sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72 };
  73
  74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  75 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
  76 static struct hammer2_mntlist hammer2_mntlist;
  77 static struct hammer2_pfslist hammer2_pfslist;
  78 static struct lock hammer2_mntlk;
  79
  80 int hammer2_debug;
  81 int hammer2_cluster_read = 4;           /* physical read-ahead */
  82 int hammer2_cluster_write = 0;          /* bdwrite() so later inval works */
  83 int hammer2_dedup_enable = 1;
  84 int hammer2_inval_enable = 0;
  85 int hammer2_flush_pipe = 100;
  86 int hammer2_synchronous_flush = 1;
  87 int hammer2_dio_count;
  88 long hammer2_chain_allocs;
  89 long hammer2_chain_frees;
  90 long hammer2_limit_dirty_chains;
  91 long hammer2_count_modified_chains;
  92 long hammer2_iod_invals;
  93 long hammer2_iod_file_read;
  94 long hammer2_iod_meta_read;
  95 long hammer2_iod_indr_read;
  96 long hammer2_iod_fmap_read;
  97 long hammer2_iod_volu_read;
  98 long hammer2_iod_file_write;
  99 long hammer2_iod_file_wembed;
 100 long hammer2_iod_file_wzero;
 101 long hammer2_iod_file_wdedup;
 102 long hammer2_iod_meta_write;
 103 long hammer2_iod_indr_write;
 104 long hammer2_iod_fmap_write;
 105 long hammer2_iod_volu_write;
 106
 107 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 108 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 109                 "Buffer used for compression.");
 110
 111 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 112 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 113                 "Buffer used for decompression.");
 114
 115 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 116
 117 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 118            &hammer2_debug, 0, "");
 119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW,
 120            &hammer2_cluster_read, 0, "");
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 122            &hammer2_cluster_write, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 124            &hammer2_dedup_enable, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 126            &hammer2_inval_enable, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 128            &hammer2_flush_pipe, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 130            &hammer2_synchronous_flush, 0, "");
 131 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 132            &hammer2_chain_allocs, 0, "");
 133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 134            &hammer2_chain_frees, 0, "");
 135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 136            &hammer2_limit_dirty_chains, 0, "");
 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 138            &hammer2_count_modified_chains, 0, "");
 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 140            &hammer2_dio_count, 0, "");
 141
 142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 143            &hammer2_iod_invals, 0, "");
 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 145            &hammer2_iod_file_read, 0, "");
 146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 147            &hammer2_iod_meta_read, 0, "");
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 149            &hammer2_iod_indr_read, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 151            &hammer2_iod_fmap_read, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 153            &hammer2_iod_volu_read, 0, "");
 154
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 156            &hammer2_iod_file_write, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 158            &hammer2_iod_file_wembed, 0, "");
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 160            &hammer2_iod_file_wzero, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 162            &hammer2_iod_file_wdedup, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 164            &hammer2_iod_meta_write, 0, "");
 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 166            &hammer2_iod_indr_write, 0, "");
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 168            &hammer2_iod_fmap_write, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 170            &hammer2_iod_volu_write, 0, "");
 171
 172 long hammer2_check_icrc32;
 173 long hammer2_check_xxhash64;
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 175            &hammer2_check_icrc32, 0, "");
 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 177            &hammer2_check_xxhash64, 0, "");
 178
 179 static int hammer2_vfs_init(struct vfsconf *conf);
 180 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 181 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 182                                 struct ucred *cred);
 183 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 184                                 struct vnode *, struct ucred *);
 185 static int hammer2_recovery(hammer2_dev_t *hmp);
 186 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 187 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 188 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 189                                 struct ucred *cred);
 190 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 191                                 struct ucred *cred);
 192 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 193                                 struct fid *fhp, struct vnode **vpp);
 194 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 195 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 196                                 int *exflagsp, struct ucred **credanonp);
 197
 198 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 199 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 200
 201 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 202
 203 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 204 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 205                                 hammer2_dev_t *hmp);
 206
 207 /*
 208  * HAMMER2 vfs operations.
 209  */
 210 static struct vfsops hammer2_vfsops = {
 211         .vfs_init       = hammer2_vfs_init,
 212         .vfs_uninit     = hammer2_vfs_uninit,
 213         .vfs_sync       = hammer2_vfs_sync,
 214         .vfs_mount      = hammer2_vfs_mount,
 215         .vfs_unmount    = hammer2_vfs_unmount,
 216         .vfs_root       = hammer2_vfs_root,
 217         .vfs_statfs     = hammer2_vfs_statfs,
 218         .vfs_statvfs    = hammer2_vfs_statvfs,
 219         .vfs_vget       = hammer2_vfs_vget,
 220         .vfs_vptofh     = hammer2_vfs_vptofh,
 221         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 222         .vfs_checkexp   = hammer2_vfs_checkexp
 223 };
 224
 225 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 226
 227 VFS_SET(hammer2_vfsops, hammer2, 0);
 228 MODULE_VERSION(hammer2, 1);
 229
 230 static
 231 int
 232 hammer2_vfs_init(struct vfsconf *conf)
 233 {
 234         static struct objcache_malloc_args margs_read;
 235         static struct objcache_malloc_args margs_write;
 236         static struct objcache_malloc_args margs_vop;
 237
 238         int error;
 239
 240         error = 0;
 241
 242         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 243                 error = EINVAL;
 244         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 245                 error = EINVAL;
 246         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 247                 error = EINVAL;
 248
 249         if (error)
 250                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 251
 252         margs_read.objsize = 65536;
 253         margs_read.mtype = M_HAMMER2_DEBUFFER;
 254
 255         margs_write.objsize = 32768;
 256         margs_write.mtype = M_HAMMER2_CBUFFER;
 257
 258         margs_vop.objsize = sizeof(hammer2_xop_t);
 259         margs_vop.mtype = M_HAMMER2;
 260
 261         /*
 262          * Note thaht for the XOPS cache we want backing store allocations
 263          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 264          * confusion), so use the backing store function that does it.  This
 265          * means that initial XOPS objects are zerod but REUSED objects are
 266          * not.  So we are responsible for cleaning the object up sufficiently
 267          * for our needs before objcache_put()ing it back (typically just the
 268          * FIFO indices).
 269          */
 270         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 271                                 0, 1, NULL, NULL, NULL,
 272                                 objcache_malloc_alloc,
 273                                 objcache_malloc_free,
 274                                 &margs_read);
 275         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 276                                 0, 1, NULL, NULL, NULL,
 277                                 objcache_malloc_alloc,
 278                                 objcache_malloc_free,
 279                                 &margs_write);
 280         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 281                                 0, 1, NULL, NULL, NULL,
 282                                 objcache_malloc_alloc_zero,
 283                                 objcache_malloc_free,
 284                                 &margs_vop);
 285
 286
 287         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 288         TAILQ_INIT(&hammer2_mntlist);
 289         TAILQ_INIT(&hammer2_pfslist);
 290
 291         hammer2_limit_dirty_chains = maxvnodes / 10;
 292         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 293                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 294
 295         return (error);
 296 }
 297
 298 static
 299 int
 300 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 301 {
 302         objcache_destroy(cache_buffer_read);
 303         objcache_destroy(cache_buffer_write);
 304         objcache_destroy(cache_xops);
 305         return 0;
 306 }
 307
 308 /*
 309  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
 310  * mounts and the spmp structure for media (hmp) structures.
 311  *
 312  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 313  * transactions.  Note that synchronization does not use this field.
 314  * (typically frontend operations and synchronization cannot run on the
 315  * same PFS node at the same time).
 316  *
 317  * XXX check locking
 318  */
 319 hammer2_pfs_t *
 320 hammer2_pfsalloc(hammer2_chain_t *chain, const hammer2_inode_data_t *ripdata,
 321                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 322 {
 323         hammer2_inode_t *iroot;
 324         hammer2_pfs_t *pmp;
 325         int count;
 326         int i;
 327         int j;
 328
 329         /*
 330          * Locate or create the PFS based on the cluster id.  If ripdata
 331          * is NULL this is a spmp which is unique and is always allocated.
 332          *
 333          * If the device is mounted in local mode all PFSs are considered
 334          * independent and not part of any cluster (for debugging only).
 335          */
 336         if (ripdata) {
 337                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 338                         if (force_local != pmp->force_local)
 339                                 continue;
 340                         if (force_local == NULL &&
 341                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 342                                  sizeof(pmp->pfs_clid)) == 0) {
 343                                         break;
 344                         } else if (force_local && pmp->pfs_names[0] &&
 345                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 346                                         break;
 347                         }
 348                 }
 349         } else {
 350                 pmp = NULL;
 351         }
 352
 353         if (pmp == NULL) {
 354                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 355                 pmp->force_local = force_local;
 356                 hammer2_trans_manage_init(pmp);
 357                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 358                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 359                 lockinit(&pmp->lock, "pfslk", 0, 0);
 360                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 361                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 362                 spin_init(&pmp->xop_spin, "h2xop");
 363                 spin_init(&pmp->lru_spin, "h2lru");
 364                 RB_INIT(&pmp->inum_tree);
 365                 TAILQ_INIT(&pmp->sideq);
 366                 TAILQ_INIT(&pmp->lru_list);
 367                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 368
 369                 /*
 370                  * Distribute backend operations to threads
 371                  */
 372                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 373                         for (j = 0; j < HAMMER2_XOPGROUPS +
 374                                         HAMMER2_SPECTHREADS; ++j) {
 375                                 TAILQ_INIT(&pmp->xopq[i][j]);
 376                         }
 377                 }
 378                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 379                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 380
 381                 /*
 382                  * Save the last media transaction id for the flusher.  Set
 383                  * initial
 384                  */
 385                 if (ripdata)
 386                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 387                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 388
 389                 /*
 390                  * The synchronization thread may start too early, make
 391                  * sure it stays frozen until we are ready to let it go.
