usr.sbin/makefs/hammer2/hammer2_vfsops.c

   1 /*
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 2022 Tomohiro Kusumi <tkusumi@netbsd.org>
   5  * Copyright (c) 2011-2022 The DragonFly Project.  All rights reserved.
   6  *
   7  * This code is derived from software contributed to The DragonFly Project
   8  * by Matthew Dillon <dillon@dragonflybsd.org>
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions
  12  * are met:
  13  *
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in
  18  *    the documentation and/or other materials provided with the
  19  *    distribution.
  20  * 3. Neither the name of The DragonFly Project nor the names of its
  21  *    contributors may be used to endorse or promote products derived
  22  *    from this software without specific, prior written permission.
  23  *
  24  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  26  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  27  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  28  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  29  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  30  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  31  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  32  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  33  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  34  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  35  * SUCH DAMAGE.
  36  */
  37 /*
  38 #include <sys/param.h>
  39 #include <sys/systm.h>
  40 #include <sys/kernel.h>
  41 #include <sys/nlookup.h>
  42 #include <sys/vnode.h>
  43 #include <sys/mount.h>
  44 #include <sys/fcntl.h>
  45 #include <sys/vfsops.h>
  46 #include <sys/sysctl.h>
  47 #include <sys/socket.h>
  48 #include <sys/objcache.h>
  49 #include <sys/proc.h>
  50 #include <sys/lock.h>
  51 #include <sys/file.h>
  52 */
  53
  54 #include "hammer2.h"
  55
  56 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  57 static struct hammer2_mntlist hammer2_mntlist;
  58
  59 struct hammer2_pfslist hammer2_pfslist;
  60 struct hammer2_pfslist hammer2_spmplist;
  61 struct lock hammer2_mntlk;
  62
  63 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
  64 int hammer2_debug;
  65 int hammer2_aux_flags;
  66 int hammer2_xop_nthreads;
  67 int hammer2_xop_sgroups;
  68 int hammer2_xop_xgroups;
  69 int hammer2_xop_xbase;
  70 int hammer2_xop_mod;
  71 long hammer2_debug_inode;
  72 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
  73 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
  74 int hammer2_cluster_write = 0;          /* physical write clustering */
  75 int hammer2_dedup_enable = 1;
  76 int hammer2_always_compress = 0;        /* always try to compress */
  77 int hammer2_flush_pipe = 100;
  78 int hammer2_dio_count;
  79 int hammer2_dio_limit = 256;
  80 int hammer2_bulkfree_tps = 5000;
  81 int hammer2_spread_workers;
  82 int hammer2_limit_saved_depth;
  83 long hammer2_chain_allocs;
  84 long hammer2_limit_saved_chains;
  85 long hammer2_limit_dirty_chains;
  86 long hammer2_limit_dirty_inodes;
  87 long hammer2_count_modified_chains;
  88 long hammer2_iod_file_read;
  89 long hammer2_iod_meta_read;
  90 long hammer2_iod_indr_read;
  91 long hammer2_iod_fmap_read;
  92 long hammer2_iod_volu_read;
  93 long hammer2_iod_file_write;
  94 long hammer2_iod_file_wembed;
  95 long hammer2_iod_file_wzero;
  96 long hammer2_iod_file_wdedup;
  97 long hammer2_iod_meta_write;
  98 long hammer2_iod_indr_write;
  99 long hammer2_iod_fmap_write;
 100 long hammer2_iod_volu_write;
 101 static long hammer2_iod_inode_creates;
 102 static long hammer2_iod_inode_deletes;
 103
 104 long hammer2_process_icrc32;
 105 long hammer2_process_xxhash64;
 106
 107 int hz;
 108 int ticks;
 109 int64_t vnode_count;
 110
 111 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 112 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 113                 "Buffer used for compression.");
 114
 115 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 116 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 117                 "Buffer used for decompression.");
 118
 119 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 120
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
 122            &hammer2_supported_version, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, aux_flags, CTLFLAG_RW,
 124            &hammer2_aux_flags, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 126            &hammer2_debug, 0, "");
 127 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW,
 128            &hammer2_debug_inode, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, spread_workers, CTLFLAG_RW,
 130            &hammer2_spread_workers, 0, "");
 131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 132            &hammer2_cluster_meta_read, 0, "");
 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 134            &hammer2_cluster_data_read, 0, "");
 135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 136            &hammer2_cluster_write, 0, "");
 137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 138            &hammer2_dedup_enable, 0, "");
 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 140            &hammer2_always_compress, 0, "");
 141 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 142            &hammer2_flush_pipe, 0, "");
 143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
 144            &hammer2_bulkfree_tps, 0, "");
 145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 146            &hammer2_chain_allocs, 0, "");
 147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_saved_chains, CTLFLAG_RW,
 148            &hammer2_limit_saved_chains, 0, "");
 149 SYSCTL_INT(_vfs_hammer2, OID_AUTO, limit_saved_depth, CTLFLAG_RW,
 150            &hammer2_limit_saved_depth, 0, "");
 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 152            &hammer2_limit_dirty_chains, 0, "");
 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
 154            &hammer2_limit_dirty_inodes, 0, "");
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 156            &hammer2_count_modified_chains, 0, "");
 157 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 158            &hammer2_dio_count, 0, "");
 159 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
 160            &hammer2_dio_limit, 0, "");
 161
 162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 163            &hammer2_iod_file_read, 0, "");
 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 165            &hammer2_iod_meta_read, 0, "");
 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 167            &hammer2_iod_indr_read, 0, "");
 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 169            &hammer2_iod_fmap_read, 0, "");
 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 171            &hammer2_iod_volu_read, 0, "");
 172
 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 174            &hammer2_iod_file_write, 0, "");
 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 176            &hammer2_iod_file_wembed, 0, "");
 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 178            &hammer2_iod_file_wzero, 0, "");
 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 180            &hammer2_iod_file_wdedup, 0, "");
 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 182            &hammer2_iod_meta_write, 0, "");
 183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 184            &hammer2_iod_indr_write, 0, "");
 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 186            &hammer2_iod_fmap_write, 0, "");
 187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 188            &hammer2_iod_volu_write, 0, "");
 189 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW,
 190            &hammer2_iod_inode_creates, 0, "");
 191 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW,
 192            &hammer2_iod_inode_deletes, 0, "");
 193
 194 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
 195            &hammer2_process_icrc32, 0, "");
 196 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
 197            &hammer2_process_xxhash64, 0, "");
 198
 199 /*
 200 static int hammer2_vfs_init(struct vfsconf *conf);
 201 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 202 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 203                                 struct ucred *cred);
 204 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 205                                 struct ucred *);
 206 */
 207 static int hammer2_recovery(hammer2_dev_t *hmp);
 208 /*
 209 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 210 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 211 */
 212 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 213                                 struct ucred *cred);
 214 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 215                                 struct ucred *cred);
 216 /*
 217 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 218                                 struct fid *fhp, struct vnode **vpp);
 219 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 220 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 221                                 int *exflagsp, struct ucred **credanonp);
 222 static int hammer2_vfs_modifying(struct mount *mp);
 223 */
 224
 225 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 226
 227 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 228 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 229                                 hammer2_dev_t *hmp);
 230 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
 231
 232 /*
 233  * HAMMER2 vfs operations.
 234  */
 235 /*
 236 static struct vfsops hammer2_vfsops = {
 237         .vfs_flags      = 0,
 238         .vfs_init       = hammer2_vfs_init,
 239         .vfs_uninit     = hammer2_vfs_uninit,
 240         .vfs_sync       = hammer2_vfs_sync,
 241         .vfs_mount      = hammer2_vfs_mount,
 242         .vfs_unmount    = hammer2_vfs_unmount,
 243         .vfs_root       = hammer2_vfs_root,
 244         .vfs_statfs     = hammer2_vfs_statfs,
 245         .vfs_statvfs    = hammer2_vfs_statvfs,
 246         .vfs_vget       = hammer2_vfs_vget,
 247         .vfs_vptofh     = hammer2_vfs_vptofh,
 248         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 249         .vfs_checkexp   = hammer2_vfs_checkexp,
 250         .vfs_modifying  = hammer2_vfs_modifying
 251 };
 252 */
 253
 254 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 255
 256 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 257 MODULE_VERSION(hammer2, 1);
 258
 259 int
 260 hammer2_vfs_init(void)
 261 {
 262         /*
 263         static struct objcache_malloc_args margs_read;
 264         static struct objcache_malloc_args margs_write;
 265         static struct objcache_malloc_args margs_vop;
 266         */
 267
 268         int error;
 269         int mod;
 270
 271         error = 0;
 272         kmalloc_raise_limit(M_HAMMER2, 0);      /* unlimited */
 273
 274         /*
 275          * hammer2_xop_nthreads must be a multiple of ncpus,
 276          * minimum 2 * ncpus.
 277          */
 278         const int ncpus = 1;
 279         mod = ncpus;
 280         hammer2_xop_mod = mod;
 281         hammer2_xop_nthreads = mod * 2;
 282         /*
 283         while (hammer2_xop_nthreads / mod < HAMMER2_XOPGROUPS_MIN ||
 284                hammer2_xop_nthreads < HAMMER2_XOPTHREADS_MIN)
 285         {
 286                 hammer2_xop_nthreads += mod;
 287         }
 288         hammer2_xop_sgroups = hammer2_xop_nthreads / mod / 2;
 289         hammer2_xop_xgroups = hammer2_xop_nthreads / mod - hammer2_xop_sgroups;
 290         hammer2_xop_xbase = hammer2_xop_sgroups * mod;
 291         */
 292
 293         /*
 294          * A large DIO cache is needed to retain dedup enablement masks.
 295          * The bulkfree code clears related masks as part of the disk block
 296          * recycling algorithm, preventing it from being used for a later
 297          * dedup.
 298          *
 299          * NOTE: A large buffer cache can actually interfere with dedup
 300          *       operation because we dedup based on media physical buffers
 301          *       and not logical buffers.  Try to make the DIO case large
 302          *       enough to avoid this problem, but also cap it.
 303          */
 304         const long nbuf = 100000; /* XXX */
 305         hammer2_dio_limit = nbuf * 2;
 306         if (hammer2_dio_limit > 100000)
 307                 hammer2_dio_limit = 100000;
 308
 309         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 310                 error = EINVAL;
 311         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 312                 error = EINVAL;
 313         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 314                 error = EINVAL;
 315
 316         if (error) {
 317                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 318                 return (error);
 319         }
 320
 321 #if 0
 322         margs_read.objsize = 65536;
 323         margs_read.mtype = M_HAMMER2_DEBUFFER;
 324
 325         margs_write.objsize = 32768;
 326         margs_write.mtype = M_HAMMER2_CBUFFER;
 327
 328         margs_vop.objsize = sizeof(hammer2_xop_t);
 329         margs_vop.mtype = M_HAMMER2;
 330
 331         /*
 332          * Note thaht for the XOPS cache we want backing store allocations
 333          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 334          * confusion), so use the backing store function that does it.  This
 335          * means that initial XOPS objects are zerod but REUSED objects are
 336          * not.  So we are responsible for cleaning the object up sufficiently
 337          * for our needs before objcache_put()ing it back (typically just the
 338          * FIFO indices).
 339          */
 340         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 341                                 0, 1, NULL, NULL, NULL,
 342                                 objcache_malloc_alloc,
 343                                 objcache_malloc_free,
 344                                 &margs_read);
 345         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 346                                 0, 1, NULL, NULL, NULL,
 347                                 objcache_malloc_alloc,
 348                                 objcache_malloc_free,
 349                                 &margs_write);
 350         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 351                                 0, 1, NULL, NULL, NULL,
 352                                 objcache_malloc_alloc_zero,
 353                                 objcache_malloc_free,
 354                                 &margs_vop);
 355 #endif
 356
 357
 358         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 359         TAILQ_INIT(&hammer2_mntlist);
 360         TAILQ_INIT(&hammer2_pfslist);
 361         TAILQ_INIT(&hammer2_spmplist);
 362
 363         const int maxvnodes = 100000; /* XXX */
 364         hammer2_limit_dirty_chains = maxvnodes / 10;
 365         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 366                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 367         if (hammer2_limit_dirty_chains < 1000)
 368                 hammer2_limit_dirty_chains = 1000;
 369
 370         hammer2_limit_dirty_inodes = maxvnodes / 25;
 371         if (hammer2_limit_dirty_inodes < 100)
 372                 hammer2_limit_dirty_inodes = 100;
 373         if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
 374                 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
 375
 376         hammer2_limit_saved_chains = hammer2_limit_dirty_chains * 5;
 377
 378         return (error);
 379 }
 380
 381 int
 382 hammer2_vfs_uninit(void)
 383 {
 384         /*
 385         objcache_destroy(cache_buffer_read);
 386         objcache_destroy(cache_buffer_write);
 387         objcache_destroy(cache_xops);
 388         */
 389         return 0;
 390 }
 391
 392 /*
 393  * Core PFS allocator.  Used to allocate or reference the pmp structure
 394  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 395  * The pmp can be passed in or loaded by this function using the chain and
 396  * inode data.
