sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72 };
  73
  74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  75 static struct hammer2_mntlist hammer2_mntlist;
  76
  77 struct hammer2_pfslist hammer2_pfslist;
  78 struct lock hammer2_mntlk;
  79
  80 int hammer2_debug;
  81 int hammer2_cluster_read = 4;           /* physical read-ahead */
  82 int hammer2_cluster_write = 0;          /* bdwrite() so later inval works */
  83 int hammer2_dedup_enable = 1;
  84 int hammer2_inval_enable = 0;
  85 int hammer2_flush_pipe = 100;
  86 int hammer2_synchronous_flush = 1;
  87 int hammer2_dio_count;
  88 long hammer2_chain_allocs;
  89 long hammer2_chain_frees;
  90 long hammer2_limit_dirty_chains;
  91 long hammer2_count_modified_chains;
  92 long hammer2_iod_invals;
  93 long hammer2_iod_file_read;
  94 long hammer2_iod_meta_read;
  95 long hammer2_iod_indr_read;
  96 long hammer2_iod_fmap_read;
  97 long hammer2_iod_volu_read;
  98 long hammer2_iod_file_write;
  99 long hammer2_iod_file_wembed;
 100 long hammer2_iod_file_wzero;
 101 long hammer2_iod_file_wdedup;
 102 long hammer2_iod_meta_write;
 103 long hammer2_iod_indr_write;
 104 long hammer2_iod_fmap_write;
 105 long hammer2_iod_volu_write;
 106
 107 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 108 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 109                 "Buffer used for compression.");
 110
 111 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 112 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 113                 "Buffer used for decompression.");
 114
 115 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 116
 117 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 118            &hammer2_debug, 0, "");
 119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW,
 120            &hammer2_cluster_read, 0, "");
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 122            &hammer2_cluster_write, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 124            &hammer2_dedup_enable, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 126            &hammer2_inval_enable, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 128            &hammer2_flush_pipe, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 130            &hammer2_synchronous_flush, 0, "");
 131 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 132            &hammer2_chain_allocs, 0, "");
 133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 134            &hammer2_chain_frees, 0, "");
 135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 136            &hammer2_limit_dirty_chains, 0, "");
 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 138            &hammer2_count_modified_chains, 0, "");
 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 140            &hammer2_dio_count, 0, "");
 141
 142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 143            &hammer2_iod_invals, 0, "");
 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 145            &hammer2_iod_file_read, 0, "");
 146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 147            &hammer2_iod_meta_read, 0, "");
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 149            &hammer2_iod_indr_read, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 151            &hammer2_iod_fmap_read, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 153            &hammer2_iod_volu_read, 0, "");
 154
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 156            &hammer2_iod_file_write, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 158            &hammer2_iod_file_wembed, 0, "");
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 160            &hammer2_iod_file_wzero, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 162            &hammer2_iod_file_wdedup, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 164            &hammer2_iod_meta_write, 0, "");
 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 166            &hammer2_iod_indr_write, 0, "");
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 168            &hammer2_iod_fmap_write, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 170            &hammer2_iod_volu_write, 0, "");
 171
 172 long hammer2_check_icrc32;
 173 long hammer2_check_xxhash64;
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 175            &hammer2_check_icrc32, 0, "");
 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 177            &hammer2_check_xxhash64, 0, "");
 178
 179 static int hammer2_vfs_init(struct vfsconf *conf);
 180 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 181 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 182                                 struct ucred *cred);
 183 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 184                                 struct vnode *, struct ucred *);
 185 static int hammer2_recovery(hammer2_dev_t *hmp);
 186 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 187 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 188 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 189                                 struct ucred *cred);
 190 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 191                                 struct ucred *cred);
 192 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 193                                 struct fid *fhp, struct vnode **vpp);
 194 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 195 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 196                                 int *exflagsp, struct ucred **credanonp);
 197
 198 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 199 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 200
 201 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 202
 203 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 204 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 205                                 hammer2_dev_t *hmp);
 206
 207 /*
 208  * HAMMER2 vfs operations.
 209  */
 210 static struct vfsops hammer2_vfsops = {
 211         .vfs_init       = hammer2_vfs_init,
 212         .vfs_uninit     = hammer2_vfs_uninit,
 213         .vfs_sync       = hammer2_vfs_sync,
 214         .vfs_mount      = hammer2_vfs_mount,
 215         .vfs_unmount    = hammer2_vfs_unmount,
 216         .vfs_root       = hammer2_vfs_root,
 217         .vfs_statfs     = hammer2_vfs_statfs,
 218         .vfs_statvfs    = hammer2_vfs_statvfs,
 219         .vfs_vget       = hammer2_vfs_vget,
 220         .vfs_vptofh     = hammer2_vfs_vptofh,
 221         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 222         .vfs_checkexp   = hammer2_vfs_checkexp
 223 };
 224
 225 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 226
 227 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 228 MODULE_VERSION(hammer2, 1);
 229
 230 static
 231 int
 232 hammer2_vfs_init(struct vfsconf *conf)
 233 {
 234         static struct objcache_malloc_args margs_read;
 235         static struct objcache_malloc_args margs_write;
 236         static struct objcache_malloc_args margs_vop;
 237
 238         int error;
 239
 240         error = 0;
 241
 242         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 243                 error = EINVAL;
 244         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 245                 error = EINVAL;
 246         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 247                 error = EINVAL;
 248
 249         if (error)
 250                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 251
 252         margs_read.objsize = 65536;
 253         margs_read.mtype = M_HAMMER2_DEBUFFER;
 254
 255         margs_write.objsize = 32768;
 256         margs_write.mtype = M_HAMMER2_CBUFFER;
 257
 258         margs_vop.objsize = sizeof(hammer2_xop_t);
 259         margs_vop.mtype = M_HAMMER2;
 260
 261         /*
 262          * Note thaht for the XOPS cache we want backing store allocations
 263          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 264          * confusion), so use the backing store function that does it.  This
 265          * means that initial XOPS objects are zerod but REUSED objects are
 266          * not.  So we are responsible for cleaning the object up sufficiently
 267          * for our needs before objcache_put()ing it back (typically just the
 268          * FIFO indices).
 269          */
 270         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 271                                 0, 1, NULL, NULL, NULL,
 272                                 objcache_malloc_alloc,
 273                                 objcache_malloc_free,
 274                                 &margs_read);
 275         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 276                                 0, 1, NULL, NULL, NULL,
 277                                 objcache_malloc_alloc,
 278                                 objcache_malloc_free,
 279                                 &margs_write);
 280         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 281                                 0, 1, NULL, NULL, NULL,
 282                                 objcache_malloc_alloc_zero,
 283                                 objcache_malloc_free,
 284                                 &margs_vop);
 285
 286
 287         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 288         TAILQ_INIT(&hammer2_mntlist);
 289         TAILQ_INIT(&hammer2_pfslist);
 290
 291         hammer2_limit_dirty_chains = maxvnodes / 10;
 292         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 293                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 294
 295         return (error);
 296 }
 297
 298 static
 299 int
 300 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 301 {
 302         objcache_destroy(cache_buffer_read);
 303         objcache_destroy(cache_buffer_write);
 304         objcache_destroy(cache_xops);
 305         return 0;
 306 }
 307
 308 /*
 309  * Core PFS allocator.  Used to allocate or reference the pmp structure
 310  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 311  * The pmp can be passed in or loaded by this function using the chain and
 312  * inode data.
 313  *
 314  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 315  * transactions.  Note that synchronization does not use this field.
 316  * (typically frontend operations and synchronization cannot run on the
 317  * same PFS node at the same time).
 318  *
 319  * XXX check locking
 320  */
 321 hammer2_pfs_t *
 322 hammer2_pfsalloc(hammer2_chain_t *chain,
 323                  const hammer2_inode_data_t *ripdata,
 324                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 325 {
 326         hammer2_pfs_t *pmp;
 327         hammer2_inode_t *iroot;
 328         int count;
 329         int i;
 330         int j;
 331
 332         pmp = NULL;
 333
 334         /*
 335          * Locate or create the PFS based on the cluster id.  If ripdata
 336          * is NULL this is a spmp which is unique and is always allocated.
 337          *
 338          * If the device is mounted in local mode all PFSs are considered
 339          * independent and not part of any cluster (for debugging only).
 340          */
 341         if (ripdata) {
 342                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 343                         if (force_local != pmp->force_local)
 344                                 continue;
 345                         if (force_local == NULL &&
 346                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 347                                  sizeof(pmp->pfs_clid)) == 0) {
 348                                         break;
 349                         } else if (force_local && pmp->pfs_names[0] &&
 350                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 351                                         break;
 352                         }
 353                 }
 354         }
 355
 356         if (pmp == NULL) {
 357                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 358                 pmp->force_local = force_local;
 359                 hammer2_trans_manage_init(pmp);
 360                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 361                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 362                 lockinit(&pmp->lock, "pfslk", 0, 0);
 363                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 364                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 365                 spin_init(&pmp->xop_spin, "h2xop");
 366                 spin_init(&pmp->lru_spin, "h2lru");
 367                 RB_INIT(&pmp->inum_tree);
 368                 TAILQ_INIT(&pmp->sideq);
 369                 TAILQ_INIT(&pmp->lru_list);
 370                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 371
 372                 /*
 373                  * Distribute backend operations to threads
 374                  */
 375                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 376                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 377
 378                 /*
 379                  * Save the last media transaction id for the flusher.  Set
 380                  * initial
 381                  */
 382                 if (ripdata)
 383                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 384                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 385
 386                 /*
 387                  * The synchronization thread may start too early, make
 388                  * sure it stays frozen until we are ready to let it go.