 392                  * XXX
 393                  */
 394                 /*
 395                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 396                                          HAMMER2_THREAD_REMASTER;
 397                 */
 398         }
 399
 400         /*
 401          * Create the PFS's root inode.
 402          */
 403         if ((iroot = pmp->iroot) == NULL) {
 404                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 405                 pmp->iroot = iroot;
 406                 hammer2_inode_ref(iroot);
 407                 hammer2_inode_unlock(iroot);
 408         }
 409
 410         /*
 411          * Stop here if no chain is passed in.
 412          */
 413         if (chain == NULL)
 414                 goto done;
 415
 416         /*
 417          * When a chain is passed in we must add it to the PFS's root
 418          * inode, update pmp->pfs_types[], and update the syncronization
 419          * threads.
 420          *
 421          * When forcing local mode, mark the PFS as a MASTER regardless.
 422          *
 423          * At the moment empty spots can develop due to removals or failures.
 424          * Ultimately we want to re-fill these spots but doing so might
 425          * confused running code. XXX
 426          */
 427         hammer2_inode_ref(iroot);
 428         hammer2_mtx_ex(&iroot->lock);
 429         j = iroot->cluster.nchains;
 430
 431         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 432
 433         if (j == HAMMER2_MAXCLUSTER) {
 434                 kprintf("hammer2_mount: cluster full!\n");
 435                 /* XXX fatal error? */
 436         } else {
 437                 KKASSERT(chain->pmp == NULL);
 438                 chain->pmp = pmp;
 439                 hammer2_chain_ref(chain);
 440                 iroot->cluster.array[j].chain = chain;
 441                 if (force_local)
 442                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 443                 else
 444                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 445                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 446                 pmp->pfs_hmps[j] = chain->hmp;
 447
 448                 /*
 449                  * If the PFS is already mounted we must account
 450                  * for the mount_count here.
 451                  */
 452                 if (pmp->mp)
 453                         ++chain->hmp->mount_count;
 454
 455                 /*
 456                  * May have to fixup dirty chain tracking.  Previous
 457                  * pmp was NULL so nothing to undo.
 458                  */
 459                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 460                         hammer2_pfs_memory_inc(pmp);
 461                 ++j;
 462         }
 463         iroot->cluster.nchains = j;
 464
 465         /*
 466          * Update nmasters from any PFS inode which is part of the cluster.
 467          * It is possible that this will result in a value which is too
 468          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 469          * override this value later on.
 470          *
 471          * (This informs us of masters that might not currently be
 472          *  discoverable by this mount).
 473          */
 474         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 475                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 476         }
 477
 478         /*
 479          * Count visible masters.  Masters are usually added with
 480          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 481          * are more (XXX and must update the master inodes).
 482          */
 483         count = 0;
 484         for (i = 0; i < iroot->cluster.nchains; ++i) {
 485                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 486                         ++count;
 487         }
 488         if (pmp->pfs_nmasters < count)
 489                 pmp->pfs_nmasters = count;
 490
 491         /*
 492          * Create missing synchronization and support threads.
 493          *
 494          * Single-node masters (including snapshots) have nothing to
 495          * synchronize and do not require this thread.
 496          *
 497          * Multi-node masters or any number of soft masters, slaves, copy,
 498          * or other PFS types need the thread.
 499          *
 500          * Each thread is responsible for its particular cluster index.
 501          * We use independent threads so stalls or mismatches related to
 502          * any given target do not affect other targets.
 503          */
 504         for (i = 0; i < iroot->cluster.nchains; ++i) {
 505                 /*
 506                  * Single-node masters (including snapshots) have nothing
 507                  * to synchronize and will make direct xops support calls,
 508                  * thus they do not require this thread.
 509                  *
 510                  * Note that there can be thousands of snapshots.  We do not
 511                  * want to create thousands of threads.
 512                  */
 513                 if (pmp->pfs_nmasters <= 1 &&
 514                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 515                         continue;
 516                 }
 517
 518                 /*
 519                  * Sync support thread
 520                  */
 521                 if (pmp->sync_thrs[i].td == NULL) {
 522                         hammer2_thr_create(&pmp->sync_thrs[i], pmp,
 523                                            "h2nod", i, -1,
 524                                            hammer2_primary_sync_thread);
 525                 }
 526         }
 527
 528         /*
 529          * Create missing Xop threads
 530          */
 531         if (pmp->mp)
 532                 hammer2_xop_helper_create(pmp);
 533
 534         hammer2_mtx_unlock(&iroot->lock);
 535         hammer2_inode_drop(iroot);
 536 done:
 537         return pmp;
 538 }
 539
 540 /*
 541  * Destroy a PFS, typically only occurs after the last mount on a device
 542  * has gone away.
 543  */
 544 static void
 545 hammer2_pfsfree(hammer2_pfs_t *pmp)
 546 {
 547         hammer2_inode_t *iroot;
 548         hammer2_chain_t *chain;
 549         int i;
 550         int j;
 551
 552         /*
 553          * Cleanup our reference on iroot.  iroot is (should) not be needed
 554          * by the flush code.
 555          */
 556         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 557
 558         iroot = pmp->iroot;
 559         if (iroot) {
 560                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 561                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 562                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 563                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 564                 }
 565 #if REPORT_REFS_ERRORS
 566                 if (pmp->iroot->refs != 1)
 567                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 568                                 pmp->iroot, pmp->iroot->refs);
 569 #else
 570                 KKASSERT(pmp->iroot->refs == 1);
 571 #endif
 572                 /* ref for pmp->iroot */
 573                 hammer2_inode_drop(pmp->iroot);
 574                 pmp->iroot = NULL;
 575         }
 576
 577         /*
 578          * Cleanup chains remaining on LRU list.
 579          */
 580         kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
 581         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 582                 hammer2_chain_ref(chain);
 583                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 584                 hammer2_chain_drop(chain);
 585         }
 586
 587         /*
 588          * Free remaining pmp resources
 589          */
 590         kmalloc_destroy(&pmp->mmsg);
 591         kmalloc_destroy(&pmp->minode);
 592
 593         kfree(pmp, M_HAMMER2);
 594 }
 595
 596 /*
 597  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 598  * empty is terminated and freed.
 599  *
 600  * XXX inefficient.
 601  */
 602 static void
 603 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 604 {
 605         hammer2_pfs_t *pmp;
 606         hammer2_inode_t *iroot;
 607         hammer2_chain_t *rchain;
 608         int didfreeze;
 609         int i;
 610         int j;
 611
 612 again:
 613         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 614                 if ((iroot = pmp->iroot) == NULL)
 615                         continue;
 616                 if (hmp->spmp == pmp) {
 617                         kprintf("unmount hmp %p remove spmp %p\n",
 618                                 hmp, pmp);
 619                         hmp->spmp = NULL;
 620                 }
 621
 622                 /*
 623                  * Determine if this PFS is affected.  If it is we must
 624                  * freeze all management threads and lock its iroot.
 625                  *
 626                  * Freezing a management thread forces it idle, operations
 627                  * in-progress will be aborted and it will have to start
 628                  * over again when unfrozen, or exit if told to exit.
 629                  */
 630                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 631                         if (pmp->pfs_hmps[i] == hmp)
 632                                 break;
 633                 }
 634                 if (i != HAMMER2_MAXCLUSTER) {
 635                         /*
 636                          * Make sure all synchronization threads are locked
 637                          * down.
 638                          */
 639                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 640                                 if (pmp->pfs_hmps[i] == NULL)
 641                                         continue;
 642                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 643                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 644                                         hammer2_thr_freeze_async(
 645                                                 &pmp->xop_groups[j].thrs[i]);
 646                                 }
 647                         }
 648                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 649                                 if (pmp->pfs_hmps[i] == NULL)
 650                                         continue;
 651                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 652                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 653                                         hammer2_thr_freeze(
 654                                                 &pmp->xop_groups[j].thrs[i]);
 655                                 }
 656                         }
 657
 658                         /*
 659                          * Lock the inode and clean out matching chains.
 660                          * Note that we cannot use hammer2_inode_lock_*()
 661                          * here because that would attempt to validate the
 662                          * cluster that we are in the middle of ripping
 663                          * apart.
 664                          *
 665                          * WARNING! We are working directly on the inodes
 666                          *          embedded cluster.
 667                          */
 668                         hammer2_mtx_ex(&iroot->lock);
 669
 670                         /*
 671                          * Remove the chain from matching elements of the PFS.
