sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2018 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72         int pass;
  73 };
  74
  75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  76 static struct hammer2_mntlist hammer2_mntlist;
  77
  78 struct hammer2_pfslist hammer2_pfslist;
  79 struct hammer2_pfslist hammer2_spmplist;
  80 struct lock hammer2_mntlk;
  81
  82 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
  83 int hammer2_debug;
  84 long hammer2_debug_inode;
  85 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
  86 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
  87 int hammer2_cluster_write = 0;          /* physical write clustering */
  88 int hammer2_dedup_enable = 1;
  89 int hammer2_always_compress = 0;        /* always try to compress */
  90 int hammer2_inval_enable = 0;
  91 int hammer2_flush_pipe = 100;
  92 int hammer2_dio_count;
  93 int hammer2_dio_limit = 256;
  94 int hammer2_bulkfree_tps = 5000;
  95 long hammer2_chain_allocs;
  96 long hammer2_chain_frees;
  97 long hammer2_limit_dirty_chains;
  98 long hammer2_limit_dirty_inodes;
  99 long hammer2_count_modified_chains;
 100 long hammer2_iod_invals;
 101 long hammer2_iod_file_read;
 102 long hammer2_iod_meta_read;
 103 long hammer2_iod_indr_read;
 104 long hammer2_iod_fmap_read;
 105 long hammer2_iod_volu_read;
 106 long hammer2_iod_file_write;
 107 long hammer2_iod_file_wembed;
 108 long hammer2_iod_file_wzero;
 109 long hammer2_iod_file_wdedup;
 110 long hammer2_iod_meta_write;
 111 long hammer2_iod_indr_write;
 112 long hammer2_iod_fmap_write;
 113 long hammer2_iod_volu_write;
 114 long hammer2_iod_inode_creates;
 115 long hammer2_iod_inode_deletes;
 116
 117 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 118 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 119                 "Buffer used for compression.");
 120
 121 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 122 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 123                 "Buffer used for decompression.");
 124
 125 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 126
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
 128            &hammer2_supported_version, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 130            &hammer2_debug, 0, "");
 131 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW,
 132            &hammer2_debug_inode, 0, "");
 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
 134            &hammer2_cluster_meta_read, 0, "");
 135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
 136            &hammer2_cluster_data_read, 0, "");
 137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 138            &hammer2_cluster_write, 0, "");
 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 140            &hammer2_dedup_enable, 0, "");
 141 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
 142            &hammer2_always_compress, 0, "");
 143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 144            &hammer2_inval_enable, 0, "");
 145 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 146            &hammer2_flush_pipe, 0, "");
 147 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
 148            &hammer2_bulkfree_tps, 0, "");
 149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 150            &hammer2_chain_allocs, 0, "");
 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 152            &hammer2_chain_frees, 0, "");
 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 154            &hammer2_limit_dirty_chains, 0, "");
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
 156            &hammer2_limit_dirty_inodes, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 158            &hammer2_count_modified_chains, 0, "");
 159 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 160            &hammer2_dio_count, 0, "");
 161 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
 162            &hammer2_dio_limit, 0, "");
 163
 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 165            &hammer2_iod_invals, 0, "");
 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 167            &hammer2_iod_file_read, 0, "");
 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 169            &hammer2_iod_meta_read, 0, "");
 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 171            &hammer2_iod_indr_read, 0, "");
 172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 173            &hammer2_iod_fmap_read, 0, "");
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 175            &hammer2_iod_volu_read, 0, "");
 176
 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 178            &hammer2_iod_file_write, 0, "");
 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 180            &hammer2_iod_file_wembed, 0, "");
 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 182            &hammer2_iod_file_wzero, 0, "");
 183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 184            &hammer2_iod_file_wdedup, 0, "");
 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 186            &hammer2_iod_meta_write, 0, "");
 187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 188            &hammer2_iod_indr_write, 0, "");
 189 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 190            &hammer2_iod_fmap_write, 0, "");
 191 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 192            &hammer2_iod_volu_write, 0, "");
 193 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW,
 194            &hammer2_iod_inode_creates, 0, "");
 195 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW,
 196            &hammer2_iod_inode_deletes, 0, "");
 197
 198 long hammer2_process_icrc32;
 199 long hammer2_process_xxhash64;
 200 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
 201            &hammer2_process_icrc32, 0, "");
 202 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
 203            &hammer2_process_xxhash64, 0, "");
 204
 205 static int hammer2_vfs_init(struct vfsconf *conf);
 206 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 207 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 208                                 struct ucred *cred);
 209 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 210                                 struct vnode *, struct ucred *);
 211 static int hammer2_recovery(hammer2_dev_t *hmp);
 212 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 213 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 214 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 215                                 struct ucred *cred);
 216 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 217                                 struct ucred *cred);
 218 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 219                                 struct fid *fhp, struct vnode **vpp);
 220 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 221 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 222                                 int *exflagsp, struct ucred **credanonp);
 223 static void hammer2_vfs_modifying(struct mount *mp);
 224
 225 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 226 #if 0
 227 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 228 #endif
 229
 230 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 231
 232 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 233 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 234                                 hammer2_dev_t *hmp);
 235 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
 236
 237 /*
 238  * HAMMER2 vfs operations.
 239  */
 240 static struct vfsops hammer2_vfsops = {
 241         .vfs_init       = hammer2_vfs_init,
 242         .vfs_uninit     = hammer2_vfs_uninit,
 243         .vfs_sync       = hammer2_vfs_sync,
 244         .vfs_mount      = hammer2_vfs_mount,
 245         .vfs_unmount    = hammer2_vfs_unmount,
 246         .vfs_root       = hammer2_vfs_root,
 247         .vfs_statfs     = hammer2_vfs_statfs,
 248         .vfs_statvfs    = hammer2_vfs_statvfs,
 249         .vfs_vget       = hammer2_vfs_vget,
 250         .vfs_vptofh     = hammer2_vfs_vptofh,
 251         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 252         .vfs_checkexp   = hammer2_vfs_checkexp,
 253         .vfs_modifying  = hammer2_vfs_modifying
 254 };
 255
 256 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 257
 258 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 259 MODULE_VERSION(hammer2, 1);
 260
 261 static
 262 int
 263 hammer2_vfs_init(struct vfsconf *conf)
 264 {
 265         static struct objcache_malloc_args margs_read;
 266         static struct objcache_malloc_args margs_write;
 267         static struct objcache_malloc_args margs_vop;
 268
 269         int error;
 270
 271         error = 0;
 272
 273         /*
 274          * A large DIO cache is needed to retain dedup enablement masks.
 275          * The bulkfree code clears related masks as part of the disk block
 276          * recycling algorithm, preventing it from being used for a later
 277          * dedup.
 278          *
 279          * NOTE: A large buffer cache can actually interfere with dedup
 280          *       operation because we dedup based on media physical buffers
 281          *       and not logical buffers.  Try to make the DIO case large
 282          *       enough to avoid this problem, but also cap it.
 283          */
 284         hammer2_dio_limit = nbuf * 2;
 285         if (hammer2_dio_limit > 100000)
 286                 hammer2_dio_limit = 100000;
 287
 288         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 289                 error = EINVAL;
 290         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 291                 error = EINVAL;
 292         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 293                 error = EINVAL;
 294
 295         if (error)
 296                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 297
 298         margs_read.objsize = 65536;
 299         margs_read.mtype = M_HAMMER2_DEBUFFER;
 300
 301         margs_write.objsize = 32768;
 302         margs_write.mtype = M_HAMMER2_CBUFFER;
 303
 304         margs_vop.objsize = sizeof(hammer2_xop_t);
 305         margs_vop.mtype = M_HAMMER2;
 306
 307         /*
 308          * Note thaht for the XOPS cache we want backing store allocations
 309          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 310          * confusion), so use the backing store function that does it.  This
 311          * means that initial XOPS objects are zerod but REUSED objects are
 312          * not.  So we are responsible for cleaning the object up sufficiently
 313          * for our needs before objcache_put()ing it back (typically just the
 314          * FIFO indices).
 315          */
 316         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 317                                 0, 1, NULL, NULL, NULL,
 318                                 objcache_malloc_alloc,
 319                                 objcache_malloc_free,
 320                                 &margs_read);
 321         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 322                                 0, 1, NULL, NULL, NULL,
 323                                 objcache_malloc_alloc,
 324                                 objcache_malloc_free,
 325                                 &margs_write);
 326         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 327                                 0, 1, NULL, NULL, NULL,
 328                                 objcache_malloc_alloc_zero,
 329                                 objcache_malloc_free,
 330                                 &margs_vop);
 331
 332
 333         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 334         TAILQ_INIT(&hammer2_mntlist);
 335         TAILQ_INIT(&hammer2_pfslist);
 336         TAILQ_INIT(&hammer2_spmplist);
 337
 338         hammer2_limit_dirty_chains = maxvnodes / 10;
 339         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 340                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 341         if (hammer2_limit_dirty_chains < 1000)
 342                 hammer2_limit_dirty_chains = 1000;
 343
 344         hammer2_limit_dirty_inodes = maxvnodes / 25;
 345         if (hammer2_limit_dirty_inodes < 100)
 346                 hammer2_limit_dirty_inodes = 100;
 347         if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
 348                 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
 349
 350         return (error);
 351 }
 352
 353 static
 354 int
 355 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 356 {
 357         objcache_destroy(cache_buffer_read);
 358         objcache_destroy(cache_buffer_write);
 359         objcache_destroy(cache_xops);
 360         return 0;
 361 }
 362
 363 /*
 364  * Core PFS allocator.  Used to allocate or reference the pmp structure
 365  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 366  * The pmp can be passed in or loaded by this function using the chain and
 367  * inode data.
 368  *
 369  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 370  * transactions.  Note that synchronization does not use this field.
 371  * (typically frontend operations and synchronization cannot run on the
 372  * same PFS node at the same time).
 373  *
 374  * XXX check locking
 375  */
 376 hammer2_pfs_t *
 377 hammer2_pfsalloc(hammer2_chain_t *chain,
 378                  const hammer2_inode_data_t *ripdata,
 379                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 380 {
 381         hammer2_pfs_t *pmp;
 382         hammer2_inode_t *iroot;
 383         int count;
 384         int i;
 385         int j;
 386
 387         pmp = NULL;
 388
 389         /*
 390          * Locate or create the PFS based on the cluster id.  If ripdata
 391          * is NULL this is a spmp which is unique and is always allocated.
 392          *
 393          * If the device is mounted in local mode all PFSs are considered
 394          * independent and not part of any cluster (for debugging only).
 395          */
 396         if (ripdata) {
 397                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 398                         if (force_local != pmp->force_local)
 399                                 continue;
 400                         if (force_local == NULL &&
 401                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 402                                  sizeof(pmp->pfs_clid)) == 0) {
 403                                         break;
 404                         } else if (force_local && pmp->pfs_names[0] &&
 405                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 406                                         break;
 407                         }
 408                 }
 409         }
 410
 411         if (pmp == NULL) {
 412                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 413                 pmp->force_local = force_local;
 414                 hammer2_trans_manage_init(pmp);
 415                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 416                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 417                 lockinit(&pmp->lock, "pfslk", 0, 0);
 418                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 419                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 420                 spin_init(&pmp->xop_spin, "h2xop");
 421                 spin_init(&pmp->lru_spin, "h2lru");
 422                 RB_INIT(&pmp->inum_tree);
 423                 TAILQ_INIT(&pmp->syncq);
 424                 TAILQ_INIT(&pmp->depq);
 425                 TAILQ_INIT(&pmp->lru_list);
 426                 spin_init(&pmp->list_spin, "h2pfsalloc_list");
 427
 428                 /*
 429                  * Distribute backend operations to threads
 430                  */
 431                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 432                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 433
 434                 /*
 435                  * Save the last media transaction id for the flusher.  Set
 436                  * initial
 437                  */
 438                 if (ripdata) {
 439                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 440                         TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 441                 } else {
 442                         pmp->flags |= HAMMER2_PMPF_SPMP;
 443                         TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
 444                 }
 445
 446                 /*
 447                  * The synchronization thread may start too early, make
 448                  * sure it stays frozen until we are ready to let it go.
 449                  * XXX
 450                  */
 451                 /*
 452                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 453                                          HAMMER2_THREAD_REMASTER;
 454                 */
 455         }
 456
 457         /*
 458          * Create the PFS's root inode and any missing XOP helper threads.
 459          */
 460         if ((iroot = pmp->iroot) == NULL) {
 461                 iroot = hammer2_inode_get(pmp, NULL, 1, -1);
 462                 if (ripdata)
 463                         iroot->meta = ripdata->meta;
 464                 pmp->iroot = iroot;
 465                 hammer2_inode_ref(iroot);
 466                 hammer2_inode_unlock(iroot);
 467         }
 468
 469         /*
 470          * Stop here if no chain is passed in.