 389                  * XXX
 390                  */
 391                 /*
 392                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 393                                          HAMMER2_THREAD_REMASTER;
 394                 */
 395         }
 396
 397         /*
 398          * Create the PFS's root inode and any missing XOP helper threads.
 399          */
 400         if ((iroot = pmp->iroot) == NULL) {
 401                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 402                 pmp->iroot = iroot;
 403                 hammer2_inode_ref(iroot);
 404                 hammer2_inode_unlock(iroot);
 405         }
 406
 407         /*
 408          * Stop here if no chain is passed in.
 409          */
 410         if (chain == NULL)
 411                 goto done;
 412
 413         /*
 414          * When a chain is passed in we must add it to the PFS's root
 415          * inode, update pmp->pfs_types[], and update the syncronization
 416          * threads.
 417          *
 418          * When forcing local mode, mark the PFS as a MASTER regardless.
 419          *
 420          * At the moment empty spots can develop due to removals or failures.
 421          * Ultimately we want to re-fill these spots but doing so might
 422          * confused running code. XXX
 423          */
 424         hammer2_inode_ref(iroot);
 425         hammer2_mtx_ex(&iroot->lock);
 426         j = iroot->cluster.nchains;
 427
 428         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 429
 430         if (j == HAMMER2_MAXCLUSTER) {
 431                 kprintf("hammer2_mount: cluster full!\n");
 432                 /* XXX fatal error? */
 433         } else {
 434                 KKASSERT(chain->pmp == NULL);
 435                 chain->pmp = pmp;
 436                 hammer2_chain_ref(chain);
 437                 iroot->cluster.array[j].chain = chain;
 438                 if (force_local)
 439                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 440                 else
 441                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 442                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 443                 pmp->pfs_hmps[j] = chain->hmp;
 444
 445                 /*
 446                  * If the PFS is already mounted we must account
 447                  * for the mount_count here.
 448                  */
 449                 if (pmp->mp)
 450                         ++chain->hmp->mount_count;
 451
 452                 /*
 453                  * May have to fixup dirty chain tracking.  Previous
 454                  * pmp was NULL so nothing to undo.
 455                  */
 456                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 457                         hammer2_pfs_memory_inc(pmp);
 458                 ++j;
 459         }
 460         iroot->cluster.nchains = j;
 461
 462         /*
 463          * Update nmasters from any PFS inode which is part of the cluster.
 464          * It is possible that this will result in a value which is too
 465          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 466          * override this value later on.
 467          *
 468          * (This informs us of masters that might not currently be
 469          *  discoverable by this mount).
 470          */
 471         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 472                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 473         }
 474
 475         /*
 476          * Count visible masters.  Masters are usually added with
 477          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 478          * are more (XXX and must update the master inodes).
 479          */
 480         count = 0;
 481         for (i = 0; i < iroot->cluster.nchains; ++i) {
 482                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 483                         ++count;
 484         }
 485         if (pmp->pfs_nmasters < count)
 486                 pmp->pfs_nmasters = count;
 487
 488         /*
 489          * Create missing synchronization and support threads.
 490          *
 491          * Single-node masters (including snapshots) have nothing to
 492          * synchronize and do not require this thread.
 493          *
 494          * Multi-node masters or any number of soft masters, slaves, copy,
 495          * or other PFS types need the thread.
 496          *
 497          * Each thread is responsible for its particular cluster index.
 498          * We use independent threads so stalls or mismatches related to
 499          * any given target do not affect other targets.
 500          */
 501         for (i = 0; i < iroot->cluster.nchains; ++i) {
 502                 /*
 503                  * Single-node masters (including snapshots) have nothing
 504                  * to synchronize and will make direct xops support calls,
 505                  * thus they do not require this thread.
 506                  *
 507                  * Note that there can be thousands of snapshots.  We do not
 508                  * want to create thousands of threads.
 509                  */
 510                 if (pmp->pfs_nmasters <= 1 &&
 511                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 512                         continue;
 513                 }
 514
 515                 /*
 516                  * Sync support thread
 517                  */
 518                 if (pmp->sync_thrs[i].td == NULL) {
 519                         hammer2_thr_create(&pmp->sync_thrs[i], pmp,
 520                                            "h2nod", i, -1,
 521                                            hammer2_primary_sync_thread);
 522                 }
 523         }
 524
 525         /*
 526          * Create missing Xop threads
 527          *
 528          * NOTE: We create helper threads for all mounted PFSs or any
 529          *       PFSs with 2+ nodes (so the sync thread can update them,
 530          *       even if not mounted).
 531          */
 532         if (pmp->mp || iroot->cluster.nchains >= 2)
 533                 hammer2_xop_helper_create(pmp);
 534
 535         hammer2_mtx_unlock(&iroot->lock);
 536         hammer2_inode_drop(iroot);
 537 done:
 538         return pmp;
 539 }
 540
 541 /*
 542  * Deallocate an element of a probed PFS.  If destroying and this is a
 543  * MASTER, adjust nmasters.
 544  *
 545  * This function does not physically destroy the PFS element in its device
 546  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 547  */
 548 void
 549 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 550 {
 551         hammer2_inode_t *iroot;
 552         hammer2_chain_t *chain;
 553         int j;
 554
 555         /*
 556          * Cleanup our reference on iroot.  iroot is (should) not be needed
 557          * by the flush code.
 558          */
 559         iroot = pmp->iroot;
 560         if (iroot) {
 561                 /*
 562                  * Stop synchronizing
 563                  *
 564                  * XXX flush after acquiring the iroot lock.
 565                  * XXX clean out the cluster index from all inode structures.
 566                  */
 567                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 568
 569                 /*
 570                  * Remove the cluster index from the group.  If destroying
 571                  * the PFS and this is a master, adjust pfs_nmasters.
 572                  */
 573                 hammer2_mtx_ex(&iroot->lock);
 574                 chain = iroot->cluster.array[clindex].chain;
 575                 iroot->cluster.array[clindex].chain = NULL;
 576
 577                 switch(pmp->pfs_types[clindex]) {
 578                 case HAMMER2_PFSTYPE_MASTER:
 579                         if (destroying && pmp->pfs_nmasters > 0)
 580                                 --pmp->pfs_nmasters;
 581                         /* XXX adjust ripdata->meta.pfs_nmasters */
 582                         break;
 583                 default:
 584                         break;
 585                 }
 586                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 587
 588                 hammer2_mtx_unlock(&iroot->lock);
 589
 590                 /*
 591                  * Release the chain.
 592                  */
 593                 if (chain) {
 594                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 595                         hammer2_chain_drop(chain);
 596                 }
 597
 598                 /*
 599                  * Terminate all XOP threads for the cluster index.
 600                  */
 601                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 602                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 603         }
 604 }
 605
 606 /*
 607  * Destroy a PFS, typically only occurs after the last mount on a device
 608  * has gone away.
 609  */
 610 static void
 611 hammer2_pfsfree(hammer2_pfs_t *pmp)
 612 {
 613         hammer2_inode_t *iroot;
 614         hammer2_chain_t *chain;
 615         int i;
 616         int j;
 617
 618         /*
 619          * Cleanup our reference on iroot.  iroot is (should) not be needed
 620          * by the flush code.
 621          */
 622         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 623
 624         iroot = pmp->iroot;
 625         if (iroot) {
 626                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 627                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 628                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 629                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 630                 }
 631 #if REPORT_REFS_ERRORS
 632                 if (pmp->iroot->refs != 1)
 633                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 634                                 pmp->iroot, pmp->iroot->refs);
 635 #else
 636                 KKASSERT(pmp->iroot->refs == 1);
 637 #endif
 638                 /* ref for pmp->iroot */
 639                 hammer2_inode_drop(pmp->iroot);
 640                 pmp->iroot = NULL;
 641         }
 642
 643         /*
 644          * Cleanup chains remaining on LRU list.
 645          */
 646         kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
 647         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 648                 hammer2_chain_ref(chain);
 649                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 650                 hammer2_chain_drop(chain);
 651         }
 652
 653         /*
 654          * Free remaining pmp resources
 655          */
 656         kmalloc_destroy(&pmp->mmsg);
 657         kmalloc_destroy(&pmp->minode);
 658
 659         kfree(pmp, M_HAMMER2);
 660 }
 661
 662 /*
 663  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 664  * empty is terminated and freed.
 665  *
 666  * XXX inefficient.
 667  */
 668 static void
 669 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 670 {
 671         hammer2_pfs_t *pmp;
 672         hammer2_inode_t *iroot;
 673         hammer2_chain_t *rchain;
 674         int didfreeze;
 675         int i;
 676         int j;
 677
 678 again:
 679         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 680                 if ((iroot = pmp->iroot) == NULL)
 681                         continue;
 682                 if (hmp->spmp == pmp) {
 683                         kprintf("unmount hmp %p remove spmp %p\n",
 684                                 hmp, pmp);
 685                         hmp->spmp = NULL;
 686                 }
 687
 688                 /*
 689                  * Determine if this PFS is affected.  If it is we must
 690                  * freeze all management threads and lock its iroot.
 691                  *
 692                  * Freezing a management thread forces it idle, operations
 693                  * in-progress will be aborted and it will have to start
 694                  * over again when unfrozen, or exit if told to exit.
 695                  */
 696                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 697                         if (pmp->pfs_hmps[i] == hmp)
 698                                 break;
 699                 }
 700                 if (i != HAMMER2_MAXCLUSTER) {
 701                         /*
 702                          * Make sure all synchronization threads are locked
 703                          * down.