 672                          */
 673                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 674                                 if (pmp->pfs_hmps[i] != hmp)
 675                                         continue;
 676                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 677                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 678                                         hammer2_thr_delete(
 679                                                 &pmp->xop_groups[j].thrs[i]);
 680                                 }
 681                                 rchain = iroot->cluster.array[i].chain;
 682                                 iroot->cluster.array[i].chain = NULL;
 683                                 pmp->pfs_types[i] = 0;
 684                                 if (pmp->pfs_names[i]) {
 685                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 686                                         pmp->pfs_names[i] = NULL;
 687                                 }
 688                                 if (rchain) {
 689                                         hammer2_chain_drop(rchain);
 690                                         /* focus hint */
 691                                         if (iroot->cluster.focus == rchain)
 692                                                 iroot->cluster.focus = NULL;
 693                                 }
 694                                 pmp->pfs_hmps[i] = NULL;
 695                         }
 696                         hammer2_mtx_unlock(&iroot->lock);
 697                         didfreeze = 1;  /* remaster, unfreeze down below */
 698                 } else {
 699                         didfreeze = 0;
 700                 }
 701
 702                 /*
 703                  * Cleanup trailing chains.  Gaps may remain.
 704                  */
 705                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 706                         if (pmp->pfs_hmps[i])
 707                                 break;
 708                 }
 709                 iroot->cluster.nchains = i + 1;
 710
 711                 /*
 712                  * If the PMP has no elements remaining we can destroy it.
 713                  * (this will transition management threads from frozen->exit).
 714                  */
 715                 if (iroot->cluster.nchains == 0) {
 716                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 717                                 hmp, pmp);
 718                         hammer2_pfsfree(pmp);
 719                         goto again;
 720                 }
 721
 722                 /*
 723                  * If elements still remain we need to set the REMASTER
 724                  * flag and unfreeze it.
 725                  */
 726                 if (didfreeze) {
 727                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 728                                 if (pmp->pfs_hmps[i] == NULL)
 729                                         continue;
 730                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 731                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 732                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 733                                         hammer2_thr_remaster(
 734                                                 &pmp->xop_groups[j].thrs[i]);
 735                                         hammer2_thr_unfreeze(
 736                                                 &pmp->xop_groups[j].thrs[i]);
 737                                 }
 738                         }
 739                 }
 740         }
 741 }
 742
 743 /*
 744  * Mount or remount HAMMER2 fileystem from physical media
 745  *
 746  *      mountroot
 747  *              mp              mount point structure
 748  *              path            NULL
 749  *              data            <unused>
 750  *              cred            <unused>
 751  *
 752  *      mount
 753  *              mp              mount point structure
 754  *              path            path to mount point
 755  *              data            pointer to argument structure in user space
 756  *                      volume  volume path (device@LABEL form)
 757  *                      hflags  user mount flags
 758  *              cred            user credentials
 759  *
 760  * RETURNS:     0       Success
 761  *              !0      error number
 762  */
 763 static
 764 int
 765 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 766                   struct ucred *cred)
 767 {
 768         struct hammer2_mount_info info;
 769         hammer2_pfs_t *pmp;
 770         hammer2_pfs_t *spmp;
 771         hammer2_dev_t *hmp;
 772         hammer2_dev_t *force_local;
 773         hammer2_key_t key_next;
 774         hammer2_key_t key_dummy;
 775         hammer2_key_t lhc;
 776         struct vnode *devvp;
 777         struct nlookupdata nd;
 778         hammer2_chain_t *parent;
 779         hammer2_chain_t *chain;
 780         hammer2_cluster_t *cluster;
 781         const hammer2_inode_data_t *ripdata;
 782         hammer2_blockref_t bref;
 783         struct file *fp;
 784         char devstr[MNAMELEN];
 785         size_t size;
 786         size_t done;
 787         char *dev;
 788         char *label;
 789         int ronly = 1;
 790         int error;
 791         int cache_index;
 792         int i;
 793
 794         hmp = NULL;
 795         pmp = NULL;
 796         dev = NULL;
 797         label = NULL;
 798         devvp = NULL;
 799         cache_index = -1;
 800
 801         kprintf("hammer2_mount\n");
 802
 803         if (path == NULL) {
 804                 /*
 805                  * Root mount
 806                  */
 807                 bzero(&info, sizeof(info));
 808                 info.cluster_fd = -1;
 809                 ksnprintf(devstr, sizeof(devstr), "%s",
 810                           mp->mnt_stat.f_mntfromname);
 811                 kprintf("hammer2_mount: root '%s'\n", devstr);
 812         } else {
 813                 /*
 814                  * Non-root mount or updating a mount
 815                  */
 816                 error = copyin(data, &info, sizeof(info));
 817                 if (error)
 818                         return (error);
 819
 820                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 821                 if (error)
 822                         return (error);
 823         }
 824
 825         /* Extract device and label */
 826         dev = devstr;
 827         label = strchr(devstr, '@');
 828         if (label == NULL ||
 829             ((label + 1) - dev) > done) {
 830                 return (EINVAL);
 831         }
 832         *label = '\0';
 833         label++;
 834         if (*label == '\0')
 835                 return (EINVAL);
 836
 837         if (mp->mnt_flag & MNT_UPDATE) {
 838                 /*
 839                  * Update mount.  Note that pmp->iroot->cluster is
 840                  * an inode-embedded cluster and thus cannot be
 841                  * directly locked.
 842                  *
 843                  * XXX HAMMER2 needs to implement NFS export via
 844                  *     mountctl.
 845                  */
 846                 pmp = MPTOPMP(mp);
 847                 pmp->hflags = info.hflags;
 848                 cluster = &pmp->iroot->cluster;
 849                 for (i = 0; i < cluster->nchains; ++i) {
 850                         if (cluster->array[i].chain == NULL)
 851                                 continue;
 852                         hmp = cluster->array[i].chain->hmp;
 853                         devvp = hmp->devvp;
 854                         error = hammer2_remount(hmp, mp, path,
 855                                                 devvp, cred);
 856                         if (error)
 857                                 break;
 858                 }
 859
 860                 return error;
 861         }
 862
 863         /*
 864          * HMP device mount
 865          *
 866          * Lookup name and verify it refers to a block device.
 867          */
 868         if (path) {
 869                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 870                 if (error == 0)
 871                         error = nlookup(&nd);
 872                 if (error == 0)
 873                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 874                 nlookup_done(&nd);
 875         } else {
 876                 /* root mount */
 877                 cdev_t cdev = kgetdiskbyname(dev);
 878                 error = bdevvp(cdev, &devvp);
 879                 if (error)
 880                         kprintf("hammer2: cannot find '%s'\n", dev);
 881         }
 882
 883         if (error == 0) {
 884                 if (vn_isdisk(devvp, &error))
 885                         error = vfs_mountedon(devvp);
 886         }
 887
 888         /*
 889          * Determine if the device has already been mounted.  After this
 890          * check hmp will be non-NULL if we are doing the second or more
 891          * hammer2 mounts from the same device.
 892          */
 893         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 894         TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 895                 if (hmp->devvp == devvp)
 896                         break;
 897         }
 898
 899         /*
 900          * Open the device if this isn't a secondary mount and construct
 901          * the H2 device mount (hmp).
 902          */
 903         if (hmp == NULL) {
 904                 hammer2_chain_t *schain;
 905                 hammer2_xid_t xid;
 906
 907                 if (error == 0 && vcount(devvp) > 0)
 908                         error = EBUSY;
 909
 910                 /*
 911                  * Now open the device
 912                  */
 913                 if (error == 0) {
 914                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
 915                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
 916                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
 917                         if (error == 0) {
 918                                 error = VOP_OPEN(devvp,
 919                                                  ronly ? FREAD : FREAD | FWRITE,
 920                                                  FSCRED, NULL);
 921                         }
 922                         vn_unlock(devvp);
 923                 }
 924                 if (error && devvp) {
 925                         vrele(devvp);
 926                         devvp = NULL;
 927                 }
 928                 if (error) {
 929                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 930                         return error;
 931                 }
 932                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
 933                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
 934                 hmp->ronly = ronly;
 935                 hmp->devvp = devvp;
 936                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
 937                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
 938                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
 939                 RB_INIT(&hmp->iotree);
 940                 spin_init(&hmp->io_spin, "hm2mount_io");
 941                 spin_init(&hmp->list_spin, "hm2mount_list");
 942                 TAILQ_INIT(&hmp->flushq);
 943
 944                 lockinit(&hmp->vollk, "h2vol", 0, 0);
 945                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
 946
 947                 /*
 948                  * vchain setup. vchain.data is embedded.
 949                  * vchain.refs is initialized and will never drop to 0.
 950                  *
 951                  * NOTE! voldata is not yet loaded.
 952                  */
 953                 hmp->vchain.hmp = hmp;
 954                 hmp->vchain.refs = 1;
 955                 hmp->vchain.data = (void *)&hmp->voldata;
 956                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
 957                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 958                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 959
 960                 hammer2_chain_core_init(&hmp->vchain);
 961                 /* hmp->vchain.u.xxx is left NULL */
 962
 963                 /*
 964                  * fchain setup.  fchain.data is embedded.
 965                  * fchain.refs is initialized and will never drop to 0.
 966                  *
 967                  * The data is not used but needs to be initialized to
 968                  * pass assertion muster.  We use this chain primarily
 969                  * as a placeholder for the freemap's top-level RBTREE
 970                  * so it does not interfere with the volume's topology
 971                  * RBTREE.
 972                  */
 973                 hmp->fchain.hmp = hmp;
 974                 hmp->fchain.refs = 1;
 975                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
 976                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
 977                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 978                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 979                 hmp->fchain.bref.methods =
 980                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
 981                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
 982
 983                 hammer2_chain_core_init(&hmp->fchain);
 984                 /* hmp->fchain.u.xxx is left NULL */
 985
 986                 /*
 987                  * Install the volume header and initialize fields from
 988                  * voldata.
