sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72 };
  73
  74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  75 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
  76 static struct hammer2_mntlist hammer2_mntlist;
  77 static struct hammer2_pfslist hammer2_pfslist;
  78 static struct lock hammer2_mntlk;
  79
  80 int hammer2_debug;
  81 int hammer2_cluster_enable = 1;
  82 int hammer2_hardlink_enable = 1;
  83 int hammer2_flush_pipe = 100;
  84 int hammer2_synchronous_flush = 1;
  85 int hammer2_dio_count;
  86 long hammer2_limit_dirty_chains;
  87 long hammer2_iod_file_read;
  88 long hammer2_iod_meta_read;
  89 long hammer2_iod_indr_read;
  90 long hammer2_iod_fmap_read;
  91 long hammer2_iod_volu_read;
  92 long hammer2_iod_file_write;
  93 long hammer2_iod_meta_write;
  94 long hammer2_iod_indr_write;
  95 long hammer2_iod_fmap_write;
  96 long hammer2_iod_volu_write;
  97 long hammer2_ioa_file_read;
  98 long hammer2_ioa_meta_read;
  99 long hammer2_ioa_indr_read;
 100 long hammer2_ioa_fmap_read;
 101 long hammer2_ioa_volu_read;
 102 long hammer2_ioa_fmap_write;
 103 long hammer2_ioa_file_write;
 104 long hammer2_ioa_meta_write;
 105 long hammer2_ioa_indr_write;
 106 long hammer2_ioa_volu_write;
 107
 108 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 109 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 110                 "Buffer used for compression.");
 111
 112 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 113 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 114                 "Buffer used for decompression.");
 115
 116 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 117
 118 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 119            &hammer2_debug, 0, "");
 120 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW,
 121            &hammer2_cluster_enable, 0, "");
 122 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
 123            &hammer2_hardlink_enable, 0, "");
 124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 125            &hammer2_flush_pipe, 0, "");
 126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 127            &hammer2_synchronous_flush, 0, "");
 128 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 129            &hammer2_limit_dirty_chains, 0, "");
 130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 131            &hammer2_dio_count, 0, "");
 132
 133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 134            &hammer2_iod_file_read, 0, "");
 135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 136            &hammer2_iod_meta_read, 0, "");
 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 138            &hammer2_iod_indr_read, 0, "");
 139 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 140            &hammer2_iod_fmap_read, 0, "");
 141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 142            &hammer2_iod_volu_read, 0, "");
 143
 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 145            &hammer2_iod_file_write, 0, "");
 146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 147            &hammer2_iod_meta_write, 0, "");
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 149            &hammer2_iod_indr_write, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 151            &hammer2_iod_fmap_write, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 153            &hammer2_iod_volu_write, 0, "");
 154
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
 156            &hammer2_ioa_file_read, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
 158            &hammer2_ioa_meta_read, 0, "");
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
 160            &hammer2_ioa_indr_read, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
 162            &hammer2_ioa_fmap_read, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
 164            &hammer2_ioa_volu_read, 0, "");
 165
 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
 167            &hammer2_ioa_file_write, 0, "");
 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
 169            &hammer2_ioa_meta_write, 0, "");
 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
 171            &hammer2_ioa_indr_write, 0, "");
 172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
 173            &hammer2_ioa_fmap_write, 0, "");
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
 175            &hammer2_ioa_volu_write, 0, "");
 176
 177 static int hammer2_vfs_init(struct vfsconf *conf);
 178 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 179 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 180                                 struct ucred *cred);
 181 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 182                                 struct vnode *, struct ucred *);
 183 static int hammer2_recovery(hammer2_dev_t *hmp);
 184 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 185 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 186 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 187                                 struct ucred *cred);
 188 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 189                                 struct ucred *cred);
 190 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
 191                                 ino_t ino, struct vnode **vpp);
 192 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 193                                 struct fid *fhp, struct vnode **vpp);
 194 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 195 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 196                                 int *exflagsp, struct ucred **credanonp);
 197
 198 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 199 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 200
 201 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 202
 203 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 204 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 205                                 hammer2_dev_t *hmp);
 206
 207 /*
 208  * HAMMER2 vfs operations.