 704                          */
 705                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 706                                 if (pmp->pfs_hmps[i] == NULL)
 707                                         continue;
 708                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 709                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 710                                         hammer2_thr_freeze_async(
 711                                                 &pmp->xop_groups[j].thrs[i]);
 712                                 }
 713                         }
 714                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 715                                 if (pmp->pfs_hmps[i] == NULL)
 716                                         continue;
 717                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 718                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 719                                         hammer2_thr_freeze(
 720                                                 &pmp->xop_groups[j].thrs[i]);
 721                                 }
 722                         }
 723
 724                         /*
 725                          * Lock the inode and clean out matching chains.
 726                          * Note that we cannot use hammer2_inode_lock_*()
 727                          * here because that would attempt to validate the
 728                          * cluster that we are in the middle of ripping
 729                          * apart.
 730                          *
 731                          * WARNING! We are working directly on the inodes
 732                          *          embedded cluster.
 733                          */
 734                         hammer2_mtx_ex(&iroot->lock);
 735
 736                         /*
 737                          * Remove the chain from matching elements of the PFS.
 738                          */
 739                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 740                                 if (pmp->pfs_hmps[i] != hmp)
 741                                         continue;
 742                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 743                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 744                                         hammer2_thr_delete(
 745                                                 &pmp->xop_groups[j].thrs[i]);
 746                                 }
 747                                 rchain = iroot->cluster.array[i].chain;
 748                                 iroot->cluster.array[i].chain = NULL;
 749                                 pmp->pfs_types[i] = 0;
 750                                 if (pmp->pfs_names[i]) {
 751                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 752                                         pmp->pfs_names[i] = NULL;
 753                                 }
 754                                 if (rchain) {
 755                                         hammer2_chain_drop(rchain);
 756                                         /* focus hint */
 757                                         if (iroot->cluster.focus == rchain)
 758                                                 iroot->cluster.focus = NULL;
 759                                 }
 760                                 pmp->pfs_hmps[i] = NULL;
 761                         }
 762                         hammer2_mtx_unlock(&iroot->lock);
 763                         didfreeze = 1;  /* remaster, unfreeze down below */
 764                 } else {
 765                         didfreeze = 0;
 766                 }
 767
 768                 /*
 769                  * Cleanup trailing chains.  Gaps may remain.
 770                  */
 771                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 772                         if (pmp->pfs_hmps[i])
 773                                 break;
 774                 }
 775                 iroot->cluster.nchains = i + 1;
 776
 777                 /*
 778                  * If the PMP has no elements remaining we can destroy it.
 779                  * (this will transition management threads from frozen->exit).
 780                  */
 781                 if (iroot->cluster.nchains == 0) {
 782                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 783                                 hmp, pmp);
 784                         hammer2_pfsfree(pmp);
 785                         goto again;
 786                 }
 787
 788                 /*
 789                  * If elements still remain we need to set the REMASTER
 790                  * flag and unfreeze it.
 791                  */
 792                 if (didfreeze) {
 793                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 794                                 if (pmp->pfs_hmps[i] == NULL)
 795                                         continue;
 796                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 797                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 798                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 799                                         hammer2_thr_remaster(
 800                                                 &pmp->xop_groups[j].thrs[i]);
 801                                         hammer2_thr_unfreeze(
 802                                                 &pmp->xop_groups[j].thrs[i]);
 803                                 }
 804                         }
 805                 }
 806         }
 807 }
 808
 809 /*
 810  * Mount or remount HAMMER2 fileystem from physical media
 811  *
 812  *      mountroot
 813  *              mp              mount point structure
 814  *              path            NULL
 815  *              data            <unused>
 816  *              cred            <unused>
 817  *
 818  *      mount
 819  *              mp              mount point structure
 820  *              path            path to mount point
 821  *              data            pointer to argument structure in user space
 822  *                      volume  volume path (device@LABEL form)
 823  *                      hflags  user mount flags
 824  *              cred            user credentials
 825  *
 826  * RETURNS:     0       Success
 827  *              !0      error number
 828  */
 829 static
 830 int
 831 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 832                   struct ucred *cred)
 833 {
 834         struct hammer2_mount_info info;
 835         hammer2_pfs_t *pmp;
 836         hammer2_pfs_t *spmp;
 837         hammer2_dev_t *hmp;
 838         hammer2_dev_t *force_local;
 839         hammer2_key_t key_next;
 840         hammer2_key_t key_dummy;
 841         hammer2_key_t lhc;
 842         struct vnode *devvp;
 843         struct nlookupdata nd;
 844         hammer2_chain_t *parent;
 845         hammer2_chain_t *chain;
 846         hammer2_cluster_t *cluster;
 847         const hammer2_inode_data_t *ripdata;
 848         hammer2_blockref_t bref;
 849         struct file *fp;
 850         char devstr[MNAMELEN];
 851         size_t size;
 852         size_t done;
 853         char *dev;
 854         char *label;
 855         int ronly = 1;
 856         int error;
 857         int cache_index;
 858         int i;
 859
 860         hmp = NULL;
 861         pmp = NULL;
 862         dev = NULL;
 863         label = NULL;
 864         devvp = NULL;
 865         cache_index = -1;
 866
 867         kprintf("hammer2_mount\n");
 868
 869         if (path == NULL) {
 870                 /*
 871                  * Root mount
 872                  */
 873                 bzero(&info, sizeof(info));
 874                 info.cluster_fd = -1;
 875                 ksnprintf(devstr, sizeof(devstr), "%s",
 876                           mp->mnt_stat.f_mntfromname);
 877                 kprintf("hammer2_mount: root '%s'\n", devstr);
 878         } else {
 879                 /*
 880                  * Non-root mount or updating a mount
 881                  */
 882                 error = copyin(data, &info, sizeof(info));
 883                 if (error)
 884                         return (error);
 885
 886                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 887                 if (error)
 888                         return (error);
 889         }
 890
 891         /* Extract device and label */
 892         dev = devstr;
 893         label = strchr(devstr, '@');
 894         if (label == NULL ||
 895             ((label + 1) - dev) > done) {
 896                 return (EINVAL);
 897         }
 898         *label = '\0';
 899         label++;
 900         if (*label == '\0')
 901                 return (EINVAL);
 902
 903         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\"\n",
 904                 dev, label);
 905
 906         if (mp->mnt_flag & MNT_UPDATE) {
 907                 /*
 908                  * Update mount.  Note that pmp->iroot->cluster is
 909                  * an inode-embedded cluster and thus cannot be
 910                  * directly locked.
 911                  *
 912                  * XXX HAMMER2 needs to implement NFS export via
 913                  *     mountctl.
 914                  */
 915                 pmp = MPTOPMP(mp);
 916                 pmp->hflags = info.hflags;
 917                 cluster = &pmp->iroot->cluster;
 918                 for (i = 0; i < cluster->nchains; ++i) {
 919                         if (cluster->array[i].chain == NULL)
 920                                 continue;
 921                         hmp = cluster->array[i].chain->hmp;
 922                         devvp = hmp->devvp;
 923                         error = hammer2_remount(hmp, mp, path,
 924                                                 devvp, cred);
 925                         if (error)
 926                                 break;
 927                 }
 928
 929                 return error;
 930         }
 931
 932         /*
 933          * HMP device mount
 934          *
 935          * If a path is specified and dev is not an empty string, lookup the
 936          * name and verify that it referes to a block device.
 937          *
 938          * If a path is specified and dev is an empty string we fall through
 939          * and locate the label in the hmp search.
 940          */
 941         if (path && *dev != 0) {
 942                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 943                 if (error == 0)
 944                         error = nlookup(&nd);
 945                 if (error == 0)
 946                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 947                 nlookup_done(&nd);
 948         } else if (path == NULL) {
 949                 /* root mount */
 950                 cdev_t cdev = kgetdiskbyname(dev);
 951                 error = bdevvp(cdev, &devvp);
 952                 if (error)
 953                         kprintf("hammer2: cannot find '%s'\n", dev);
 954         } else {
 955                 /*
 956                  * We will locate the hmp using the label in the hmp loop.
 957                  */
 958                 error = 0;
 959         }
 960
 961         if (error == 0 && devvp) {
 962                 if (vn_isdisk(devvp, &error))
 963                         error = vfs_mountedon(devvp);
 964         }
 965
 966         /*
 967          * Determine if the device has already been mounted.  After this
 968          * check hmp will be non-NULL if we are doing the second or more
 969          * hammer2 mounts from the same device.
 970          */
 971         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 972         if (devvp) {
 973                 /*
 974                  * Match the device
 975                  */
 976                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 977                         if (hmp->devvp == devvp)
 978                                 break;
 979                 }
 980         } else if (error == 0) {
 981                 /*
 982                  * Match the label to a pmp already probed.
 983                  */
 984                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 985                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 986                                 if (pmp->pfs_names[i] &&
 987                                     strcmp(pmp->pfs_names[i], label) == 0) {
 988                                         hmp = pmp->pfs_hmps[i];
 989                                         break;
 990                                 }
 991                         }
 992                         if (hmp)
 993                                 break;
 994                 }
 995                 if (hmp == NULL)
 996                         error = ENOENT;
 997         }
 998
 999         /*
1000          * Open the device if this isn't a secondary mount and construct
1001          * the H2 device mount (hmp).