 989                  */
 990                 error = hammer2_install_volume_header(hmp);
 991                 if (error) {
 992                         hammer2_unmount_helper(mp, NULL, hmp);
 993                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 994                         hammer2_vfs_unmount(mp, MNT_FORCE);
 995                         return error;
 996                 }
 997
 998                 /*
 999                  * Really important to get these right or flush will get
1000                  * confused.
1001                  */
1002                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1003                 kprintf("alloc spmp %p tid %016jx\n",
1004                         hmp->spmp, hmp->voldata.mirror_tid);
1005                 spmp = hmp->spmp;
1006
1007                 /*
1008                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1009                  * is inherited from the volume header.
1010                  */
1011                 xid = 0;
1012                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1013                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1014                 hmp->vchain.pmp = spmp;
1015                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1016                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1017                 hmp->fchain.pmp = spmp;
1018
1019                 /*
1020                  * First locate the super-root inode, which is key 0
1021                  * relative to the volume header's blockset.
1022                  *
1023                  * Then locate the root inode by scanning the directory keyspace
1024                  * represented by the label.
1025                  */
1026                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1027                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1028                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1029                                       &cache_index, 0);
1030                 hammer2_chain_lookup_done(parent);
1031                 if (schain == NULL) {
1032                         kprintf("hammer2_mount: invalid super-root\n");
1033                         hammer2_unmount_helper(mp, NULL, hmp);
1034                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1035                         hammer2_vfs_unmount(mp, MNT_FORCE);
1036                         return EINVAL;
1037                 }
1038                 if (schain->error) {
1039                         kprintf("hammer2_mount: error %s reading super-root\n",
1040                                 hammer2_error_str(schain->error));
1041                         hammer2_chain_unlock(schain);
1042                         hammer2_chain_drop(schain);
1043                         schain = NULL;
1044                         hammer2_unmount_helper(mp, NULL, hmp);
1045                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1046                         hammer2_vfs_unmount(mp, MNT_FORCE);
1047                         return EINVAL;
1048                 }
1049
1050                 /*
1051                  * The super-root always uses an inode_tid of 1 when
1052                  * creating PFSs.
1053                  */
1054                 spmp->inode_tid = 1;
1055                 spmp->modify_tid = schain->bref.modify_tid + 1;
1056
1057                 /*
1058                  * Sanity-check schain's pmp and finish initialization.
1059                  * Any chain belonging to the super-root topology should
1060                  * have a NULL pmp (not even set to spmp).
1061                  */
1062                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1063                 KKASSERT(schain->pmp == NULL);
1064                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1065
1066                 /*
1067                  * Replace the dummy spmp->iroot with a real one.  It's
1068                  * easier to just do a wholesale replacement than to try
1069                  * to update the chain and fixup the iroot fields.
1070                  *
1071                  * The returned inode is locked with the supplied cluster.
1072                  */
1073                 cluster = hammer2_cluster_from_chain(schain);
1074                 hammer2_inode_drop(spmp->iroot);
1075                 spmp->iroot = NULL;
1076                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1077                 spmp->spmp_hmp = hmp;
1078                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1079                 spmp->pfs_hmps[0] = hmp;
1080                 hammer2_inode_ref(spmp->iroot);
1081                 hammer2_inode_unlock(spmp->iroot);
1082                 hammer2_cluster_unlock(cluster);
1083                 hammer2_cluster_drop(cluster);
1084                 schain = NULL;
1085                 /* leave spmp->iroot with one ref */
1086
1087                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1088                         error = hammer2_recovery(hmp);
1089                         /* XXX do something with error */
1090                 }
1091                 hammer2_update_pmps(hmp);
1092                 hammer2_iocom_init(hmp);
1093
1094                 /*
1095                  * Ref the cluster management messaging descriptor.  The mount
1096                  * program deals with the other end of the communications pipe.
1097                  *
1098                  * Root mounts typically do not supply one.
1099                  */
1100                 if (info.cluster_fd >= 0) {
1101                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1102                         if (fp) {
1103                                 hammer2_cluster_reconnect(hmp, fp);
1104                         } else {
1105                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1106                         }
1107                 }
1108         } else {
1109                 spmp = hmp->spmp;
1110                 if (info.hflags & HMNT2_DEVFLAGS) {
1111                         kprintf("hammer2: Warning: mount flags pertaining "
1112                                 "to the whole device may only be specified "
1113                                 "on the first mount of the device: %08x\n",
1114                                 info.hflags & HMNT2_DEVFLAGS);
1115                 }
1116         }
1117
1118         /*
1119          * Force local mount (disassociate all PFSs from their clusters).
1120          * Used primarily for debugging.
1121          */
1122         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1123
1124         /*
1125          * Lookup the mount point under the media-localized super-root.
1126          * Scanning hammer2_pfslist doesn't help us because it represents
1127          * PFS cluster ids which can aggregate several named PFSs together.
1128          *
1129          * cluster->pmp will incorrectly point to spmp and must be fixed
1130          * up later on.
1131          */
1132         hammer2_inode_lock(spmp->iroot, 0);
1133         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1134         lhc = hammer2_dirhash(label, strlen(label));
1135         chain = hammer2_chain_lookup(&parent, &key_next,
1136                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1137                                      &cache_index, 0);
1138         while (chain) {
1139                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1140                     strcmp(label, chain->data->ipdata.filename) == 0) {
1141                         break;
1142                 }
1143                 chain = hammer2_chain_next(&parent, chain, &key_next,
1144                                             key_next,
1145                                             lhc + HAMMER2_DIRHASH_LOMASK,
1146                                             &cache_index, 0);
1147         }
1148         if (parent) {
1149                 hammer2_chain_unlock(parent);
1150                 hammer2_chain_drop(parent);
1151         }
1152         hammer2_inode_unlock(spmp->iroot);
1153
1154         /*
1155          * PFS could not be found?
1156          */
1157         if (chain == NULL) {
1158                 kprintf("hammer2_mount: PFS label not found\n");
1159                 hammer2_unmount_helper(mp, NULL, hmp);
1160                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1161                 hammer2_vfs_unmount(mp, MNT_FORCE);
1162
1163                 return EINVAL;
1164         }
1165
1166         /*
1167          * Acquire the pmp structure (it should have already been allocated
1168          * via hammer2_update_pmps() so do not pass cluster in to add to
1169          * available chains).
1170          *
1171          * Check if the cluster has already been mounted.  A cluster can
1172          * only be mounted once, use null mounts to mount additional copies.
1173          */
1174         ripdata = &chain->data->ipdata;
1175         bref = chain->bref;
1176         pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid, force_local);
1177         hammer2_chain_unlock(chain);
1178         hammer2_chain_drop(chain);
1179
1180         if (pmp->mp) {
1181                 kprintf("hammer2_mount: PFS already mounted!\n");
1182                 hammer2_unmount_helper(mp, NULL, hmp);
1183                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1184                 hammer2_vfs_unmount(mp, MNT_FORCE);
1185
1186                 return EBUSY;
1187         }
1188
1189         /*
1190          * Finish the mount
1191          */
1192         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1193
1194         pmp->hflags = info.hflags;
1195         mp->mnt_flag = MNT_LOCAL;
1196         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1197         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1198
1199         /*
1200          * required mount structure initializations
1201          */
1202         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1203         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1204
1205         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1206         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1207
1208         /*
1209          * Optional fields
1210          */
1211         mp->mnt_iosize_max = MAXPHYS;
1212
1213         /*
1214          * Connect up mount pointers.
1215          */
1216         hammer2_mount_helper(mp, pmp);
1217
1218         lockmgr(&hammer2_mntlk, LK_RELEASE);
1219
1220         /*
1221          * Finish setup
1222          */
1223         vfs_getnewfsid(mp);
1224         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1225         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1226         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1227
1228         if (path) {
1229                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1230                           MNAMELEN - 1, &size);
1231                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1232         } /* else root mount, already in there */
1233
1234         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1235         if (path) {
1236                 copyinstr(path, mp->mnt_stat.f_mntonname,
1237                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1238                           &size);
1239         } else {
1240                 /* root mount */
1241                 mp->mnt_stat.f_mntonname[0] = '/';
1242         }
1243
1244         /*
1245          * Initial statfs to prime mnt_stat.
1246          */
1247         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1248
1249         return 0;
1250 }
1251
1252 /*
1253  * Scan PFSs under the super-root and create hammer2_pfs structures.
1254  */
1255 static
1256 void
1257 hammer2_update_pmps(hammer2_dev_t *hmp)
1258 {
1259         const hammer2_inode_data_t *ripdata;
1260         hammer2_chain_t *parent;
1261         hammer2_chain_t *chain;
1262         hammer2_blockref_t bref;
1263         hammer2_dev_t *force_local;
1264         hammer2_pfs_t *spmp;
1265         hammer2_pfs_t *pmp;
1266         hammer2_key_t key_next;
1267         int cache_index = -1;
1268
1269         /*
1270          * Force local mount (disassociate all PFSs from their clusters).
1271          * Used primarily for debugging.
1272          */
1273         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1274
1275         /*
1276          * Lookup mount point under the media-localized super-root.