 397  *
 398  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 399  * transactions.  Note that synchronization does not use this field.
 400  * (typically frontend operations and synchronization cannot run on the
 401  * same PFS node at the same time).
 402  *
 403  * XXX check locking
 404  */
 405 hammer2_pfs_t *
 406 hammer2_pfsalloc(hammer2_chain_t *chain,
 407                  const hammer2_inode_data_t *ripdata,
 408                  hammer2_dev_t *force_local)
 409 {
 410         hammer2_pfs_t *pmp;
 411         hammer2_inode_t *iroot;
 412         int count;
 413         int i;
 414         int j;
 415
 416         pmp = NULL;
 417
 418         /*
 419          * Locate or create the PFS based on the cluster id.  If ripdata
 420          * is NULL this is a spmp which is unique and is always allocated.
 421          *
 422          * If the device is mounted in local mode all PFSs are considered
 423          * independent and not part of any cluster (for debugging only).
 424          */
 425         if (ripdata) {
 426                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 427                         if (force_local != pmp->force_local)
 428                                 continue;
 429                         if (force_local == NULL &&
 430                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 431                                  sizeof(pmp->pfs_clid)) == 0) {
 432                                 break;
 433                         } else if (force_local && pmp->pfs_names[0] &&
 434                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 435                                 break;
 436                         }
 437                 }
 438         }
 439
 440         if (pmp == NULL) {
 441                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 442                 pmp->force_local = force_local;
 443                 hammer2_trans_manage_init(pmp);
 444                 kmalloc_create_obj(&pmp->minode, "HAMMER2-inodes",
 445                                    sizeof(struct hammer2_inode));
 446                 lockinit(&pmp->lock, "pfslk", 0, 0);
 447                 hammer2_spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 448                 hammer2_spin_init(&pmp->xop_spin, "h2xop");
 449                 hammer2_spin_init(&pmp->lru_spin, "h2lru");
 450                 RB_INIT(&pmp->inum_tree);
 451                 TAILQ_INIT(&pmp->syncq);
 452                 TAILQ_INIT(&pmp->depq);
 453                 TAILQ_INIT(&pmp->lru_list);
 454                 hammer2_spin_init(&pmp->list_spin, "h2pfsalloc_list");
 455
 456                 /*
 457                  * Save the last media transaction id for the flusher.  Set
 458                  * initial
 459                  */
 460                 if (ripdata) {
 461                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 462                         TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 463                 } else {
 464                         pmp->flags |= HAMMER2_PMPF_SPMP;
 465                         TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
 466                 }
 467
 468                 /*
 469                  * The synchronization thread may start too early, make
 470                  * sure it stays frozen until we are ready to let it go.
 471                  * XXX
 472                  */
 473                 /*
 474                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 475                                          HAMMER2_THREAD_REMASTER;
 476                 */
 477         }
 478
 479         /*
 480          * Create the PFS's root inode and any missing XOP helper threads.
 481          */
 482         if ((iroot = pmp->iroot) == NULL) {
 483                 iroot = hammer2_inode_get(pmp, NULL, 1, -1);
 484                 if (ripdata)
 485                         iroot->meta = ripdata->meta;
 486                 pmp->iroot = iroot;
 487                 hammer2_inode_ref(iroot);
 488                 hammer2_inode_unlock(iroot);
 489         }
 490
 491         /*
 492          * Stop here if no chain is passed in.
 493          */
 494         if (chain == NULL)
 495                 goto done;
 496
 497         /*
 498          * When a chain is passed in we must add it to the PFS's root
 499          * inode, update pmp->pfs_types[], and update the syncronization
 500          * threads.
 501          *
 502          * When forcing local mode, mark the PFS as a MASTER regardless.
 503          *
 504          * At the moment empty spots can develop due to removals or failures.
 505          * Ultimately we want to re-fill these spots but doing so might
 506          * confused running code. XXX
 507          */
 508         hammer2_inode_ref(iroot);
 509         hammer2_mtx_ex(&iroot->lock);
 510         j = iroot->cluster.nchains;
 511
 512         if (j == HAMMER2_MAXCLUSTER) {
 513                 kprintf("hammer2_pfsalloc: cluster full!\n");
 514                 /* XXX fatal error? */
 515         } else {
 516                 KKASSERT(chain->pmp == NULL);
 517                 chain->pmp = pmp;
 518                 hammer2_chain_ref(chain);
 519                 iroot->cluster.array[j].chain = chain;
 520                 if (force_local)
 521                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 522                 else
 523                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 524                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 525                 pmp->pfs_hmps[j] = chain->hmp;
 526                 hammer2_spin_ex(&pmp->inum_spin);
 527                 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset;
 528                 hammer2_spin_unex(&pmp->inum_spin);
 529
 530                 /*
 531                  * If the PFS is already mounted we must account
 532                  * for the mount_count here.
 533                  */
 534                 if (pmp->mp)
 535                         ++chain->hmp->mount_count;
 536
 537                 /*
 538                  * May have to fixup dirty chain tracking.  Previous
 539                  * pmp was NULL so nothing to undo.
 540                  */
 541                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 542                         hammer2_pfs_memory_inc(pmp);
 543                 ++j;
 544         }
 545         iroot->cluster.nchains = j;
 546
 547         /*
 548          * Update nmasters from any PFS inode which is part of the cluster.
 549          * It is possible that this will result in a value which is too
 550          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 551          * override this value later on.
 552          *
 553          * (This informs us of masters that might not currently be
 554          *  discoverable by this mount).
 555          */
 556         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 557                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 558         }
 559
 560         /*
 561          * Count visible masters.  Masters are usually added with
 562          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 563          * are more (XXX and must update the master inodes).
 564          */
 565         count = 0;
 566         for (i = 0; i < iroot->cluster.nchains; ++i) {
 567                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 568                         ++count;
 569         }
 570         if (pmp->pfs_nmasters < count)
 571                 pmp->pfs_nmasters = count;
 572
 573         /*
 574          * Create missing synchronization and support threads.
 575          *
 576          * Single-node masters (including snapshots) have nothing to
 577          * synchronize and do not require this thread.
 578          *
 579          * Multi-node masters or any number of soft masters, slaves, copy,
 580          * or other PFS types need the thread.
 581          *
 582          * Each thread is responsible for its particular cluster index.
 583          * We use independent threads so stalls or mismatches related to
 584          * any given target do not affect other targets.
 585          */
 586         for (i = 0; i < iroot->cluster.nchains; ++i) {
 587                 /*
 588                  * Single-node masters (including snapshots) have nothing
 589                  * to synchronize and will make direct xops support calls,
 590                  * thus they do not require this thread.
 591                  *
 592                  * Note that there can be thousands of snapshots.  We do not
 593                  * want to create thousands of threads.
 594                  */
 595                 if (pmp->pfs_nmasters <= 1 &&
 596                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 597                         continue;
 598                 }
 599
 600                 /*
 601                  * Sync support thread
 602                  */
 603                 /*
 604                 if (pmp->sync_thrs[i].td == NULL) {
 605                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 606                                            "h2nod", i, -1,
 607                                            hammer2_primary_sync_thread);
 608                 }
 609                 */
 610         }
 611
 612         /*
 613          * Create missing Xop threads
 614          *
 615          * NOTE: We create helper threads for all mounted PFSs or any
 616          *       PFSs with 2+ nodes (so the sync thread can update them,
 617          *       even if not mounted).
 618          */
 619         if (pmp->mp || iroot->cluster.nchains >= 2)
 620                 hammer2_xop_helper_create(pmp);
 621
 622         hammer2_mtx_unlock(&iroot->lock);
 623         hammer2_inode_drop(iroot);
 624 done:
 625         return pmp;
 626 }
 627
 628 /*
 629  * Deallocate an element of a probed PFS.  If destroying and this is a
 630  * MASTER, adjust nmasters.
 631  *
 632  * This function does not physically destroy the PFS element in its device
 633  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 634  */
 635 void
 636 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 637 {
 638         hammer2_inode_t *iroot;
 639         hammer2_chain_t *chain;
 640         int j;
 641
 642         /*
 643          * Cleanup our reference on iroot.  iroot is (should) not be needed
 644          * by the flush code.
 645          */
 646         iroot = pmp->iroot;
 647         if (iroot) {
 648                 /*
 649                  * Stop synchronizing
 650                  *
 651                  * XXX flush after acquiring the iroot lock.
 652                  * XXX clean out the cluster index from all inode structures.
 653                  */
 654                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 655
 656                 /*
 657                  * Remove the cluster index from the group.  If destroying
 658                  * the PFS and this is a master, adjust pfs_nmasters.
 659                  */
 660                 hammer2_mtx_ex(&iroot->lock);
 661                 chain = iroot->cluster.array[clindex].chain;
 662                 iroot->cluster.array[clindex].chain = NULL;
 663
 664                 switch(pmp->pfs_types[clindex]) {
 665                 case HAMMER2_PFSTYPE_MASTER:
 666                         if (destroying && pmp->pfs_nmasters > 0)
 667                                 --pmp->pfs_nmasters;
 668                         /* XXX adjust ripdata->meta.pfs_nmasters */
 669                         break;
 670                 default:
 671                         break;
 672                 }
 673                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 674
 675                 hammer2_mtx_unlock(&iroot->lock);
 676
 677                 /*
 678                  * Release the chain.
 679                  */
 680                 if (chain) {
 681                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 682                         hammer2_chain_drop(chain);
 683                 }
 684
 685                 /*
 686                  * Terminate all XOP threads for the cluster index.
 687                  */
 688                 if (pmp->xop_groups) {
 689                         for (j = 0; j < hammer2_xop_nthreads; ++j) {
 690                                 hammer2_thr_delete(
 691                                         &pmp->xop_groups[j].thrs[clindex]);
 692                         }
 693                 }
 694         }
 695 }
 696
 697 /*
 698  * Destroy a PFS, typically only occurs after the last mount on a device
 699  * has gone away.
 700  */
 701 static void
 702 hammer2_pfsfree(hammer2_pfs_t *pmp)
 703 {
 704         hammer2_inode_t *iroot;
 705         hammer2_chain_t *chain;
 706         int chains_still_present = 0;
 707         int i;
 708         //int j;
 709
 710         /*
 711          * Cleanup our reference on iroot.  iroot is (should) not be needed
 712          * by the flush code.
 713          */
 714         if (pmp->flags & HAMMER2_PMPF_SPMP)
 715                 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
 716         else
 717                 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 718
 719         /*
 720          * Cleanup chains remaining on LRU list.
 721          */
 722         hammer2_spin_ex(&pmp->lru_spin);
 723         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 724                 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
 725                 atomic_add_int(&pmp->lru_count, -1);
 726                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
 727                 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
 728                 hammer2_chain_ref(chain);
 729                 hammer2_spin_unex(&pmp->lru_spin);
 730                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 731                 hammer2_chain_drop(chain);
 732                 hammer2_spin_ex(&pmp->lru_spin);
 733         }
 734         hammer2_spin_unex(&pmp->lru_spin);
 735
 736         /*
 737          * Clean up iroot
 738          */
 739         iroot = pmp->iroot;
 740         if (iroot) {
 741                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 742                         /*
 743                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 744                         if (pmp->xop_groups) {
 745                                 for (j = 0; j < hammer2_xop_nthreads; ++j)
 746                                         hammer2_thr_delete(
 747                                                 &pmp->xop_groups[j].thrs[i]);
 748                         }
 749                         */
 750                         chain = iroot->cluster.array[i].chain;
 751                         if (chain && !RB_EMPTY(&chain->core.rbtree)) {
 752                                 kprintf("hammer2: Warning pmp %p still "
 753                                         "has active chains\n", pmp);
 754                                 chains_still_present = 1;
 755                         }
 756                 }
 757                 KASSERT(iroot->refs == 1,
 758                         ("PMP->IROOT %p REFS WRONG %d", iroot, iroot->refs));
 759
 760                 /* ref for iroot */
 761                 hammer2_inode_drop(iroot);
 762                 pmp->iroot = NULL;
 763         }
 764
 765         /*
 766          * Free remaining pmp resources
 767          */
 768         if (chains_still_present) {
 769                 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp);
 770         } else {
 771                 /*
 772                  * In makefs HAMMER2, all inodes must be gone at this point.
 773                  * XXX vnode_count may not be 0 at this point.
 774                  */
 775                 assert(hammer2_pfs_inode_count(pmp) == 0);
 776
 777                 kmalloc_destroy_obj(&pmp->minode);
 778                 kfree(pmp, M_HAMMER2);
 779         }
 780 }
 781
 782 /*
 783  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 784  * empty is terminated and freed.
 785  *
 786  * XXX inefficient.