1002          */
1003         if (hmp == NULL) {
1004                 hammer2_chain_t *schain;
1005                 hammer2_xid_t xid;
1006
1007                 if (error == 0 && vcount(devvp) > 0)
1008                         error = EBUSY;
1009
1010                 /*
1011                  * Now open the device
1012                  */
1013                 if (error == 0) {
1014                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1015                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1016                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1017                         if (error == 0) {
1018                                 error = VOP_OPEN(devvp,
1019                                                  ronly ? FREAD : FREAD | FWRITE,
1020                                                  FSCRED, NULL);
1021                         }
1022                         vn_unlock(devvp);
1023                 }
1024                 if (error && devvp) {
1025                         vrele(devvp);
1026                         devvp = NULL;
1027                 }
1028                 if (error) {
1029                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1030                         return error;
1031                 }
1032                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1033                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1034                 hmp->ronly = ronly;
1035                 hmp->devvp = devvp;
1036                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1037                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1038                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1039                 RB_INIT(&hmp->iotree);
1040                 spin_init(&hmp->io_spin, "hm2mount_io");
1041                 spin_init(&hmp->list_spin, "hm2mount_list");
1042                 TAILQ_INIT(&hmp->flushq);
1043
1044                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1045                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1046
1047                 /*
1048                  * vchain setup. vchain.data is embedded.
1049                  * vchain.refs is initialized and will never drop to 0.
1050                  *
1051                  * NOTE! voldata is not yet loaded.
1052                  */
1053                 hmp->vchain.hmp = hmp;
1054                 hmp->vchain.refs = 1;
1055                 hmp->vchain.data = (void *)&hmp->voldata;
1056                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1057                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1058                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1059
1060                 hammer2_chain_core_init(&hmp->vchain);
1061                 /* hmp->vchain.u.xxx is left NULL */
1062
1063                 /*
1064                  * fchain setup.  fchain.data is embedded.
1065                  * fchain.refs is initialized and will never drop to 0.
1066                  *
1067                  * The data is not used but needs to be initialized to
1068                  * pass assertion muster.  We use this chain primarily
1069                  * as a placeholder for the freemap's top-level RBTREE
1070                  * so it does not interfere with the volume's topology
1071                  * RBTREE.
1072                  */
1073                 hmp->fchain.hmp = hmp;
1074                 hmp->fchain.refs = 1;
1075                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1076                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1077                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1078                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1079                 hmp->fchain.bref.methods =
1080                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1081                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1082
1083                 hammer2_chain_core_init(&hmp->fchain);
1084                 /* hmp->fchain.u.xxx is left NULL */
1085
1086                 /*
1087                  * Install the volume header and initialize fields from
1088                  * voldata.
1089                  */
1090                 error = hammer2_install_volume_header(hmp);
1091                 if (error) {
1092                         hammer2_unmount_helper(mp, NULL, hmp);
1093                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1094                         hammer2_vfs_unmount(mp, MNT_FORCE);
1095                         return error;
1096                 }
1097
1098                 /*
1099                  * Really important to get these right or flush will get
1100                  * confused.
1101                  */
1102                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1103                 kprintf("alloc spmp %p tid %016jx\n",
1104                         hmp->spmp, hmp->voldata.mirror_tid);
1105                 spmp = hmp->spmp;
1106
1107                 /*
1108                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1109                  * is inherited from the volume header.
1110                  */
1111                 xid = 0;
1112                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1113                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1114                 hmp->vchain.pmp = spmp;
1115                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1116                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1117                 hmp->fchain.pmp = spmp;
1118
1119                 /*
1120                  * First locate the super-root inode, which is key 0
1121                  * relative to the volume header's blockset.
1122                  *
1123                  * Then locate the root inode by scanning the directory keyspace
1124                  * represented by the label.
1125                  */
1126                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1127                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1128                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1129                                       &cache_index, 0);
1130                 hammer2_chain_lookup_done(parent);
1131                 if (schain == NULL) {
1132                         kprintf("hammer2_mount: invalid super-root\n");
1133                         hammer2_unmount_helper(mp, NULL, hmp);
1134                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1135                         hammer2_vfs_unmount(mp, MNT_FORCE);
1136                         return EINVAL;
1137                 }
1138                 if (schain->error) {
1139                         kprintf("hammer2_mount: error %s reading super-root\n",
1140                                 hammer2_error_str(schain->error));
1141                         hammer2_chain_unlock(schain);
1142                         hammer2_chain_drop(schain);
1143                         schain = NULL;
1144                         hammer2_unmount_helper(mp, NULL, hmp);
1145                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1146                         hammer2_vfs_unmount(mp, MNT_FORCE);
1147                         return EINVAL;
1148                 }
1149
1150                 /*
1151                  * The super-root always uses an inode_tid of 1 when
1152                  * creating PFSs.
1153                  */
1154                 spmp->inode_tid = 1;
1155                 spmp->modify_tid = schain->bref.modify_tid + 1;
1156
1157                 /*
1158                  * Sanity-check schain's pmp and finish initialization.
1159                  * Any chain belonging to the super-root topology should
1160                  * have a NULL pmp (not even set to spmp).
1161                  */
1162                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1163                 KKASSERT(schain->pmp == NULL);
1164                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1165
1166                 /*
1167                  * Replace the dummy spmp->iroot with a real one.  It's
1168                  * easier to just do a wholesale replacement than to try
1169                  * to update the chain and fixup the iroot fields.
1170                  *
1171                  * The returned inode is locked with the supplied cluster.
1172                  */
1173                 cluster = hammer2_cluster_from_chain(schain);
1174                 hammer2_inode_drop(spmp->iroot);
1175                 spmp->iroot = NULL;
1176                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1177                 spmp->spmp_hmp = hmp;
1178                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1179                 spmp->pfs_hmps[0] = hmp;
1180                 hammer2_inode_ref(spmp->iroot);
1181                 hammer2_inode_unlock(spmp->iroot);
1182                 hammer2_cluster_unlock(cluster);
1183                 hammer2_cluster_drop(cluster);
1184                 schain = NULL;
1185                 /* leave spmp->iroot with one ref */
1186
1187                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1188                         error = hammer2_recovery(hmp);
1189                         /* XXX do something with error */
1190                 }
1191                 hammer2_update_pmps(hmp);
1192                 hammer2_iocom_init(hmp);
1193
1194                 /*
1195                  * Ref the cluster management messaging descriptor.  The mount
1196                  * program deals with the other end of the communications pipe.
1197                  *
1198                  * Root mounts typically do not supply one.
1199                  */
1200                 if (info.cluster_fd >= 0) {
1201                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1202                         if (fp) {
1203                                 hammer2_cluster_reconnect(hmp, fp);
1204                         } else {
1205                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1206                         }
1207                 }
1208         } else {
1209                 spmp = hmp->spmp;
1210                 if (info.hflags & HMNT2_DEVFLAGS) {
1211                         kprintf("hammer2: Warning: mount flags pertaining "
1212                                 "to the whole device may only be specified "
1213                                 "on the first mount of the device: %08x\n",
1214                                 info.hflags & HMNT2_DEVFLAGS);
1215                 }
1216         }
1217
1218         /*
1219          * Force local mount (disassociate all PFSs from their clusters).
1220          * Used primarily for debugging.
1221          */
1222         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1223
1224         /*
1225          * Lookup the mount point under the media-localized super-root.
1226          * Scanning hammer2_pfslist doesn't help us because it represents
1227          * PFS cluster ids which can aggregate several named PFSs together.
1228          *
1229          * cluster->pmp will incorrectly point to spmp and must be fixed
1230          * up later on.
1231          */
1232         hammer2_inode_lock(spmp->iroot, 0);
1233         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1234         lhc = hammer2_dirhash(label, strlen(label));
1235         chain = hammer2_chain_lookup(&parent, &key_next,
1236                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1237                                      &cache_index, 0);
1238         while (chain) {
1239                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1240                     strcmp(label, chain->data->ipdata.filename) == 0) {
1241                         break;
1242                 }
1243                 chain = hammer2_chain_next(&parent, chain, &key_next,
1244                                             key_next,
1245                                             lhc + HAMMER2_DIRHASH_LOMASK,
1246                                             &cache_index, 0);
1247         }
1248         if (parent) {
1249                 hammer2_chain_unlock(parent);
1250                 hammer2_chain_drop(parent);
1251         }
1252         hammer2_inode_unlock(spmp->iroot);
1253
1254         /*
1255          * PFS could not be found?
1256          */
1257         if (chain == NULL) {
1258                 kprintf("hammer2_mount: PFS label not found\n");
1259                 hammer2_unmount_helper(mp, NULL, hmp);
1260                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1261                 hammer2_vfs_unmount(mp, MNT_FORCE);
1262
1263                 return EINVAL;
1264         }
1265
1266         /*
1267          * Acquire the pmp structure (it should have already been allocated
1268          * via hammer2_update_pmps() so do not pass cluster in to add to
1269          * available chains).
1270          *
1271          * Check if the cluster has already been mounted.  A cluster can
1272          * only be mounted once, use null mounts to mount additional copies.
1273          */
1274         ripdata = &chain->data->ipdata;
1275         bref = chain->bref;
1276         pmp = hammer2_pfsalloc(NULL, ripdata,
1277                                bref.modify_tid, force_local);
1278         hammer2_chain_unlock(chain);
1279         hammer2_chain_drop(chain);
1280
1281         /*
1282          * Finish the mount
1283          */
1284         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1285
1286         if (pmp->mp) {
1287                 kprintf("hammer2_mount: PFS already mounted!\n");
1288                 hammer2_unmount_helper(mp, NULL, hmp);
1289                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1290                 hammer2_vfs_unmount(mp, MNT_FORCE);
1291
1292                 return EBUSY;
1293         }
1294
1295         pmp->hflags = info.hflags;
1296         mp->mnt_flag = MNT_LOCAL;
1297         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1298         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1299
1300         /*
1301          * required mount structure initializations
1302          */
1303         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1304         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1305
1306         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1307         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1308
1309         /*
1310          * Optional fields
1311          */
1312         mp->mnt_iosize_max = MAXPHYS;
1313
1314         /*
1315          * Connect up mount pointers.
