmp);
1831
1832         /*
1833          * We can't acquire locks on existing vnodes while in a transaction
1834          * without risking a deadlock.  This assumes that vfsync() can be
1835          * called without the vnode locked (which it can in DragonFly).
1836          * Otherwise we'd have to implement a multi-pass or flag the lock
1837          * failures and retry.
1838          *
1839          * The reclamation code interlocks with the sync list's token
1840          * (by removing the vnode from the scan list) before unlocking
1841          * the inode, giving us time to ref the inode.
1842          */
1843         /*flags = VMSC_GETVP;*/
1844         flags = 0;
1845         if (waitfor & MNT_LAZY)
1846                 flags |= VMSC_ONEPASS;
1847
1848 #if 0
1849         /*
1850          * Preflush the vnodes using a normal transaction before interlocking
1851          * with a flush transaction.
1852          */
1853         hammer2_trans_init(pmp, 0);
1854         info.error = 0;
1855         info.waitfor = MNT_NOWAIT;
1856         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1857         hammer2_trans_done(pmp);
1858 #endif
1859
1860         /*
1861          * Start our flush transaction.  This does not return until all
1862          * concurrent transactions have completed and will prevent any
1863          * new transactions from running concurrently, except for the
1864          * buffer cache transactions.
1865          *
1866          * For efficiency do an async pass before making sure with a
1867          * synchronous pass on all related buffer cache buffers.  It
1868          * should theoretically not be possible for any new file buffers
1869          * to be instantiated during this sequence.
1870          */
1871         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH |
1872                                 HAMMER2_TRANS_PREFLUSH);
1873         hammer2_run_unlinkq(pmp);
1874
1875         info.error = 0;
1876         info.waitfor = MNT_NOWAIT;
1877         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1878         info.waitfor = MNT_WAIT;
1879         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
1880
1881         /*
1882          * Clear PREFLUSH.  This prevents (or asserts on) any new logical
1883          * buffer cache flushes which occur during the flush.  Device buffers
1884          * are not affected.
1885          */
1886         hammer2_bioq_sync(pmp);
1887         atomic_clear_int(&pmp->trans.flags, HAMMER2_TRANS_PREFLUSH);
1888
1889         /*
1890          * Use the XOP interface to concurrently flush all nodes to
1891          * synchronize the PFSROOT subtopology to the media.  A standard
1892          * end-of-scan ENOENT error indicates cluster sufficiency.
1893          *
1894          * Note that this flush will not be visible on crash recovery until
1895          * we flush the super-root topology in the next loop.
1896          *
1897          * XXX For now wait for all flushes to complete.
1898          */
1899         if (iroot) {
1900                 xop = &hammer2_xop_alloc(iroot)->xop_flush;
1901                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
1902                 error = hammer2_xop_collect(&xop->head,
1903                                             HAMMER2_XOP_COLLECT_WAITALL);
1904                 if (error == ENOENT)
1905                         error = 0;
1906         } else {
1907                 error = 0;
1908         }
1909         hammer2_trans_done(pmp);
1910
1911         return (error);
1912 }
1913
1914 /*
1915  * Sync passes.
1916  */
1917 static int
1918 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1919 {
1920         struct hammer2_sync_info *info = data;
1921         hammer2_inode_t *ip;
1922         int error;
1923
1924         /*
1925          * Degenerate cases.  Note that ip == NULL typically means the
1926          * syncer vnode itself and we don't want to vclrisdirty() in that
1927          * situation.
1928          */
1929         ip = VTOI(vp);
1930         if (ip == NULL) {
1931                 return(0);
1932         }
1933         if (vp->v_type == VNON || vp->v_type == VBAD) {
1934                 vclrisdirty(vp);
1935                 return(0);
1936         }
1937
1938         /*
1939          * VOP_FSYNC will start a new transaction so replicate some code
1940          * here to do it inline (see hammer2_vop_fsync()).
1941          *
1942          * WARNING: The vfsync interacts with the buffer cache and might
1943          *          block, we can't hold the inode lock at that time.
1944          *          However, we MUST ref ip before blocking to ensure that
1945          *          it isn't ripped out from under us (since we do not
1946          *          hold a lock on the vnode).