 209  */
 210 static struct vfsops hammer2_vfsops = {
 211         .vfs_init       = hammer2_vfs_init,
 212         .vfs_uninit     = hammer2_vfs_uninit,
 213         .vfs_sync       = hammer2_vfs_sync,
 214         .vfs_mount      = hammer2_vfs_mount,
 215         .vfs_unmount    = hammer2_vfs_unmount,
 216         .vfs_root       = hammer2_vfs_root,
 217         .vfs_statfs     = hammer2_vfs_statfs,
 218         .vfs_statvfs    = hammer2_vfs_statvfs,
 219         .vfs_vget       = hammer2_vfs_vget,
 220         .vfs_vptofh     = hammer2_vfs_vptofh,
 221         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 222         .vfs_checkexp   = hammer2_vfs_checkexp
 223 };
 224
 225 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 226
 227 VFS_SET(hammer2_vfsops, hammer2, 0);
 228 MODULE_VERSION(hammer2, 1);
 229
 230 static
 231 int
 232 hammer2_vfs_init(struct vfsconf *conf)
 233 {
 234         static struct objcache_malloc_args margs_read;
 235         static struct objcache_malloc_args margs_write;
 236         static struct objcache_malloc_args margs_vop;
 237
 238         int error;
 239
 240         error = 0;
 241
 242         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 243                 error = EINVAL;
 244         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 245                 error = EINVAL;
 246         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 247                 error = EINVAL;
 248
 249         if (error)
 250                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 251
 252         margs_read.objsize = 65536;
 253         margs_read.mtype = M_HAMMER2_DEBUFFER;
 254
 255         margs_write.objsize = 32768;
 256         margs_write.mtype = M_HAMMER2_CBUFFER;
 257
 258         margs_vop.objsize = sizeof(hammer2_xop_t);
 259         margs_vop.mtype = M_HAMMER2;
 260
 261         /*
 262          * Note thaht for the XOPS cache we want backing store allocations
 263          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 264          * confusion), so use the backing store function that does it.  This
 265          * means that initial XOPS objects are zerod but REUSED objects are
 266          * not.  So we are responsible for cleaning the object up sufficiently
 267          * for our needs before objcache_put()ing it back (typically just the
 268          * FIFO indices).
 269          */
 270         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 271                                 0, 1, NULL, NULL, NULL,
 272                                 objcache_malloc_alloc,
 273                                 objcache_malloc_free,
 274                                 &margs_read);
 275         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 276                                 0, 1, NULL, NULL, NULL,
 277                                 objcache_malloc_alloc,
 278                                 objcache_malloc_free,
 279                                 &margs_write);
 280         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 281                                 0, 1, NULL, NULL, NULL,
 282                                 objcache_malloc_alloc_zero,
 283                                 objcache_malloc_free,
 284                                 &margs_vop);
 285
 286
 287         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 288         TAILQ_INIT(&hammer2_mntlist);
 289         TAILQ_INIT(&hammer2_pfslist);
 290
 291         hammer2_limit_dirty_chains = desiredvnodes / 10;
 292
 293         return (error);
 294 }
 295
 296 static
 297 int
 298 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 299 {
 300         objcache_destroy(cache_buffer_read);
 301         objcache_destroy(cache_buffer_write);
 302         objcache_destroy(cache_xops);
 303         return 0;
 304 }
 305
 306 /*
 307  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
 308  * mounts and the spmp structure for media (hmp) structures.
 309  *
 310  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 311  * transactions.  Note that synchronization does not use this field.
 312  * (typically frontend operations and synchronization cannot run on the
 313  * same PFS node at the same time).
 314  *
 315  * XXX check locking
 316  */
 317 hammer2_pfs_t *
 318 hammer2_pfsalloc(hammer2_chain_t *chain, const hammer2_inode_data_t *ripdata,
 319                  hammer2_tid_t modify_tid)
 320 {
 321         hammer2_inode_t *iroot;
 322         hammer2_pfs_t *pmp;
 323         int count;
 324         int i;
 325         int j;
 326
 327         /*
 328          * Locate or create the PFS based on the cluster id.  If ripdata
 329          * is NULL this is a spmp which is unique and is always allocated.