1316          */
1317         hammer2_mount_helper(mp, pmp);
1318
1319         lockmgr(&hammer2_mntlk, LK_RELEASE);
1320
1321         /*
1322          * Finish setup
1323          */
1324         vfs_getnewfsid(mp);
1325         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1326         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1327         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1328
1329         if (path) {
1330                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1331                           MNAMELEN - 1, &size);
1332                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1333         } /* else root mount, already in there */
1334
1335         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1336         if (path) {
1337                 copyinstr(path, mp->mnt_stat.f_mntonname,
1338                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1339                           &size);
1340         } else {
1341                 /* root mount */
1342                 mp->mnt_stat.f_mntonname[0] = '/';
1343         }
1344
1345         /*
1346          * Initial statfs to prime mnt_stat.
1347          */
1348         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1349
1350         return 0;
1351 }
1352
1353 /*
1354  * Scan PFSs under the super-root and create hammer2_pfs structures.
1355  */
1356 static
1357 void
1358 hammer2_update_pmps(hammer2_dev_t *hmp)
1359 {
1360         const hammer2_inode_data_t *ripdata;
1361         hammer2_chain_t *parent;
1362         hammer2_chain_t *chain;
1363         hammer2_blockref_t bref;
1364         hammer2_dev_t *force_local;
1365         hammer2_pfs_t *spmp;
1366         hammer2_pfs_t *pmp;
1367         hammer2_key_t key_next;
1368         int cache_index = -1;
1369
1370         /*
1371          * Force local mount (disassociate all PFSs from their clusters).
1372          * Used primarily for debugging.
1373          */
1374         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1375
1376         /*
1377          * Lookup mount point under the media-localized super-root.
1378          *
1379          * cluster->pmp will incorrectly point to spmp and must be fixed
1380          * up later on.
1381          */
1382         spmp = hmp->spmp;
1383         hammer2_inode_lock(spmp->iroot, 0);
1384         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1385         chain = hammer2_chain_lookup(&parent, &key_next,
1386                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1387                                          &cache_index, 0);
1388         while (chain) {
1389                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1390                         continue;
1391                 ripdata = &chain->data->ipdata;
1392                 bref = chain->bref;
1393                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1394
1395                 pmp = hammer2_pfsalloc(chain, ripdata,
1396                                        bref.modify_tid, force_local);
1397                 chain = hammer2_chain_next(&parent, chain, &key_next,
1398                                            key_next, HAMMER2_KEY_MAX,
1399                                            &cache_index, 0);
1400         }
1401         if (parent) {
1402                 hammer2_chain_unlock(parent);
1403                 hammer2_chain_drop(parent);
1404         }
1405         hammer2_inode_unlock(spmp->iroot);
1406 }
1407
1408 static
1409 int
1410 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1411                 struct vnode *devvp, struct ucred *cred)
1412 {
1413         int error;
1414
1415         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1416                 error = hammer2_recovery(hmp);
1417         } else {
1418                 error = 0;
1419         }
1420         return error;
1421 }
1422
1423 static
1424 int
1425 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1426 {
1427         hammer2_pfs_t *pmp;
1428         int flags;
1429         int error = 0;
1430
1431         pmp = MPTOPMP(mp);
1432
1433         if (pmp == NULL)
1434                 return(0);
1435
1436         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1437
1438         /*
1439          * If mount initialization proceeded far enough we must flush
1440          * its vnodes and sync the underlying mount points.  Three syncs
1441          * are required to fully flush the filesystem (freemap updates lag
1442          * by one flush, and one extra for safety).
1443          */
1444         if (mntflags & MNT_FORCE)
1445                 flags = FORCECLOSE;
1446         else
1447                 flags = 0;
1448         if (pmp->iroot) {
1449                 error = vflush(mp, 0, flags);
1450                 if (error)
1451                         goto failed;
1452                 hammer2_vfs_sync(mp, MNT_WAIT);
1453                 hammer2_vfs_sync(mp, MNT_WAIT);
1454                 hammer2_vfs_sync(mp, MNT_WAIT);
1455         }
1456
1457         /*
1458          * Cleanup the frontend support XOPS threads
1459          */
1460         hammer2_xop_helper_cleanup(pmp);
1461
1462         if (pmp->mp)
1463                 hammer2_unmount_helper(mp, pmp, NULL);
1464
1465         error = 0;
1466 failed:
1467         lockmgr(&hammer2_mntlk, LK_RELEASE);
1468
1469         return (error);
1470 }
1471
1472 /*
1473  * Mount helper, hook the system mount into our PFS.
1474  * The mount lock is held.
1475  *
1476  * We must bump the mount_count on related devices for any
1477  * mounted PFSs.
1478  */
1479 static
1480 void
1481 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1482 {
1483         hammer2_cluster_t *cluster;
1484         hammer2_chain_t *rchain;
1485         int i;
1486
1487         mp->mnt_data = (qaddr_t)pmp;
1488         pmp->mp = mp;
1489
1490         /*
1491          * After pmp->mp is set we have to adjust hmp->mount_count.
1492          */
1493         cluster = &pmp->iroot->cluster;
1494         for (i = 0; i < cluster->nchains; ++i) {
1495                 rchain = cluster->array[i].chain;
1496                 if (rchain == NULL)
1497                         continue;
1498                 ++rchain->hmp->mount_count;
1499                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1500                         rchain->hmp, rchain->hmp->mount_count);
1501         }
1502
1503         /*
1504          * Create missing Xop threads
1505          */
1506         hammer2_xop_helper_create(pmp);
1507 }
1508
1509 /*
1510  * Mount helper, unhook the system mount from our PFS.
1511  * The mount lock is held.
1512  *
1513  * If hmp is supplied a mount responsible for being the first to open
1514  * the block device failed and the block device and all PFSs using the
1515  * block device must be cleaned up.
1516  *
1517  * If pmp is supplied multiple devices might be backing the PFS and each
1518  * must be disconnected.  This might not be the last PFS using some of the
1519  * underlying devices.  Also, we have to adjust our hmp->mount_count
1520  * accounting for the devices backing the pmp which is now undergoing an
1521  * unmount.
1522  */
1523 static
1524 void
1525 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1526 {
1527         hammer2_cluster_t *cluster;
1528         hammer2_chain_t *rchain;
1529         struct vnode *devvp;
1530         int dumpcnt;
1531         int ronly = 0;
1532         int i;
1533
1534         /*
1535          * If no device supplied this is a high-level unmount and we have to
1536          * to disconnect the mount, adjust mount_count, and locate devices
1537          * that might now have no mounts.
1538          */
1539         if (pmp) {
1540                 KKASSERT(hmp == NULL);
1541                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1542                 pmp->mp = NULL;
1543                 mp->mnt_data = NULL;
1544
1545                 /*
1546                  * After pmp->mp is cleared we have to account for
1547                  * mount_count.
1548                  */
1549                 cluster = &pmp->iroot->cluster;
1550                 for (i = 0; i < cluster->nchains; ++i) {
1551                         rchain = cluster->array[i].chain;
1552                         if (rchain == NULL)
1553                                 continue;
1554                         --rchain->hmp->mount_count;
1555                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1556                                 rchain->hmp, rchain->hmp->mount_count);
1557                         /* scrapping hmp now may invalidate the pmp */
1558                 }
1559 again:
1560                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1561                         if (hmp->mount_count == 0) {
1562                                 hammer2_unmount_helper(NULL, NULL, hmp);
1563                                 goto again;
1564                         }
1565                 }
1566                 return;
1567         }
1568
1569         /*
1570          * Try to terminate the block device.  We can't terminate it if
1571          * there are still PFSs referencing it.
1572          */
1573         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1574                 hmp, hmp->mount_count);
1575         if (hmp->mount_count)
1576                 return;
1577
1578         hammer2_pfsfree_scan(hmp);
1579         hammer2_dev_exlock(hmp);        /* XXX order */
1580
1581         /*
1582          * Cycle the volume data lock as a safety (probably not needed any
1583          * more).  To ensure everything is out we need to flush at least
1584          * three times.  (1) The running of the sideq can dirty the
1585          * filesystem, (2) A normal flush can dirty the freemap, and
1586          * (3) ensure that the freemap is fully synchronized.
1587          *
1588          * The next mount's recovery scan can clean everything up but we want
1589          * to leave the filesystem in a 100% clean state on a normal unmount.
1590          */
1591 #if 0
1592         hammer2_voldata_lock(hmp);
1593         hammer2_voldata_unlock(hmp);
1594 #endif
1595         hammer2_iocom_uninit(hmp);
1596
1597         if ((hmp->vchain.flags | hmp->fchain.flags) &
1598             HAMMER2_CHAIN_FLUSH_MASK) {
1599                 kprintf("hammer2_unmount: chains left over "
1600                         "after final sync\n");
1601                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1602                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1603
1604                 if (hammer2_debug & 0x0010)
1605                         Debugger("entered debugger");
1606         }
1607
1608         KKASSERT(hmp->spmp == NULL);
1609
1610         /*
1611          * Finish up with the device vnode
1612          */
1613         if ((devvp = hmp->devvp) != NULL) {
1614                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1615                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1616                 hmp->devvp = NULL;
1617                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1618                 vn_unlock(devvp);
1619                 vrele(devvp);
1620                 devvp = NULL;
1621         }
1622
1623         /*
1624          * Clear vchain/fchain flags that might prevent final cleanup
1625          * of these chains.