 471          */
 472         if (chain == NULL)
 473                 goto done;
 474
 475         /*
 476          * When a chain is passed in we must add it to the PFS's root
 477          * inode, update pmp->pfs_types[], and update the syncronization
 478          * threads.
 479          *
 480          * When forcing local mode, mark the PFS as a MASTER regardless.
 481          *
 482          * At the moment empty spots can develop due to removals or failures.
 483          * Ultimately we want to re-fill these spots but doing so might
 484          * confused running code. XXX
 485          */
 486         hammer2_inode_ref(iroot);
 487         hammer2_mtx_ex(&iroot->lock);
 488         j = iroot->cluster.nchains;
 489
 490         if (j == HAMMER2_MAXCLUSTER) {
 491                 kprintf("hammer2_mount: cluster full!\n");
 492                 /* XXX fatal error? */
 493         } else {
 494                 KKASSERT(chain->pmp == NULL);
 495                 chain->pmp = pmp;
 496                 hammer2_chain_ref(chain);
 497                 iroot->cluster.array[j].chain = chain;
 498                 if (force_local)
 499                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 500                 else
 501                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 502                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 503                 pmp->pfs_hmps[j] = chain->hmp;
 504                 hammer2_spin_ex(&pmp->inum_spin);
 505                 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset;
 506                 hammer2_spin_unex(&pmp->inum_spin);
 507
 508                 /*
 509                  * If the PFS is already mounted we must account
 510                  * for the mount_count here.
 511                  */
 512                 if (pmp->mp)
 513                         ++chain->hmp->mount_count;
 514
 515                 /*
 516                  * May have to fixup dirty chain tracking.  Previous
 517                  * pmp was NULL so nothing to undo.
 518                  */
 519                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 520                         hammer2_pfs_memory_inc(pmp);
 521                 ++j;
 522         }
 523         iroot->cluster.nchains = j;
 524
 525         /*
 526          * Update nmasters from any PFS inode which is part of the cluster.
 527          * It is possible that this will result in a value which is too
 528          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 529          * override this value later on.
 530          *
 531          * (This informs us of masters that might not currently be
 532          *  discoverable by this mount).
 533          */
 534         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 535                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 536         }
 537
 538         /*
 539          * Count visible masters.  Masters are usually added with
 540          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 541          * are more (XXX and must update the master inodes).
 542          */
 543         count = 0;
 544         for (i = 0; i < iroot->cluster.nchains; ++i) {
 545                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 546                         ++count;
 547         }
 548         if (pmp->pfs_nmasters < count)
 549                 pmp->pfs_nmasters = count;
 550
 551         /*
 552          * Create missing synchronization and support threads.
 553          *
 554          * Single-node masters (including snapshots) have nothing to
 555          * synchronize and do not require this thread.
 556          *
 557          * Multi-node masters or any number of soft masters, slaves, copy,
 558          * or other PFS types need the thread.
 559          *
 560          * Each thread is responsible for its particular cluster index.
 561          * We use independent threads so stalls or mismatches related to
 562          * any given target do not affect other targets.
 563          */
 564         for (i = 0; i < iroot->cluster.nchains; ++i) {
 565                 /*
 566                  * Single-node masters (including snapshots) have nothing
 567                  * to synchronize and will make direct xops support calls,
 568                  * thus they do not require this thread.
 569                  *
 570                  * Note that there can be thousands of snapshots.  We do not
 571                  * want to create thousands of threads.
 572                  */
 573                 if (pmp->pfs_nmasters <= 1 &&
 574                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 575                         continue;
 576                 }
 577
 578                 /*
 579                  * Sync support thread
 580                  */
 581                 if (pmp->sync_thrs[i].td == NULL) {
 582                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
 583                                            "h2nod", i, -1,
 584                                            hammer2_primary_sync_thread);
 585                 }
 586         }
 587
 588         /*
 589          * Create missing Xop threads
 590          *
 591          * NOTE: We create helper threads for all mounted PFSs or any
 592          *       PFSs with 2+ nodes (so the sync thread can update them,
 593          *       even if not mounted).
 594          */
 595         if (pmp->mp || iroot->cluster.nchains >= 2)
 596                 hammer2_xop_helper_create(pmp);
 597
 598         hammer2_mtx_unlock(&iroot->lock);
 599         hammer2_inode_drop(iroot);
 600 done:
 601         return pmp;
 602 }
 603
 604 /*
 605  * Deallocate an element of a probed PFS.  If destroying and this is a
 606  * MASTER, adjust nmasters.
 607  *
 608  * This function does not physically destroy the PFS element in its device
 609  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 610  */
 611 void
 612 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 613 {
 614         hammer2_inode_t *iroot;
 615         hammer2_chain_t *chain;
 616         int j;
 617
 618         /*
 619          * Cleanup our reference on iroot.  iroot is (should) not be needed
 620          * by the flush code.
 621          */
 622         iroot = pmp->iroot;
 623         if (iroot) {
 624                 /*
 625                  * Stop synchronizing
 626                  *
 627                  * XXX flush after acquiring the iroot lock.
 628                  * XXX clean out the cluster index from all inode structures.
 629                  */
 630                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 631
 632                 /*
 633                  * Remove the cluster index from the group.  If destroying
 634                  * the PFS and this is a master, adjust pfs_nmasters.
 635                  */
 636                 hammer2_mtx_ex(&iroot->lock);
 637                 chain = iroot->cluster.array[clindex].chain;
 638                 iroot->cluster.array[clindex].chain = NULL;
 639
 640                 switch(pmp->pfs_types[clindex]) {
 641                 case HAMMER2_PFSTYPE_MASTER:
 642                         if (destroying && pmp->pfs_nmasters > 0)
 643                                 --pmp->pfs_nmasters;
 644                         /* XXX adjust ripdata->meta.pfs_nmasters */
 645                         break;
 646                 default:
 647                         break;
 648                 }
 649                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 650
 651                 hammer2_mtx_unlock(&iroot->lock);
 652
 653                 /*
 654                  * Release the chain.
 655                  */
 656                 if (chain) {
 657                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 658                         hammer2_chain_drop(chain);
 659                 }
 660
 661                 /*
 662                  * Terminate all XOP threads for the cluster index.
 663                  */
 664                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 665                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 666         }
 667 }
 668
 669 /*
 670  * Destroy a PFS, typically only occurs after the last mount on a device
 671  * has gone away.
 672  */
 673 static void
 674 hammer2_pfsfree(hammer2_pfs_t *pmp)
 675 {
 676         hammer2_inode_t *iroot;
 677         hammer2_chain_t *chain;
 678         int chains_still_present = 0;
 679         int i;
 680         int j;
 681
 682         /*
 683          * Cleanup our reference on iroot.  iroot is (should) not be needed
 684          * by the flush code.
 685          */
 686         if (pmp->flags & HAMMER2_PMPF_SPMP)
 687                 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
 688         else
 689                 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 690
 691         /*
 692          * Cleanup chains remaining on LRU list.
 693          */
 694         hammer2_spin_ex(&pmp->lru_spin);
 695         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 696                 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
 697                 atomic_add_int(&pmp->lru_count, -1);
 698                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
 699                 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
 700                 hammer2_chain_ref(chain);
 701                 hammer2_spin_unex(&pmp->lru_spin);
 702                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 703                 hammer2_chain_drop(chain);
 704                 hammer2_spin_ex(&pmp->lru_spin);
 705         }
 706         hammer2_spin_unex(&pmp->lru_spin);
 707
 708         /*
 709          * Clean up iroot
 710          */
 711         iroot = pmp->iroot;
 712         if (iroot) {
 713                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 714                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 715                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 716                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 717                         chain = iroot->cluster.array[i].chain;
 718                         if (chain && !RB_EMPTY(&chain->core.rbtree)) {
 719                                 kprintf("hammer2: Warning pmp %p still "
 720                                         "has active chains\n", pmp);
 721                                 chains_still_present = 1;
 722                         }
 723                 }
 724 #if REPORT_REFS_ERRORS
 725                 if (iroot->refs != 1)
 726                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 727                                 iroot, iroot->refs);
 728 #else
 729                 KKASSERT(iroot->refs == 1);
 730 #endif
 731                 /* ref for iroot */
 732                 hammer2_inode_drop(iroot);
 733                 pmp->iroot = NULL;
 734         }
 735
 736         /*
 737          * Free remaining pmp resources
 738          */
 739         if (chains_still_present) {
 740                 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp);
 741         } else {
 742                 kmalloc_destroy(&pmp->mmsg);
 743                 kmalloc_destroy(&pmp->minode);
 744                 kfree(pmp, M_HAMMER2);
 745         }
 746 }
 747
 748 /*
 749  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 750  * empty is terminated and freed.
 751  *
 752  * XXX inefficient.
 753  */
 754 static void
 755 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
 756 {
 757         hammer2_pfs_t *pmp;
 758         hammer2_inode_t *iroot;
 759         hammer2_chain_t *rchain;
 760         int i;
 761         int j;
 762         struct hammer2_pfslist *wlist;
 763
 764         if (which == 0)
 765                 wlist = &hammer2_pfslist;
 766         else
 767                 wlist = &hammer2_spmplist;
 768 again:
 769         TAILQ_FOREACH(pmp, wlist, mntentry) {
 770                 if ((iroot = pmp->iroot) == NULL)
 771                         continue;
 772
 773                 /*
 774                  * Determine if this PFS is affected.  If it is we must
 775                  * freeze all management threads and lock its iroot.
 776                  *
 777                  * Freezing a management thread forces it idle, operations
 778                  * in-progress will be aborted and it will have to start
 779                  * over again when unfrozen, or exit if told to exit.
 780                  */
 781                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 782                         if (pmp->pfs_hmps[i] == hmp)
 783                                 break;
 784                 }
 785                 if (i == HAMMER2_MAXCLUSTER)
 786                         continue;
 787
 788                 hammer2_vfs_sync_pmp(pmp, MNT_WAIT);
 789
 790                 /*
 791                  * Make sure all synchronization threads are locked
 792                  * down.
 793                  */
 794                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 795                         if (pmp->pfs_hmps[i] == NULL)
 796                                 continue;
 797                         hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 798                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 799                                 hammer2_thr_freeze_async(
 800                                         &pmp->xop_groups[j].thrs[i]);
 801                         }
 802                 }
 803                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 804                         if (pmp->pfs_hmps[i] == NULL)
 805                                 continue;
 806                         hammer2_thr_freeze(&pmp->sync_thrs[i]);
 807                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 808                                 hammer2_thr_freeze(
 809                                         &pmp->xop_groups[j].thrs[i]);
 810                         }
 811                 }
 812
 813                 /*
 814                  * Lock the inode and clean out matching chains.
 815                  * Note that we cannot use hammer2_inode_lock_*()
 816                  * here because that would attempt to validate the
 817                  * cluster that we are in the middle of ripping
 818                  * apart.
 819                  *
 820                  * WARNING! We are working directly on the inodes
 821                  *          embedded cluster.
 822                  */
 823                 hammer2_mtx_ex(&iroot->lock);
 824
 825                 /*
 826                  * Remove the chain from matching elements of the PFS.
 827                  */
 828                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 829                         if (pmp->pfs_hmps[i] != hmp)
 830                                 continue;
 831                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 832                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 833                                 hammer2_thr_delete(
 834                                         &pmp->xop_groups[j].thrs[i]);
 835                         }
 836                         rchain = iroot->cluster.array[i].chain;
 837                         iroot->cluster.array[i].chain = NULL;
 838                         pmp->pfs_types[i] = 0;
 839                         if (pmp->pfs_names[i]) {
 840                                 kfree(pmp->pfs_names[i], M_HAMMER2);
 841                                 pmp->pfs_names[i] = NULL;
 842                         }
 843                         if (rchain) {
 844                                 hammer2_chain_drop(rchain);
 845                                 /* focus hint */
 846                                 if (iroot->cluster.focus == rchain)
 847                                         iroot->cluster.focus = NULL;
 848                         }
 849                         pmp->pfs_hmps[i] = NULL;
 850                 }
 851                 hammer2_mtx_unlock(&iroot->lock);
 852
 853                 /*
 854                  * Cleanup trailing chains.  Gaps may remain.
 855                  */
 856                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 857                         if (pmp->pfs_hmps[i])
 858                                 break;
 859                 }
 860                 iroot->cluster.nchains = i + 1;
 861
 862                 /*
 863                  * If the PMP has no elements remaining we can destroy it.
 864                  * (this will transition management threads from frozen->exit).