1947          */
1948         hammer2_inode_ref(ip);
1949         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
1950             !RB_EMPTY(&vp->v_rbdirty_tree)) {
1951                 vfsync(vp, info->waitfor, 1, NULL, NULL);
1952                 hammer2_inode_fsync(ip, NULL);
1953         }
1954         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
1955             RB_EMPTY(&vp->v_rbdirty_tree)) {
1956                 vclrisdirty(vp);
1957         }
1958
1959         hammer2_inode_drop(ip);
1960 #if 1
1961         error = 0;
1962         if (error)
1963                 info->error = error;
1964 #endif
1965         return(0);
1966 }
1967
1968 static
1969 int
1970 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
1971 {
1972         return (0);
1973 }
1974
1975 static
1976 int
1977 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
1978                struct fid *fhp, struct vnode **vpp)
1979 {
1980         return (0);
1981 }
1982
1983 static
1984 int
1985 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
1986                  int *exflagsp, struct ucred **credanonp)
1987 {
1988         return (0);
1989 }
1990
1991 /*
1992  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
1993  * header into the HMP
1994  *
1995  * XXX read four volhdrs and use the one with the highest TID whos CRC
1996  *     matches.
1997  *
1998  * XXX check iCRCs.
1999  *
2000  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2001  *     nonexistant locations.
2002  *
2003  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2004  */
2005 static
2006 int
2007 hammer2_install_volume_header(hammer2_dev_t *hmp)
2008 {
2009         hammer2_volume_data_t *vd;
2010         struct buf *bp;
2011         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2012         int error_reported;
2013         int error;
2014         int valid;
2015         int i;
2016
2017         error_reported = 0;
2018         error = 0;
2019         valid = 0;
2020         bp = NULL;
2021
2022         /*
2023          * There are up to 4 copies of the volume header (syncs iterate
2024          * between them so there is no single master).  We don't trust the
2025          * volu_size field so we don't know precisely how large the filesystem
2026          * is, so depend on the OS to return an error if we go beyond the
2027          * block device's EOF.
2028          */
2029         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2030                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2031                               HAMMER2_VOLUME_BYTES, &bp);
2032                 if (error) {
2033                         brelse(bp);
2034                         bp = NULL;
2035                         continue;
2036                 }
2037
2038                 vd = (struct hammer2_volume_data *) bp->b_data;
2039                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2040                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2041                         brelse(bp);
2042                         bp = NULL;
2043                         continue;
2044                 }
2045
2046                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2047                         /* XXX: Reversed-endianness filesystem */
2048                         kprintf("hammer2: reverse-endian filesystem detected");
2049                         brelse(bp);
2050                         bp = NULL;
2051                         continue;
2052                 }
2053
2054                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2055                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2056                                       HAMMER2_VOLUME_ICRC0_SIZE);
2057                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2058                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2059                                        HAMMER2_VOLUME_ICRC1_SIZE);
2060                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2061                         kprintf("hammer2 volume header crc "
2062                                 "mismatch copy #%d %08x/%08x\n",
2063                                 i, crc0, crc);
2064                         error_reported = 1;
2065                         brelse(bp);
2066                         bp = NULL;
2067                         continue;
2068                 }
2069                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2070                         valid = 1;
2071                         hmp->voldata = *vd;
2072                         hmp->volhdrno = i;
2073                 }
2074                 brelse(bp);
2075                 bp = NULL;
2076         }
2077         if (valid) {
2078                 hmp->volsync = hmp->voldata;
2079                 error = 0;
2080                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2081                         kprintf("hammer2: using volume header #%d\n",
2082                                 hmp->volhdrno);
2083                 }
2084         } else {
2085                 error = EINVAL;
2086                 kprintf("hammer2: no valid volume headers found!\n");
2087         }
2088         return (error);
2089 }
2090
2091 /*
2092  * This handles hysteresis on regular file flushes.  Because the BIOs are
2093  * routed to a thread it is possible for an excessive number to build up
2094  * and cause long front-end stalls long before the runningbuffspace limit
2095  * is hit, so we implement hammer2_flush_pipe to control the
2096  * hysteresis.
2097  *
2098  * This is a particular problem when compression is used.