1626          */
1627         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1628                 atomic_add_long(&hammer2_count_modified_chains, -1);
1629                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1630                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1631         }
1632         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1633                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1634         }
1635
1636         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1637                 atomic_add_long(&hammer2_count_modified_chains, -1);
1638                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1639                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1640         }
1641         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1642                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1643         }
1644
1645         /*
1646          * Final drop of embedded freemap root chain to
1647          * clean up fchain.core (fchain structure is not
1648          * flagged ALLOCATED so it is cleaned out and then
1649          * left to rot).
1650          */
1651         hammer2_chain_drop(&hmp->fchain);
1652
1653         /*
1654          * Final drop of embedded volume root chain to clean
1655          * up vchain.core (vchain structure is not flagged
1656          * ALLOCATED so it is cleaned out and then left to
1657          * rot).
1658          */
1659         dumpcnt = 50;
1660         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1661         dumpcnt = 50;
1662         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1663         hammer2_dev_unlock(hmp);
1664         hammer2_chain_drop(&hmp->vchain);
1665
1666         hammer2_io_cleanup(hmp, &hmp->iotree);
1667         if (hmp->iofree_count) {
1668                 kprintf("io_cleanup: %d I/O's left hanging\n",
1669                         hmp->iofree_count);
1670         }
1671
1672         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1673         kmalloc_destroy(&hmp->mchain);
1674         kfree(hmp, M_HAMMER2);
1675 }
1676
1677 int
1678 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1679                  ino_t ino, struct vnode **vpp)
1680 {
1681         hammer2_xop_lookup_t *xop;
1682         hammer2_pfs_t *pmp;
1683         hammer2_inode_t *ip;
1684         hammer2_tid_t inum;
1685         int error;
1686
1687         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1688
1689         error = 0;
1690         pmp = MPTOPMP(mp);
1691
1692         /*
1693          * Easy if we already have it cached
1694          */
1695         ip = hammer2_inode_lookup(pmp, inum);
1696         if (ip) {
1697                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1698                 *vpp = hammer2_igetv(ip, &error);
1699                 hammer2_inode_unlock(ip);
1700                 hammer2_inode_drop(ip);         /* from lookup */
1701
1702                 return error;
1703         }
1704
1705         /*
1706          * Otherwise we have to find the inode
1707          */
1708         xop = hammer2_xop_alloc(pmp->iroot, 0);
1709         xop->lhc = inum;
1710         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1711         error = hammer2_xop_collect(&xop->head, 0);
1712
1713         if (error == 0) {
1714                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1715                         kprintf("vget: no collect error but also no rdata\n");
1716                         kprintf("xop %p\n", xop);
1717                         while ((hammer2_debug & 0x80000) == 0) {
1718                                 tsleep(xop, PCATCH, "wait", hz * 10);
1719                         }
1720                         ip = NULL;
1721                 } else {
1722                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1723                 }
1724         }
1725         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1726
1727         if (ip) {
1728                 *vpp = hammer2_igetv(ip, &error);
1729                 hammer2_inode_unlock(ip);
1730         } else {
1731                 *vpp = NULL;
1732                 error = ENOENT;
1733         }
1734         return (error);
1735 }
1736
1737 static
1738 int
1739 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1740 {
1741         hammer2_pfs_t *pmp;
1742         struct vnode *vp;
1743         int error;
1744
1745         pmp = MPTOPMP(mp);
1746         if (pmp->iroot == NULL) {
1747                 *vpp = NULL;
1748                 return EINVAL;
1749         }
1750
1751         error = 0;
1752         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1753
1754         while (pmp->inode_tid == 0) {
1755                 hammer2_xop_ipcluster_t *xop;
1756                 hammer2_inode_meta_t *meta;
1757
1758                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1759                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1760                 error = hammer2_xop_collect(&xop->head, 0);
1761
1762                 if (error == 0) {
1763                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1764                         pmp->iroot->meta = *meta;
1765                         pmp->inode_tid = meta->pfs_inum + 1;
1766                         if (pmp->inode_tid < HAMMER2_INODE_START)
1767                                 pmp->inode_tid = HAMMER2_INODE_START;
1768                         pmp->modify_tid =
1769                                 xop->head.cluster.focus->bref.modify_tid + 1;
1770                         kprintf("PFS: Starting inode %jd\n",
1771                                 (intmax_t)pmp->inode_tid);
1772                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1773                                 pmp->inode_tid, pmp->modify_tid);
1774                         wakeup(&pmp->iroot);
1775
1776                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1777
1778                         /*
1779                          * Prime the mount info.
1780                          */
1781                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1782                         break;
1783                 }
1784
1785                 /*
1786                  * Loop, try again
1787                  */
1788                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1789                 hammer2_inode_unlock(pmp->iroot);
1790                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1791                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1792                 if (error == EINTR)
1793                         break;
1794         }
1795
1796         if (error) {
1797                 hammer2_inode_unlock(pmp->iroot);
1798                 *vpp = NULL;
1799         } else {
1800                 vp = hammer2_igetv(pmp->iroot, &error);
1801                 hammer2_inode_unlock(pmp->iroot);
1802                 *vpp = vp;
1803         }
1804
1805         return (error);
1806 }
1807
1808 /*
1809  * Filesystem status
1810  *
1811  * XXX incorporate ipdata->meta.inode_quota and data_quota
1812  */
1813 static
1814 int
1815 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1816 {
1817         hammer2_pfs_t *pmp;
1818         hammer2_dev_t *hmp;
1819         hammer2_blockref_t bref;
1820         int i;
1821
1822         /*
1823          * NOTE: iroot might not have validated the cluster yet.
1824          */
1825         pmp = MPTOPMP(mp);
1826
1827         mp->mnt_stat.f_files = 0;
1828         mp->mnt_stat.f_ffree = 0;
1829         mp->mnt_stat.f_blocks = 0;
1830         mp->mnt_stat.f_bfree = 0;
1831         mp->mnt_stat.f_bavail = 0;
1832
1833         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1834                 hmp = pmp->pfs_hmps[i];
1835                 if (hmp == NULL)
1836                         continue;
1837                 if (pmp->iroot->cluster.array[i].chain)
1838                         bref = pmp->iroot->cluster.array[i].chain->bref;
1839                 else
1840                         bzero(&bref, sizeof(bref));
1841
1842                 mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1843                 mp->mnt_stat.f_ffree = 0;
1844                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1845                                         mp->mnt_vstat.f_bsize;
1846                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1847                                         mp->mnt_vstat.f_bsize;
1848                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1849
1850                 *sbp = mp->mnt_stat;
1851         }
1852         return (0);
1853 }
1854
1855 static
1856 int
1857 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1858 {
1859         hammer2_pfs_t *pmp;
1860         hammer2_dev_t *hmp;
1861         hammer2_blockref_t bref;
1862         int i;
1863
1864         /*
1865          * NOTE: iroot might not have validated the cluster yet.
1866          */
1867         pmp = MPTOPMP(mp);
1868
1869         mp->mnt_vstat.f_bsize = 0;
1870         mp->mnt_vstat.f_files = 0;
1871         mp->mnt_vstat.f_ffree = 0;
1872         mp->mnt_vstat.f_blocks = 0;
1873         mp->mnt_vstat.f_bfree = 0;
1874         mp->mnt_vstat.f_bavail = 0;
1875
1876         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1877                 hmp = pmp->pfs_hmps[i];
1878                 if (hmp == NULL)
1879                         continue;
1880                 if (pmp->iroot->cluster.array[i].chain)
1881                         bref = pmp->iroot->cluster.array[i].chain->bref;
1882                 else
1883                         bzero(&bref, sizeof(bref));
1884
1885                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1886                 mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
1887                 mp->mnt_vstat.f_ffree = 0;
1888                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1889                                         mp->mnt_vstat.f_bsize;
1890                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1891                                         mp->mnt_vstat.f_bsize;
1892                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1893
1894                 *sbp = mp->mnt_vstat;
1895         }
1896         return (0);
1897 }
1898
1899 /*
1900  * Mount-time recovery (RW mounts)
1901  *
1902  * Updates to the free block table are allowed to lag flushes by one
1903  * transaction.  In case of a crash, then on a fresh mount we must do an
1904  * incremental scan of the last committed transaction id and make sure that
1905  * all related blocks have been marked allocated.
1906  *
1907  * The super-root topology and each PFS has its own transaction id domain,
1908  * so we must track PFS boundary transitions.