 787  */
 788 static void
 789 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
 790 {
 791         hammer2_pfs_t *pmp;
 792         hammer2_inode_t *iroot;
 793         hammer2_chain_t *rchain;
 794         int i;
 795         //int j;
 796         struct hammer2_pfslist *wlist;
 797
 798         if (which == 0)
 799                 wlist = &hammer2_pfslist;
 800         else
 801                 wlist = &hammer2_spmplist;
 802 again:
 803         TAILQ_FOREACH(pmp, wlist, mntentry) {
 804                 if ((iroot = pmp->iroot) == NULL)
 805                         continue;
 806
 807                 /*
 808                  * Determine if this PFS is affected.  If it is we must
 809                  * freeze all management threads and lock its iroot.
 810                  *
 811                  * Freezing a management thread forces it idle, operations
 812                  * in-progress will be aborted and it will have to start
 813                  * over again when unfrozen, or exit if told to exit.
 814                  */
 815                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 816                         if (pmp->pfs_hmps[i] == hmp)
 817                                 break;
 818                 }
 819                 if (i == HAMMER2_MAXCLUSTER)
 820                         continue;
 821
 822                 hammer2_vfs_sync_pmp(pmp, MNT_WAIT);
 823
 824                 /*
 825                  * Make sure all synchronization threads are locked
 826                  * down.
 827                  */
 828                 /*
 829                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 830                         if (pmp->pfs_hmps[i] == NULL)
 831                                 continue;
 832                         hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 833                         if (pmp->xop_groups) {
 834                                 for (j = 0; j < hammer2_xop_nthreads; ++j) {
 835                                         hammer2_thr_freeze_async(
 836                                                 &pmp->xop_groups[j].thrs[i]);
 837                                 }
 838                         }
 839                 }
 840                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 841                         if (pmp->pfs_hmps[i] == NULL)
 842                                 continue;
 843                         hammer2_thr_freeze(&pmp->sync_thrs[i]);
 844                         if (pmp->xop_groups) {
 845                                 for (j = 0; j < hammer2_xop_nthreads; ++j) {
 846                                         hammer2_thr_freeze(
 847                                                 &pmp->xop_groups[j].thrs[i]);
 848                                 }
 849                         }
 850                 }
 851                 */
 852
 853                 /*
 854                  * Lock the inode and clean out matching chains.
 855                  * Note that we cannot use hammer2_inode_lock_*()
 856                  * here because that would attempt to validate the
 857                  * cluster that we are in the middle of ripping
 858                  * apart.
 859                  *
 860                  * WARNING! We are working directly on the inodes
 861                  *          embedded cluster.
 862                  */
 863                 hammer2_mtx_ex(&iroot->lock);
 864
 865                 /*
 866                  * Remove the chain from matching elements of the PFS.
 867                  */
 868                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 869                         if (pmp->pfs_hmps[i] != hmp)
 870                                 continue;
 871                         /*
 872                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 873                         if (pmp->xop_groups) {
 874                                 for (j = 0; j < hammer2_xop_nthreads; ++j) {
 875                                         hammer2_thr_delete(
 876                                                 &pmp->xop_groups[j].thrs[i]);
 877                                 }
 878                         }
 879                         */
 880                         rchain = iroot->cluster.array[i].chain;
 881                         iroot->cluster.array[i].chain = NULL;
 882                         pmp->pfs_types[i] = 0;
 883                         if (pmp->pfs_names[i]) {
 884                                 kfree(pmp->pfs_names[i], M_HAMMER2);
 885                                 pmp->pfs_names[i] = NULL;
 886                         }
 887                         if (rchain) {
 888                                 hammer2_chain_drop(rchain);
 889                                 /* focus hint */
 890                                 if (iroot->cluster.focus == rchain)
 891                                         iroot->cluster.focus = NULL;
 892                         }
 893                         pmp->pfs_hmps[i] = NULL;
 894                 }
 895                 hammer2_mtx_unlock(&iroot->lock);
 896
 897                 /*
 898                  * Cleanup trailing chains.  Gaps may remain.
 899                  */
 900                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 901                         if (pmp->pfs_hmps[i])
 902                                 break;
 903                 }
 904                 iroot->cluster.nchains = i + 1;
 905
 906                 /*
 907                  * If the PMP has no elements remaining we can destroy it.
 908                  * (this will transition management threads from frozen->exit).
 909                  */
 910                 if (iroot->cluster.nchains == 0) {
 911                         /*
 912                          * If this was the hmp's spmp, we need to clean
 913                          * a little more stuff out.
 914                          */
 915                         if (hmp->spmp == pmp) {
 916                                 hmp->spmp = NULL;
 917                                 hmp->vchain.pmp = NULL;
 918                                 hmp->fchain.pmp = NULL;
 919                         }
 920
 921                         /*
 922                          * Free the pmp and restart the loop
 923                          */
 924                         KKASSERT(TAILQ_EMPTY(&pmp->syncq));
 925                         KKASSERT(TAILQ_EMPTY(&pmp->depq));
 926                         hammer2_pfsfree(pmp);
 927                         goto again;
 928                 }
 929
 930                 /*
 931                  * If elements still remain we need to set the REMASTER
 932                  * flag and unfreeze it.
 933                  */
 934                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 935                         if (pmp->pfs_hmps[i] == NULL)
 936                                 continue;
 937                         /*
 938                         hammer2_thr_remaster(&pmp->sync_thrs[i]);
 939                         hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 940                         if (pmp->xop_groups) {
 941                                 for (j = 0; j < hammer2_xop_nthreads; ++j) {
 942                                         hammer2_thr_remaster(
 943                                                 &pmp->xop_groups[j].thrs[i]);
 944                                         hammer2_thr_unfreeze(
 945                                                 &pmp->xop_groups[j].thrs[i]);
 946                                 }
 947                         }
 948                         */
 949                 }
 950         }
 951 }
 952
 953 /*
 954  * Mount or remount HAMMER2 fileystem from physical media
 955  *
 956  *      mountroot
 957  *              mp              mount point structure
 958  *              path            NULL
 959  *              data            <unused>
 960  *              cred            <unused>
 961  *
 962  *      mount
 963  *              mp              mount point structure
 964  *              path            path to mount point
 965  *              data            pointer to argument structure in user space
 966  *                      volume  volume path (device@LABEL form)
 967  *                      hflags  user mount flags
 968  *              cred            user credentials
 969  *
 970  * RETURNS:     0       Success
 971  *              !0      error number
 972  */
 973 int
 974 hammer2_vfs_mount(struct vnode *makefs_devvp, struct mount *mp,
 975                   const char *label, const struct hammer2_mount_info *mi)
 976 {
 977         struct hammer2_mount_info info = *mi;
 978         hammer2_pfs_t *pmp;
 979         hammer2_pfs_t *spmp;
 980         hammer2_dev_t *hmp, *hmp_tmp;
 981         hammer2_dev_t *force_local;
 982         hammer2_key_t key_next;
 983         hammer2_key_t key_dummy;
 984         hammer2_key_t lhc;
 985         hammer2_chain_t *parent;
 986         hammer2_chain_t *chain;
 987         const hammer2_inode_data_t *ripdata;
 988         hammer2_devvp_list_t devvpl;
 989         hammer2_devvp_t *e, *e_tmp;
 990         char *devstr;
 991         int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
 992         int error;
 993         int i;
 994
 995         hmp = NULL;
 996         pmp = NULL;
 997         devstr = NULL;
 998
 999         kprintf("hammer2_mount: device=\"%s\" label=\"%s\" rdonly=%d\n",
1000                 devstr, label, ronly);
1001
1002         /*
1003          * Initialize all device vnodes.
1004          */
1005         TAILQ_INIT(&devvpl);
1006         error = hammer2_init_devvp(makefs_devvp, &devvpl);
1007         if (error) {
1008                 kprintf("hammer2: failed to initialize devvp in %s\n", devstr);
1009                 hammer2_cleanup_devvp(&devvpl);
1010                 return error;
1011         }
1012
1013         /*
1014          * Determine if the device has already been mounted.  After this
1015          * check hmp will be non-NULL if we are doing the second or more
1016          * hammer2 mounts from the same device.
1017          */
1018         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1019         if (!TAILQ_EMPTY(&devvpl)) {
1020                 /*
1021                  * Match the device.  Due to the way devfs works,
1022                  * we may not be able to directly match the vnode pointer,
1023                  * so also check to see if the underlying device matches.
1024                  */
1025                 TAILQ_FOREACH(hmp_tmp, &hammer2_mntlist, mntentry) {
1026                         TAILQ_FOREACH(e_tmp, &hmp_tmp->devvpl, entry) {
1027                                 int devvp_found = 0;
1028                                 TAILQ_FOREACH(e, &devvpl, entry) {
1029                                         KKASSERT(e->devvp);
1030                                         if (e_tmp->devvp == e->devvp)
1031                                                 devvp_found = 1;
1032                                         /*
1033                                         if (e_tmp->devvp->v_rdev &&
1034                                             e_tmp->devvp->v_rdev == e->devvp->v_rdev)
1035                                                 devvp_found = 1;
1036                                         */
1037                                 }
1038                                 if (!devvp_found)
1039                                         goto next_hmp;
1040                         }
1041                         hmp = hmp_tmp;
1042                         kprintf("hammer2_mount: hmp=%p matched\n", hmp);
1043                         break;
1044 next_hmp:
1045                         continue;
1046                 }
1047
1048                 /*
1049                  * If no match this may be a fresh H2 mount, make sure
1050                  * the device is not mounted on anything else.
1051                  */
1052                 if (hmp == NULL) {
1053                         TAILQ_FOREACH(e, &devvpl, entry) {
1054                                 struct vnode *devvp = e->devvp;
1055                                 KKASSERT(devvp);
1056                                 error = vfs_mountedon(devvp);
1057                                 if (error) {
1058                                         kprintf("hammer2_mount: %s mounted %d\n",
1059                                                 e->path, error);
1060                                         hammer2_cleanup_devvp(&devvpl);
1061                                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1062                                         return error;
1063                                 }
1064                         }
1065                 }
1066         } else {
1067                 /*
1068                  * Match the label to a pmp already probed.
1069                  */
1070                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1071                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1072                                 if (pmp->pfs_names[i] &&
1073                                     strcmp(pmp->pfs_names[i], label) == 0) {
1074                                         hmp = pmp->pfs_hmps[i];
1075                                         break;
1076                                 }
1077                         }
1078                         if (hmp)
1079                                 break;
1080                 }
1081                 if (hmp == NULL) {
1082                         kprintf("hammer2_mount: PFS label \"%s\" not found\n",
1083                                 label);
1084                         hammer2_cleanup_devvp(&devvpl);
1085                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1086                         return ENOENT;
1087                 }
1088         }
1089
1090         /*
1091          * Open the device if this isn't a secondary mount and construct
1092          * the H2 device mount (hmp).
1093          */
1094         if (hmp == NULL) {
1095                 hammer2_chain_t *schain;
1096                 hammer2_xop_head_t xop;
1097
1098                 /*
1099                  * Now open the device
1100                  */
1101                 KKASSERT(!TAILQ_EMPTY(&devvpl));
1102                 if (error == 0) {
1103                         error = hammer2_open_devvp(&devvpl, ronly);
1104                         if (error) {
1105                                 hammer2_close_devvp(&devvpl, ronly);
1106                                 hammer2_cleanup_devvp(&devvpl);
1107                                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1108                                 return error;
1109                         }
1110                 }
1111
1112                 /*
1113                  * Construct volumes and link with device vnodes.
1114                  */
1115                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1116                 hmp->devvp = NULL;
1117                 error = hammer2_init_vfsvolumes(mp, &devvpl, hmp->volumes,
1118                                              &hmp->voldata, &hmp->volhdrno,
1119                                              &hmp->devvp);
1120                 if (error) {
1121                         hammer2_close_devvp(&devvpl, ronly);
1122                         hammer2_cleanup_devvp(&devvpl);
1123                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1124                         kfree(hmp, M_HAMMER2);
1125                         return error;
1126                 }
1127                 if (!hmp->devvp) {
1128                         kprintf("hammer2: failed to initialize root volume\n");
1129                         hammer2_unmount_helper(mp, NULL, hmp);
1130                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1131                         hammer2_vfs_unmount(mp, MNT_FORCE);
1132                         return EINVAL;
1133                 }
1134
1135                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", devstr);
1136                 hmp->ronly = ronly;
1137                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1138                 kmalloc_create_obj(&hmp->mchain, "HAMMER2-chains",
1139                                    sizeof(struct hammer2_chain));
1140                 kmalloc_create_obj(&hmp->mio, "HAMMER2-dio",
1141                                    sizeof(struct hammer2_io));
1142                 kmalloc_create(&hmp->mmsg, "HAMMER2-msg");
1143                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1144                 RB_INIT(&hmp->iotree);
1145                 hammer2_spin_init(&hmp->io_spin, "h2mount_io");
1146                 hammer2_spin_init(&hmp->list_spin, "h2mount_list");
1147
1148                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1149                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1150                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1151
1152                 /*
1153                  * vchain setup. vchain.data is embedded.
1154                  * vchain.refs is initialized and will never drop to 0.
1155                  *
1156                  * NOTE! voldata is not yet loaded.