1277          *
1278          * cluster->pmp will incorrectly point to spmp and must be fixed
1279          * up later on.
1280          */
1281         spmp = hmp->spmp;
1282         hammer2_inode_lock(spmp->iroot, 0);
1283         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1284         chain = hammer2_chain_lookup(&parent, &key_next,
1285                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1286                                          &cache_index, 0);
1287         while (chain) {
1288                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1289                         continue;
1290                 ripdata = &chain->data->ipdata;
1291                 bref = chain->bref;
1292                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1293
1294                 pmp = hammer2_pfsalloc(chain, ripdata,
1295                                        bref.modify_tid, force_local);
1296                 chain = hammer2_chain_next(&parent, chain, &key_next,
1297                                            key_next, HAMMER2_KEY_MAX,
1298                                            &cache_index, 0);
1299         }
1300         if (parent) {
1301                 hammer2_chain_unlock(parent);
1302                 hammer2_chain_drop(parent);
1303         }
1304         hammer2_inode_unlock(spmp->iroot);
1305 }
1306
1307 static
1308 int
1309 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1310                 struct vnode *devvp, struct ucred *cred)
1311 {
1312         int error;
1313
1314         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1315                 error = hammer2_recovery(hmp);
1316         } else {
1317                 error = 0;
1318         }
1319         return error;
1320 }
1321
1322 static
1323 int
1324 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1325 {
1326         hammer2_pfs_t *pmp;
1327         int flags;
1328         int error = 0;
1329
1330         pmp = MPTOPMP(mp);
1331
1332         if (pmp == NULL)
1333                 return(0);
1334
1335         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1336
1337         /*
1338          * If mount initialization proceeded far enough we must flush
1339          * its vnodes and sync the underlying mount points.  Three syncs
1340          * are required to fully flush the filesystem (freemap updates lag
1341          * by one flush, and one extra for safety).
1342          */
1343         if (mntflags & MNT_FORCE)
1344                 flags = FORCECLOSE;
1345         else
1346                 flags = 0;
1347         if (pmp->iroot) {
1348                 error = vflush(mp, 0, flags);
1349                 if (error)
1350                         goto failed;
1351                 hammer2_vfs_sync(mp, MNT_WAIT);
1352                 hammer2_vfs_sync(mp, MNT_WAIT);
1353                 hammer2_vfs_sync(mp, MNT_WAIT);
1354         }
1355
1356         /*
1357          * Cleanup the frontend support XOPS threads
1358          */
1359         hammer2_xop_helper_cleanup(pmp);
1360
1361         if (pmp->mp)
1362                 hammer2_unmount_helper(mp, pmp, NULL);
1363
1364         error = 0;
1365 failed:
1366         lockmgr(&hammer2_mntlk, LK_RELEASE);
1367
1368         return (error);
1369 }
1370
1371 /*
1372  * Mount helper, hook the system mount into our PFS.
1373  * The mount lock is held.
1374  *
1375  * We must bump the mount_count on related devices for any
1376  * mounted PFSs.
1377  */
1378 static
1379 void
1380 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1381 {
1382         hammer2_cluster_t *cluster;
1383         hammer2_chain_t *rchain;
1384         int i;
1385
1386         mp->mnt_data = (qaddr_t)pmp;
1387         pmp->mp = mp;
1388
1389         /*
1390          * After pmp->mp is set we have to adjust hmp->mount_count.
1391          */
1392         cluster = &pmp->iroot->cluster;
1393         for (i = 0; i < cluster->nchains; ++i) {
1394                 rchain = cluster->array[i].chain;
1395                 if (rchain == NULL)
1396                         continue;
1397                 ++rchain->hmp->mount_count;
1398                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1399                         rchain->hmp, rchain->hmp->mount_count);
1400         }
1401
1402         /*
1403          * Create missing Xop threads
1404          */
1405         hammer2_xop_helper_create(pmp);
1406 }
1407
1408 /*
1409  * Mount helper, unhook the system mount from our PFS.
1410  * The mount lock is held.
1411  *
1412  * If hmp is supplied a mount responsible for being the first to open
1413  * the block device failed and the block device and all PFSs using the
1414  * block device must be cleaned up.
1415  *
1416  * If pmp is supplied multiple devices might be backing the PFS and each
1417  * must be disconnected.  This might not be the last PFS using some of the
1418  * underlying devices.  Also, we have to adjust our hmp->mount_count
1419  * accounting for the devices backing the pmp which is now undergoing an
1420  * unmount.
1421  */
1422 static
1423 void
1424 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1425 {
1426         hammer2_cluster_t *cluster;
1427         hammer2_chain_t *rchain;
1428         struct vnode *devvp;
1429         int dumpcnt;
1430         int ronly = 0;
1431         int i;
1432
1433         /*
1434          * If no device supplied this is a high-level unmount and we have to
1435          * to disconnect the mount, adjust mount_count, and locate devices
1436          * that might now have no mounts.
1437          */
1438         if (pmp) {
1439                 KKASSERT(hmp == NULL);
1440                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1441                 pmp->mp = NULL;
1442                 mp->mnt_data = NULL;
1443
1444                 /*
1445                  * After pmp->mp is cleared we have to account for
1446                  * mount_count.
1447                  */
1448                 cluster = &pmp->iroot->cluster;
1449                 for (i = 0; i < cluster->nchains; ++i) {
1450                         rchain = cluster->array[i].chain;
1451                         if (rchain == NULL)
1452                                 continue;
1453                         --rchain->hmp->mount_count;
1454                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1455                                 rchain->hmp, rchain->hmp->mount_count);
1456                         /* scrapping hmp now may invalidate the pmp */
1457                 }
1458 again:
1459                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1460                         if (hmp->mount_count == 0) {
1461                                 hammer2_unmount_helper(NULL, NULL, hmp);
1462                                 goto again;
1463                         }
1464                 }
1465                 return;
1466         }
1467
1468         /*
1469          * Try to terminate the block device.  We can't terminate it if
1470          * there are still PFSs referencing it.
1471          */
1472         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1473                 hmp, hmp->mount_count);
1474         if (hmp->mount_count)
1475                 return;
1476
1477         hammer2_pfsfree_scan(hmp);
1478         hammer2_dev_exlock(hmp);        /* XXX order */
1479
1480         /*
1481          * Cycle the volume data lock as a safety (probably not needed any
1482          * more).  To ensure everything is out we need to flush at least
1483          * three times.  (1) The running of the sideq can dirty the
1484          * filesystem, (2) A normal flush can dirty the freemap, and
1485          * (3) ensure that the freemap is fully synchronized.
1486          *
1487          * The next mount's recovery scan can clean everything up but we want
1488          * to leave the filesystem in a 100% clean state on a normal unmount.
1489          */
1490 #if 0
1491         hammer2_voldata_lock(hmp);
1492         hammer2_voldata_unlock(hmp);
1493 #endif
1494         hammer2_iocom_uninit(hmp);
1495
1496         if ((hmp->vchain.flags | hmp->fchain.flags) &
1497             HAMMER2_CHAIN_FLUSH_MASK) {
1498                 kprintf("hammer2_unmount: chains left over "
1499                         "after final sync\n");
1500                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1501                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1502
1503                 if (hammer2_debug & 0x0010)
1504                         Debugger("entered debugger");
1505         }
1506
1507         KKASSERT(hmp->spmp == NULL);
1508
1509         /*
1510          * Finish up with the device vnode
1511          */
1512         if ((devvp = hmp->devvp) != NULL) {
1513                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1514                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1515                 hmp->devvp = NULL;
1516                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1517                 vn_unlock(devvp);
1518                 vrele(devvp);
1519                 devvp = NULL;
1520         }
1521
1522         /*
1523          * Clear vchain/fchain flags that might prevent final cleanup
1524          * of these chains.
1525          */
1526         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1527                 atomic_add_long(&hammer2_count_modified_chains, -1);
1528                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1529                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1530         }
1531         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1532                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1533         }
1534
1535         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1536                 atomic_add_long(&hammer2_count_modified_chains, -1);
1537                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1538                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1539         }
1540         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1541                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1542         }
1543
1544         /*
1545          * Final drop of embedded freemap root chain to
1546          * clean up fchain.core (fchain structure is not
1547          * flagged ALLOCATED so it is cleaned out and then
1548          * left to rot).
1549          */
1550         hammer2_chain_drop(&hmp->fchain);
1551
1552         /*
1553          * Final drop of embedded volume root chain to clean
1554          * up vchain.core (vchain structure is not flagged
1555          * ALLOCATED so it is cleaned out and then left to
1556          * rot).