 865                  */
 866                 if (iroot->cluster.nchains == 0) {
 867                         /*
 868                          * If this was the hmp's spmp, we need to clean
 869                          * a little more stuff out.
 870                          */
 871                         if (hmp->spmp == pmp) {
 872                                 hmp->spmp = NULL;
 873                                 hmp->vchain.pmp = NULL;
 874                                 hmp->fchain.pmp = NULL;
 875                         }
 876
 877                         /*
 878                          * Free the pmp and restart the loop
 879                          */
 880                         KKASSERT(TAILQ_EMPTY(&pmp->syncq));
 881                         KKASSERT(TAILQ_EMPTY(&pmp->depq));
 882                         hammer2_pfsfree(pmp);
 883                         goto again;
 884                 }
 885
 886                 /*
 887                  * If elements still remain we need to set the REMASTER
 888                  * flag and unfreeze it.
 889                  */
 890                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 891                         if (pmp->pfs_hmps[i] == NULL)
 892                                 continue;
 893                         hammer2_thr_remaster(&pmp->sync_thrs[i]);
 894                         hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 895                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 896                                 hammer2_thr_remaster(
 897                                         &pmp->xop_groups[j].thrs[i]);
 898                                 hammer2_thr_unfreeze(
 899                                         &pmp->xop_groups[j].thrs[i]);
 900                         }
 901                 }
 902         }
 903 }
 904
 905 /*
 906  * Mount or remount HAMMER2 fileystem from physical media
 907  *
 908  *      mountroot
 909  *              mp              mount point structure
 910  *              path            NULL
 911  *              data            <unused>
 912  *              cred            <unused>
 913  *
 914  *      mount
 915  *              mp              mount point structure
 916  *              path            path to mount point
 917  *              data            pointer to argument structure in user space
 918  *                      volume  volume path (device@LABEL form)
 919  *                      hflags  user mount flags
 920  *              cred            user credentials
 921  *
 922  * RETURNS:     0       Success
 923  *              !0      error number
 924  */
 925 static
 926 int
 927 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 928                   struct ucred *cred)
 929 {
 930         struct hammer2_mount_info info;
 931         hammer2_pfs_t *pmp;
 932         hammer2_pfs_t *spmp;
 933         hammer2_dev_t *hmp;
 934         hammer2_dev_t *force_local;
 935         hammer2_key_t key_next;
 936         hammer2_key_t key_dummy;
 937         hammer2_key_t lhc;
 938         struct vnode *devvp;
 939         struct nlookupdata nd;
 940         hammer2_chain_t *parent;
 941         hammer2_chain_t *chain;
 942         const hammer2_inode_data_t *ripdata;
 943         hammer2_blockref_t bref;
 944         struct file *fp;
 945         char devstr[MNAMELEN];
 946         size_t size;
 947         size_t done;
 948         char *dev;
 949         char *label;
 950         int ronly = 1;
 951         int error;
 952         int i;
 953
 954         hmp = NULL;
 955         pmp = NULL;
 956         dev = NULL;
 957         label = NULL;
 958         devvp = NULL;
 959
 960         if (path == NULL) {
 961                 /*
 962                  * Root mount
 963                  */
 964                 bzero(&info, sizeof(info));
 965                 info.cluster_fd = -1;
 966                 ksnprintf(devstr, sizeof(devstr), "%s",
 967                           mp->mnt_stat.f_mntfromname);
 968                 kprintf("hammer2_mount: root '%s'\n", devstr);
 969                 done = strlen(devstr) + 1;
 970         } else {
 971                 /*
 972                  * Non-root mount or updating a mount
 973                  */
 974                 error = copyin(data, &info, sizeof(info));
 975                 if (error)
 976                         return (error);
 977
 978                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 979                 if (error)
 980                         return (error);
 981                 kprintf("hammer2_mount: '%s'\n", devstr);
 982         }
 983
 984         /*
 985          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
 986          * if no label specified, based on the partition id.  Error out if no
 987          * label or device (with partition id) is specified.  This is strictly
 988          * a convenience to match the default label created by newfs_hammer2,
 989          * our preference is that a label always be specified.
 990          *
 991          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
 992          *       that does not specify a device, as long as some H2 label
 993          *       has already been mounted from that device.  This makes
 994          *       mounting snapshots a lot easier.
 995          */
 996         dev = devstr;
 997         label = strchr(devstr, '@');
 998         if (label && ((label + 1) - dev) > done) {
 999                 kprintf("hammer2: mount: bad label %s/%zd\n",
1000                         devstr, done);
1001                 return (EINVAL);
1002         }
1003         if (label == NULL || label[1] == 0) {
1004                 char slice;
1005
1006                 if (label == NULL)
1007                         label = devstr + strlen(devstr);
1008                 else
1009                         *label = '\0';          /* clean up trailing @ */
1010
1011                 slice = label[-1];
1012                 switch(slice) {
1013                 case 'a':
1014                         label = "BOOT";
1015                         break;
1016                 case 'd':
1017                         label = "ROOT";
1018                         break;
1019                 default:
1020                         label = "DATA";
1021                         break;
1022                 }
1023         } else {
1024                 *label = '\0';
1025                 label++;
1026         }
1027
1028         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
1029                 dev, label, (mp->mnt_flag & MNT_RDONLY));
1030
1031         if (mp->mnt_flag & MNT_UPDATE) {
1032                 /*
1033                  * Update mount.  Note that pmp->iroot->cluster is
1034                  * an inode-embedded cluster and thus cannot be
1035                  * directly locked.
1036                  *
1037                  * XXX HAMMER2 needs to implement NFS export via
1038                  *     mountctl.
1039                  */
1040                 hammer2_cluster_t *cluster;
1041
1042                 pmp = MPTOPMP(mp);
1043                 pmp->hflags = info.hflags;
1044                 cluster = &pmp->iroot->cluster;
1045                 for (i = 0; i < cluster->nchains; ++i) {
1046                         if (cluster->array[i].chain == NULL)
1047                                 continue;
1048                         hmp = cluster->array[i].chain->hmp;
1049                         devvp = hmp->devvp;
1050                         error = hammer2_remount(hmp, mp, path,
1051                                                 devvp, cred);
1052                         if (error)
1053                                 break;
1054                 }
1055
1056                 return error;
1057         }
1058
1059         /*
1060          * HMP device mount
1061          *
1062          * If a path is specified and dev is not an empty string, lookup the
1063          * name and verify that it referes to a block device.
1064          *
1065          * If a path is specified and dev is an empty string we fall through
1066          * and locate the label in the hmp search.
1067          */
1068         if (path && *dev != 0) {
1069                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1070                 if (error == 0)
1071                         error = nlookup(&nd);
1072                 if (error == 0)
1073                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1074                 nlookup_done(&nd);
1075         } else if (path == NULL) {
1076                 /* root mount */
1077                 cdev_t cdev = kgetdiskbyname(dev);
1078                 error = bdevvp(cdev, &devvp);
1079                 if (error)
1080                         kprintf("hammer2: cannot find '%s'\n", dev);
1081         } else {
1082                 /*
1083                  * We will locate the hmp using the label in the hmp loop.
1084                  */
1085                 error = 0;
1086         }
1087
1088         /*
1089          * Make sure its a block device.  Do not check to see if it is
1090          * already mounted until we determine that its a fresh H2 device.
1091          */
1092         if (error == 0 && devvp) {
1093                 vn_isdisk(devvp, &error);
1094         }
1095
1096         /*
1097          * Determine if the device has already been mounted.  After this
1098          * check hmp will be non-NULL if we are doing the second or more
1099          * hammer2 mounts from the same device.
1100          */
1101         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1102         if (devvp) {
1103                 /*
1104                  * Match the device.  Due to the way devfs works,
1105                  * we may not be able to directly match the vnode pointer,
1106                  * so also check to see if the underlying device matches.
1107                  */
1108                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1109                         if (hmp->devvp == devvp)
1110                                 break;
1111                         if (devvp->v_rdev &&
1112                             hmp->devvp->v_rdev == devvp->v_rdev) {
1113                                 break;
1114                         }
1115                 }
1116
1117                 /*
1118                  * If no match this may be a fresh H2 mount, make sure
1119                  * the device is not mounted on anything else.
1120                  */
1121                 if (hmp == NULL)
1122                         error = vfs_mountedon(devvp);
1123         } else if (error == 0) {
1124                 /*
1125                  * Match the label to a pmp already probed.
1126                  */
1127                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1128                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1129                                 if (pmp->pfs_names[i] &&
1130                                     strcmp(pmp->pfs_names[i], label) == 0) {
1131                                         hmp = pmp->pfs_hmps[i];
1132                                         break;
1133                                 }
1134                         }
1135                         if (hmp)
1136                                 break;
1137                 }
1138                 if (hmp == NULL)
1139                         error = ENOENT;
1140         }
1141
1142         /*
1143          * Open the device if this isn't a secondary mount and construct
1144          * the H2 device mount (hmp).
1145          */
1146         if (hmp == NULL) {
1147                 hammer2_chain_t *schain;
1148                 hammer2_xid_t xid;
1149                 hammer2_xop_head_t xop;
1150
1151                 if (error == 0 && vcount(devvp) > 0) {
1152                         kprintf("Primary device already has references\n");
1153                         error = EBUSY;
1154                 }
1155
1156                 /*
1157                  * Now open the device
1158                  */
1159                 if (error == 0) {
1160                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1161                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1162                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1163                         if (error == 0) {
1164                                 error = VOP_OPEN(devvp,
1165                                              (ronly ? FREAD : FREAD | FWRITE),
1166                                              FSCRED, NULL);
1167                         }
1168                         vn_unlock(devvp);
1169                 }
1170                 if (error && devvp) {
1171                         vrele(devvp);
1172                         devvp = NULL;
1173                 }
1174                 if (error) {
1175                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1176                         return error;
1177                 }
1178                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1179                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1180                 hmp->ronly = ronly;
1181                 hmp->devvp = devvp;
1182                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1183                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1184                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1185                 RB_INIT(&hmp->iotree);
1186                 spin_init(&hmp->io_spin, "h2mount_io");
1187                 spin_init(&hmp->list_spin, "h2mount_list");
1188
1189                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1190                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1191                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1192
1193                 /*
1194                  * vchain setup. vchain.data is embedded.
1195                  * vchain.refs is initialized and will never drop to 0.
1196                  *
1197                  * NOTE! voldata is not yet loaded.
1198                  */
1199                 hmp->vchain.hmp = hmp;
1200                 hmp->vchain.refs = 1;
1201                 hmp->vchain.data = (void *)&hmp->voldata;
1202                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1203                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1204                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1205
1206                 hammer2_chain_core_init(&hmp->vchain);
1207                 /* hmp->vchain.u.xxx is left NULL */
1208
1209                 /*
1210                  * fchain setup.  fchain.data is embedded.
1211                  * fchain.refs is initialized and will never drop to 0.
1212                  *
1213                  * The data is not used but needs to be initialized to
1214                  * pass assertion muster.  We use this chain primarily
1215                  * as a placeholder for the freemap's top-level RBTREE
1216                  * so it does not interfere with the volume's topology
1217                  * RBTREE.
1218                  */
1219                 hmp->fchain.hmp = hmp;
1220                 hmp->fchain.refs = 1;
1221                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1222                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1223                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1224                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1225                 hmp->fchain.bref.methods =
1226                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1227                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1228
1229                 hammer2_chain_core_init(&hmp->fchain);
1230                 /* hmp->fchain.u.xxx is left NULL */
1231
1232                 /*
1233                  * Install the volume header and initialize fields from
1234                  * voldata.
1235                  */
1236                 error = hammer2_install_volume_header(hmp);
1237                 if (error) {
1238                         hammer2_unmount_helper(mp, NULL, hmp);
1239                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1240                         hammer2_vfs_unmount(mp, MNT_FORCE);
1241                         return error;
1242                 }
1243
1244                 /*
1245                  * Really important to get these right or flush will get
1246                  * confused.
1247                  */
1248                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1249                 spmp = hmp->spmp;
1250
1251                 /*
1252                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1253                  * is inherited from the volume header.
1254                  */
1255                 xid = 0;
1256                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1257                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1258                 hmp->vchain.pmp = spmp;
1259                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1260                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1261                 hmp->fchain.pmp = spmp;
1262
1263                 /*
1264                  * First locate the super-root inode, which is key 0
1265                  * relative to the volume header's blockset.
1266                  *
1267                  * Then locate the root inode by scanning the directory keyspace
1268                  * represented by the label.