1157                  */
1158                 hmp->vchain.hmp = hmp;
1159                 hmp->vchain.refs = 1;
1160                 hmp->vchain.data = (void *)&hmp->voldata;
1161                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1162                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1163                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1164                 hammer2_chain_core_init(&hmp->vchain);
1165
1166                 /*
1167                  * fchain setup.  fchain.data is embedded.
1168                  * fchain.refs is initialized and will never drop to 0.
1169                  *
1170                  * The data is not used but needs to be initialized to
1171                  * pass assertion muster.  We use this chain primarily
1172                  * as a placeholder for the freemap's top-level radix tree
1173                  * so it does not interfere with the volume's topology
1174                  * radix tree.
1175                  */
1176                 hmp->fchain.hmp = hmp;
1177                 hmp->fchain.refs = 1;
1178                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1179                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1180                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1181                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1182                 hmp->fchain.bref.methods =
1183                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1184                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1185                 hammer2_chain_core_init(&hmp->fchain);
1186
1187                 /*
1188                  * Initialize volume header related fields.
1189                  */
1190                 KKASSERT(hmp->voldata.magic == HAMMER2_VOLUME_ID_HBO ||
1191                          hmp->voldata.magic == HAMMER2_VOLUME_ID_ABO);
1192                 hmp->volsync = hmp->voldata;
1193                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
1194                 /*
1195                  * Must use hmp instead of volume header for these two
1196                  * in order to handle volume versions transparently.
1197                  */
1198                 if (hmp->voldata.version >= HAMMER2_VOL_VERSION_MULTI_VOLUMES) {
1199                         hmp->nvolumes = hmp->voldata.nvolumes;
1200                         hmp->total_size = hmp->voldata.total_size;
1201                 } else {
1202                         hmp->nvolumes = 1;
1203                         hmp->total_size = hmp->voldata.volu_size;
1204                 }
1205                 KKASSERT(hmp->nvolumes > 0);
1206
1207                 /*
1208                  * Move devvpl entries to hmp.
1209                  */
1210                 TAILQ_INIT(&hmp->devvpl);
1211                 while ((e = TAILQ_FIRST(&devvpl)) != NULL) {
1212                         TAILQ_REMOVE(&devvpl, e, entry);
1213                         TAILQ_INSERT_TAIL(&hmp->devvpl, e, entry);
1214                 }
1215                 KKASSERT(TAILQ_EMPTY(&devvpl));
1216                 KKASSERT(!TAILQ_EMPTY(&hmp->devvpl));
1217
1218                 /*
1219                  * Really important to get these right or the flush and
1220                  * teardown code will get confused.
1221                  */
1222                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, NULL);
1223                 spmp = hmp->spmp;
1224                 spmp->pfs_hmps[0] = hmp;
1225
1226                 /*
1227                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1228                  * is inherited from the volume header.
1229                  */
1230                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1231                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1232                 hmp->vchain.pmp = spmp;
1233                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1234                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1235                 hmp->fchain.pmp = spmp;
1236
1237                 /*
1238                  * First locate the super-root inode, which is key 0
1239                  * relative to the volume header's blockset.
1240                  *
1241                  * Then locate the root inode by scanning the directory keyspace
1242                  * represented by the label.
1243                  */
1244                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1245                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1246                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1247                                       &error, 0);
1248                 hammer2_chain_lookup_done(parent);
1249                 if (schain == NULL) {
1250                         kprintf("hammer2_mount: invalid super-root\n");
1251                         hammer2_unmount_helper(mp, NULL, hmp);
1252                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1253                         hammer2_vfs_unmount(mp, MNT_FORCE);
1254                         return EINVAL;
1255                 }
1256                 if (schain->error) {
1257                         kprintf("hammer2_mount: error %s reading super-root\n",
1258                                 hammer2_error_str(schain->error));
1259                         hammer2_chain_unlock(schain);
1260                         hammer2_chain_drop(schain);
1261                         schain = NULL;
1262                         hammer2_unmount_helper(mp, NULL, hmp);
1263                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1264                         hammer2_vfs_unmount(mp, MNT_FORCE);
1265                         return EINVAL;
1266                 }
1267
1268                 /*
1269                  * The super-root always uses an inode_tid of 1 when
1270                  * creating PFSs.
1271                  */
1272                 spmp->inode_tid = 1;
1273                 spmp->modify_tid = schain->bref.modify_tid + 1;
1274
1275                 /*
1276                  * Sanity-check schain's pmp and finish initialization.
1277                  * Any chain belonging to the super-root topology should
1278                  * have a NULL pmp (not even set to spmp).
1279                  */
1280                 ripdata = &schain->data->ipdata;
1281                 KKASSERT(schain->pmp == NULL);
1282                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1283
1284                 /*
1285                  * Replace the dummy spmp->iroot with a real one.  It's
1286                  * easier to just do a wholesale replacement than to try
1287                  * to update the chain and fixup the iroot fields.
1288                  *
1289                  * The returned inode is locked with the supplied cluster.
1290                  */
1291                 hammer2_dummy_xop_from_chain(&xop, schain);
1292                 hammer2_inode_drop(spmp->iroot);
1293                 spmp->iroot = NULL;
1294                 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1);
1295                 spmp->spmp_hmp = hmp;
1296                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1297                 spmp->pfs_hmps[0] = hmp;
1298                 hammer2_inode_ref(spmp->iroot);
1299                 hammer2_inode_unlock(spmp->iroot);
1300                 hammer2_cluster_unlock(&xop.cluster);
1301                 hammer2_chain_drop(schain);
1302                 /* do not call hammer2_cluster_drop() on an embedded cluster */
1303                 schain = NULL;  /* now invalid */
1304                 /* leave spmp->iroot with one ref */
1305
1306                 if (!hmp->ronly) {
1307                         error = hammer2_recovery(hmp);
1308                         if (error == 0)
1309                                 error |= hammer2_fixup_pfses(hmp);
1310                         /* XXX do something with error */
1311                 }
1312                 hammer2_update_pmps(hmp);
1313                 hammer2_iocom_init(hmp);
1314                 hammer2_bulkfree_init(hmp);
1315
1316                 /*
1317                  * Ref the cluster management messaging descriptor.  The mount
1318                  * program deals with the other end of the communications pipe.
1319                  *
1320                  * Root mounts typically do not supply one.
1321                  */
1322                 /*
1323                 if (info.cluster_fd >= 0) {
1324                         fp = holdfp(curthread, info.cluster_fd, -1);
1325                         if (fp) {
1326                                 hammer2_cluster_reconnect(hmp, fp);
1327                         } else {
1328                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1329                         }
1330                 }
1331                 */
1332         } else {
1333                 spmp = hmp->spmp;
1334                 if (info.hflags & HMNT2_DEVFLAGS) {
1335                         kprintf("hammer2_mount: Warning: mount flags pertaining "
1336                                 "to the whole device may only be specified "
1337                                 "on the first mount of the device: %08x\n",
1338                                 info.hflags & HMNT2_DEVFLAGS);
1339                 }
1340         }
1341
1342         /*
1343          * Force local mount (disassociate all PFSs from their clusters).
1344          * Used primarily for debugging.
1345          */
1346         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1347
1348         /*
1349          * Lookup the mount point under the media-localized super-root.
1350          * Scanning hammer2_pfslist doesn't help us because it represents
1351          * PFS cluster ids which can aggregate several named PFSs together.
1352          *
1353          * cluster->pmp will incorrectly point to spmp and must be fixed
1354          * up later on.
1355          */
1356         hammer2_inode_lock(spmp->iroot, 0);
1357         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1358         lhc = hammer2_dirhash(label, strlen(label));
1359         chain = hammer2_chain_lookup(&parent, &key_next,
1360                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1361                                      &error, 0);
1362         while (chain) {
1363                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1364                     strcmp(label, chain->data->ipdata.filename) == 0) {
1365                         break;
1366                 }
1367                 chain = hammer2_chain_next(&parent, chain, &key_next,
1368                                             key_next,
1369                                             lhc + HAMMER2_DIRHASH_LOMASK,
1370                                             &error, 0);
1371         }
1372         if (parent) {
1373                 hammer2_chain_unlock(parent);
1374                 hammer2_chain_drop(parent);
1375         }
1376         hammer2_inode_unlock(spmp->iroot);
1377
1378         /*
1379          * PFS could not be found?
1380          */
1381         if (chain == NULL) {
1382                 hammer2_unmount_helper(mp, NULL, hmp);
1383                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1384                 hammer2_vfs_unmount(mp, MNT_FORCE);
1385
1386                 if (error) {
1387                         kprintf("hammer2_mount: PFS label I/O error\n");
1388                         return EINVAL;
1389                 } else {
1390                         kprintf("hammer2_mount: PFS label \"%s\" not found\n",
1391                                 label);
1392                         return ENOENT;
1393                 }
1394         }
1395
1396         /*
1397          * Acquire the pmp structure (it should have already been allocated
1398          * via hammer2_update_pmps() so do not pass cluster in to add to
1399          * available chains).
1400          *
1401          * Check if the cluster has already been mounted.  A cluster can
1402          * only be mounted once, use null mounts to mount additional copies.
1403          */
1404         if (chain->error) {
1405                 kprintf("hammer2_mount: PFS label I/O error\n");
1406         } else {
1407                 ripdata = &chain->data->ipdata;
1408                 pmp = hammer2_pfsalloc(NULL, ripdata, force_local);
1409         }
1410         hammer2_chain_unlock(chain);
1411         hammer2_chain_drop(chain);
1412
1413         /*
1414          * Finish the mount
1415          */
1416         kprintf("hammer2_mount: hmp=%p pmp=%p\n", hmp, pmp);
1417
1418         if (pmp->mp) {
1419                 kprintf("hammer2_mount: PFS already mounted!\n");
1420                 hammer2_unmount_helper(mp, NULL, hmp);
1421                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1422                 hammer2_vfs_unmount(mp, MNT_FORCE);
1423
1424                 return EBUSY;
1425         }
1426
1427         pmp->hflags = info.hflags;
1428         mp->mnt_flag |= MNT_LOCAL;
1429         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1430         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1431
1432         /*
1433          * required mount structure initializations
1434          */
1435         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1436         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1437
1438         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1439         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1440
1441         /*
1442          * Optional fields
1443          */
1444         mp->mnt_iosize_max = MAXPHYS;
1445
1446         /*
1447          * Connect up mount pointers.
1448          */
1449         hammer2_mount_helper(mp, pmp);
1450         lockmgr(&hammer2_mntlk, LK_RELEASE);
1451
1452 #if 0
1453         /*
1454          * Finish setup
1455          */
1456         vfs_getnewfsid(mp);
1457         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1458         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1459         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1460
1461         if (path) {
1462                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1463                           MNAMELEN - 1, &size);
1464                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1465         } /* else root mount, already in there */
1466
1467         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1468         if (path) {
1469                 copyinstr(path, mp->mnt_stat.f_mntonname,
1470                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1471                           &size);
1472         } else {
1473                 /* root mount */
1474                 mp->mnt_stat.f_mntonname[0] = '/';
1475         }
1476 #endif
1477
1478         /*
1479          * Initial statfs to prime mnt_stat.
1480          */
1481         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1482         hammer2_vfs_statvfs(mp, &mp->mnt_vstat, NULL);
1483
1484         return 0;
1485 }
1486
1487 /*
1488  * Scan PFSs under the super-root and create hammer2_pfs structures.
1489  */
1490 static
1491 void
1492 hammer2_update_pmps(hammer2_dev_t *hmp)
1493 {
1494         const hammer2_inode_data_t *ripdata;
1495         hammer2_chain_t *parent;
1496         hammer2_chain_t *chain;
1497         hammer2_dev_t *force_local;
1498         hammer2_pfs_t *spmp;
1499         hammer2_key_t key_next;
1500         int error;
1501
1502         /*
1503          * Force local mount (disassociate all PFSs from their clusters).
1504          * Used primarily for debugging.
1505          */
1506         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1507
1508         /*
1509          * Lookup mount point under the media-localized super-root.
1510          *
1511          * cluster->pmp will incorrectly point to spmp and must be fixed
1512          * up later on.