2099  */
2100 void
2101 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2102 {
2103         atomic_add_int(&pmp->count_lwinprog, 1);
2104 }
2105
2106 void
2107 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2108 {
2109         int lwinprog;
2110
2111         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2112         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2113             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2114                 atomic_clear_int(&pmp->count_lwinprog,
2115                                  HAMMER2_LWINPROG_WAITING);
2116                 wakeup(&pmp->count_lwinprog);
2117         }
2118         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2119             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2120                 atomic_clear_int(&pmp->count_lwinprog,
2121                                  HAMMER2_LWINPROG_WAITING0);
2122                 wakeup(&pmp->count_lwinprog);
2123         }
2124 }
2125
2126 void
2127 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2128 {
2129         int lwinprog;
2130         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2131                                     HAMMER2_LWINPROG_WAITING0;
2132
2133         for (;;) {
2134                 lwinprog = pmp->count_lwinprog;
2135                 cpu_ccfence();
2136                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2137                         break;
2138                 tsleep_interlock(&pmp->count_lwinprog, 0);
2139                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2140                 lwinprog = pmp->count_lwinprog;
2141                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2142                         break;
2143                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2144         }
2145 }
2146
2147 /*
2148  * Manage excessive memory resource use for chain and related
2149  * structures.
2150  */
2151 void
2152 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2153 {
2154         uint32_t waiting;
2155         uint32_t count;
2156         uint32_t limit;
2157 #if 0
2158         static int zzticks;
2159 #endif
2160
2161         /*
2162          * Atomic check condition and wait.  Also do an early speedup of
2163          * the syncer to try to avoid hitting the wait.
2164          */
2165         for (;;) {
2166                 waiting = pmp->inmem_dirty_chains;
2167                 cpu_ccfence();
2168                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2169
2170                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2171                 if (limit < hammer2_limit_dirty_chains)
2172                         limit = hammer2_limit_dirty_chains;
2173                 if (limit < 1000)
2174                         limit = 1000;
2175
2176 #if 0
2177                 if ((int)(ticks - zzticks) > hz) {
2178                         zzticks = ticks;
2179                         kprintf("count %ld %ld\n", count, limit);
2180                 }
2181 #endif
2182
2183                 /*
2184                  * Block if there are too many dirty chains present, wait
2185                  * for the flush to clean some out.
2186                  */
2187                 if (count > limit) {
2188                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2189                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2190                                                waiting,
2191                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2192                                 speedup_syncer(pmp->mp);
2193                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2194                                        "chnmem", hz);
2195                         }
2196                         continue;       /* loop on success or fail */
2197                 }
2198
2199                 /*
2200                  * Try to start an early flush before we are forced to block.
2201                  */
2202                 if (count > limit * 7 / 10)
2203                         speedup_syncer(pmp->mp);
2204                 break;
2205         }
2206 }
2207
2208 void
2209 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2210 {
2211         if (pmp) {
2212                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2213         }
2214 }
2215
2216 void
2217 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2218 {
2219         uint32_t waiting;
2220
2221         if (pmp == NULL)
2222                 return;
2223
2224         for (;;) {
2225                 waiting = pmp->inmem_dirty_chains;
2226                 cpu_ccfence();
2227                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2228                                        waiting,
2229                                        (waiting - 1) &
2230                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2231                         break;
2232                 }
2233         }
2234
2235         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2236                 wakeup(&pmp->inmem_dirty_chains);
2237 }
2238
2239 /*
2240  * Debugging
2241  */
2242 void
2243 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2244 {
2245         hammer2_chain_t *scan;
2246         hammer2_chain_t *parent;
2247
2248         --*countp;
2249         if (*countp == 0) {
2250                 kprintf("%*.*s...\n", tab, tab, "");
2251                 return;
2252         }
2253         if (*countp < 0)
2254                 return;
2255         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2256                 tab, tab, "", pfx,
2257                 chain, chain->bref.type,
2258                 chain->bref.key, chain->bref.keybits,
2259                 chain->bref.mirror_tid);
2260
2261         kprintf("%*.*s      [%08x] (%s) refs=%d",
2262                 tab, tab, "",
2263                 chain->flags,
2264                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2265                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2266                 chain->refs);
2267
2268         parent = chain->parent;
2269         if (parent)
2270                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2271                         tab, tab, "",
2272                         parent, parent->flags, parent->refs);
2273         if (RB_EMPTY(&chain->core.rbtree)) {
2274                 kprintf("\n");
2275         } else {
2276                 kprintf(" {\n");
2277                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2278                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2279                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2280                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2281                                 chain->data->ipdata.filename);
2282                 else
2283                         kprintf("%*.*s}\n", tab, tab, "");
2284         }
2285 }