1269                  */
1270                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1271                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1272                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1273                                       &error, 0);
1274                 hammer2_chain_lookup_done(parent);
1275                 if (schain == NULL) {
1276                         kprintf("hammer2_mount: invalid super-root\n");
1277                         hammer2_unmount_helper(mp, NULL, hmp);
1278                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1279                         hammer2_vfs_unmount(mp, MNT_FORCE);
1280                         return EINVAL;
1281                 }
1282                 if (schain->error) {
1283                         kprintf("hammer2_mount: error %s reading super-root\n",
1284                                 hammer2_error_str(schain->error));
1285                         hammer2_chain_unlock(schain);
1286                         hammer2_chain_drop(schain);
1287                         schain = NULL;
1288                         hammer2_unmount_helper(mp, NULL, hmp);
1289                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1290                         hammer2_vfs_unmount(mp, MNT_FORCE);
1291                         return EINVAL;
1292                 }
1293
1294                 /*
1295                  * The super-root always uses an inode_tid of 1 when
1296                  * creating PFSs.
1297                  */
1298                 spmp->inode_tid = 1;
1299                 spmp->modify_tid = schain->bref.modify_tid + 1;
1300
1301                 /*
1302                  * Sanity-check schain's pmp and finish initialization.
1303                  * Any chain belonging to the super-root topology should
1304                  * have a NULL pmp (not even set to spmp).
1305                  */
1306                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1307                 KKASSERT(schain->pmp == NULL);
1308                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1309
1310                 /*
1311                  * Replace the dummy spmp->iroot with a real one.  It's
1312                  * easier to just do a wholesale replacement than to try
1313                  * to update the chain and fixup the iroot fields.
1314                  *
1315                  * The returned inode is locked with the supplied cluster.
1316                  */
1317                 hammer2_dummy_xop_from_chain(&xop, schain);
1318                 hammer2_inode_drop(spmp->iroot);
1319                 spmp->iroot = NULL;
1320                 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1);
1321                 spmp->spmp_hmp = hmp;
1322                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1323                 spmp->pfs_hmps[0] = hmp;
1324                 hammer2_inode_ref(spmp->iroot);
1325                 hammer2_inode_unlock(spmp->iroot);
1326                 hammer2_cluster_unlock(&xop.cluster);
1327                 hammer2_chain_drop(schain);
1328                 /* do not call hammer2_cluster_drop() on an embedded cluster */
1329                 schain = NULL;  /* now invalid */
1330                 /* leave spmp->iroot with one ref */
1331
1332                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1333                         error = hammer2_recovery(hmp);
1334                         if (error == 0)
1335                                 error |= hammer2_fixup_pfses(hmp);
1336                         /* XXX do something with error */
1337                 }
1338                 hammer2_update_pmps(hmp);
1339                 hammer2_iocom_init(hmp);
1340                 hammer2_bulkfree_init(hmp);
1341
1342                 /*
1343                  * Ref the cluster management messaging descriptor.  The mount
1344                  * program deals with the other end of the communications pipe.
1345                  *
1346                  * Root mounts typically do not supply one.
1347                  */
1348                 if (info.cluster_fd >= 0) {
1349                         fp = holdfp(curthread, info.cluster_fd, -1);
1350                         if (fp) {
1351                                 hammer2_cluster_reconnect(hmp, fp);
1352                         } else {
1353                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1354                         }
1355                 }
1356         } else {
1357                 spmp = hmp->spmp;
1358                 if (info.hflags & HMNT2_DEVFLAGS) {
1359                         kprintf("hammer2: Warning: mount flags pertaining "
1360                                 "to the whole device may only be specified "
1361                                 "on the first mount of the device: %08x\n",
1362                                 info.hflags & HMNT2_DEVFLAGS);
1363                 }
1364         }
1365
1366         /*
1367          * Force local mount (disassociate all PFSs from their clusters).
1368          * Used primarily for debugging.
1369          */
1370         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1371
1372         /*
1373          * Lookup the mount point under the media-localized super-root.
1374          * Scanning hammer2_pfslist doesn't help us because it represents
1375          * PFS cluster ids which can aggregate several named PFSs together.
1376          *
1377          * cluster->pmp will incorrectly point to spmp and must be fixed
1378          * up later on.
1379          */
1380         hammer2_inode_lock(spmp->iroot, 0);
1381         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1382         lhc = hammer2_dirhash(label, strlen(label));
1383         chain = hammer2_chain_lookup(&parent, &key_next,
1384                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1385                                      &error, 0);
1386         while (chain) {
1387                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1388                     strcmp(label, chain->data->ipdata.filename) == 0) {
1389                         break;
1390                 }
1391                 chain = hammer2_chain_next(&parent, chain, &key_next,
1392                                             key_next,
1393                                             lhc + HAMMER2_DIRHASH_LOMASK,
1394                                             &error, 0);
1395         }
1396         if (parent) {
1397                 hammer2_chain_unlock(parent);
1398                 hammer2_chain_drop(parent);
1399         }
1400         hammer2_inode_unlock(spmp->iroot);
1401
1402         /*
1403          * PFS could not be found?
1404          */
1405         if (chain == NULL) {
1406                 if (error)
1407                         kprintf("hammer2_mount: PFS label I/O error\n");
1408                 else
1409                         kprintf("hammer2_mount: PFS label not found\n");
1410                 hammer2_unmount_helper(mp, NULL, hmp);
1411                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1412                 hammer2_vfs_unmount(mp, MNT_FORCE);
1413
1414                 return EINVAL;
1415         }
1416
1417         /*
1418          * Acquire the pmp structure (it should have already been allocated
1419          * via hammer2_update_pmps() so do not pass cluster in to add to
1420          * available chains).
1421          *
1422          * Check if the cluster has already been mounted.  A cluster can
1423          * only be mounted once, use null mounts to mount additional copies.
1424          */
1425         if (chain->error) {
1426                 kprintf("hammer2_mount: PFS label I/O error\n");
1427         } else {
1428                 ripdata = &chain->data->ipdata;
1429                 bref = chain->bref;
1430                 pmp = hammer2_pfsalloc(NULL, ripdata,
1431                                        bref.modify_tid, force_local);
1432         }
1433         hammer2_chain_unlock(chain);
1434         hammer2_chain_drop(chain);
1435
1436         /*
1437          * Finish the mount
1438          */
1439         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1440
1441         if (pmp->mp) {
1442                 kprintf("hammer2_mount: PFS already mounted!\n");
1443                 hammer2_unmount_helper(mp, NULL, hmp);
1444                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1445                 hammer2_vfs_unmount(mp, MNT_FORCE);
1446
1447                 return EBUSY;
1448         }
1449
1450         pmp->hflags = info.hflags;
1451         mp->mnt_flag |= MNT_LOCAL;
1452         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1453         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1454
1455         /*
1456          * required mount structure initializations
1457          */
1458         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1459         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1460
1461         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1462         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1463
1464         /*
1465          * Optional fields
1466          */
1467         mp->mnt_iosize_max = MAXPHYS;
1468
1469         /*
1470          * Connect up mount pointers.
1471          */
1472         hammer2_mount_helper(mp, pmp);
1473
1474         lockmgr(&hammer2_mntlk, LK_RELEASE);
1475
1476         /*
1477          * Finish setup
1478          */
1479         vfs_getnewfsid(mp);
1480         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1481         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1482         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1483
1484         if (path) {
1485                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1486                           MNAMELEN - 1, &size);
1487                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1488         } /* else root mount, already in there */
1489
1490         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1491         if (path) {
1492                 copyinstr(path, mp->mnt_stat.f_mntonname,
1493                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1494                           &size);
1495         } else {
1496                 /* root mount */
1497                 mp->mnt_stat.f_mntonname[0] = '/';
1498         }
1499
1500         /*
1501          * Initial statfs to prime mnt_stat.
1502          */
1503         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1504
1505         return 0;
1506 }
1507
1508 /*
1509  * Scan PFSs under the super-root and create hammer2_pfs structures.
1510  */
1511 static
1512 void
1513 hammer2_update_pmps(hammer2_dev_t *hmp)
1514 {
1515         const hammer2_inode_data_t *ripdata;
1516         hammer2_chain_t *parent;
1517         hammer2_chain_t *chain;
1518         hammer2_blockref_t bref;
1519         hammer2_dev_t *force_local;
1520         hammer2_pfs_t *spmp;
1521         hammer2_pfs_t *pmp;
1522         hammer2_key_t key_next;
1523         int error;
1524
1525         /*
1526          * Force local mount (disassociate all PFSs from their clusters).
1527          * Used primarily for debugging.
1528          */
1529         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1530
1531         /*
1532          * Lookup mount point under the media-localized super-root.
1533          *
1534          * cluster->pmp will incorrectly point to spmp and must be fixed
1535          * up later on.
1536          */
1537         spmp = hmp->spmp;
1538         hammer2_inode_lock(spmp->iroot, 0);
1539         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1540         chain = hammer2_chain_lookup(&parent, &key_next,
1541                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1542                                          &error, 0);
1543         while (chain) {
1544                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1545                         continue;
1546                 if (chain->error) {
1547                         kprintf("I/O error scanning PFS labels\n");
1548                 } else {
1549                         ripdata = &chain->data->ipdata;
1550                         bref = chain->bref;
1551
1552                         pmp = hammer2_pfsalloc(chain, ripdata,
1553                                                bref.modify_tid, force_local);
1554                 }
1555                 chain = hammer2_chain_next(&parent, chain, &key_next,
1556                                            key_next, HAMMER2_KEY_MAX,
1557                                            &error, 0);
1558         }
1559         if (parent) {
1560                 hammer2_chain_unlock(parent);
1561                 hammer2_chain_drop(parent);
1562         }
1563         hammer2_inode_unlock(spmp->iroot);
1564 }
1565
1566 static
1567 int
1568 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1569                 struct vnode *devvp, struct ucred *cred)
1570 {
1571         int error;
1572
1573         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1574                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1575                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1576                 vn_unlock(devvp);
1577                 error = hammer2_recovery(hmp);
1578                 if (error == 0)
1579                         error |= hammer2_fixup_pfses(hmp);
1580                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1581                 if (error == 0) {
1582                         VOP_CLOSE(devvp, FREAD, NULL);
1583                         hmp->ronly = 0;
1584                 } else {
1585                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1586                 }
1587                 vn_unlock(devvp);
1588         } else {
1589                 error = 0;
1590         }
1591         return error;
1592 }
1593
1594 static
1595 int
1596 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1597 {
1598         hammer2_pfs_t *pmp;
1599         int flags;
1600         int error = 0;
1601
1602         pmp = MPTOPMP(mp);
1603
1604         if (pmp == NULL)
1605                 return(0);
1606
1607         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1608
1609         /*
1610          * If mount initialization proceeded far enough we must flush
1611          * its vnodes and sync the underlying mount points.  Three syncs
1612          * are required to fully flush the filesystem (freemap updates lag
1613          * by one flush, and one extra for safety).
1614          */
1615         if (mntflags & MNT_FORCE)
1616                 flags = FORCECLOSE;
1617         else
1618                 flags = 0;
1619         if (pmp->iroot) {
1620                 error = vflush(mp, 0, flags);
1621                 if (error)
1622                         goto failed;
1623                 hammer2_vfs_sync(mp, MNT_WAIT);
1624                 hammer2_vfs_sync(mp, MNT_WAIT);
1625                 hammer2_vfs_sync(mp, MNT_WAIT);
1626         }
1627
1628         /*
1629          * Cleanup the frontend support XOPS threads
1630          */
1631         hammer2_xop_helper_cleanup(pmp);
1632
1633         if (pmp->mp)
1634                 hammer2_unmount_helper(mp, pmp, NULL);
1635
1636         error = 0;
1637 failed:
1638         lockmgr(&hammer2_mntlk, LK_RELEASE);
1639
1640         return (error);
1641 }
1642
1643 /*
1644  * Mount helper, hook the system mount into our PFS.
1645  * The mount lock is held.
1646  *
1647  * We must bump the mount_count on related devices for any
1648  * mounted PFSs.
1649  */
1650 static
1651 void
1652 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1653 {
1654         hammer2_cluster_t *cluster;
1655         hammer2_chain_t *rchain;
1656         int i;
1657
1658         mp->mnt_data = (qaddr_t)pmp;
1659         pmp->mp = mp;
1660
1661         /*
1662          * After pmp->mp is set we have to adjust hmp->mount_count.
1663          */
1664         cluster = &pmp->iroot->cluster;
1665         for (i = 0; i < cluster->nchains; ++i) {
1666                 rchain = cluster->array[i].chain;
1667                 if (rchain == NULL)
1668                         continue;
1669                 ++rchain->hmp->mount_count;
1670         }
1671
1672         /*
1673          * Create missing Xop threads
1674          */
1675         hammer2_xop_helper_create(pmp);
1676 }
1677
1678 /*
1679  * Mount helper, unhook the system mount from our PFS.