1513          */
1514         spmp = hmp->spmp;
1515         hammer2_inode_lock(spmp->iroot, 0);
1516         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1517         chain = hammer2_chain_lookup(&parent, &key_next,
1518                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1519                                          &error, 0);
1520         while (chain) {
1521                 if (chain->error) {
1522                         kprintf("I/O error scanning PFS labels\n");
1523                 } else if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
1524                         kprintf("Non inode chain type %d under super-root\n",
1525                                 chain->bref.type);
1526                 } else {
1527                         ripdata = &chain->data->ipdata;
1528                         hammer2_pfsalloc(chain, ripdata, force_local);
1529                 }
1530                 chain = hammer2_chain_next(&parent, chain, &key_next,
1531                                            key_next, HAMMER2_KEY_MAX,
1532                                            &error, 0);
1533         }
1534         if (parent) {
1535                 hammer2_chain_unlock(parent);
1536                 hammer2_chain_drop(parent);
1537         }
1538         hammer2_inode_unlock(spmp->iroot);
1539 }
1540
1541 #if 0
1542 static
1543 int
1544 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1545                 struct ucred *cred)
1546 {
1547         hammer2_volume_t *vol;
1548         struct vnode *devvp;
1549         int i, error, result = 0;
1550
1551         if (!(hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)))
1552                 return 0;
1553
1554         for (i = 0; i < hmp->nvolumes; ++i) {
1555                 vol = &hmp->volumes[i];
1556                 devvp = vol->dev->devvp;
1557                 KKASSERT(devvp);
1558                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1559                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1560                 vn_unlock(devvp);
1561                 error = 0;
1562                 if (vol->id == HAMMER2_ROOT_VOLUME) {
1563                         error = hammer2_recovery(hmp);
1564                         if (error == 0)
1565                                 error |= hammer2_fixup_pfses(hmp);
1566                 }
1567                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1568                 if (error == 0) {
1569                         VOP_CLOSE(devvp, FREAD, NULL);
1570                 } else {
1571                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1572                 }
1573                 vn_unlock(devvp);
1574                 result |= error;
1575         }
1576         if (result == 0) {
1577                 kprintf("hammer2: enable read/write\n");
1578                 hmp->ronly = 0;
1579         }
1580
1581         return result;
1582 }
1583 #endif
1584
1585 int
1586 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1587 {
1588         hammer2_pfs_t *pmp;
1589         int flags;
1590         int error = 0;
1591
1592         pmp = MPTOPMP(mp);
1593
1594         if (pmp == NULL)
1595                 return(0);
1596
1597         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1598
1599         /*
1600          * If mount initialization proceeded far enough we must flush
1601          * its vnodes and sync the underlying mount points.  Three syncs
1602          * are required to fully flush the filesystem (freemap updates lag
1603          * by one flush, and one extra for safety).
1604          */
1605         if (mntflags & MNT_FORCE)
1606                 flags = FORCECLOSE;
1607         else
1608                 flags = 0;
1609         if (pmp->iroot) {
1610                 error = vflush(mp, 0, flags);
1611                 if (error)
1612                         goto failed;
1613                 hammer2_vfs_sync(mp, MNT_WAIT);
1614                 hammer2_vfs_sync(mp, MNT_WAIT);
1615                 hammer2_vfs_sync(mp, MNT_WAIT);
1616         }
1617
1618         /*
1619          * Cleanup the frontend support XOPS threads
1620          */
1621         hammer2_xop_helper_cleanup(pmp);
1622
1623         if (pmp->mp)
1624                 hammer2_unmount_helper(mp, pmp, NULL);
1625
1626         error = 0;
1627 failed:
1628         lockmgr(&hammer2_mntlk, LK_RELEASE);
1629
1630         return (error);
1631 }
1632
1633 /*
1634  * Mount helper, hook the system mount into our PFS.
1635  * The mount lock is held.
1636  *
1637  * We must bump the mount_count on related devices for any
1638  * mounted PFSs.
1639  */
1640 static
1641 void
1642 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1643 {
1644         hammer2_cluster_t *cluster;
1645         hammer2_chain_t *rchain;
1646         int i;
1647
1648         mp->mnt_data = (qaddr_t)pmp;
1649         pmp->mp = mp;
1650
1651         /*
1652          * After pmp->mp is set we have to adjust hmp->mount_count.
1653          */
1654         cluster = &pmp->iroot->cluster;
1655         for (i = 0; i < cluster->nchains; ++i) {
1656                 rchain = cluster->array[i].chain;
1657                 if (rchain == NULL)
1658                         continue;
1659                 ++rchain->hmp->mount_count;
1660         }
1661
1662         /*
1663          * Create missing Xop threads
1664          */
1665         hammer2_xop_helper_create(pmp);
1666 }
1667
1668 /*
1669  * Unmount helper, unhook the system mount from our PFS.
1670  * The mount lock is held.
1671  *
1672  * If hmp is supplied a mount responsible for being the first to open
1673  * the block device failed and the block device and all PFSs using the
1674  * block device must be cleaned up.
1675  *
1676  * If pmp is supplied multiple devices might be backing the PFS and each
1677  * must be disconnected.  This might not be the last PFS using some of the
1678  * underlying devices.  Also, we have to adjust our hmp->mount_count
1679  * accounting for the devices backing the pmp which is now undergoing an
1680  * unmount.
1681  */
1682 static
1683 void
1684 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1685 {
1686         hammer2_cluster_t *cluster;
1687         hammer2_chain_t *rchain;
1688         int dumpcnt;
1689         int i;
1690
1691         /*
1692          * If no device supplied this is a high-level unmount and we have to
1693          * to disconnect the mount, adjust mount_count, and locate devices
1694          * that might now have no mounts.
1695          */
1696         if (pmp) {
1697                 KKASSERT(hmp == NULL);
1698                 KKASSERT(MPTOPMP(mp) == pmp);
1699                 pmp->mp = NULL;
1700                 mp->mnt_data = NULL;
1701
1702                 /*
1703                  * After pmp->mp is cleared we have to account for
1704                  * mount_count.
1705                  */
1706                 cluster = &pmp->iroot->cluster;
1707                 for (i = 0; i < cluster->nchains; ++i) {
1708                         rchain = cluster->array[i].chain;
1709                         if (rchain == NULL)
1710                                 continue;
1711                         --rchain->hmp->mount_count;
1712                         /* scrapping hmp now may invalidate the pmp */
1713                 }
1714 again:
1715                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1716                         if (hmp->mount_count == 0) {
1717                                 hammer2_unmount_helper(NULL, NULL, hmp);
1718                                 goto again;
1719                         }
1720                 }
1721                 return;
1722         }
1723
1724         /*
1725          * Try to terminate the block device.  We can't terminate it if
1726          * there are still PFSs referencing it.
1727          */
1728         if (hmp->mount_count)
1729                 return;
1730
1731         /*
1732          * Decomission the network before we start messing with the
1733          * device and PFS.
1734          */
1735         hammer2_iocom_uninit(hmp);
1736
1737         hammer2_bulkfree_uninit(hmp);
1738         hammer2_pfsfree_scan(hmp, 0);
1739
1740         /*
1741          * Cycle the volume data lock as a safety (probably not needed any
1742          * more).  To ensure everything is out we need to flush at least
1743          * three times.  (1) The running of the sideq can dirty the
1744          * filesystem, (2) A normal flush can dirty the freemap, and
1745          * (3) ensure that the freemap is fully synchronized.
1746          *
1747          * The next mount's recovery scan can clean everything up but we want
1748          * to leave the filesystem in a 100% clean state on a normal unmount.
1749          */
1750 #if 0
1751         hammer2_voldata_lock(hmp);
1752         hammer2_voldata_unlock(hmp);
1753 #endif
1754
1755         /*
1756          * Flush whatever is left.  Unmounted but modified PFS's might still
1757          * have some dirty chains on them.
1758          */
1759         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1760         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1761
1762         if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1763                 hammer2_voldata_modify(hmp);
1764                 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1765                                             HAMMER2_FLUSH_ALL);
1766         }
1767         hammer2_chain_unlock(&hmp->fchain);
1768
1769         if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1770                 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1771                                             HAMMER2_FLUSH_ALL);
1772         }
1773         hammer2_chain_unlock(&hmp->vchain);
1774
1775         if ((hmp->vchain.flags | hmp->fchain.flags) &
1776             HAMMER2_CHAIN_FLUSH_MASK) {
1777                 kprintf("hammer2_unmount: chains left over after final sync\n");
1778                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1779                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1780
1781                 if (hammer2_debug & 0x0010)
1782                         Debugger("entered debugger");
1783         }
1784
1785         hammer2_pfsfree_scan(hmp, 1);
1786
1787         KKASSERT(hmp->spmp == NULL);
1788
1789         /*
1790          * Finish up with the device vnode
1791          */
1792         if (!TAILQ_EMPTY(&hmp->devvpl)) {
1793                 hammer2_close_devvp(&hmp->devvpl, hmp->ronly);
1794                 hammer2_cleanup_devvp(&hmp->devvpl);
1795         }
1796         KKASSERT(TAILQ_EMPTY(&hmp->devvpl));
1797
1798         /*
1799          * Clear vchain/fchain flags that might prevent final cleanup
1800          * of these chains.
1801          */
1802         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1803                 atomic_add_long(&hammer2_count_modified_chains, -1);
1804                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1805                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp, -1);
1806         }
1807         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1808                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1809         }
1810
1811         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1812                 atomic_add_long(&hammer2_count_modified_chains, -1);
1813                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1814                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp, -1);
1815         }
1816         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1817                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1818         }
1819
1820         /*
1821          * Final drop of embedded freemap root chain to
1822          * clean up fchain.core (fchain structure is not
1823          * flagged ALLOCATED so it is cleaned out and then
1824          * left to rot).
1825          */
1826         hammer2_chain_drop(&hmp->fchain);
1827
1828         /*
1829          * Final drop of embedded volume root chain to clean
1830          * up vchain.core (vchain structure is not flagged
1831          * ALLOCATED so it is cleaned out and then left to
1832          * rot).
1833          */
1834         dumpcnt = 50;
1835         hammer2_dump_chain(&hmp->vchain, 0, 0, &dumpcnt, 'v', (u_int)-1);
1836         dumpcnt = 50;
1837         hammer2_dump_chain(&hmp->fchain, 0, 0, &dumpcnt, 'f', (u_int)-1);
1838
1839         hammer2_chain_drop(&hmp->vchain);
1840
1841         hammer2_io_cleanup(hmp, &hmp->iotree);
1842         if (hmp->iofree_count) {
1843                 kprintf("io_cleanup: %d I/O's left hanging\n",
1844                         hmp->iofree_count);
1845         }
1846
1847         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1848         kmalloc_destroy_obj(&hmp->mchain);
1849         kmalloc_destroy_obj(&hmp->mio);
1850         kmalloc_destroy(&hmp->mmsg);
1851         kfree(hmp, M_HAMMER2);
1852 }
1853
1854 int
1855 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1856                  ino_t ino, struct vnode **vpp)
1857 {
1858         hammer2_xop_lookup_t *xop;
1859         hammer2_pfs_t *pmp;
1860         hammer2_inode_t *ip;
1861         hammer2_tid_t inum;
1862         int error;
1863
1864         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1865
1866         error = 0;
1867         pmp = MPTOPMP(mp);
1868
1869         /*
1870          * Easy if we already have it cached
1871          */
1872         ip = hammer2_inode_lookup(pmp, inum);
1873         if (ip) {
1874                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1875                 *vpp = hammer2_igetv(ip, &error);
1876                 hammer2_inode_unlock(ip);
1877                 hammer2_inode_drop(ip);         /* from lookup */
1878
1879                 return error;
1880         }
1881
1882         /*
1883          * Otherwise we have to find the inode
1884          */
1885         xop = hammer2_xop_alloc(pmp->iroot, 0);
1886         xop->lhc = inum;
1887         hammer2_xop_start(&xop->head, &hammer2_lookup_desc);
1888         error = hammer2_xop_collect(&xop->head, 0);
1889
1890         if (error == 0)
1891                 ip = hammer2_inode_get(pmp, &xop->head, -1, -1);
1892         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1893
1894         if (ip) {
1895                 *vpp = hammer2_igetv(ip, &error);
1896                 hammer2_inode_unlock(ip);
1897         } else {
1898                 *vpp = NULL;
1899                 error = ENOENT;
1900         }
1901         return (error);
1902 }
1903
1904 int
1905 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1906 {
1907         hammer2_pfs_t *pmp;
1908         struct vnode *vp;
1909         int error;
1910
1911         pmp = MPTOPMP(mp);
1912         if (pmp->iroot == NULL) {
1913                 kprintf("hammer2 (%s): no root inode\n",
1914                         mp->mnt_stat.f_mntfromname);
1915                 *vpp = NULL;
1916                 return EINVAL;
1917         }
1918
1919         error = 0;
1920         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1921
1922         while (pmp->inode_tid == 0) {
1923                 hammer2_xop_ipcluster_t *xop;
1924                 const hammer2_inode_meta_t *meta;
1925
1926                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1927                 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc);
1928                 error = hammer2_xop_collect(&xop->head, 0);
1929
1930                 if (error == 0) {
1931                         meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1932                         pmp->iroot->meta = *meta;
1933                         pmp->inode_tid = meta->pfs_inum + 1;
1934                         hammer2_xop_pdata(&xop->head);
1935                         /* meta invalid */
1936
1937                         if (pmp->inode_tid < HAMMER2_INODE_START)
1938                                 pmp->inode_tid = HAMMER2_INODE_START;
1939                         pmp->modify_tid =
1940                                 xop->head.cluster.focus->bref.modify_tid + 1;
1941 #if 0
1942                         kprintf("PFS: Starting inode %jd\n",
1943                                 (intmax_t)pmp->inode_tid);
1944                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1945                                 pmp->inode_tid, pmp->modify_tid);
1946 #endif
1947                         //wakeup(&pmp->iroot); XXX
1948
1949                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1950
1951                         /*
1952                          * Prime the mount info.