 330          */
 331         if (ripdata) {
 332                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 333                         if (bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 334                                  sizeof(pmp->pfs_clid)) == 0) {
 335                                         break;
 336                         }
 337                 }
 338         } else {
 339                 pmp = NULL;
 340         }
 341
 342         if (pmp == NULL) {
 343                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 344                 hammer2_trans_manage_init(pmp);
 345                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 346                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 347                 lockinit(&pmp->lock, "pfslk", 0, 0);
 348                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 349                 RB_INIT(&pmp->inum_tree);
 350                 TAILQ_INIT(&pmp->unlinkq);
 351                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 352
 353                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 354                         hammer2_xop_group_init(pmp, &pmp->xop_groups[j]);
 355
 356                 /*
 357                  * Save the last media transaction id for the flusher.  Set
 358                  * initial
 359                  */
 360                 if (ripdata)
 361                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 362                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 363
 364                 /*
 365                  * The synchronization thread may start too early, make
 366                  * sure it stays frozen until we are ready to let it go.
 367                  * XXX
 368                  */
 369                 /*
 370                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 371                                          HAMMER2_THREAD_REMASTER;
 372                 */
 373         }
 374
 375         /*
 376          * Create the PFS's root inode.
 377          */
 378         if ((iroot = pmp->iroot) == NULL) {
 379                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 380                 pmp->iroot = iroot;
 381                 hammer2_inode_ref(iroot);
 382                 hammer2_inode_unlock(iroot);
 383         }
 384
 385         /*
 386          * Stop here if no chain is passed in.
 387          */
 388         if (chain == NULL)
 389                 goto done;
 390
 391         /*
 392          * When a chain is passed in we must add it to the PFS's root
 393          * inode, update pmp->pfs_types[], and update the syncronization
 394          * threads.
 395          *
 396          * At the moment empty spots can develop due to removals or failures.
 397          * Ultimately we want to re-fill these spots but doing so might
 398          * confused running code. XXX
 399          */
 400         hammer2_inode_ref(iroot);
 401         hammer2_mtx_ex(&iroot->lock);
 402         j = iroot->cluster.nchains;
 403
 404         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 405
 406         if (j == HAMMER2_MAXCLUSTER) {
 407                 kprintf("hammer2_mount: cluster full!\n");
 408                 /* XXX fatal error? */
 409         } else {
 410                 KKASSERT(chain->pmp == NULL);
 411                 chain->pmp = pmp;
 412                 hammer2_chain_ref(chain);
 413                 iroot->cluster.array[j].chain = chain;
 414                 pmp->pfs_types[j] = ripdata->meta.pfs_type;
 415                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 416                 pmp->pfs_hmps[j] = chain->hmp;
 417
 418                 /*
 419                  * If the PFS is already mounted we must account
 420                  * for the mount_count here.
 421                  */
 422                 if (pmp->mp)
 423                         ++chain->hmp->mount_count;
 424
 425                 /*
 426                  * May have to fixup dirty chain tracking.  Previous
 427                  * pmp was NULL so nothing to undo.
 428                  */
 429                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 430                         hammer2_pfs_memory_inc(pmp);
 431                 ++j;
 432         }
 433         iroot->cluster.nchains = j;
 434
 435         /*
 436          * Update nmasters from any PFS inode which is part of the cluster.
 437          * It is possible that this will result in a value which is too
 438          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 439          * override this value later on.
 440          *
 441          * (This informs us of masters that might not currently be
 442          *  discoverable by this mount).
 443          */
 444         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 445                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 446         }
 447
 448         /*
 449          * Count visible masters.  Masters are usually added with
 450          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 451          * are more (XXX and must update the master inodes).
 452          */
 453         count = 0;
 454         for (i = 0; i < iroot->cluster.nchains; ++i) {
 455                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 456                         ++count;
 457         }
 458         if (pmp->pfs_nmasters < count)
 459                 pmp->pfs_nmasters = count;
 460
 461         /*
 462          * Create missing synchronization and support threads.
 463          *
 464          * Single-node masters (including snapshots) have nothing to
 465          * synchronize and do not require this thread.
 466          *
 467          * Multi-node masters or any number of soft masters, slaves, copy,
 468          * or other PFS types need the thread.
 469          *
 470          * Each thread is responsible for its particular cluster index.
 471          * We use independent threads so stalls or mismatches related to
 472          * any given target do not affect other targets.
 473          */
 474         for (i = 0; i < iroot->cluster.nchains; ++i) {
 475                 /*
 476                  * Single-node masters (including snapshots) have nothing
 477                  * to synchronize and will make direct xops support calls,
 478                  * thus they do not require this thread.
 479                  *
 480                  * Note that there can be thousands of snapshots.  We do not
 481                  * want to create thousands of threads.