1680  * The mount lock is held.
1681  *
1682  * If hmp is supplied a mount responsible for being the first to open
1683  * the block device failed and the block device and all PFSs using the
1684  * block device must be cleaned up.
1685  *
1686  * If pmp is supplied multiple devices might be backing the PFS and each
1687  * must be disconnected.  This might not be the last PFS using some of the
1688  * underlying devices.  Also, we have to adjust our hmp->mount_count
1689  * accounting for the devices backing the pmp which is now undergoing an
1690  * unmount.
1691  */
1692 static
1693 void
1694 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1695 {
1696         hammer2_cluster_t *cluster;
1697         hammer2_chain_t *rchain;
1698         struct vnode *devvp;
1699         int dumpcnt;
1700         int ronly;
1701         int i;
1702
1703         /*
1704          * If no device supplied this is a high-level unmount and we have to
1705          * to disconnect the mount, adjust mount_count, and locate devices
1706          * that might now have no mounts.
1707          */
1708         if (pmp) {
1709                 KKASSERT(hmp == NULL);
1710                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1711                 pmp->mp = NULL;
1712                 mp->mnt_data = NULL;
1713
1714                 /*
1715                  * After pmp->mp is cleared we have to account for
1716                  * mount_count.
1717                  */
1718                 cluster = &pmp->iroot->cluster;
1719                 for (i = 0; i < cluster->nchains; ++i) {
1720                         rchain = cluster->array[i].chain;
1721                         if (rchain == NULL)
1722                                 continue;
1723                         --rchain->hmp->mount_count;
1724                         /* scrapping hmp now may invalidate the pmp */
1725                 }
1726 again:
1727                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1728                         if (hmp->mount_count == 0) {
1729                                 hammer2_unmount_helper(NULL, NULL, hmp);
1730                                 goto again;
1731                         }
1732                 }
1733                 return;
1734         }
1735
1736         /*
1737          * Try to terminate the block device.  We can't terminate it if
1738          * there are still PFSs referencing it.
1739          */
1740         if (hmp->mount_count)
1741                 return;
1742
1743         /*
1744          * Decomission the network before we start messing with the
1745          * device and PFS.
1746          */
1747         hammer2_iocom_uninit(hmp);
1748
1749         hammer2_bulkfree_uninit(hmp);
1750         hammer2_pfsfree_scan(hmp, 0);
1751 #if 0
1752         hammer2_dev_exlock(hmp);        /* XXX order */
1753 #endif
1754
1755         /*
1756          * Cycle the volume data lock as a safety (probably not needed any
1757          * more).  To ensure everything is out we need to flush at least
1758          * three times.  (1) The running of the sideq can dirty the
1759          * filesystem, (2) A normal flush can dirty the freemap, and
1760          * (3) ensure that the freemap is fully synchronized.
1761          *
1762          * The next mount's recovery scan can clean everything up but we want
1763          * to leave the filesystem in a 100% clean state on a normal unmount.
1764          */
1765 #if 0
1766         hammer2_voldata_lock(hmp);
1767         hammer2_voldata_unlock(hmp);
1768 #endif
1769
1770         /*
1771          * Flush whatever is left.  Unmounted but modified PFS's might still
1772          * have some dirty chains on them.
1773          */
1774         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1775         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1776
1777         if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1778                 hammer2_voldata_modify(hmp);
1779                 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1780                                             HAMMER2_FLUSH_ALL);
1781         }
1782         hammer2_chain_unlock(&hmp->fchain);
1783
1784         if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1785                 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1786                                             HAMMER2_FLUSH_ALL);
1787         }
1788         hammer2_chain_unlock(&hmp->vchain);
1789
1790         if ((hmp->vchain.flags | hmp->fchain.flags) &
1791             HAMMER2_CHAIN_FLUSH_MASK) {
1792                 kprintf("hammer2_unmount: chains left over "
1793                         "after final sync\n");
1794                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1795                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1796
1797                 if (hammer2_debug & 0x0010)
1798                         Debugger("entered debugger");
1799         }
1800
1801         hammer2_pfsfree_scan(hmp, 1);
1802
1803         KKASSERT(hmp->spmp == NULL);
1804
1805         /*
1806          * Finish up with the device vnode
1807          */
1808         if ((devvp = hmp->devvp) != NULL) {
1809                 ronly = hmp->ronly;
1810                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1811                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1812                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1813                         ronly);
1814                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1815                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1816                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1817                 hmp->devvp = NULL;
1818                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1819                 vn_unlock(devvp);
1820                 vrele(devvp);
1821                 devvp = NULL;
1822         }
1823
1824         /*
1825          * Clear vchain/fchain flags that might prevent final cleanup
1826          * of these chains.
1827          */
1828         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1829                 atomic_add_long(&hammer2_count_modified_chains, -1);
1830                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1831                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1832         }
1833         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1834                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1835         }
1836
1837         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1838                 atomic_add_long(&hammer2_count_modified_chains, -1);
1839                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1840                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1841         }
1842         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1843                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1844         }
1845
1846         /*
1847          * Final drop of embedded freemap root chain to
1848          * clean up fchain.core (fchain structure is not
1849          * flagged ALLOCATED so it is cleaned out and then
1850          * left to rot).
1851          */
1852         hammer2_chain_drop(&hmp->fchain);
1853
1854         /*
1855          * Final drop of embedded volume root chain to clean
1856          * up vchain.core (vchain structure is not flagged
1857          * ALLOCATED so it is cleaned out and then left to
1858          * rot).
1859          */
1860         dumpcnt = 50;
1861         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1862         dumpcnt = 50;
1863         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1864 #if 0
1865         hammer2_dev_unlock(hmp);
1866 #endif
1867         hammer2_chain_drop(&hmp->vchain);
1868
1869         hammer2_io_cleanup(hmp, &hmp->iotree);
1870         if (hmp->iofree_count) {
1871                 kprintf("io_cleanup: %d I/O's left hanging\n",
1872                         hmp->iofree_count);
1873         }
1874
1875         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1876         kmalloc_destroy(&hmp->mchain);
1877         kfree(hmp, M_HAMMER2);
1878 }
1879
1880 int
1881 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1882                  ino_t ino, struct vnode **vpp)
1883 {
1884         hammer2_xop_lookup_t *xop;
1885         hammer2_pfs_t *pmp;
1886         hammer2_inode_t *ip;
1887         hammer2_tid_t inum;
1888         int error;
1889
1890         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1891
1892         error = 0;
1893         pmp = MPTOPMP(mp);
1894
1895         /*
1896          * Easy if we already have it cached
1897          */
1898         ip = hammer2_inode_lookup(pmp, inum);
1899         if (ip) {
1900                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1901                 *vpp = hammer2_igetv(ip, &error);
1902                 hammer2_inode_unlock(ip);
1903                 hammer2_inode_drop(ip);         /* from lookup */
1904
1905                 return error;
1906         }
1907
1908         /*
1909          * Otherwise we have to find the inode
1910          */
1911         xop = hammer2_xop_alloc(pmp->iroot, 0);
1912         xop->lhc = inum;
1913         hammer2_xop_start(&xop->head, &hammer2_lookup_desc);
1914         error = hammer2_xop_collect(&xop->head, 0);
1915
1916         if (error == 0)
1917                 ip = hammer2_inode_get(pmp, &xop->head, -1, -1);
1918         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1919
1920         if (ip) {
1921                 *vpp = hammer2_igetv(ip, &error);
1922                 hammer2_inode_unlock(ip);
1923         } else {
1924                 *vpp = NULL;
1925                 error = ENOENT;
1926         }
1927         return (error);
1928 }
1929
1930 static
1931 int
1932 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1933 {
1934         hammer2_pfs_t *pmp;
1935         struct vnode *vp;
1936         int error;
1937
1938         pmp = MPTOPMP(mp);
1939         if (pmp->iroot == NULL) {
1940                 kprintf("hammer2 (%s): no root inode\n",
1941                         mp->mnt_stat.f_mntfromname);
1942                 *vpp = NULL;
1943                 return EINVAL;
1944         }
1945
1946         error = 0;
1947         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1948
1949         while (pmp->inode_tid == 0) {
1950                 hammer2_xop_ipcluster_t *xop;
1951                 const hammer2_inode_meta_t *meta;
1952
1953                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1954                 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc);
1955                 error = hammer2_xop_collect(&xop->head, 0);
1956
1957                 if (error == 0) {
1958                         meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1959                         pmp->iroot->meta = *meta;
1960                         pmp->inode_tid = meta->pfs_inum + 1;
1961                         hammer2_xop_pdata(&xop->head);
1962                         /* meta invalid */
1963
1964                         if (pmp->inode_tid < HAMMER2_INODE_START)
1965                                 pmp->inode_tid = HAMMER2_INODE_START;
1966                         pmp->modify_tid =
1967                                 xop->head.cluster.focus->bref.modify_tid + 1;
1968 #if 0
1969                         kprintf("PFS: Starting inode %jd\n",
1970                                 (intmax_t)pmp->inode_tid);
1971                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1972                                 pmp->inode_tid, pmp->modify_tid);
1973 #endif
1974                         wakeup(&pmp->iroot);
1975
1976                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1977
1978                         /*
1979                          * Prime the mount info.
1980                          */
1981                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1982                         break;
1983                 }
1984
1985                 /*
1986                  * Loop, try again
1987                  */
1988                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1989                 hammer2_inode_unlock(pmp->iroot);
1990                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1991                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1992                 if (error == EINTR)
1993                         break;
1994         }
1995
1996         if (error) {
1997                 hammer2_inode_unlock(pmp->iroot);
1998                 *vpp = NULL;
1999         } else {
2000                 vp = hammer2_igetv(pmp->iroot, &error);
2001                 hammer2_inode_unlock(pmp->iroot);
2002                 *vpp = vp;
2003         }
2004
2005         return (error);
2006 }
2007
2008 /*
2009  * Filesystem status
2010  *
2011  * XXX incorporate ipdata->meta.inode_quota and data_quota
2012  */
2013 static
2014 int
2015 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
2016 {
2017         hammer2_pfs_t *pmp;
2018         hammer2_dev_t *hmp;
2019         hammer2_blockref_t bref;
2020         struct statfs tmp;
2021         int i;
2022
2023         /*
2024          * NOTE: iroot might not have validated the cluster yet.
2025          */
2026         pmp = MPTOPMP(mp);
2027
2028         bzero(&tmp, sizeof(tmp));
2029
2030         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2031                 hmp = pmp->pfs_hmps[i];
2032                 if (hmp == NULL)
2033                         continue;
2034                 if (pmp->iroot->cluster.array[i].chain)
2035                         bref = pmp->iroot->cluster.array[i].chain->bref;
2036                 else
2037                         bzero(&bref, sizeof(bref));
2038
2039                 tmp.f_files = bref.embed.stats.inode_count;
2040                 tmp.f_ffree = 0;
2041                 tmp.f_blocks = hmp->voldata.allocator_size /
2042                                mp->mnt_vstat.f_bsize;
2043                 tmp.f_bfree = hmp->voldata.allocator_free /
2044                               mp->mnt_vstat.f_bsize;
2045                 tmp.f_bavail = tmp.f_bfree;
2046
2047                 if (cred && cred->cr_uid != 0) {
2048                         uint64_t adj;
2049
2050                         /* 5% */
2051                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2052                         tmp.f_blocks -= adj;
2053                         tmp.f_bfree -= adj;
2054                         tmp.f_bavail -= adj;
2055                 }
2056
2057                 mp->mnt_stat.f_blocks = tmp.f_blocks;
2058                 mp->mnt_stat.f_bfree = tmp.f_bfree;
2059                 mp->mnt_stat.f_bavail = tmp.f_bavail;
2060                 mp->mnt_stat.f_files = tmp.f_files;
2061                 mp->mnt_stat.f_ffree = tmp.f_ffree;
2062
2063                 *sbp = mp->mnt_stat;
2064         }
2065         return (0);
2066 }
2067
2068 static
2069 int
2070 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2071 {
2072         hammer2_pfs_t *pmp;
2073         hammer2_dev_t *hmp;
2074         hammer2_blockref_t bref;
2075         struct statvfs tmp;
2076         int i;
2077
2078         /*
2079          * NOTE: iroot might not have validated the cluster yet.