1953                          */
1954                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1955                         break;
1956                 }
1957
1958                 /*
1959                  * Loop, try again
1960                  */
1961                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1962                 hammer2_inode_unlock(pmp->iroot);
1963                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1964                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1965                 if (error == EINTR)
1966                         break;
1967         }
1968
1969         if (error) {
1970                 hammer2_inode_unlock(pmp->iroot);
1971                 *vpp = NULL;
1972         } else {
1973                 vp = hammer2_igetv(pmp->iroot, &error);
1974                 hammer2_inode_unlock(pmp->iroot);
1975                 *vpp = vp;
1976         }
1977
1978         return (error);
1979 }
1980
1981 /*
1982  * Filesystem status
1983  *
1984  * XXX incorporate ipdata->meta.inode_quota and data_quota
1985  */
1986 static
1987 int
1988 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1989 {
1990         hammer2_pfs_t *pmp;
1991         hammer2_dev_t *hmp;
1992         hammer2_blockref_t bref;
1993         struct statfs tmp;
1994         int i;
1995
1996         /*
1997          * NOTE: iroot might not have validated the cluster yet.
1998          */
1999         pmp = MPTOPMP(mp);
2000
2001         bzero(&tmp, sizeof(tmp));
2002
2003         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2004                 hmp = pmp->pfs_hmps[i];
2005                 if (hmp == NULL)
2006                         continue;
2007                 if (pmp->iroot->cluster.array[i].chain)
2008                         bref = pmp->iroot->cluster.array[i].chain->bref;
2009                 else
2010                         bzero(&bref, sizeof(bref));
2011
2012                 tmp.f_files = bref.embed.stats.inode_count;
2013                 tmp.f_ffree = 0;
2014                 tmp.f_blocks = hmp->voldata.allocator_size /
2015                                mp->mnt_vstat.f_bsize;
2016                 tmp.f_bfree = hmp->voldata.allocator_free /
2017                               mp->mnt_vstat.f_bsize;
2018                 tmp.f_bavail = tmp.f_bfree;
2019
2020                 if (cred && cred->cr_uid != 0) {
2021                         uint64_t adj;
2022
2023                         /* 5% */
2024                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2025                         tmp.f_blocks -= adj;
2026                         tmp.f_bfree -= adj;
2027                         tmp.f_bavail -= adj;
2028                 }
2029
2030                 mp->mnt_stat.f_blocks = tmp.f_blocks;
2031                 mp->mnt_stat.f_bfree = tmp.f_bfree;
2032                 mp->mnt_stat.f_bavail = tmp.f_bavail;
2033                 mp->mnt_stat.f_files = tmp.f_files;
2034                 mp->mnt_stat.f_ffree = tmp.f_ffree;
2035
2036                 *sbp = mp->mnt_stat;
2037         }
2038         return (0);
2039 }
2040
2041 static
2042 int
2043 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2044 {
2045         hammer2_pfs_t *pmp;
2046         hammer2_dev_t *hmp;
2047         hammer2_blockref_t bref;
2048         struct statvfs tmp;
2049         int i;
2050
2051         /*
2052          * NOTE: iroot might not have validated the cluster yet.
2053          */
2054         pmp = MPTOPMP(mp);
2055         bzero(&tmp, sizeof(tmp));
2056
2057         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2058                 hmp = pmp->pfs_hmps[i];
2059                 if (hmp == NULL)
2060                         continue;
2061                 if (pmp->iroot->cluster.array[i].chain)
2062                         bref = pmp->iroot->cluster.array[i].chain->bref;
2063                 else
2064                         bzero(&bref, sizeof(bref));
2065
2066                 tmp.f_files = bref.embed.stats.inode_count;
2067                 tmp.f_ffree = 0;
2068                 tmp.f_blocks = hmp->voldata.allocator_size /
2069                                mp->mnt_vstat.f_bsize;
2070                 tmp.f_bfree = hmp->voldata.allocator_free /
2071                               mp->mnt_vstat.f_bsize;
2072                 tmp.f_bavail = tmp.f_bfree;
2073
2074                 if (cred && cred->cr_uid != 0) {
2075                         uint64_t adj;
2076
2077                         /* 5% */
2078                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2079                         tmp.f_blocks -= adj;
2080                         tmp.f_bfree -= adj;
2081                         tmp.f_bavail -= adj;
2082                 }
2083
2084                 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2085                 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2086                 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2087                 mp->mnt_vstat.f_files = tmp.f_files;
2088                 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2089
2090                 *sbp = mp->mnt_vstat;
2091         }
2092         return (0);
2093 }
2094
2095 /*
2096  * Mount-time recovery (RW mounts)
2097  *
2098  * Updates to the free block table are allowed to lag flushes by one
2099  * transaction.  In case of a crash, then on a fresh mount we must do an
2100  * incremental scan of the last committed transaction id and make sure that
2101  * all related blocks have been marked allocated.
2102  */
2103 struct hammer2_recovery_elm {
2104         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2105         hammer2_chain_t *chain;
2106         hammer2_tid_t sync_tid;
2107 };
2108
2109 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2110
2111 struct hammer2_recovery_info {
2112         struct hammer2_recovery_list list;
2113         hammer2_tid_t   mtid;
2114         int     depth;
2115 };
2116
2117 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2118                         hammer2_chain_t *parent,
2119                         struct hammer2_recovery_info *info,
2120                         hammer2_tid_t sync_tid);
2121
2122 #define HAMMER2_RECOVERY_MAXDEPTH       10
2123
2124 static
2125 int
2126 hammer2_recovery(hammer2_dev_t *hmp)
2127 {
2128         struct hammer2_recovery_info info;
2129         struct hammer2_recovery_elm *elm;
2130         hammer2_chain_t *parent;
2131         hammer2_tid_t sync_tid;
2132         hammer2_tid_t mirror_tid;
2133         int error;
2134
2135         hammer2_trans_init(hmp->spmp, 0);
2136
2137         sync_tid = hmp->voldata.freemap_tid;
2138         mirror_tid = hmp->voldata.mirror_tid;
2139
2140         kprintf("hammer2_mount: \"%s\": ", hmp->devrepname);
2141         if (sync_tid >= mirror_tid) {
2142                 kprintf("no recovery needed\n");
2143         } else {
2144                 kprintf("freemap recovery %016jx-%016jx\n",
2145                         sync_tid + 1, mirror_tid);
2146         }
2147
2148         TAILQ_INIT(&info.list);
2149         info.depth = 0;
2150         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2151         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2152         hammer2_chain_lookup_done(parent);
2153
2154         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2155                 TAILQ_REMOVE(&info.list, elm, entry);
2156                 parent = elm->chain;
2157                 sync_tid = elm->sync_tid;
2158                 kfree(elm, M_HAMMER2);
2159
2160                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2161                 error |= hammer2_recovery_scan(hmp, parent, &info,
2162                                               hmp->voldata.freemap_tid);
2163                 hammer2_chain_unlock(parent);
2164                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2165         }
2166
2167         hammer2_trans_done(hmp->spmp, 0);
2168
2169         return error;
2170 }
2171
2172 static
2173 int
2174 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2175                       struct hammer2_recovery_info *info,
2176                       hammer2_tid_t sync_tid)
2177 {
2178         const hammer2_inode_data_t *ripdata;
2179         hammer2_chain_t *chain;
2180         hammer2_blockref_t bref;
2181         int tmp_error;
2182         int rup_error;
2183         int error;
2184         int first;
2185
2186         /*
2187          * Adjust freemap to ensure that the block(s) are marked allocated.
2188          */
2189         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2190                 hammer2_freemap_adjust(hmp, &parent->bref,
2191                                        HAMMER2_FREEMAP_DORECOVER);
2192         }
2193
2194         /*
2195          * Check type for recursive scan
2196          */
2197         switch(parent->bref.type) {
2198         case HAMMER2_BREF_TYPE_VOLUME:
2199                 /* data already instantiated */
2200                 break;
2201         case HAMMER2_BREF_TYPE_INODE:
2202                 /*
2203                  * Must instantiate data for DIRECTDATA test and also
2204                  * for recursion.
2205                  */
2206                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2207                 ripdata = &parent->data->ipdata;
2208                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2209                         /* not applicable to recovery scan */
2210                         hammer2_chain_unlock(parent);
2211                         return 0;
2212                 }
2213                 hammer2_chain_unlock(parent);
2214                 break;
2215         case HAMMER2_BREF_TYPE_INDIRECT:
2216                 /*
2217                  * Must instantiate data for recursion
2218                  */
2219                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2220                 hammer2_chain_unlock(parent);
2221                 break;
2222         case HAMMER2_BREF_TYPE_DIRENT:
2223         case HAMMER2_BREF_TYPE_DATA:
2224         case HAMMER2_BREF_TYPE_FREEMAP:
2225         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2226         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2227                 /* not applicable to recovery scan */
2228                 return 0;
2229                 break;
2230         default:
2231                 return HAMMER2_ERROR_BADBREF;
2232         }
2233
2234         /*
2235          * Defer operation if depth limit reached.
2236          */
2237         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2238                 struct hammer2_recovery_elm *elm;
2239
2240                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2241                 elm->chain = parent;
2242                 elm->sync_tid = sync_tid;
2243                 hammer2_chain_ref(parent);
2244                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2245                 /* unlocked by caller */
2246
2247                 return(0);
2248         }
2249
2250
2251         /*
2252          * Recursive scan of the last flushed transaction only.  We are
2253          * doing this without pmp assignments so don't leave the chains
2254          * hanging around after we are done with them.
2255          *
2256          * error        Cumulative error this level only
2257          * rup_error    Cumulative error for recursion
2258          * tmp_error    Specific non-cumulative recursion error
2259          */
2260         chain = NULL;
2261         first = 1;
2262         rup_error = 0;
2263         error = 0;
2264
2265         for (;;) {
2266                 error |= hammer2_chain_scan(parent, &chain, &bref,
2267                                             &first,
2268                                             HAMMER2_LOOKUP_NODATA);
2269
2270                 /*
2271                  * Problem during scan or EOF
2272                  */
2273                 if (error)
2274                         break;
2275
2276                 /*
2277                  * If this is a leaf
2278                  */
2279                 if (chain == NULL) {
2280                         if (bref.mirror_tid > sync_tid) {
2281                                 hammer2_freemap_adjust(hmp, &bref,
2282                                                      HAMMER2_FREEMAP_DORECOVER);
2283                         }
2284                         continue;
2285                 }
2286
2287                 /*
2288                  * This may or may not be a recursive node.
2289                  */
2290                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2291                 if (bref.mirror_tid > sync_tid) {
2292                         ++info->depth;
2293                         tmp_error = hammer2_recovery_scan(hmp, chain,
2294                                                            info, sync_tid);
2295                         --info->depth;
2296                 } else {
2297                         tmp_error = 0;
2298                 }
2299
2300                 /*
2301                  * Flush the recovery at the PFS boundary to stage it for
2302                  * the final flush of the super-root topology.
2303                  */
2304                 if (tmp_error == 0 &&
2305                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2306                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2307                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2308                                              HAMMER2_FLUSH_ALL);
2309                 }
2310                 rup_error |= tmp_error;
2311         }
2312         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2313 }
2314
2315 /*
2316  * This fixes up an error introduced in earlier H2 implementations where
2317  * moving a PFS inode into an indirect block wound up causing the
2318  * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2319  */
2320 static
2321 int
2322 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2323 {
2324         const hammer2_inode_data_t *ripdata;
2325         hammer2_chain_t *parent;
2326         hammer2_chain_t *chain;
2327         hammer2_key_t key_next;
2328         hammer2_pfs_t *spmp;
2329         int error;
2330
2331         error = 0;
2332
2333         /*
2334          * Lookup mount point under the media-localized super-root.
2335          *
2336          * cluster->pmp will incorrectly point to spmp and must be fixed
2337          * up later on.
2338          */
2339         spmp = hmp->spmp;
2340         hammer2_inode_lock(spmp->iroot, 0);
2341         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2342         chain = hammer2_chain_lookup(&parent, &key_next,
2343                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2344                                          &error, 0);
2345         while (chain) {
2346                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2347                         continue;
2348                 if (chain->error) {
2349                         kprintf("I/O error scanning PFS labels\n");
2350                         error |= chain->error;
2351                 } else if ((chain->bref.flags &
2352                             HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2353                         int error2;
2354
2355                         ripdata = &chain->data->ipdata;
2356                         hammer2_trans_init(hmp->spmp, 0);
2357                         error2 = hammer2_chain_modify(chain,
2358                                                       chain->bref.modify_tid,
2359                                                       0, 0);
2360                         if (error2 == 0) {
2361                                 kprintf("hammer2: Correct mis-flagged PFS %s\n",
2362                                         ripdata->filename);
2363                                 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2364                         } else {
2365                                 error |= error2;
2366                         }
2367                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2368                                              HAMMER2_FLUSH_ALL);
2369                         hammer2_trans_done(hmp->spmp, 0);
2370                 }
2371                 chain = hammer2_chain_next(&parent, chain, &key_next,
2372                                            key_next, HAMMER2_KEY_MAX,
2373                                            &error, 0);
2374         }
2375         if (parent) {
2376                 hammer2_chain_unlock(parent);
2377                 hammer2_chain_drop(parent);
2378         }
2379         hammer2_inode_unlock(spmp->iroot);
2380
2381         return error;
2382 }
2383
2384 /*
2385  * Sync a mount point; this is called periodically on a per-mount basis from
2386  * the filesystem syncer, and whenever a user issues a sync.