1557          */
1558         dumpcnt = 50;
1559         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1560         dumpcnt = 50;
1561         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1562         hammer2_dev_unlock(hmp);
1563         hammer2_chain_drop(&hmp->vchain);
1564
1565         hammer2_io_cleanup(hmp, &hmp->iotree);
1566         if (hmp->iofree_count) {
1567                 kprintf("io_cleanup: %d I/O's left hanging\n",
1568                         hmp->iofree_count);
1569         }
1570
1571         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1572         kmalloc_destroy(&hmp->mchain);
1573         kfree(hmp, M_HAMMER2);
1574 }
1575
1576 int
1577 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1578                  ino_t ino, struct vnode **vpp)
1579 {
1580         hammer2_xop_lookup_t *xop;
1581         hammer2_pfs_t *pmp;
1582         hammer2_inode_t *ip;
1583         hammer2_tid_t inum;
1584         int error;
1585
1586         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1587
1588         error = 0;
1589         pmp = MPTOPMP(mp);
1590
1591         /*
1592          * Easy if we already have it cached
1593          */
1594         ip = hammer2_inode_lookup(pmp, inum);
1595         if (ip) {
1596                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1597                 *vpp = hammer2_igetv(ip, &error);
1598                 hammer2_inode_unlock(ip);
1599                 hammer2_inode_drop(ip);         /* from lookup */
1600
1601                 return error;
1602         }
1603
1604         /*
1605          * Otherwise we have to find the inode
1606          */
1607         xop = hammer2_xop_alloc(pmp->iroot, 0);
1608         xop->lhc = inum;
1609         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1610         error = hammer2_xop_collect(&xop->head, 0);
1611
1612         if (error == 0) {
1613                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1614                         kprintf("vget: no collect error but also no rdata\n");
1615                         kprintf("xop %p\n", xop);
1616                         while ((hammer2_debug & 0x80000) == 0) {
1617                                 tsleep(xop, PCATCH, "wait", hz * 10);
1618                         }
1619                         ip = NULL;
1620                 } else {
1621                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1622                 }
1623         }
1624         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1625
1626         if (ip) {
1627                 *vpp = hammer2_igetv(ip, &error);
1628                 hammer2_inode_unlock(ip);
1629         } else {
1630                 *vpp = NULL;
1631                 error = ENOENT;
1632         }
1633         return (error);
1634 }
1635
1636 static
1637 int
1638 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1639 {
1640         hammer2_pfs_t *pmp;
1641         struct vnode *vp;
1642         int error;
1643
1644         pmp = MPTOPMP(mp);
1645         if (pmp->iroot == NULL) {
1646                 *vpp = NULL;
1647                 return EINVAL;
1648         }
1649
1650         error = 0;
1651         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1652
1653         while (pmp->inode_tid == 0) {
1654                 hammer2_xop_ipcluster_t *xop;
1655                 hammer2_inode_meta_t *meta;
1656
1657                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1658                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1659                 error = hammer2_xop_collect(&xop->head, 0);
1660
1661                 if (error == 0) {
1662                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1663                         pmp->iroot->meta = *meta;
1664                         pmp->inode_tid = meta->pfs_inum + 1;
1665                         if (pmp->inode_tid < HAMMER2_INODE_START)
1666                                 pmp->inode_tid = HAMMER2_INODE_START;
1667                         pmp->modify_tid =
1668                                 xop->head.cluster.focus->bref.modify_tid + 1;
1669                         kprintf("PFS: Starting inode %jd\n",
1670                                 (intmax_t)pmp->inode_tid);
1671                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1672                                 pmp->inode_tid, pmp->modify_tid);
1673                         wakeup(&pmp->iroot);
1674
1675                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1676
1677                         /*
1678                          * Prime the mount info.
1679                          */
1680                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1681                         break;
1682                 }
1683
1684                 /*
1685                  * Loop, try again
1686                  */
1687                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1688                 hammer2_inode_unlock(pmp->iroot);
1689                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1690                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1691                 if (error == EINTR)
1692                         break;
1693         }
1694
1695         if (error) {
1696                 hammer2_inode_unlock(pmp->iroot);
1697                 *vpp = NULL;
1698         } else {
1699                 vp = hammer2_igetv(pmp->iroot, &error);
1700                 hammer2_inode_unlock(pmp->iroot);
1701                 *vpp = vp;
1702         }
1703
1704         return (error);
1705 }
1706
1707 /*
1708  * Filesystem status
1709  *
1710  * XXX incorporate ipdata->meta.inode_quota and data_quota
1711  */
1712 static
1713 int
1714 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1715 {
1716         hammer2_pfs_t *pmp;
1717         hammer2_dev_t *hmp;
1718         hammer2_blockref_t bref;
1719         int i;
1720
1721         /*
1722          * NOTE: iroot might not have validated the cluster yet.
1723          */
1724         pmp = MPTOPMP(mp);
1725
1726         mp->mnt_stat.f_files = 0;
1727         mp->mnt_stat.f_ffree = 0;
1728         mp->mnt_stat.f_blocks = 0;
1729         mp->mnt_stat.f_bfree = 0;
1730         mp->mnt_stat.f_bavail = 0;
1731
1732         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1733                 hmp = pmp->pfs_hmps[i];
1734                 if (hmp == NULL)
1735                         continue;
1736                 if (pmp->iroot->cluster.array[i].chain)
1737                         bref = pmp->iroot->cluster.array[i].chain->bref;
1738                 else
1739                         bzero(&bref, sizeof(bref));
1740
1741                 mp->mnt_stat.f_files = bref.inode_count;
1742                 mp->mnt_stat.f_ffree = 0;
1743 #if 0
1744                 mp->mnt_stat.f_blocks = (bref.data_count +
1745                                          hmp->voldata.allocator_free) /
1746                                         mp->mnt_vstat.f_bsize;
1747                 mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free /
1748                                         mp->mnt_vstat.f_bsize;
1749 #endif
1750                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1751                                         mp->mnt_vstat.f_bsize;
1752                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1753                                         mp->mnt_vstat.f_bsize;
1754                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1755
1756                 *sbp = mp->mnt_stat;
1757         }
1758         return (0);
1759 }
1760
1761 static
1762 int
1763 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1764 {
1765         hammer2_pfs_t *pmp;
1766         hammer2_dev_t *hmp;
1767         hammer2_blockref_t bref;
1768         int i;
1769
1770         /*
1771          * NOTE: iroot might not have validated the cluster yet.
1772          */
1773         pmp = MPTOPMP(mp);
1774
1775         mp->mnt_vstat.f_bsize = 0;
1776         mp->mnt_vstat.f_files = 0;
1777         mp->mnt_vstat.f_ffree = 0;
1778         mp->mnt_vstat.f_blocks = 0;
1779         mp->mnt_vstat.f_bfree = 0;
1780         mp->mnt_vstat.f_bavail = 0;
1781
1782         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1783                 hmp = pmp->pfs_hmps[i];
1784                 if (hmp == NULL)
1785                         continue;
1786                 if (pmp->iroot->cluster.array[i].chain)
1787                         bref = pmp->iroot->cluster.array[i].chain->bref;
1788                 else
1789                         bzero(&bref, sizeof(bref));
1790
1791                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1792                 mp->mnt_vstat.f_files = bref.inode_count;
1793                 mp->mnt_vstat.f_ffree = 0;
1794 #if 0
1795                 mp->mnt_vstat.f_blocks = (bref.data_count +
1796                                          hmp->voldata.allocator_free) /
1797                                         mp->mnt_vstat.f_bsize;
1798                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1799                                         mp->mnt_vstat.f_bsize;
1800 #endif
1801                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1802                                         mp->mnt_vstat.f_bsize;
1803                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1804                                         mp->mnt_vstat.f_bsize;
1805                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1806
1807                 *sbp = mp->mnt_vstat;
1808         }
1809         return (0);
1810 }
1811
1812 /*
1813  * Mount-time recovery (RW mounts)
1814  *
1815  * Updates to the free block table are allowed to lag flushes by one
1816  * transaction.  In case of a crash, then on a fresh mount we must do an
1817  * incremental scan of the last committed transaction id and make sure that
1818  * all related blocks have been marked allocated.
1819  *
1820  * The super-root topology and each PFS has its own transaction id domain,
1821  * so we must track PFS boundary transitions.
1822  */
1823 struct hammer2_recovery_elm {
1824         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1825         hammer2_chain_t *chain;
1826         hammer2_tid_t sync_tid;
1827 };
1828
1829 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1830
1831 struct hammer2_recovery_info {
1832         struct hammer2_recovery_list list;
1833         hammer2_tid_t   mtid;
1834         int     depth;
1835 };
1836
1837 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1838                         hammer2_chain_t *parent,
1839                         struct hammer2_recovery_info *info,
1840                         hammer2_tid_t sync_tid);
1841
1842 #define HAMMER2_RECOVERY_MAXDEPTH       10
1843
1844 static
1845 int
1846 hammer2_recovery(hammer2_dev_t *hmp)
1847 {
1848         struct hammer2_recovery_info info;
1849         struct hammer2_recovery_elm *elm;
1850         hammer2_chain_t *parent;
1851         hammer2_tid_t sync_tid;
1852         hammer2_tid_t mirror_tid;
1853         int error;
1854         int cumulative_error = 0;
1855
1856         hammer2_trans_init(hmp->spmp, 0);
1857
1858         sync_tid = hmp->voldata.freemap_tid;
1859         mirror_tid = hmp->voldata.mirror_tid;
1860
1861         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1862         if (sync_tid >= mirror_tid) {
1863                 kprintf(" no recovery needed\n");
1864         } else {
1865                 kprintf(" freemap recovery %016jx-%016jx\n",
1866                         sync_tid + 1, mirror_tid);
1867         }
1868
1869         TAILQ_INIT(&info.list);
1870         info.depth = 0;
1871         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1872         cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
1873         hammer2_chain_lookup_done(parent);
1874
1875         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1876                 TAILQ_REMOVE(&info.list, elm, entry);
1877                 parent = elm->chain;
1878                 sync_tid = elm->sync_tid;
1879                 kfree(elm, M_HAMMER2);
1880
1881                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1882                 error = hammer2_recovery_scan(hmp, parent, &info,
1883                                               hmp->voldata.freemap_tid);
1884                 hammer2_chain_unlock(parent);
1885                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
1886                 if (error)
1887                         cumulative_error = error;
1888         }
1889         hammer2_trans_done(hmp->spmp);
1890
1891         return cumulative_error;
1892 }
1893
1894 static
1895 int
1896 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1897                       struct hammer2_recovery_info *info,
1898                       hammer2_tid_t sync_tid)
1899 {
1900         const hammer2_inode_data_t *ripdata;
1901         hammer2_chain_t *chain;
1902         hammer2_blockref_t bref;
1903         int cache_index;
1904         int cumulative_error = 0;
1905         int error;
1906         int first;
1907
1908         /*
1909          * Adjust freemap to ensure that the block(s) are marked allocated.