2080          */
2081         pmp = MPTOPMP(mp);
2082         bzero(&tmp, sizeof(tmp));
2083
2084         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2085                 hmp = pmp->pfs_hmps[i];
2086                 if (hmp == NULL)
2087                         continue;
2088                 if (pmp->iroot->cluster.array[i].chain)
2089                         bref = pmp->iroot->cluster.array[i].chain->bref;
2090                 else
2091                         bzero(&bref, sizeof(bref));
2092
2093                 tmp.f_files = bref.embed.stats.inode_count;
2094                 tmp.f_ffree = 0;
2095                 tmp.f_blocks = hmp->voldata.allocator_size /
2096                                mp->mnt_vstat.f_bsize;
2097                 tmp.f_bfree = hmp->voldata.allocator_free /
2098                               mp->mnt_vstat.f_bsize;
2099                 tmp.f_bavail = tmp.f_bfree;
2100
2101                 if (cred && cred->cr_uid != 0) {
2102                         uint64_t adj;
2103
2104                         /* 5% */
2105                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2106                         tmp.f_blocks -= adj;
2107                         tmp.f_bfree -= adj;
2108                         tmp.f_bavail -= adj;
2109                 }
2110
2111                 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2112                 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2113                 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2114                 mp->mnt_vstat.f_files = tmp.f_files;
2115                 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2116
2117                 *sbp = mp->mnt_vstat;
2118         }
2119         return (0);
2120 }
2121
2122 /*
2123  * Mount-time recovery (RW mounts)
2124  *
2125  * Updates to the free block table are allowed to lag flushes by one
2126  * transaction.  In case of a crash, then on a fresh mount we must do an
2127  * incremental scan of the last committed transaction id and make sure that
2128  * all related blocks have been marked allocated.
2129  *
2130  * The super-root topology and each PFS has its own transaction id domain,
2131  * so we must track PFS boundary transitions.
2132  */
2133 struct hammer2_recovery_elm {
2134         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2135         hammer2_chain_t *chain;
2136         hammer2_tid_t sync_tid;
2137 };
2138
2139 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2140
2141 struct hammer2_recovery_info {
2142         struct hammer2_recovery_list list;
2143         hammer2_tid_t   mtid;
2144         int     depth;
2145 };
2146
2147 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2148                         hammer2_chain_t *parent,
2149                         struct hammer2_recovery_info *info,
2150                         hammer2_tid_t sync_tid);
2151
2152 #define HAMMER2_RECOVERY_MAXDEPTH       10
2153
2154 static
2155 int
2156 hammer2_recovery(hammer2_dev_t *hmp)
2157 {
2158         struct hammer2_recovery_info info;
2159         struct hammer2_recovery_elm *elm;
2160         hammer2_chain_t *parent;
2161         hammer2_tid_t sync_tid;
2162         hammer2_tid_t mirror_tid;
2163         int error;
2164
2165         hammer2_trans_init(hmp->spmp, 0);
2166
2167         sync_tid = hmp->voldata.freemap_tid;
2168         mirror_tid = hmp->voldata.mirror_tid;
2169
2170         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2171         if (sync_tid >= mirror_tid) {
2172                 kprintf(" no recovery needed\n");
2173         } else {
2174                 kprintf(" freemap recovery %016jx-%016jx\n",
2175                         sync_tid + 1, mirror_tid);
2176         }
2177
2178         TAILQ_INIT(&info.list);
2179         info.depth = 0;
2180         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2181         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2182         hammer2_chain_lookup_done(parent);
2183
2184         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2185                 TAILQ_REMOVE(&info.list, elm, entry);
2186                 parent = elm->chain;
2187                 sync_tid = elm->sync_tid;
2188                 kfree(elm, M_HAMMER2);
2189
2190                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2191                 error |= hammer2_recovery_scan(hmp, parent, &info,
2192                                               hmp->voldata.freemap_tid);
2193                 hammer2_chain_unlock(parent);
2194                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2195         }
2196
2197         hammer2_trans_done(hmp->spmp, 0);
2198
2199         return error;
2200 }
2201
2202 static
2203 int
2204 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2205                       struct hammer2_recovery_info *info,
2206                       hammer2_tid_t sync_tid)
2207 {
2208         const hammer2_inode_data_t *ripdata;
2209         hammer2_chain_t *chain;
2210         hammer2_blockref_t bref;
2211         int tmp_error;
2212         int rup_error;
2213         int error;
2214         int first;
2215
2216         /*
2217          * Adjust freemap to ensure that the block(s) are marked allocated.
2218          */
2219         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2220                 hammer2_freemap_adjust(hmp, &parent->bref,
2221                                        HAMMER2_FREEMAP_DORECOVER);
2222         }
2223
2224         /*
2225          * Check type for recursive scan
2226          */
2227         switch(parent->bref.type) {
2228         case HAMMER2_BREF_TYPE_VOLUME:
2229                 /* data already instantiated */
2230                 break;
2231         case HAMMER2_BREF_TYPE_INODE:
2232                 /*
2233                  * Must instantiate data for DIRECTDATA test and also
2234                  * for recursion.
2235                  */
2236                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2237                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2238                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2239                         /* not applicable to recovery scan */
2240                         hammer2_chain_unlock(parent);
2241                         return 0;
2242                 }
2243                 hammer2_chain_unlock(parent);
2244                 break;
2245         case HAMMER2_BREF_TYPE_INDIRECT:
2246                 /*
2247                  * Must instantiate data for recursion
2248                  */
2249                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2250                 hammer2_chain_unlock(parent);
2251                 break;
2252         case HAMMER2_BREF_TYPE_DIRENT:
2253         case HAMMER2_BREF_TYPE_DATA:
2254         case HAMMER2_BREF_TYPE_FREEMAP:
2255         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2256         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2257                 /* not applicable to recovery scan */
2258                 return 0;
2259                 break;
2260         default:
2261                 return HAMMER2_ERROR_BADBREF;
2262         }
2263
2264         /*
2265          * Defer operation if depth limit reached or if we are crossing a
2266          * PFS boundary.
2267          */
2268         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2269                 struct hammer2_recovery_elm *elm;
2270
2271                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2272                 elm->chain = parent;
2273                 elm->sync_tid = sync_tid;
2274                 hammer2_chain_ref(parent);
2275                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2276                 /* unlocked by caller */
2277
2278                 return(0);
2279         }
2280
2281
2282         /*
2283          * Recursive scan of the last flushed transaction only.  We are
2284          * doing this without pmp assignments so don't leave the chains
2285          * hanging around after we are done with them.
2286          *
2287          * error        Cumulative error this level only
2288          * rup_error    Cumulative error for recursion
2289          * tmp_error    Specific non-cumulative recursion error
2290          */
2291         chain = NULL;
2292         first = 1;
2293         rup_error = 0;
2294         error = 0;
2295
2296         for (;;) {
2297                 error |= hammer2_chain_scan(parent, &chain, &bref,
2298                                             &first,
2299                                             HAMMER2_LOOKUP_NODATA);
2300
2301                 /*
2302                  * Problem during scan or EOF
2303                  */
2304                 if (error)
2305                         break;
2306
2307                 /*
2308                  * If this is a leaf
2309                  */
2310                 if (chain == NULL) {
2311                         if (bref.mirror_tid > sync_tid) {
2312                                 hammer2_freemap_adjust(hmp, &bref,
2313                                                      HAMMER2_FREEMAP_DORECOVER);
2314                         }
2315                         continue;
2316                 }
2317
2318                 /*
2319                  * This may or may not be a recursive node.
2320                  */
2321                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2322                 if (bref.mirror_tid > sync_tid) {
2323                         ++info->depth;
2324                         tmp_error = hammer2_recovery_scan(hmp, chain,
2325                                                            info, sync_tid);
2326                         --info->depth;
2327                 } else {
2328                         tmp_error = 0;
2329                 }
2330
2331                 /*
2332                  * Flush the recovery at the PFS boundary to stage it for
2333                  * the final flush of the super-root topology.
2334                  */
2335                 if (tmp_error == 0 &&
2336                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2337                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2338                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2339                                              HAMMER2_FLUSH_ALL);
2340                 }
2341                 rup_error |= tmp_error;
2342         }
2343         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2344 }
2345
2346 /*
2347  * This fixes up an error introduced in earlier H2 implementations where
2348  * moving a PFS inode into an indirect block wound up causing the
2349  * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2350  */
2351 static
2352 int
2353 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2354 {
2355         const hammer2_inode_data_t *ripdata;
2356         hammer2_chain_t *parent;
2357         hammer2_chain_t *chain;
2358         hammer2_key_t key_next;
2359         hammer2_pfs_t *spmp;
2360         int error;
2361
2362         error = 0;
2363
2364         /*
2365          * Lookup mount point under the media-localized super-root.
2366          *
2367          * cluster->pmp will incorrectly point to spmp and must be fixed
2368          * up later on.
2369          */
2370         spmp = hmp->spmp;
2371         hammer2_inode_lock(spmp->iroot, 0);
2372         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2373         chain = hammer2_chain_lookup(&parent, &key_next,
2374                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2375                                          &error, 0);
2376         while (chain) {
2377                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2378                         continue;
2379                 if (chain->error) {
2380                         kprintf("I/O error scanning PFS labels\n");
2381                         error |= chain->error;
2382                 } else if ((chain->bref.flags &
2383                             HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2384                         int error2;
2385
2386                         ripdata = &chain->data->ipdata;
2387                         hammer2_trans_init(hmp->spmp, 0);
2388                         error2 = hammer2_chain_modify(chain,
2389                                                       chain->bref.modify_tid,
2390                                                       0, 0);
2391                         if (error2 == 0) {
2392                                 kprintf("hammer2: Correct mis-flagged PFS %s\n",
2393                                         ripdata->filename);
2394                                 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2395                         } else {
2396                                 error |= error2;
2397                         }
2398                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2399                                              HAMMER2_FLUSH_ALL);
2400                         hammer2_trans_done(hmp->spmp, 0);
2401                 }
2402                 chain = hammer2_chain_next(&parent, chain, &key_next,
2403                                            key_next, HAMMER2_KEY_MAX,
2404                                            &error, 0);
2405         }
2406         if (parent) {
2407                 hammer2_chain_unlock(parent);
2408                 hammer2_chain_drop(parent);
2409         }
2410         hammer2_inode_unlock(spmp->iroot);
2411
2412         return error;
2413 }
2414
2415 /*
2416  * Sync a mount point; this is called periodically on a per-mount basis from
2417  * the filesystem syncer, and whenever a user issues a sync.
2418  */
2419 int
2420 hammer2_vfs_sync(struct mount *mp, int waitfor)
2421 {
2422         int error;
2423
2424         error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor);
2425
2426         return error;
2427 }
2428
2429 /*
2430  * Because frontend operations lock vnodes before we get a chance to
2431  * lock the related inode, we can't just acquire a vnode lock without
2432  * risking a deadlock.  The frontend may be holding a vnode lock while
2433  * also blocked on our SYNCQ flag while trying to get the inode lock.
2434  *
2435  * To deal with this situation we can check the vnode lock situation
2436  * after locking the inode and perform a work-around.
2437  */
2438 int
2439 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor)
2440 {
2441         struct mount *mp;
2442         /*hammer2_xop_flush_t *xop;*/
2443         /*struct hammer2_sync_info info;*/
2444         hammer2_inode_t *ip;
2445         hammer2_depend_t *depend;
2446         hammer2_depend_t *depend_next;
2447         struct vnode *vp;
2448         uint32_t pass2;
2449         int error;
2450         int dorestart;
2451
2452         mp = pmp->mp;
2453
2454         /*
2455          * Move all inodes on sideq to syncq.  This will clear sideq.
2456          * This should represent all flushable inodes.  These inodes
2457          * will already have refs due to being on syncq or sideq.  We
2458          * must do this all at once with the spinlock held to ensure that
2459          * all inode dependencies are part of the same flush.
2460          *
2461          * We should be able to do this asynchronously from frontend
2462          * operations because we will be locking the inodes later on
2463          * to actually flush them, and that will partition any frontend
2464          * op using the same inode.  Either it has already locked the
2465          * inode and we will block, or it has not yet locked the inode
2466          * and it will block until we are finished flushing that inode.
2467          *
2468          * When restarting, only move the inodes flagged as PASS2 from
2469          * SIDEQ to SYNCQ.  PASS2 propagation by inode_lock4() and
2470          * inode_depend() are atomic with the spin-lock.
2471          */
2472         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2473 #ifdef HAMMER2_DEBUG_SYNC
2474         kprintf("FILESYSTEM SYNC BOUNDARY\n");
2475 #endif
2476         dorestart = 0;
2477
2478         /*
2479          * Move inodes from depq to syncq, releasing the related
2480          * depend structures.
2481          */
2482 restart:
2483 #ifdef HAMMER2_DEBUG_SYNC
2484         kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart);
2485 #endif
2486         hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/);
2487         hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN);
2488
2489         /*
2490          * Move inodes from depq to syncq.  When restarting, only depq's
2491          * marked pass2 are moved.