2387  */
2388 int
2389 hammer2_vfs_sync(struct mount *mp, int waitfor)
2390 {
2391         int error;
2392
2393         error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor);
2394
2395         return error;
2396 }
2397
2398 /*
2399  * Because frontend operations lock vnodes before we get a chance to
2400  * lock the related inode, we can't just acquire a vnode lock without
2401  * risking a deadlock.  The frontend may be holding a vnode lock while
2402  * also blocked on our SYNCQ flag while trying to get the inode lock.
2403  *
2404  * To deal with this situation we can check the vnode lock situation
2405  * after locking the inode and perform a work-around.
2406  */
2407 int
2408 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor)
2409 {
2410         hammer2_inode_t *ip;
2411         hammer2_depend_t *depend;
2412         hammer2_depend_t *depend_next;
2413         struct vnode *vp;
2414         uint32_t pass2;
2415         int error;
2416         int wakecount;
2417         int dorestart;
2418
2419         /*
2420          * Move all inodes on sideq to syncq.  This will clear sideq.
2421          * This should represent all flushable inodes.  These inodes
2422          * will already have refs due to being on syncq or sideq.  We
2423          * must do this all at once with the spinlock held to ensure that
2424          * all inode dependencies are part of the same flush.
2425          *
2426          * We should be able to do this asynchronously from frontend
2427          * operations because we will be locking the inodes later on
2428          * to actually flush them, and that will partition any frontend
2429          * op using the same inode.  Either it has already locked the
2430          * inode and we will block, or it has not yet locked the inode
2431          * and it will block until we are finished flushing that inode.
2432          *
2433          * When restarting, only move the inodes flagged as PASS2 from
2434          * SIDEQ to SYNCQ.  PASS2 propagation by inode_lock4() and
2435          * inode_depend() are atomic with the spin-lock.
2436          */
2437         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2438 #ifdef HAMMER2_DEBUG_SYNC
2439         kprintf("FILESYSTEM SYNC BOUNDARY\n");
2440 #endif
2441         dorestart = 0;
2442
2443         /*
2444          * Move inodes from depq to syncq, releasing the related
2445          * depend structures.
2446          */
2447 restart:
2448 #ifdef HAMMER2_DEBUG_SYNC
2449         kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart);
2450 #endif
2451         hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/);
2452         hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN);
2453
2454         /*
2455          * Move inodes from depq to syncq.  When restarting, only depq's
2456          * marked pass2 are moved.
2457          */
2458         hammer2_spin_ex(&pmp->list_spin);
2459         depend_next = TAILQ_FIRST(&pmp->depq);
2460         wakecount = 0;
2461
2462         while ((depend = depend_next) != NULL) {
2463                 depend_next = TAILQ_NEXT(depend, entry);
2464                 if (dorestart && depend->pass2 == 0)
2465                         continue;
2466                 TAILQ_FOREACH(ip, &depend->sideq, entry) {
2467                         KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ);
2468                         atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ);
2469                         atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ);
2470                         ip->depend = NULL;
2471                 }
2472
2473                 /*
2474                  * NOTE: pmp->sideq_count includes both sideq and syncq
2475                  */
2476                 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry);
2477
2478                 depend->count = 0;
2479                 depend->pass2 = 0;
2480                 TAILQ_REMOVE(&pmp->depq, depend, entry);
2481         }
2482
2483         hammer2_spin_unex(&pmp->list_spin);
2484         hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/
2485                                       HAMMER2_TRANS_WAITING);
2486         dorestart = 0;
2487
2488         /*
2489          * sideq_count may have dropped enough to allow us to unstall
2490          * the frontend.
2491          */
2492         hammer2_pfs_memory_wakeup(pmp, 0);
2493
2494         /*
2495          * Now run through all inodes on syncq.
2496          *
2497          * Flush transactions only interlock with other flush transactions.
2498          * Any conflicting frontend operations will block on the inode, but
2499          * may hold a vnode lock while doing so.
2500          */
2501         hammer2_spin_ex(&pmp->list_spin);
2502         while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) {
2503                 /*
2504                  * Remove the inode from the SYNCQ, transfer the syncq ref
2505                  * to us.  We must clear SYNCQ to allow any potential
2506                  * front-end deadlock to proceed.  We must set PASS2 so
2507                  * the dependency code knows what to do.
2508                  */
2509                 pass2 = ip->flags;
2510                 cpu_ccfence();
2511                 if (atomic_cmpset_int(&ip->flags,
2512                               pass2,
2513                               (pass2 & ~(HAMMER2_INODE_SYNCQ |
2514                                          HAMMER2_INODE_SYNCQ_WAKEUP)) |
2515                               HAMMER2_INODE_SYNCQ_PASS2) == 0)
2516                 {
2517                         continue;
2518                 }
2519                 TAILQ_REMOVE(&pmp->syncq, ip, entry);
2520                 --pmp->sideq_count;
2521                 hammer2_spin_unex(&pmp->list_spin);
2522
2523                 /*
2524                  * Tickle anyone waiting on ip->flags or the hysteresis
2525                  * on the dirty inode count.
2526                  */
2527                 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP)
2528                         wakeup(&ip->flags);
2529                 if (++wakecount >= hammer2_limit_dirty_inodes / 20 + 1) {
2530                         wakecount = 0;
2531                         hammer2_pfs_memory_wakeup(pmp, 0);
2532                 }
2533
2534                 /*
2535                  * Relock the inode, and we inherit a ref from the above.
2536                  * We will check for a race after we acquire the vnode.
2537                  */
2538                 hammer2_mtx_ex(&ip->lock);
2539
2540                 /*
2541                  * We need the vp in order to vfsync() dirty buffers, so if
2542                  * one isn't attached we can skip it.
2543                  *
2544                  * Ordering the inode lock and then the vnode lock has the
2545                  * potential to deadlock.  If we had left SYNCQ set that could
2546                  * also deadlock us against the frontend even if we don't hold
2547                  * any locks, but the latter is not a problem now since we
2548                  * cleared it.  igetv will temporarily release the inode lock
2549                  * in a safe manner to work-around the deadlock.
2550                  *
2551                  * Unfortunately it is still possible to deadlock when the
2552                  * frontend obtains multiple inode locks, because all the
2553                  * related vnodes are already locked (nor can the vnode locks
2554                  * be released and reacquired without messing up RECLAIM and
2555                  * INACTIVE sequencing).
2556                  *
2557                  * The solution for now is to move the vp back onto SIDEQ
2558                  * and set dorestart, which will restart the flush after we
2559                  * exhaust the current SYNCQ.  Note that additional
2560                  * dependencies may build up, so we definitely need to move
2561                  * the whole SIDEQ back to SYNCQ when we restart.
2562                  */
2563                 vp = ip->vp;
2564                 if (vp) {
2565                         if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) {
2566                                 /*
2567                                  * Failed to get the vnode, requeue the inode
2568                                  * (PASS2 is already set so it will be found
2569                                  * again on the restart).
2570                                  *
2571                                  * Then unlock, possibly sleep, and retry
2572                                  * later.  We sleep if PASS2 was *previously*
2573                                  * set, before we set it again above.
2574                                  */
2575                                 vp = NULL;
2576                                 dorestart = 1;
2577 #ifdef HAMMER2_DEBUG_SYNC
2578                                 kprintf("inum %ld (sync delayed by vnode)\n",
2579                                         (long)ip->meta.inum);
2580 #endif
2581                                 hammer2_inode_delayed_sideq(ip);
2582
2583                                 hammer2_mtx_unlock(&ip->lock);
2584                                 hammer2_inode_drop(ip);
2585
2586                                 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) {
2587                                         tsleep(&dorestart, 0, "h2syndel", 2);
2588                                 }
2589                                 hammer2_spin_ex(&pmp->list_spin);
2590                                 continue;
2591                         }
2592                 } else {
2593                         vp = NULL;
2594                 }
2595
2596                 /*
2597                  * If the inode wound up on a SIDEQ again it will already be
2598                  * prepped for another PASS2.  In this situation if we flush
2599                  * it now we will just wind up flushing it again in the same
2600                  * syncer run, so we might as well not flush it now.
2601                  */
2602                 if (ip->flags & HAMMER2_INODE_SIDEQ) {
2603                         hammer2_mtx_unlock(&ip->lock);
2604                         hammer2_inode_drop(ip);
2605                         if (vp)
2606                                 vput(vp);
2607                         dorestart = 1;
2608                         hammer2_spin_ex(&pmp->list_spin);
2609                         continue;
2610                 }
2611
2612                 /*
2613                  * Ok we have the inode exclusively locked and if vp is
2614                  * not NULL that will also be exclusively locked.  Do the
2615                  * meat of the flush.
2616                  *
2617                  * vp token needed for v_rbdirty_tree check / vclrisdirty
2618                  * sequencing.  Though we hold the vnode exclusively so
2619                  * we shouldn't need to hold the token also in this case.
2620                  */
2621                 if (vp) {
2622                         vfsync(vp, MNT_WAIT, 1, NULL, NULL);
2623                         bio_track_wait(NULL, 0, 0); /* XXX */
2624                 }
2625
2626                 /*
2627                  * If the inode has not yet been inserted into the tree
2628                  * we must do so.  Then sync and flush it.  The flush should
2629                  * update the parent.
2630                  */
2631                 if (ip->flags & HAMMER2_INODE_DELETING) {
2632 #ifdef HAMMER2_DEBUG_SYNC
2633                         kprintf("inum %ld destroy\n", (long)ip->meta.inum);
2634 #endif
2635                         hammer2_inode_chain_des(ip);
2636                         atomic_add_long(&hammer2_iod_inode_deletes, 1);
2637                 } else if (ip->flags & HAMMER2_INODE_CREATING) {
2638 #ifdef HAMMER2_DEBUG_SYNC
2639                         kprintf("inum %ld insert\n", (long)ip->meta.inum);
2640 #endif
2641                         hammer2_inode_chain_ins(ip);
2642                         atomic_add_long(&hammer2_iod_inode_creates, 1);
2643                 }
2644 #ifdef HAMMER2_DEBUG_SYNC
2645                 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum);
2646 #endif
2647
2648                 /*
2649                  * Because I kinda messed up the design and index the inodes
2650                  * under the root inode, along side the directory entries,
2651                  * we can't flush the inode index under the iroot until the
2652                  * end.  If we do it now we might miss effects created by
2653                  * other inodes on the SYNCQ.
2654                  *
2655                  * Do a normal (non-FSSYNC) flush instead, which allows the
2656                  * vnode code to work the same.  We don't want to force iroot
2657                  * back onto the SIDEQ, and we also don't want the flush code
2658                  * to update pfs_iroot_blocksets until the final flush later.
2659                  *
2660                  * XXX at the moment this will likely result in a double-flush
2661                  * of the iroot chain.
2662                  */
2663                 hammer2_inode_chain_sync(ip);
2664                 if (ip == pmp->iroot) {
2665                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP);
2666                 } else {
2667                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2668                                                       HAMMER2_XOP_FSSYNC);
2669                 }
2670                 if (vp) {
2671                         lwkt_gettoken(NULL);
2672                         if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2673                                           HAMMER2_INODE_RESIZED |
2674                                           HAMMER2_INODE_DIRTYDATA)) == 0) {
2675                             //RB_EMPTY(&vp->v_rbdirty_tree) &&
2676                             //!bio_track_active(&vp->v_track_write)) {
2677                                 vclrisdirty(vp);
2678                         } else {
2679                                 hammer2_inode_delayed_sideq(ip);
2680                         }
2681                         lwkt_reltoken(NULL);
2682                         vput(vp);
2683                         vp = NULL;      /* safety */
2684                 }
2685                 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2);
2686                 hammer2_inode_unlock(ip);       /* unlock+drop */
2687                 /* ip pointer invalid */
2688
2689                 /*
2690                  * If the inode got dirted after we dropped our locks,
2691                  * it will have already been moved back to the SIDEQ.
2692                  */
2693                 hammer2_spin_ex(&pmp->list_spin);
2694         }
2695         hammer2_spin_unex(&pmp->list_spin);
2696         hammer2_pfs_memory_wakeup(pmp, 0);
2697
2698         if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) {
2699 #ifdef HAMMER2_DEBUG_SYNC
2700                 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n");
2701                 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/
2702 #endif
2703                 dorestart = 1;
2704                 goto restart;
2705         }
2706 #ifdef HAMMER2_DEBUG_SYNC
2707         kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n");
2708         /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/
2709 #endif
2710
2711         /*
2712          * We have to flush the PFS root last, even if it does not appear to
2713          * be dirty, because all the inodes in the PFS are indexed under it.