1910          */
1911         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1912                 hammer2_freemap_adjust(hmp, &parent->bref,
1913                                        HAMMER2_FREEMAP_DORECOVER);
1914         }
1915
1916         /*
1917          * Check type for recursive scan
1918          */
1919         switch(parent->bref.type) {
1920         case HAMMER2_BREF_TYPE_VOLUME:
1921                 /* data already instantiated */
1922                 break;
1923         case HAMMER2_BREF_TYPE_INODE:
1924                 /*
1925                  * Must instantiate data for DIRECTDATA test and also
1926                  * for recursion.
1927                  */
1928                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1929                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
1930                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1931                         /* not applicable to recovery scan */
1932                         hammer2_chain_unlock(parent);
1933                         return 0;
1934                 }
1935                 hammer2_chain_unlock(parent);
1936                 break;
1937         case HAMMER2_BREF_TYPE_INDIRECT:
1938                 /*
1939                  * Must instantiate data for recursion
1940                  */
1941                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1942                 hammer2_chain_unlock(parent);
1943                 break;
1944         case HAMMER2_BREF_TYPE_DATA:
1945         case HAMMER2_BREF_TYPE_FREEMAP:
1946         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1947         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1948                 /* not applicable to recovery scan */
1949                 return 0;
1950                 break;
1951         default:
1952                 return EDOM;
1953         }
1954
1955         /*
1956          * Defer operation if depth limit reached or if we are crossing a
1957          * PFS boundary.
1958          */
1959         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1960                 struct hammer2_recovery_elm *elm;
1961
1962                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1963                 elm->chain = parent;
1964                 elm->sync_tid = sync_tid;
1965                 hammer2_chain_ref(parent);
1966                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
1967                 /* unlocked by caller */
1968
1969                 return(0);
1970         }
1971
1972
1973         /*
1974          * Recursive scan of the last flushed transaction only.  We are
1975          * doing this without pmp assignments so don't leave the chains
1976          * hanging around after we are done with them.
1977          */
1978         cache_index = 0;
1979         chain = NULL;
1980         first = 1;
1981
1982         while (hammer2_chain_scan(parent, &chain, &bref,
1983                                   &first, &cache_index,
1984                                   HAMMER2_LOOKUP_NODATA) != NULL) {
1985                 /*
1986                  * If this is a leaf
1987                  */
1988                 if (chain == NULL) {
1989                         if (bref.mirror_tid > sync_tid) {
1990                                 hammer2_freemap_adjust(hmp, &bref,
1991                                                      HAMMER2_FREEMAP_DORECOVER);
1992                         }
1993                         continue;
1994                 }
1995
1996                 /*
1997                  * This may or may not be a recursive node.
1998                  */
1999                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2000                 if (bref.mirror_tid > sync_tid) {
2001                         ++info->depth;
2002                         error = hammer2_recovery_scan(hmp, chain,
2003                                                       info, sync_tid);
2004                         --info->depth;
2005                         if (error)
2006                                 cumulative_error = error;
2007                 }
2008
2009                 /*
2010                  * Flush the recovery at the PFS boundary to stage it for
2011                  * the final flush of the super-root topology.
2012                  */
2013                 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2014                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2015                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2016                 }
2017         }
2018
2019         return cumulative_error;
2020 }
2021
2022 /*
2023  * Sync a mount point; this is called on a per-mount basis from the
2024  * filesystem syncer process periodically and whenever a user issues
2025  * a sync.
2026  */
2027 int
2028 hammer2_vfs_sync(struct mount *mp, int waitfor)
2029 {
2030         hammer2_xop_flush_t *xop;
2031         struct hammer2_sync_info info;
2032         hammer2_inode_t *iroot;
2033         hammer2_pfs_t *pmp;
2034         int flags;
2035         int error;
2036
2037         pmp = MPTOPMP(mp);
2038         iroot = pmp->iroot;
2039         KKASSERT(iroot);
2040         KKASSERT(iroot->pmp == pmp);
2041
2042         /*
2043          * We can't acquire locks on existing vnodes while in a transaction
2044          * without risking a deadlock.  This assumes that vfsync() can be
2045          * called without the vnode locked (which it can in DragonFly).
2046          * Otherwise we'd have to implement a multi-pass or flag the lock
2047          * failures and retry.
2048          *
2049          * The reclamation code interlocks with the sync list's token
2050          * (by removing the vnode from the scan list) before unlocking
2051          * the inode, giving us time to ref the inode.
2052          */
2053         /*flags = VMSC_GETVP;*/
2054         flags = 0;
2055         if (waitfor & MNT_LAZY)
2056                 flags |= VMSC_ONEPASS;
2057
2058 #if 1
2059         /*
2060          * Preflush the vnodes using a normal transaction before interlocking
2061          * with a flush transaction.  We do this to try to run as much of
2062          * the compression as possible outside the flush transaction.
2063          */
2064         hammer2_trans_init(pmp, 0);
2065         info.error = 0;
2066         info.waitfor = MNT_NOWAIT;
2067         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2068         hammer2_trans_done(pmp);
2069 #endif
2070
2071         /*
2072          * Start our flush transaction.  This does not return until all
2073          * concurrent transactions have completed and will prevent any
2074          * new transactions from running concurrently, except for the
2075          * buffer cache transactions.
2076          *
2077          * For efficiency do an async pass before making sure with a
2078          * synchronous pass on all related buffer cache buffers.  It
2079          * should theoretically not be possible for any new file buffers
2080          * to be instantiated during this sequence.
2081          */
2082         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH |
2083                                 HAMMER2_TRANS_PREFLUSH);
2084         hammer2_inode_run_sideq(pmp);
2085
2086         info.error = 0;
2087         info.waitfor = MNT_NOWAIT;
2088         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2089         info.waitfor = MNT_WAIT;
2090         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2091
2092         /*
2093          * Clear PREFLUSH.  This prevents (or asserts on) any new logical
2094          * buffer cache flushes which occur during the flush.  Device buffers
2095          * are not affected.
2096          */
2097         hammer2_bioq_sync(pmp);
2098         hammer2_trans_clear_preflush(pmp);
2099
2100         /*
2101          * Use the XOP interface to concurrently flush all nodes to
2102          * synchronize the PFSROOT subtopology to the media.  A standard
2103          * end-of-scan ENOENT error indicates cluster sufficiency.
2104          *
2105          * Note that this flush will not be visible on crash recovery until
2106          * we flush the super-root topology in the next loop.
2107          *
2108          * XXX For now wait for all flushes to complete.
2109          */
2110         if (iroot) {
2111                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2112                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2113                 error = hammer2_xop_collect(&xop->head,
2114                                             HAMMER2_XOP_COLLECT_WAITALL);
2115                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2116                 if (error == ENOENT)
2117                         error = 0;
2118         } else {
2119                 error = 0;
2120         }
2121         hammer2_trans_done(pmp);
2122
2123         return (error);
2124 }
2125
2126 /*
2127  * Sync passes.
2128  *
2129  * Note that we ignore the tranasction mtid we got above.  Instead,
2130  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2131  * transactions.
2132  */
2133 static int
2134 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2135 {
2136         struct hammer2_sync_info *info = data;
2137         hammer2_inode_t *ip;
2138         int error;
2139
2140         /*
2141          * Degenerate cases.  Note that ip == NULL typically means the
2142          * syncer vnode itself and we don't want to vclrisdirty() in that
2143          * situation.
2144          */
2145         ip = VTOI(vp);
2146         if (ip == NULL) {
2147                 return(0);
2148         }
2149         if (vp->v_type == VNON || vp->v_type == VBAD) {
2150                 vclrisdirty(vp);
2151                 return(0);
2152         }
2153
2154         /*
2155          * VOP_FSYNC will start a new transaction so replicate some code
2156          * here to do it inline (see hammer2_vop_fsync()).
2157          *
2158          * WARNING: The vfsync interacts with the buffer cache and might
2159          *          block, we can't hold the inode lock at that time.
2160          *          However, we MUST ref ip before blocking to ensure that
2161          *          it isn't ripped out from under us (since we do not
2162          *          hold a lock on the vnode).