1909  */
1910 struct hammer2_recovery_elm {
1911         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1912         hammer2_chain_t *chain;
1913         hammer2_tid_t sync_tid;
1914 };
1915
1916 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1917
1918 struct hammer2_recovery_info {
1919         struct hammer2_recovery_list list;
1920         hammer2_tid_t   mtid;
1921         int     depth;
1922 };
1923
1924 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1925                         hammer2_chain_t *parent,
1926                         struct hammer2_recovery_info *info,
1927                         hammer2_tid_t sync_tid);
1928
1929 #define HAMMER2_RECOVERY_MAXDEPTH       10
1930
1931 static
1932 int
1933 hammer2_recovery(hammer2_dev_t *hmp)
1934 {
1935         struct hammer2_recovery_info info;
1936         struct hammer2_recovery_elm *elm;
1937         hammer2_chain_t *parent;
1938         hammer2_tid_t sync_tid;
1939         hammer2_tid_t mirror_tid;
1940         int error;
1941         int cumulative_error = 0;
1942
1943         hammer2_trans_init(hmp->spmp, 0);
1944
1945         sync_tid = hmp->voldata.freemap_tid;
1946         mirror_tid = hmp->voldata.mirror_tid;
1947
1948         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1949         if (sync_tid >= mirror_tid) {
1950                 kprintf(" no recovery needed\n");
1951         } else {
1952                 kprintf(" freemap recovery %016jx-%016jx\n",
1953                         sync_tid + 1, mirror_tid);
1954         }
1955
1956         TAILQ_INIT(&info.list);
1957         info.depth = 0;
1958         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1959         cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
1960         hammer2_chain_lookup_done(parent);
1961
1962         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1963                 TAILQ_REMOVE(&info.list, elm, entry);
1964                 parent = elm->chain;
1965                 sync_tid = elm->sync_tid;
1966                 kfree(elm, M_HAMMER2);
1967
1968                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1969                 error = hammer2_recovery_scan(hmp, parent, &info,
1970                                               hmp->voldata.freemap_tid);
1971                 hammer2_chain_unlock(parent);
1972                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
1973                 if (error)
1974                         cumulative_error = error;
1975         }
1976         hammer2_trans_done(hmp->spmp);
1977
1978         return cumulative_error;
1979 }
1980
1981 static
1982 int
1983 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1984                       struct hammer2_recovery_info *info,
1985                       hammer2_tid_t sync_tid)
1986 {
1987         const hammer2_inode_data_t *ripdata;
1988         hammer2_chain_t *chain;
1989         hammer2_blockref_t bref;
1990         int cache_index;
1991         int cumulative_error = 0;
1992         int error;
1993         int first;
1994
1995         /*
1996          * Adjust freemap to ensure that the block(s) are marked allocated.
1997          */
1998         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1999                 hammer2_freemap_adjust(hmp, &parent->bref,
2000                                        HAMMER2_FREEMAP_DORECOVER);
2001         }
2002
2003         /*
2004          * Check type for recursive scan
2005          */
2006         switch(parent->bref.type) {
2007         case HAMMER2_BREF_TYPE_VOLUME:
2008                 /* data already instantiated */
2009                 break;
2010         case HAMMER2_BREF_TYPE_INODE:
2011                 /*
2012                  * Must instantiate data for DIRECTDATA test and also
2013                  * for recursion.
2014                  */
2015                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2016                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2017                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2018                         /* not applicable to recovery scan */
2019                         hammer2_chain_unlock(parent);
2020                         return 0;
2021                 }
2022                 hammer2_chain_unlock(parent);
2023                 break;
2024         case HAMMER2_BREF_TYPE_INDIRECT:
2025                 /*
2026                  * Must instantiate data for recursion
2027                  */
2028                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2029                 hammer2_chain_unlock(parent);
2030                 break;
2031         case HAMMER2_BREF_TYPE_DATA:
2032         case HAMMER2_BREF_TYPE_FREEMAP:
2033         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2034         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2035                 /* not applicable to recovery scan */
2036                 return 0;
2037                 break;
2038         default:
2039                 return EDOM;
2040         }
2041
2042         /*
2043          * Defer operation if depth limit reached or if we are crossing a
2044          * PFS boundary.
2045          */
2046         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2047                 struct hammer2_recovery_elm *elm;
2048
2049                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2050                 elm->chain = parent;
2051                 elm->sync_tid = sync_tid;
2052                 hammer2_chain_ref(parent);
2053                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2054                 /* unlocked by caller */
2055
2056                 return(0);
2057         }
2058
2059
2060         /*
2061          * Recursive scan of the last flushed transaction only.  We are
2062          * doing this without pmp assignments so don't leave the chains
2063          * hanging around after we are done with them.
2064          */
2065         cache_index = 0;
2066         chain = NULL;
2067         first = 1;
2068
2069         while (hammer2_chain_scan(parent, &chain, &bref,
2070                                   &first, &cache_index,
2071                                   HAMMER2_LOOKUP_NODATA) != NULL) {
2072                 /*
2073                  * If this is a leaf
2074                  */
2075                 if (chain == NULL) {
2076                         if (bref.mirror_tid > sync_tid) {
2077                                 hammer2_freemap_adjust(hmp, &bref,
2078                                                      HAMMER2_FREEMAP_DORECOVER);
2079                         }
2080                         continue;
2081                 }
2082
2083                 /*
2084                  * This may or may not be a recursive node.
2085                  */
2086                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2087                 if (bref.mirror_tid > sync_tid) {
2088                         ++info->depth;
2089                         error = hammer2_recovery_scan(hmp, chain,
2090                                                       info, sync_tid);
2091                         --info->depth;
2092                         if (error)
2093                                 cumulative_error = error;
2094                 }
2095
2096                 /*
2097                  * Flush the recovery at the PFS boundary to stage it for
2098                  * the final flush of the super-root topology.
2099                  */
2100                 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2101                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2102                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2103                 }
2104         }
2105
2106         return cumulative_error;
2107 }
2108
2109 /*
2110  * Sync a mount point; this is called on a per-mount basis from the
2111  * filesystem syncer process periodically and whenever a user issues
2112  * a sync.
2113  */
2114 int
2115 hammer2_vfs_sync(struct mount *mp, int waitfor)
2116 {
2117         hammer2_xop_flush_t *xop;
2118         struct hammer2_sync_info info;
2119         hammer2_inode_t *iroot;
2120         hammer2_pfs_t *pmp;
2121         int flags;
2122         int error;
2123
2124         pmp = MPTOPMP(mp);
2125         iroot = pmp->iroot;
2126         KKASSERT(iroot);
2127         KKASSERT(iroot->pmp == pmp);
2128
2129         /*
2130          * We can't acquire locks on existing vnodes while in a transaction
2131          * without risking a deadlock.  This assumes that vfsync() can be
2132          * called without the vnode locked (which it can in DragonFly).
2133          * Otherwise we'd have to implement a multi-pass or flag the lock
2134          * failures and retry.
2135          *
2136          * The reclamation code interlocks with the sync list's token
2137          * (by removing the vnode from the scan list) before unlocking
2138          * the inode, giving us time to ref the inode.
2139          */
2140         /*flags = VMSC_GETVP;*/
2141         flags = 0;
2142         if (waitfor & MNT_LAZY)
2143                 flags |= VMSC_ONEPASS;
2144
2145         /*
2146          * Preflush the vnodes using a normal transaction before interlocking
2147          * with a flush transaction.  We do this to try to run as much of
2148          * the compression as possible outside the flush transaction.
2149          *
2150          * For efficiency do an async pass before making sure with a
2151          * synchronous pass on all related buffer cache buffers.
2152          */
2153         hammer2_trans_init(pmp, 0);
2154         info.error = 0;
2155         info.waitfor = MNT_NOWAIT;
2156         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2157         info.waitfor = MNT_WAIT;
2158         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2159         hammer2_trans_done(pmp);
2160
2161         /*
2162          * Start our flush transaction.  This does not return until all
2163          * concurrent transactions have completed and will prevent any
2164          * new transactions from running concurrently, except for the
2165          * buffer cache transactions.
2166          *
2167          * NOTE!  It is still possible for the paging code to push pages
2168          *        out via a UIO_NOCOPY hammer2_vop_write() during the main
2169          *        flush.
2170          */
2171         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2172         hammer2_inode_run_sideq(pmp);
2173
2174         info.error = 0;
2175         info.waitfor = MNT_NOWAIT;
2176         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2177         info.waitfor = MNT_WAIT;
2178         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2179         hammer2_bioq_sync(pmp);
2180
2181         /*
2182          * Use the XOP interface to concurrently flush all nodes to
2183          * synchronize the PFSROOT subtopology to the media.  A standard
2184          * end-of-scan ENOENT error indicates cluster sufficiency.
2185          *
2186          * Note that this flush will not be visible on crash recovery until
2187          * we flush the super-root topology in the next loop.
2188          *
2189          * XXX For now wait for all flushes to complete.
2190          */
2191         if (iroot) {
2192                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2193                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2194                 error = hammer2_xop_collect(&xop->head,
2195                                             HAMMER2_XOP_COLLECT_WAITALL);
2196                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2197                 if (error == ENOENT)
2198                         error = 0;
2199         } else {
2200                 error = 0;
2201         }
2202         hammer2_trans_done(pmp);
2203
2204         return (error);
2205 }
2206
2207 /*
2208  * Sync passes.
2209  *
2210  * Note that we ignore the tranasction mtid we got above.  Instead,
2211  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2212  * transactions.
2213  */
2214 static int
2215 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2216 {
2217         struct hammer2_sync_info *info = data;
2218         hammer2_inode_t *ip;
2219         int error;
2220
2221         /*
2222          * Degenerate cases.  Note that ip == NULL typically means the
2223          * syncer vnode itself and we don't want to vclrisdirty() in that
2224          * situation.
2225          */
2226         ip = VTOI(vp);
2227         if (ip == NULL) {
2228                 return(0);
2229         }
2230         if (vp->v_type == VNON || vp->v_type == VBAD) {
2231                 vclrisdirty(vp);
2232                 return(0);
2233         }
2234
2235         /*
2236          * VOP_FSYNC will start a new transaction so replicate some code
2237          * here to do it inline (see hammer2_vop_fsync()).
2238          *
2239          * WARNING: The vfsync interacts with the buffer cache and might
2240          *          block, we can't hold the inode lock at that time.
2241          *          However, we MUST ref ip before blocking to ensure that
2242          *          it isn't ripped out from under us (since we do not
2243          *          hold a lock on the vnode).