2714          * The normal flushing of iroot above would only occur if directory
2715          * entries under the root were changed.
2716          *
2717          * Specifying VOLHDR will cause an additionl flush of hmp->spmp
2718          * for the media making up the cluster.
2719          */
2720         if ((ip = pmp->iroot) != NULL) {
2721                 hammer2_inode_ref(ip);
2722                 hammer2_mtx_ex(&ip->lock);
2723                 hammer2_inode_chain_sync(ip);
2724                 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2725                                               HAMMER2_XOP_FSSYNC |
2726                                               HAMMER2_XOP_VOLHDR);
2727                 hammer2_inode_unlock(ip);       /* unlock+drop */
2728         }
2729 #ifdef HAMMER2_DEBUG_SYNC
2730         kprintf("FILESYSTEM SYNC STAGE 2 DONE\n");
2731 #endif
2732
2733         /*
2734          * device bioq sync
2735          */
2736         hammer2_bioq_sync(pmp);
2737
2738         error = 0;      /* XXX */
2739         hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH);
2740
2741         return (error);
2742 }
2743
2744 #if 0
2745 static
2746 int
2747 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2748 {
2749         hammer2_inode_t *ip;
2750
2751         KKASSERT(MAXFIDSZ >= 16);
2752         ip = VTOI(vp);
2753         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2754         fhp->fid_ext = 0;
2755         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2756         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2757
2758         return 0;
2759 }
2760
2761 static
2762 int
2763 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2764                struct fid *fhp, struct vnode **vpp)
2765 {
2766         hammer2_tid_t inum;
2767         int error;
2768
2769         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2770         if (vpp) {
2771                 if (inum == 1)
2772                         error = hammer2_vfs_root(mp, vpp);
2773                 else
2774                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2775         } else {
2776                 error = 0;
2777         }
2778         return error;
2779 }
2780
2781 static
2782 int
2783 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2784                  int *exflagsp, struct ucred **credanonp)
2785 {
2786         hammer2_pfs_t *pmp;
2787         struct netcred *np;
2788         int error;
2789
2790         pmp = MPTOPMP(mp);
2791         np = vfs_export_lookup(mp, &pmp->export, nam);
2792         if (np) {
2793                 *exflagsp = np->netc_exflags;
2794                 *credanonp = &np->netc_anon;
2795                 error = 0;
2796         } else {
2797                 error = EACCES;
2798         }
2799         return error;
2800 }
2801 #endif
2802
2803 /*
2804  * This handles hysteresis on regular file flushes.  Because the BIOs are
2805  * routed to a thread it is possible for an excessive number to build up
2806  * and cause long front-end stalls long before the runningbuffspace limit
2807  * is hit, so we implement hammer2_flush_pipe to control the
2808  * hysteresis.
2809  *
2810  * This is a particular problem when compression is used.
2811  */
2812 void
2813 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2814 {
2815         atomic_add_int(&pmp->count_lwinprog, 1);
2816 }
2817
2818 void
2819 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2820 {
2821 #if 0
2822         int lwinprog;
2823
2824         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2825         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2826             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2827                 atomic_clear_int(&pmp->count_lwinprog,
2828                                  HAMMER2_LWINPROG_WAITING);
2829                 wakeup(&pmp->count_lwinprog);
2830         }
2831         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2832             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2833                 atomic_clear_int(&pmp->count_lwinprog,
2834                                  HAMMER2_LWINPROG_WAITING0);
2835                 wakeup(&pmp->count_lwinprog);
2836         }
2837 #endif
2838 }
2839
2840 void
2841 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2842 {
2843 #if 0
2844         int lwinprog;
2845         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2846                                     HAMMER2_LWINPROG_WAITING0;
2847
2848         for (;;) {
2849                 lwinprog = pmp->count_lwinprog;
2850                 cpu_ccfence();
2851                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2852                         break;
2853                 tsleep_interlock(&pmp->count_lwinprog, 0);
2854                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2855                 lwinprog = pmp->count_lwinprog;
2856                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2857                         break;
2858                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2859         }
2860 #endif
2861 }
2862
2863 #if 0
2864 /*
2865  * It is possible for an excessive number of dirty chains or dirty inodes
2866  * to build up.  When this occurs we start an asynchronous filesystem sync.
2867  * If the level continues to build up, we stall, waiting for it to drop,
2868  * with some hysteresis.
2869  *
2870  * This relies on the kernel calling hammer2_vfs_modifying() prior to
2871  * obtaining any vnode locks before making a modifying VOP call.
2872  */
2873 static int
2874 hammer2_vfs_modifying(struct mount *mp)
2875 {
2876         if (mp->mnt_flag & MNT_RDONLY)
2877                 return EROFS;
2878         hammer2_pfs_memory_wait(MPTOPMP(mp));
2879
2880         return 0;
2881 }
2882 #endif
2883
2884 /*
2885  * Initiate an asynchronous filesystem sync and, with hysteresis,
2886  * stall if the internal data structure count becomes too bloated.
2887  */
2888 void
2889 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2890 {
2891         uint32_t waiting;
2892         int pcatch;
2893         int error;
2894
2895         if (pmp == NULL || pmp->mp == NULL)
2896                 return;
2897
2898         for (;;) {
2899                 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK;
2900                 cpu_ccfence();
2901
2902                 /*
2903                  * Start the syncer running at 1/2 the limit
2904                  */
2905                 if (waiting > hammer2_limit_dirty_chains / 2 ||
2906                     pmp->sideq_count > hammer2_limit_dirty_inodes / 2) {
2907                         trigger_syncer(pmp->mp);
2908                 }
2909
2910                 /*
2911                  * Stall at the limit waiting for the counts to drop.
2912                  * This code will typically be woken up once the count
2913                  * drops below 3/4 the limit, or in one second.
2914                  */
2915                 if (waiting < hammer2_limit_dirty_chains &&
2916                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
2917                         break;
2918                 }
2919
2920                 pcatch = curthread->td_proc ? PCATCH : 0;
2921
2922                 tsleep_interlock(&pmp->inmem_dirty_chains, pcatch);
2923                 atomic_set_int(&pmp->inmem_dirty_chains,
2924                                HAMMER2_DIRTYCHAIN_WAITING);
2925                 if (waiting < hammer2_limit_dirty_chains &&
2926                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
2927                         break;
2928                 }
2929                 trigger_syncer(pmp->mp);
2930                 error = tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED | pcatch,
2931                                "h2memw", hz);
2932                 if (error == ERESTART)
2933                         break;
2934         }
2935 }
2936
2937 /*
2938  * Wake up any stalled frontend ops waiting, with hysteresis, using
2939  * 2/3 of the limit.
2940  */
2941 void
2942 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp, int count)
2943 {
2944         uint32_t waiting;
2945
2946         if (pmp) {
2947                 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, count);
2948                 /* don't need --waiting to test flag */
2949
2950                 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) &&
2951                     (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <=
2952                     hammer2_limit_dirty_chains * 2 / 3 &&
2953                     pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) {
2954                         atomic_clear_int(&pmp->inmem_dirty_chains,
2955                                          HAMMER2_DIRTYCHAIN_WAITING);
2956                         wakeup(&pmp->inmem_dirty_chains);
2957                 }
2958         }
2959 }
2960
2961 void
2962 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2963 {
2964         if (pmp) {
2965                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2966         }
2967 }
2968
2969 /*
2970  * Volume header data locks
2971  */
2972 void
2973 hammer2_voldata_lock(hammer2_dev_t *hmp)
2974 {
2975         lockmgr(&hmp->vollk, LK_EXCLUSIVE);
2976 }
2977
2978 void
2979 hammer2_voldata_unlock(hammer2_dev_t *hmp)
2980 {
2981         lockmgr(&hmp->vollk, LK_RELEASE);
2982 }
2983
2984 /*
2985  * Caller indicates that the volume header is being modified.  Flag
2986  * the related chain and adjust its transaction id.
2987  *
2988  * The transaction id is set to voldata.mirror_tid + 1, similar to
2989  * what hammer2_chain_modify() does.  Be very careful here, volume
2990  * data can be updated independently of the rest of the filesystem.
2991  */
2992 void
2993 hammer2_voldata_modify(hammer2_dev_t *hmp)
2994 {
2995         if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) {
2996                 atomic_add_long(&hammer2_count_modified_chains, 1);
2997                 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
2998                 hammer2_pfs_memory_inc(hmp->vchain.pmp);
2999                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid + 1;
3000         }
3001 }
3002
3003 /*
3004  * Returns 0 if the filesystem has tons of free space
3005  * Returns 1 if the filesystem has less than 10% remaining
3006  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
3007  */
3008 int
3009 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
3010 {
3011         hammer2_pfs_t *pmp;
3012         hammer2_dev_t *hmp;
3013         hammer2_off_t free_reserved;
3014         hammer2_off_t free_nominal;
3015         int i;
3016
3017         pmp = ip->pmp;
3018
3019         if (/*XXX*/ 1 || pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
3020                 free_reserved = HAMMER2_SEGSIZE;
3021                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
3022                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
3023                         hmp = pmp->pfs_hmps[i];
3024                         if (hmp == NULL)
3025                                 continue;
3026                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
3027                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
3028                                 continue;
3029
3030                         if (free_nominal > hmp->voldata.allocator_free)
3031                                 free_nominal = hmp->voldata.allocator_free;
3032                         if (free_reserved < hmp->free_reserved)
3033                                 free_reserved = hmp->free_reserved;
3034                 }
3035
3036                 /*
3037                  * SMP races ok
3038                  */
3039                 pmp->free_reserved = free_reserved;
3040                 pmp->free_nominal = free_nominal;
3041                 pmp->free_ticks = ticks;
3042         } else {
3043                 free_reserved = pmp->free_reserved;
3044                 free_nominal = pmp->free_nominal;
3045         }
3046         if (cred && cred->cr_uid != 0) {
3047                 if ((int64_t)(free_nominal - bytes) <
3048                     (int64_t)free_reserved) {
3049                         return 2;
3050                 }
3051         } else {
3052                 if ((int64_t)(free_nominal - bytes) <
3053                     (int64_t)free_reserved / 2) {
3054                         return 2;
3055                 }
3056         }
3057         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
3058                 return 1;
3059         return 0;
3060 }
3061
3062 /*
3063  * Debugging
3064  */
3065 void
3066 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int bi, int *countp,
3067                    char pfx, u_int flags)
3068 {
3069         hammer2_chain_t *scan;
3070         hammer2_chain_t *parent;
3071
3072         if (hammer2_debug & 0x80000000)
3073                 *countp = INT_MAX;
3074
3075         --*countp;
3076         if (*countp == 0) {
3077                 kprintf("%*.*s...\n", tab, tab, "");
3078                 return;
3079         }
3080         if (*countp < 0)
3081                 return;
3082         kprintf("%*.*s%c-chain %p %s.%-3d %016jx %016jx/%-2d mir=%016jx\n",
3083                 tab, tab, "", pfx, chain,
3084                 hammer2_bref_type_str(chain->bref.type), bi,
3085                 chain->bref.data_off, chain->bref.key, chain->bref.keybits,
3086                 chain->bref.mirror_tid);
3087
3088         kprintf("%*.*s      [%08x] (%s) refs=%d",
3089                 tab, tab, "",
3090                 chain->flags,
3091                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3092                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
3093                 chain->refs);
3094
3095         parent = chain->parent;
3096         if (parent)
3097                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d]",
3098                         tab, tab, "",
3099                         parent, parent->flags, parent->refs);
3100         if (RB_EMPTY(&chain->core.rbtree)) {
3101                 kprintf("\n");
3102         } else {
3103                 int bi = 0;
3104                 kprintf(" {\n");
3105                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3106                         if ((scan->flags & flags) || flags == (u_int)-1) {
3107                                 hammer2_dump_chain(scan, tab + 4, bi, countp,
3108                                                    'a', flags);
3109                         }
3110                         bi++;
3111                 }
3112                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3113                         kprintf("%*.*s}(%s)\n", tab, tab, "",
3114                                 chain->data->ipdata.filename);
3115                 else
3116                         kprintf("%*.*s}\n", tab, tab, "");
3117         }
3118 }
3119
3120 void
3121 hammer2_dump_chains(hammer2_dev_t *hmp, char vpfx, char fpfx)
3122 {
3123         int dumpcnt;
3124
3125         dumpcnt = 50;
3126         hammer2_dump_chain(&hmp->vchain, 0, 0, &dumpcnt, vpfx, (u_int)-1);
3127
3128         dumpcnt = 50;
3129         hammer2_dump_chain(&hmp->fchain, 0, 0, &dumpcnt, fpfx, (u_int)-1);
3130 }