2492          */
2493         hammer2_spin_ex(&pmp->list_spin);
2494         depend_next = TAILQ_FIRST(&pmp->depq);
2495
2496         while ((depend = depend_next) != NULL) {
2497                 depend_next = TAILQ_NEXT(depend, entry);
2498                 if (dorestart && depend->pass2 == 0)
2499                         continue;
2500                 TAILQ_FOREACH(ip, &depend->sideq, entry) {
2501                         KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ);
2502                         atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ);
2503                         atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ);
2504                         ip->depend = NULL;
2505                 }
2506                 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry);
2507                 pmp->sideq_count -= depend->count;
2508                 depend->count = 0;
2509                 depend->pass2 = 0;
2510                 TAILQ_REMOVE(&pmp->depq, depend, entry);
2511         }
2512
2513         hammer2_spin_unex(&pmp->list_spin);
2514         hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/
2515                                       HAMMER2_TRANS_WAITING);
2516         dorestart = 0;
2517
2518         /*
2519          * sideq_count may have dropped enough to allow us to unstall
2520          * the frontend.
2521          */
2522         hammer2_pfs_memory_inc(pmp);
2523         hammer2_pfs_memory_wakeup(pmp);
2524
2525         /*
2526          * Now run through all inodes on syncq.
2527          *
2528          * Flush transactions only interlock with other flush transactions.
2529          * Any conflicting frontend operations will block on the inode, but
2530          * may hold a vnode lock while doing so.
2531          */
2532         hammer2_spin_ex(&pmp->list_spin);
2533         while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) {
2534                 /*
2535                  * Remove the inode from the SYNCQ, transfer the syncq ref
2536                  * to us.  We must clear SYNCQ to allow any potential
2537                  * front-end deadlock to proceed.  We must set PASS2 so
2538                  * the dependency code knows what to do.
2539                  */
2540                 pass2 = ip->flags;
2541                 cpu_ccfence();
2542                 if (atomic_cmpset_int(&ip->flags,
2543                               pass2,
2544                               (pass2 & ~(HAMMER2_INODE_SYNCQ |
2545                                          HAMMER2_INODE_SYNCQ_WAKEUP)) |
2546                               HAMMER2_INODE_SYNCQ_PASS2) == 0) {
2547                         continue;
2548                 }
2549                 TAILQ_REMOVE(&pmp->syncq, ip, entry);
2550                 hammer2_spin_unex(&pmp->list_spin);
2551                 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP)
2552                         wakeup(&ip->flags);
2553
2554                 /*
2555                  * Relock the inode, and we inherit a ref from the above.
2556                  * We will check for a race after we acquire the vnode.
2557                  */
2558                 hammer2_mtx_ex(&ip->lock);
2559
2560                 /*
2561                  * We need the vp in order to vfsync() dirty buffers, so if
2562                  * one isn't attached we can skip it.
2563                  *
2564                  * Ordering the inode lock and then the vnode lock has the
2565                  * potential to deadlock.  If we had left SYNCQ set that could
2566                  * also deadlock us against the frontend even if we don't hold
2567                  * any locks, but the latter is not a problem now since we
2568                  * cleared it.  igetv will temporarily release the inode lock
2569                  * in a safe manner to work-around the deadlock.
2570                  *
2571                  * Unfortunately it is still possible to deadlock when the
2572                  * frontend obtains multiple inode locks, because all the
2573                  * related vnodes are already locked (nor can the vnode locks
2574                  * be released and reacquired without messing up RECLAIM and
2575                  * INACTIVE sequencing).
2576                  *
2577                  * The solution for now is to move the vp back onto SIDEQ
2578                  * and set dorestart, which will restart the flush after we
2579                  * exhaust the current SYNCQ.  Note that additional
2580                  * dependencies may build up, so we definitely need to move
2581                  * the whole SIDEQ back to SYNCQ when we restart.
2582                  */
2583                 vp = ip->vp;
2584                 if (vp) {
2585                         if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) {
2586                                 /*
2587                                  * Failed to get the vnode, requeue the inode
2588                                  * (PASS2 is already set so it will be found
2589                                  * again on the restart).
2590                                  *
2591                                  * Then unlock, possibly sleep, and retry
2592                                  * later.  We sleep if PASS2 was *previously*
2593                                  * set, before we set it again above.
2594                                  */
2595                                 vp = NULL;
2596                                 dorestart = 1;
2597 #ifdef HAMMER2_DEBUG_SYNC
2598                                 kprintf("inum %ld (sync delayed by vnode)\n",
2599                                         (long)ip->meta.inum);
2600 #endif
2601                                 hammer2_inode_delayed_sideq(ip);
2602
2603                                 hammer2_mtx_unlock(&ip->lock);
2604                                 hammer2_inode_drop(ip);
2605
2606                                 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) {
2607                                         tsleep(&dorestart, 0, "h2syndel", 2);
2608                                 }
2609                                 hammer2_spin_ex(&pmp->list_spin);
2610                                 continue;
2611                         }
2612                 } else {
2613                         vp = NULL;
2614                 }
2615
2616                 /*
2617                  * If the inode wound up on a SIDEQ again it will already be
2618                  * prepped for another PASS2.  In this situation if we flush
2619                  * it now we will just wind up flushing it again in the same
2620                  * syncer run, so we might as well not flush it now.
2621                  */
2622                 if (ip->flags & HAMMER2_INODE_SIDEQ) {
2623                         hammer2_mtx_unlock(&ip->lock);
2624                         hammer2_inode_drop(ip);
2625                         if (vp)
2626                                 vput(vp);
2627                         dorestart = 1;
2628                         hammer2_spin_ex(&pmp->list_spin);
2629                         continue;
2630                 }
2631
2632                 /*
2633                  * Ok we have the inode exclusively locked and if vp is
2634                  * not NULL that will also be exclusively locked.  Do the
2635                  * meat of the flush.
2636                  *
2637                  * vp token needed for v_rbdirty_tree check / vclrisdirty
2638                  * sequencing.  Though we hold the vnode exclusively so
2639                  * we shouldn't need to hold the token also in this case.
2640                  */
2641                 if (vp) {
2642                         vfsync(vp, MNT_WAIT, 1, NULL, NULL);
2643                         bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */
2644                 }
2645
2646                 /*
2647                  * If the inode has not yet been inserted into the tree
2648                  * we must do so.  Then sync and flush it.  The flush should
2649                  * update the parent.
2650                  */
2651                 if (ip->flags & HAMMER2_INODE_DELETING) {
2652 #ifdef HAMMER2_DEBUG_SYNC
2653                         kprintf("inum %ld destroy\n", (long)ip->meta.inum);
2654 #endif
2655                         hammer2_inode_chain_des(ip);
2656                         atomic_add_long(&hammer2_iod_inode_deletes, 1);
2657                 } else if (ip->flags & HAMMER2_INODE_CREATING) {
2658 #ifdef HAMMER2_DEBUG_SYNC
2659                         kprintf("inum %ld insert\n", (long)ip->meta.inum);
2660 #endif
2661                         hammer2_inode_chain_ins(ip);
2662                         atomic_add_long(&hammer2_iod_inode_creates, 1);
2663                 }
2664 #ifdef HAMMER2_DEBUG_SYNC
2665                 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum);
2666 #endif
2667
2668                 /*
2669                  * Because I kinda messed up the design and index the inodes
2670                  * under the root inode, along side the directory entries,
2671                  * we can't flush the inode index under the iroot until the
2672                  * end.  If we do it now we might miss effects created by
2673                  * other inodes on the SYNCQ.
2674                  *
2675                  * Do a normal (non-FSSYNC) flush instead, which allows the
2676                  * vnode code to work the same.  We don't want to force iroot
2677                  * back onto the SIDEQ, and we also don't want the flush code
2678                  * to update pfs_iroot_blocksets until the final flush later.
2679                  *
2680                  * XXX at the moment this will likely result in a double-flush
2681                  * of the iroot chain.
2682                  */
2683                 hammer2_inode_chain_sync(ip);
2684                 if (ip == pmp->iroot) {
2685                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP);
2686                 } else {
2687                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2688                                                       HAMMER2_XOP_FSSYNC);
2689                 }
2690                 if (vp) {
2691                         lwkt_gettoken(&vp->v_token);
2692                         if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2693                                           HAMMER2_INODE_RESIZED |
2694                                           HAMMER2_INODE_DIRTYDATA)) == 0 &&
2695                             RB_EMPTY(&vp->v_rbdirty_tree) &&
2696                             !bio_track_active(&vp->v_track_write)) {
2697                                 vclrisdirty(vp);
2698                         } else {
2699                                 hammer2_inode_delayed_sideq(ip);
2700                         }
2701                         lwkt_reltoken(&vp->v_token);
2702                         vput(vp);
2703                         vp = NULL;      /* safety */
2704                 }
2705                 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2);
2706                 hammer2_inode_unlock(ip);       /* unlock+drop */
2707                 /* ip pointer invalid */
2708
2709                 /*
2710                  * If the inode got dirted after we dropped our locks,
2711                  * it will have already been moved back to the SIDEQ.
2712                  */
2713                 hammer2_spin_ex(&pmp->list_spin);
2714         }
2715         hammer2_spin_unex(&pmp->list_spin);
2716         if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) {
2717 #ifdef HAMMER2_DEBUG_SYNC
2718                 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n");
2719                 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/
2720 #endif
2721                 dorestart = 1;
2722                 goto restart;
2723         }
2724 #ifdef HAMMER2_DEBUG_SYNC
2725         kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n");
2726         /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/
2727 #endif
2728
2729         /*
2730          * We have to flush the PFS root last, even if it does not appear to
2731          * be dirty, because all the inodes in the PFS are indexed under it.
2732          * The normal flushing of iroot above would only occur if directory
2733          * entries under the root were changed.
2734          *
2735          * Specifying VOLHDR will cause an additionl flush of hmp->spmp
2736          * for the media making up the cluster.
2737          */
2738         if ((ip = pmp->iroot) != NULL) {
2739                 hammer2_inode_ref(ip);
2740                 hammer2_mtx_ex(&ip->lock);
2741                 hammer2_inode_chain_sync(ip);
2742                 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2743                                               HAMMER2_XOP_FSSYNC |
2744                                               HAMMER2_XOP_VOLHDR);
2745                 hammer2_inode_unlock(ip);       /* unlock+drop */
2746         }
2747 #ifdef HAMMER2_DEBUG_SYNC
2748         kprintf("FILESYSTEM SYNC STAGE 2 DONE\n");
2749 #endif
2750
2751         /*
2752          * device bioq sync
2753          */
2754         hammer2_bioq_sync(pmp);
2755
2756 #if 0
2757         info.pass = 1;
2758         info.waitfor = MNT_WAIT;
2759         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2760
2761         info.pass = 2;
2762         info.waitfor = MNT_WAIT;
2763         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2764 #endif
2765 #if 0
2766         /*
2767          * Generally speaking we now want to flush the media topology from
2768          * the iroot through to the inodes.  The flush stops at any inode
2769          * boundary, which allows the frontend to continue running concurrent
2770          * modifying operations on inodes (including kernel flushes of
2771          * buffers) without interfering with the main sync.
2772          *
2773          * Use the XOP interface to concurrently flush all nodes to
2774          * synchronize the PFSROOT subtopology to the media.  A standard
2775          * end-of-scan ENOENT error indicates cluster sufficiency.
2776          *
2777          * Note that this flush will not be visible on crash recovery until
2778          * we flush the super-root topology in the next loop.
2779          *
2780          * XXX For now wait for all flushes to complete.
2781          */
2782         if (mp && (ip = pmp->iroot) != NULL) {
2783                 /*
2784                  * If unmounting try to flush everything including any
2785                  * sub-trees under inodes, just in case there is dangling
2786                  * modified data, as a safety.  Otherwise just flush up to
2787                  * the inodes in this stage.