2244          */
2245         hammer2_inode_ref(ip);
2246         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2247             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2248                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2249                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2250                                  HAMMER2_INODE_MODIFIED)) {
2251                         hammer2_inode_lock(ip, 0);
2252                         hammer2_inode_chain_sync(ip);
2253                         hammer2_inode_unlock(ip);
2254                 }
2255         }
2256         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2257             RB_EMPTY(&vp->v_rbdirty_tree)) {
2258                 vclrisdirty(vp);
2259         }
2260
2261         hammer2_inode_drop(ip);
2262 #if 1
2263         error = 0;
2264         if (error)
2265                 info->error = error;
2266 #endif
2267         return(0);
2268 }
2269
2270 static
2271 int
2272 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2273 {
2274         hammer2_inode_t *ip;
2275
2276         KKASSERT(MAXFIDSZ >= 16);
2277         ip = VTOI(vp);
2278         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2279         fhp->fid_ext = 0;
2280         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2281         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2282
2283         return 0;
2284 }
2285
2286 static
2287 int
2288 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2289                struct fid *fhp, struct vnode **vpp)
2290 {
2291         hammer2_pfs_t *pmp;
2292         hammer2_tid_t inum;
2293         int error;
2294
2295         pmp = MPTOPMP(mp);
2296         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2297         if (vpp) {
2298                 if (inum == 1)
2299                         error = hammer2_vfs_root(mp, vpp);
2300                 else
2301                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2302         } else {
2303                 error = 0;
2304         }
2305         if (error)
2306                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2307         return error;
2308 }
2309
2310 static
2311 int
2312 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2313                  int *exflagsp, struct ucred **credanonp)
2314 {
2315         hammer2_pfs_t *pmp;
2316         struct netcred *np;
2317         int error;
2318
2319         pmp = MPTOPMP(mp);
2320         np = vfs_export_lookup(mp, &pmp->export, nam);
2321         if (np) {
2322                 *exflagsp = np->netc_exflags;
2323                 *credanonp = &np->netc_anon;
2324                 error = 0;
2325         } else {
2326                 error = EACCES;
2327         }
2328         return error;
2329 }
2330
2331 /*
2332  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2333  * header into the HMP
2334  *
2335  * XXX read four volhdrs and use the one with the highest TID whos CRC
2336  *     matches.
2337  *
2338  * XXX check iCRCs.
2339  *
2340  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2341  *     nonexistant locations.
2342  *
2343  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2344  */
2345 static
2346 int
2347 hammer2_install_volume_header(hammer2_dev_t *hmp)
2348 {
2349         hammer2_volume_data_t *vd;
2350         struct buf *bp;
2351         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2352         int error_reported;
2353         int error;
2354         int valid;
2355         int i;
2356
2357         error_reported = 0;
2358         error = 0;
2359         valid = 0;
2360         bp = NULL;
2361
2362         /*
2363          * There are up to 4 copies of the volume header (syncs iterate
2364          * between them so there is no single master).  We don't trust the
2365          * volu_size field so we don't know precisely how large the filesystem
2366          * is, so depend on the OS to return an error if we go beyond the
2367          * block device's EOF.
2368          */
2369         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2370                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2371                               HAMMER2_VOLUME_BYTES, &bp);
2372                 if (error) {
2373                         brelse(bp);
2374                         bp = NULL;
2375                         continue;
2376                 }
2377
2378                 vd = (struct hammer2_volume_data *) bp->b_data;
2379                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2380                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2381                         brelse(bp);
2382                         bp = NULL;
2383                         continue;
2384                 }
2385
2386                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2387                         /* XXX: Reversed-endianness filesystem */
2388                         kprintf("hammer2: reverse-endian filesystem detected");
2389                         brelse(bp);
2390                         bp = NULL;
2391                         continue;
2392                 }
2393
2394                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2395                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2396                                       HAMMER2_VOLUME_ICRC0_SIZE);
2397                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2398                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2399                                        HAMMER2_VOLUME_ICRC1_SIZE);
2400                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2401                         kprintf("hammer2 volume header crc "
2402                                 "mismatch copy #%d %08x/%08x\n",
2403                                 i, crc0, crc);
2404                         error_reported = 1;
2405                         brelse(bp);
2406                         bp = NULL;
2407                         continue;
2408                 }
2409                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2410                         valid = 1;
2411                         hmp->voldata = *vd;
2412                         hmp->volhdrno = i;
2413                 }
2414                 brelse(bp);
2415                 bp = NULL;
2416         }
2417         if (valid) {
2418                 hmp->volsync = hmp->voldata;
2419                 error = 0;
2420                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2421                         kprintf("hammer2: using volume header #%d\n",
2422                                 hmp->volhdrno);
2423                 }
2424         } else {
2425                 error = EINVAL;
2426                 kprintf("hammer2: no valid volume headers found!\n");
2427         }
2428         return (error);
2429 }
2430
2431 /*
2432  * This handles hysteresis on regular file flushes.  Because the BIOs are
2433  * routed to a thread it is possible for an excessive number to build up
2434  * and cause long front-end stalls long before the runningbuffspace limit
2435  * is hit, so we implement hammer2_flush_pipe to control the
2436  * hysteresis.
2437  *
2438  * This is a particular problem when compression is used.
2439  */
2440 void
2441 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2442 {
2443         atomic_add_int(&pmp->count_lwinprog, 1);
2444 }
2445
2446 void
2447 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2448 {
2449         int lwinprog;
2450
2451         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2452         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2453             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2454                 atomic_clear_int(&pmp->count_lwinprog,
2455                                  HAMMER2_LWINPROG_WAITING);
2456                 wakeup(&pmp->count_lwinprog);
2457         }
2458         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2459             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2460                 atomic_clear_int(&pmp->count_lwinprog,
2461                                  HAMMER2_LWINPROG_WAITING0);
2462                 wakeup(&pmp->count_lwinprog);
2463         }
2464 }
2465
2466 void
2467 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2468 {
2469         int lwinprog;
2470         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2471                                     HAMMER2_LWINPROG_WAITING0;
2472
2473         for (;;) {
2474                 lwinprog = pmp->count_lwinprog;
2475                 cpu_ccfence();
2476                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2477                         break;
2478                 tsleep_interlock(&pmp->count_lwinprog, 0);
2479                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2480                 lwinprog = pmp->count_lwinprog;
2481                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2482                         break;
2483                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2484         }
2485 }
2486
2487 /*
2488  * Manage excessive memory resource use for chain and related
2489  * structures.
2490  */
2491 void
2492 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2493 {
2494         uint32_t waiting;
2495         uint32_t count;
2496         uint32_t limit;
2497 #if 0
2498         static int zzticks;
2499 #endif
2500
2501         /*
2502          * Atomic check condition and wait.  Also do an early speedup of
2503          * the syncer to try to avoid hitting the wait.
2504          */
2505         for (;;) {
2506                 waiting = pmp->inmem_dirty_chains;
2507                 cpu_ccfence();
2508                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2509
2510                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2511                 if (limit < hammer2_limit_dirty_chains)
2512                         limit = hammer2_limit_dirty_chains;
2513                 if (limit < 1000)
2514                         limit = 1000;
2515
2516 #if 0
2517                 if ((int)(ticks - zzticks) > hz) {
2518                         zzticks = ticks;
2519                         kprintf("count %ld %ld\n", count, limit);
2520                 }
2521 #endif
2522
2523                 /*
2524                  * Block if there are too many dirty chains present, wait
2525                  * for the flush to clean some out.
2526                  */
2527                 if (count > limit) {
2528                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2529                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2530                                                waiting,
2531                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2532                                 speedup_syncer(pmp->mp);
2533                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2534                                        "chnmem", hz);
2535                         }
2536                         continue;       /* loop on success or fail */
2537                 }
2538
2539                 /*
2540                  * Try to start an early flush before we are forced to block.
2541                  */
2542                 if (count > limit * 7 / 10)
2543                         speedup_syncer(pmp->mp);
2544                 break;
2545         }
2546 }
2547
2548 void
2549 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2550 {
2551         if (pmp) {
2552                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2553         }
2554 }
2555
2556 void
2557 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2558 {
2559         uint32_t waiting;
2560
2561         if (pmp == NULL)
2562                 return;
2563
2564         for (;;) {
2565                 waiting = pmp->inmem_dirty_chains;
2566                 cpu_ccfence();
2567                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2568                                        waiting,
2569                                        (waiting - 1) &
2570                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2571                         break;
2572                 }
2573         }
2574
2575         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2576                 wakeup(&pmp->inmem_dirty_chains);
2577 }
2578
2579 /*
2580  * Debugging
2581  */
2582 void
2583 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2584 {
2585         hammer2_chain_t *scan;
2586         hammer2_chain_t *parent;
2587
2588         --*countp;
2589         if (*countp == 0) {
2590                 kprintf("%*.*s...\n", tab, tab, "");
2591                 return;
2592         }
2593         if (*countp < 0)
2594                 return;
2595         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2596                 tab, tab, "", pfx,
2597                 chain, chain->bref.type,
2598                 chain->bref.key, chain->bref.keybits,
2599                 chain->bref.mirror_tid);
2600
2601         kprintf("%*.*s      [%08x] (%s) refs=%d",
2602                 tab, tab, "",
2603                 chain->flags,
2604                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2605                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2606                 chain->refs);
2607
2608         parent = chain->parent;
2609         if (parent)
2610                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2611                         tab, tab, "",
2612                         parent, parent->flags, parent->refs);
2613         if (RB_EMPTY(&chain->core.rbtree)) {
2614                 kprintf("\n");
2615         } else {
2616                 kprintf(" {\n");
2617                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2618                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2619                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2620                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2621                                 chain->data->ipdata.filename);
2622                 else
2623                         kprintf("%*.*s}\n", tab, tab, "");
2624         }
2625 }