2163          */
2164         hammer2_inode_ref(ip);
2165         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2166             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2167                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2168                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2169                                  HAMMER2_INODE_MODIFIED)) {
2170                         hammer2_inode_lock(ip, 0);
2171                         hammer2_inode_chain_sync(ip);
2172                         hammer2_inode_unlock(ip);
2173                 }
2174         }
2175         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2176             RB_EMPTY(&vp->v_rbdirty_tree)) {
2177                 vclrisdirty(vp);
2178         }
2179
2180         hammer2_inode_drop(ip);
2181 #if 1
2182         error = 0;
2183         if (error)
2184                 info->error = error;
2185 #endif
2186         return(0);
2187 }
2188
2189 static
2190 int
2191 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2192 {
2193         hammer2_inode_t *ip;
2194
2195         KKASSERT(MAXFIDSZ >= 16);
2196         ip = VTOI(vp);
2197         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2198         fhp->fid_ext = 0;
2199         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2200         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2201
2202         return 0;
2203 }
2204
2205 static
2206 int
2207 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2208                struct fid *fhp, struct vnode **vpp)
2209 {
2210         hammer2_pfs_t *pmp;
2211         hammer2_tid_t inum;
2212         int error;
2213
2214         pmp = MPTOPMP(mp);
2215         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2216         if (vpp) {
2217                 if (inum == 1)
2218                         error = hammer2_vfs_root(mp, vpp);
2219                 else
2220                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2221         } else {
2222                 error = 0;
2223         }
2224         if (error)
2225                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2226         return error;
2227 }
2228
2229 static
2230 int
2231 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2232                  int *exflagsp, struct ucred **credanonp)
2233 {
2234         hammer2_pfs_t *pmp;
2235         struct netcred *np;
2236         int error;
2237
2238         pmp = MPTOPMP(mp);
2239         np = vfs_export_lookup(mp, &pmp->export, nam);
2240         if (np) {
2241                 *exflagsp = np->netc_exflags;
2242                 *credanonp = &np->netc_anon;
2243                 error = 0;
2244         } else {
2245                 error = EACCES;
2246         }
2247         return error;
2248 }
2249
2250 /*
2251  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2252  * header into the HMP
2253  *
2254  * XXX read four volhdrs and use the one with the highest TID whos CRC
2255  *     matches.
2256  *
2257  * XXX check iCRCs.
2258  *
2259  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2260  *     nonexistant locations.
2261  *
2262  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2263  */
2264 static
2265 int
2266 hammer2_install_volume_header(hammer2_dev_t *hmp)
2267 {
2268         hammer2_volume_data_t *vd;
2269         struct buf *bp;
2270         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2271         int error_reported;
2272         int error;
2273         int valid;
2274         int i;
2275
2276         error_reported = 0;
2277         error = 0;
2278         valid = 0;
2279         bp = NULL;
2280
2281         /*
2282          * There are up to 4 copies of the volume header (syncs iterate
2283          * between them so there is no single master).  We don't trust the
2284          * volu_size field so we don't know precisely how large the filesystem
2285          * is, so depend on the OS to return an error if we go beyond the
2286          * block device's EOF.
2287          */
2288         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2289                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2290                               HAMMER2_VOLUME_BYTES, &bp);
2291                 if (error) {
2292                         brelse(bp);
2293                         bp = NULL;
2294                         continue;
2295                 }
2296
2297                 vd = (struct hammer2_volume_data *) bp->b_data;
2298                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2299                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2300                         brelse(bp);
2301                         bp = NULL;
2302                         continue;
2303                 }
2304
2305                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2306                         /* XXX: Reversed-endianness filesystem */
2307                         kprintf("hammer2: reverse-endian filesystem detected");
2308                         brelse(bp);
2309                         bp = NULL;
2310                         continue;
2311                 }
2312
2313                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2314                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2315                                       HAMMER2_VOLUME_ICRC0_SIZE);
2316                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2317                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2318                                        HAMMER2_VOLUME_ICRC1_SIZE);
2319                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2320                         kprintf("hammer2 volume header crc "
2321                                 "mismatch copy #%d %08x/%08x\n",
2322                                 i, crc0, crc);
2323                         error_reported = 1;
2324                         brelse(bp);
2325                         bp = NULL;
2326                         continue;
2327                 }
2328                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2329                         valid = 1;
2330                         hmp->voldata = *vd;
2331                         hmp->volhdrno = i;
2332                 }
2333                 brelse(bp);
2334                 bp = NULL;
2335         }
2336         if (valid) {
2337                 hmp->volsync = hmp->voldata;
2338                 error = 0;
2339                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2340                         kprintf("hammer2: using volume header #%d\n",
2341                                 hmp->volhdrno);
2342                 }
2343         } else {
2344                 error = EINVAL;
2345                 kprintf("hammer2: no valid volume headers found!\n");
2346         }
2347         return (error);
2348 }
2349
2350 /*
2351  * This handles hysteresis on regular file flushes.  Because the BIOs are
2352  * routed to a thread it is possible for an excessive number to build up
2353  * and cause long front-end stalls long before the runningbuffspace limit
2354  * is hit, so we implement hammer2_flush_pipe to control the
2355  * hysteresis.
2356  *
2357  * This is a particular problem when compression is used.
2358  */
2359 void
2360 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2361 {
2362         atomic_add_int(&pmp->count_lwinprog, 1);
2363 }
2364
2365 void
2366 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2367 {
2368         int lwinprog;
2369
2370         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2371         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2372             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2373                 atomic_clear_int(&pmp->count_lwinprog,
2374                                  HAMMER2_LWINPROG_WAITING);
2375                 wakeup(&pmp->count_lwinprog);
2376         }
2377         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2378             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2379                 atomic_clear_int(&pmp->count_lwinprog,
2380                                  HAMMER2_LWINPROG_WAITING0);
2381                 wakeup(&pmp->count_lwinprog);
2382         }
2383 }
2384
2385 void
2386 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2387 {
2388         int lwinprog;
2389         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2390                                     HAMMER2_LWINPROG_WAITING0;
2391
2392         for (;;) {
2393                 lwinprog = pmp->count_lwinprog;
2394                 cpu_ccfence();
2395                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2396                         break;
2397                 tsleep_interlock(&pmp->count_lwinprog, 0);
2398                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2399                 lwinprog = pmp->count_lwinprog;
2400                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2401                         break;
2402                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2403         }
2404 }
2405
2406 /*
2407  * Manage excessive memory resource use for chain and related
2408  * structures.
2409  */
2410 void
2411 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2412 {
2413         uint32_t waiting;
2414         uint32_t count;
2415         uint32_t limit;
2416 #if 0
2417         static int zzticks;
2418 #endif
2419
2420         /*
2421          * Atomic check condition and wait.  Also do an early speedup of
2422          * the syncer to try to avoid hitting the wait.
2423          */
2424         for (;;) {
2425                 waiting = pmp->inmem_dirty_chains;
2426                 cpu_ccfence();
2427                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2428
2429                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2430                 if (limit < hammer2_limit_dirty_chains)
2431                         limit = hammer2_limit_dirty_chains;
2432                 if (limit < 1000)
2433                         limit = 1000;
2434
2435 #if 0
2436                 if ((int)(ticks - zzticks) > hz) {
2437                         zzticks = ticks;
2438                         kprintf("count %ld %ld\n", count, limit);
2439                 }
2440 #endif
2441
2442                 /*
2443                  * Block if there are too many dirty chains present, wait
2444                  * for the flush to clean some out.
2445                  */
2446                 if (count > limit) {
2447                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2448                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2449                                                waiting,
2450                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2451                                 speedup_syncer(pmp->mp);
2452                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2453                                        "chnmem", hz);
2454                         }
2455                         continue;       /* loop on success or fail */
2456                 }
2457
2458                 /*
2459                  * Try to start an early flush before we are forced to block.
2460                  */
2461                 if (count > limit * 7 / 10)
2462                         speedup_syncer(pmp->mp);
2463                 break;
2464         }
2465 }
2466
2467 void
2468 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2469 {
2470         if (pmp) {
2471                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2472         }
2473 }
2474
2475 void
2476 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2477 {
2478         uint32_t waiting;
2479
2480         if (pmp == NULL)
2481                 return;
2482
2483         for (;;) {
2484                 waiting = pmp->inmem_dirty_chains;
2485                 cpu_ccfence();
2486                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2487                                        waiting,
2488                                        (waiting - 1) &
2489                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2490                         break;
2491                 }
2492         }
2493
2494         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2495                 wakeup(&pmp->inmem_dirty_chains);
2496 }
2497
2498 /*
2499  * Debugging
2500  */
2501 void
2502 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2503 {
2504         hammer2_chain_t *scan;
2505         hammer2_chain_t *parent;
2506
2507         --*countp;
2508         if (*countp == 0) {
2509                 kprintf("%*.*s...\n", tab, tab, "");
2510                 return;
2511         }
2512         if (*countp < 0)
2513                 return;
2514         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2515                 tab, tab, "", pfx,
2516                 chain, chain->bref.type,
2517                 chain->bref.key, chain->bref.keybits,
2518                 chain->bref.mirror_tid);
2519
2520         kprintf("%*.*s      [%08x] (%s) refs=%d",
2521                 tab, tab, "",
2522                 chain->flags,
2523                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2524                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2525                 chain->refs);
2526
2527         parent = chain->parent;
2528         if (parent)
2529                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2530                         tab, tab, "",
2531                         parent, parent->flags, parent->refs);
2532         if (RB_EMPTY(&chain->core.rbtree)) {
2533                 kprintf("\n");
2534         } else {
2535                 kprintf(" {\n");
2536                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2537                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2538                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2539                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2540                                 chain->data->ipdata.filename);
2541                 else
2542                         kprintf("%*.*s}\n", tab, tab, "");
2543         }
2544 }