2788                  */
2789                 kprintf("MP & IROOT\n");
2790 #ifdef HAMMER2_DEBUG_SYNC
2791                 kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n");
2792 #endif
2793                 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2794                         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2795                                                     HAMMER2_XOP_VOLHDR |
2796                                                     HAMMER2_XOP_FSSYNC |
2797                                                     HAMMER2_XOP_INODE_STOP);
2798                 } else {
2799                         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2800                                                     HAMMER2_XOP_INODE_STOP |
2801                                                     HAMMER2_XOP_VOLHDR |
2802                                                     HAMMER2_XOP_FSSYNC |
2803                                                     HAMMER2_XOP_INODE_STOP);
2804                 }
2805                 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc);
2806                 error = hammer2_xop_collect(&xop->head,
2807                                             HAMMER2_XOP_COLLECT_WAITALL);
2808                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2809 #ifdef HAMMER2_DEBUG_SYNC
2810                 kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n");
2811 #endif
2812                 if (error == HAMMER2_ERROR_ENOENT)
2813                         error = 0;
2814                 else
2815                         error = hammer2_error_to_errno(error);
2816         } else {
2817                 error = 0;
2818         }
2819 #endif
2820         error = 0;      /* XXX */
2821         hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH);
2822
2823         return (error);
2824 }
2825
2826 static
2827 int
2828 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2829 {
2830         hammer2_inode_t *ip;
2831
2832         KKASSERT(MAXFIDSZ >= 16);
2833         ip = VTOI(vp);
2834         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2835         fhp->fid_ext = 0;
2836         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2837         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2838
2839         return 0;
2840 }
2841
2842 static
2843 int
2844 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2845                struct fid *fhp, struct vnode **vpp)
2846 {
2847         hammer2_pfs_t *pmp;
2848         hammer2_tid_t inum;
2849         int error;
2850
2851         pmp = MPTOPMP(mp);
2852         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2853         if (vpp) {
2854                 if (inum == 1)
2855                         error = hammer2_vfs_root(mp, vpp);
2856                 else
2857                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2858         } else {
2859                 error = 0;
2860         }
2861         if (error)
2862                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2863         return error;
2864 }
2865
2866 static
2867 int
2868 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2869                  int *exflagsp, struct ucred **credanonp)
2870 {
2871         hammer2_pfs_t *pmp;
2872         struct netcred *np;
2873         int error;
2874
2875         pmp = MPTOPMP(mp);
2876         np = vfs_export_lookup(mp, &pmp->export, nam);
2877         if (np) {
2878                 *exflagsp = np->netc_exflags;
2879                 *credanonp = &np->netc_anon;
2880                 error = 0;
2881         } else {
2882                 error = EACCES;
2883         }
2884         return error;
2885 }
2886
2887 /*
2888  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2889  * header into the HMP
2890  *
2891  * XXX read four volhdrs and use the one with the highest TID whos CRC
2892  *     matches.
2893  *
2894  * XXX check iCRCs.
2895  *
2896  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2897  *     nonexistant locations.
2898  *
2899  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2900  */
2901 static
2902 int
2903 hammer2_install_volume_header(hammer2_dev_t *hmp)
2904 {
2905         hammer2_volume_data_t *vd;
2906         struct buf *bp;
2907         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2908         int error_reported;
2909         int error;
2910         int valid;
2911         int i;
2912
2913         error_reported = 0;
2914         error = 0;
2915         valid = 0;
2916         bp = NULL;
2917
2918         /*
2919          * There are up to 4 copies of the volume header (syncs iterate
2920          * between them so there is no single master).  We don't trust the
2921          * volu_size field so we don't know precisely how large the filesystem
2922          * is, so depend on the OS to return an error if we go beyond the
2923          * block device's EOF.
2924          */
2925         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2926                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2927                               HAMMER2_VOLUME_BYTES, &bp);
2928                 if (error) {
2929                         brelse(bp);
2930                         bp = NULL;
2931                         continue;
2932                 }
2933
2934                 vd = (struct hammer2_volume_data *) bp->b_data;
2935                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2936                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2937                         brelse(bp);
2938                         bp = NULL;
2939                         continue;
2940                 }
2941
2942                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2943                         /* XXX: Reversed-endianness filesystem */
2944                         kprintf("hammer2: reverse-endian filesystem detected");
2945                         brelse(bp);
2946                         bp = NULL;
2947                         continue;
2948                 }
2949
2950                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2951                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2952                                       HAMMER2_VOLUME_ICRC0_SIZE);
2953                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2954                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2955                                        HAMMER2_VOLUME_ICRC1_SIZE);
2956                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2957                         kprintf("hammer2 volume header crc "
2958                                 "mismatch copy #%d %08x/%08x\n",
2959                                 i, crc0, crc);
2960                         error_reported = 1;
2961                         brelse(bp);
2962                         bp = NULL;
2963                         continue;
2964                 }
2965                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2966                         valid = 1;
2967                         hmp->voldata = *vd;
2968                         hmp->volhdrno = i;
2969                 }
2970                 brelse(bp);
2971                 bp = NULL;
2972         }
2973         if (valid) {
2974                 hmp->volsync = hmp->voldata;
2975                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2976                 error = 0;
2977                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2978                         kprintf("hammer2: using volume header #%d\n",
2979                                 hmp->volhdrno);
2980                 }
2981         } else {
2982                 error = EINVAL;
2983                 kprintf("hammer2: no valid volume headers found!\n");
2984         }
2985         return (error);
2986 }
2987
2988 /*
2989  * This handles hysteresis on regular file flushes.  Because the BIOs are
2990  * routed to a thread it is possible for an excessive number to build up
2991  * and cause long front-end stalls long before the runningbuffspace limit
2992  * is hit, so we implement hammer2_flush_pipe to control the
2993  * hysteresis.
2994  *
2995  * This is a particular problem when compression is used.
2996  */
2997 void
2998 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2999 {
3000         atomic_add_int(&pmp->count_lwinprog, 1);
3001 }
3002
3003 void
3004 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
3005 {
3006         int lwinprog;
3007
3008         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
3009         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
3010             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
3011                 atomic_clear_int(&pmp->count_lwinprog,
3012                                  HAMMER2_LWINPROG_WAITING);
3013                 wakeup(&pmp->count_lwinprog);
3014         }
3015         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
3016             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
3017                 atomic_clear_int(&pmp->count_lwinprog,
3018                                  HAMMER2_LWINPROG_WAITING0);
3019                 wakeup(&pmp->count_lwinprog);
3020         }
3021 }
3022
3023 void
3024 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
3025 {
3026         int lwinprog;
3027         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
3028                                     HAMMER2_LWINPROG_WAITING0;
3029
3030         for (;;) {
3031                 lwinprog = pmp->count_lwinprog;
3032                 cpu_ccfence();
3033                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3034                         break;
3035                 tsleep_interlock(&pmp->count_lwinprog, 0);
3036                 atomic_set_int(&pmp->count_lwinprog, lwflag);
3037                 lwinprog = pmp->count_lwinprog;
3038                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3039                         break;
3040                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
3041         }
3042 }
3043
3044 /*
3045  * It is possible for an excessive number of dirty chains or dirty inodes
3046  * to build up.  When this occurs we start an asynchronous filesystem sync.
3047  * If the level continues to build up, we stall, waiting for it to drop,
3048  * with some hysteresis.
3049  *
3050  * We limit the stall to two seconds per call.
3051  *
3052  * This relies on the kernel calling hammer2_vfs_modifying() prior to
3053  * obtaining any vnode locks before making a modifying VOP call.
3054  */
3055 static void
3056 hammer2_vfs_modifying(struct mount *mp)
3057 {
3058         hammer2_pfs_memory_wait(MPTOPMP(mp));
3059 }
3060
3061 /*
3062  * Initiate an asynchronous filesystem sync and, with hysteresis,
3063  * stall if the internal data structure count becomes too bloated.
3064  */
3065 void
3066 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
3067 {
3068         uint32_t waiting;
3069         int loops;
3070
3071         if (pmp == NULL || pmp->mp == NULL)
3072                 return;
3073
3074         for (loops = 0; loops < 2; ++loops) {
3075                 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK;
3076                 cpu_ccfence();
3077
3078                 /*
3079                  * Start the syncer running at 1/2 the limit
3080                  */
3081                 if (waiting > hammer2_limit_dirty_chains / 2 ||
3082                     pmp->sideq_count > hammer2_limit_dirty_inodes / 2) {
3083                         trigger_syncer(pmp->mp);
3084                 }
3085
3086                 /*
3087                  * Stall at the limit waiting for the counts to drop.
3088                  * This code will typically be woken up once the count
3089                  * drops below 3/4 the limit, or in one second.
3090                  */
3091                 if (waiting < hammer2_limit_dirty_chains &&
3092                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
3093                         break;
3094                 }
3095                 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
3096                 atomic_set_int(&pmp->inmem_dirty_chains,
3097                                HAMMER2_DIRTYCHAIN_WAITING);
3098                 if (waiting < hammer2_limit_dirty_chains &&
3099                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
3100                         break;
3101                 }
3102                 trigger_syncer(pmp->mp);
3103                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED, "h2memw", hz);
3104 #if 0
3105                 limit = pmp->mp->mnt_nvnodelistsize / 10;
3106                 if (limit < hammer2_limit_dirty_chains)
3107                         limit = hammer2_limit_dirty_chains;
3108                 if (limit < 1000)
3109                         limit = 1000;
3110 #endif
3111         }
3112 }
3113
3114 void
3115 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
3116 {
3117         if (pmp) {
3118                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
3119         }
3120 }
3121
3122 void
3123 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
3124 {
3125         uint32_t waiting;
3126
3127         if (pmp) {
3128                 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
3129                 /* don't need --waiting to test flag */
3130
3131                 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) &&
3132                     (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <=
3133                     hammer2_limit_dirty_chains * 2 / 3 &&
3134                     pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) {
3135                         atomic_clear_int(&pmp->inmem_dirty_chains,
3136                                          HAMMER2_DIRTYCHAIN_WAITING);
3137                         wakeup(&pmp->inmem_dirty_chains);
3138                 }
3139         }
3140 }
3141
3142 /*
3143  * Returns 0 if the filesystem has tons of free space
3144  * Returns 1 if the filesystem has less than 10% remaining
3145  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
3146  */
3147 int
3148 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
3149 {
3150         hammer2_pfs_t *pmp;
3151         hammer2_dev_t *hmp;
3152         hammer2_off_t free_reserved;
3153         hammer2_off_t free_nominal;
3154         int i;
3155
3156         pmp = ip->pmp;
3157
3158         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
3159                 free_reserved = HAMMER2_SEGSIZE;
3160                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
3161                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
3162                         hmp = pmp->pfs_hmps[i];
3163                         if (hmp == NULL)
3164                                 continue;
3165                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
3166                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
3167                                 continue;
3168
3169                         if (free_nominal > hmp->voldata.allocator_free)
3170                                 free_nominal = hmp->voldata.allocator_free;
3171                         if (free_reserved < hmp->free_reserved)
3172                                 free_reserved = hmp->free_reserved;
3173                 }
3174
3175                 /*
3176                  * SMP races ok
3177                  */
3178                 pmp->free_reserved = free_reserved;
3179                 pmp->free_nominal = free_nominal;
3180                 pmp->free_ticks = ticks;
3181         } else {
3182                 free_reserved = pmp->free_reserved;
3183                 free_nominal = pmp->free_nominal;
3184         }
3185         if (cred && cred->cr_uid != 0) {
3186                 if ((int64_t)(free_nominal - bytes) <
3187                     (int64_t)free_reserved) {
3188                         return 2;
3189                 }
3190         } else {
3191                 if ((int64_t)(free_nominal - bytes) <
3192                     (int64_t)free_reserved / 2) {
3193                         return 2;
3194                 }
3195         }
3196         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
3197                 return 1;
3198         return 0;
3199 }
3200
3201 /*
3202  * Debugging
3203  */
3204 void
3205 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
3206                    u_int flags)
3207 {
3208         hammer2_chain_t *scan;
3209         hammer2_chain_t *parent;
3210
3211         --*countp;
3212         if (*countp == 0) {
3213                 kprintf("%*.*s...\n", tab, tab, "");
3214                 return;
3215         }
3216         if (*countp < 0)
3217                 return;
3218         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
3219                 tab, tab, "", pfx,
3220                 chain, chain->bref.type,
3221                 chain->bref.key, chain->bref.keybits,
3222                 chain->bref.mirror_tid);
3223
3224         kprintf("%*.*s      [%08x] (%s) refs=%d",
3225                 tab, tab, "",
3226                 chain->flags,
3227                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3228                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
3229                 chain->refs);
3230
3231         parent = chain->parent;
3232         if (parent)
3233                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
3234                         tab, tab, "",
3235                         parent, parent->flags, parent->refs);
3236         if (RB_EMPTY(&chain->core.rbtree)) {
3237                 kprintf("\n");
3238         } else {
3239                 kprintf(" {\n");
3240                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3241                         if ((scan->flags & flags) || flags == (u_int)-1) {
3242                                 hammer2_dump_chain(scan, tab + 4, countp, 'a',
3243                                                    flags);
3244                         }
3245                 }
3246                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3247                         kprintf("%*.*s}(%s)\n", tab, tab, "",
3248                                 chain->data->ipdata.filename);
3249                 else
3250                         kprintf("%*.*s}\n", tab, tab, "");
3251         }
3252 }