 482                  */
 483                 if (pmp->pfs_nmasters <= 1 &&
 484                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 485                         continue;
 486                 }
 487
 488                 /*
 489                  * Sync support thread
 490                  */
 491                 if (pmp->sync_thrs[i].td == NULL) {
 492                         hammer2_thr_create(&pmp->sync_thrs[i], pmp,
 493                                            "h2nod", i, -1,
 494                                            hammer2_primary_sync_thread);
 495                 }
 496         }
 497
 498         /*
 499          * Create missing Xop threads
 500          */
 501         if (pmp->mp)
 502                 hammer2_xop_helper_create(pmp);
 503
 504         hammer2_mtx_unlock(&iroot->lock);
 505         hammer2_inode_drop(iroot);
 506 done:
 507         return pmp;
 508 }
 509
 510 /*
 511  * Destroy a PFS, typically only occurs after the last mount on a device
 512  * has gone away.
 513  */
 514 static void
 515 hammer2_pfsfree(hammer2_pfs_t *pmp)
 516 {
 517         hammer2_inode_t *iroot;
 518         int i;
 519         int j;
 520
 521         /*
 522          * Cleanup our reference on iroot.  iroot is (should) not be needed
 523          * by the flush code.
 524          */
 525         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 526
 527         iroot = pmp->iroot;
 528         if (iroot) {
 529                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 530                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 531                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 532                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 533                 }
 534 #if REPORT_REFS_ERRORS
 535                 if (pmp->iroot->refs != 1)
 536                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 537                                 pmp->iroot, pmp->iroot->refs);
 538 #else
 539                 KKASSERT(pmp->iroot->refs == 1);
 540 #endif
 541                 /* ref for pmp->iroot */
 542                 hammer2_inode_drop(pmp->iroot);
 543                 pmp->iroot = NULL;
 544         }
 545
 546         kmalloc_destroy(&pmp->mmsg);
 547         kmalloc_destroy(&pmp->minode);
 548
 549         kfree(pmp, M_HAMMER2);
 550 }
 551
 552 /*
 553  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 554  * empty is terminated and freed.
 555  *
 556  * XXX inefficient.
 557  */
 558 static void
 559 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 560 {
 561         hammer2_pfs_t *pmp;
 562         hammer2_inode_t *iroot;
 563         hammer2_chain_t *rchain;
 564         int didfreeze;
 565         int i;
 566         int j;
 567
 568 again:
 569         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 570                 if ((iroot = pmp->iroot) == NULL)
 571                         continue;
 572                 if (hmp->spmp == pmp) {
 573                         kprintf("unmount hmp %p remove spmp %p\n",
 574                                 hmp, pmp);
 575                         hmp->spmp = NULL;
 576                 }
 577
 578                 /*
 579                  * Determine if this PFS is affected.  If it is we must
 580                  * freeze all management threads and lock its iroot.
 581                  *
 582                  * Freezing a management thread forces it idle, operations
 583                  * in-progress will be aborted and it will have to start
 584                  * over again when unfrozen, or exit if told to exit.
 585                  */
 586                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 587                         if (pmp->pfs_hmps[i] == hmp)
 588                                 break;
 589                 }
 590                 if (i != HAMMER2_MAXCLUSTER) {
 591                         /*
 592                          * Make sure all synchronization threads are locked
 593                          * down.
 594                          */
 595                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 596                                 if (pmp->pfs_hmps[i] == NULL)
 597                                         continue;
 598                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 599                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 600                                         hammer2_thr_freeze_async(
 601                                                 &pmp->xop_groups[j].thrs[i]);
 602                                 }
 603                         }
 604                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 605                                 if (pmp->pfs_hmps[i] == NULL)
 606                                         continue;
 607                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 608                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 609                                         hammer2_thr_freeze(
 610                                                 &pmp->xop_groups[j].thrs[i]);
 611                                 }
 612                         }
 613
 614                         /*
 615                          * Lock the inode and clean out matching chains.
 616                          * Note that we cannot use hammer2_inode_lock_*()
 617                          * here because that would attempt to validate the
 618                          * cluster that we are in the middle of ripping
 619                          * apart.
 620                          *
 621                          * WARNING! We are working directly on the inodes
 622                          *          embedded cluster.
 623                          */
 624                         hammer2_mtx_ex(&iroot->lock);
 625
 626                         /*
 627                          * Remove the chain from matching elements of the PFS.
 628                          */
 629                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 630                                 if (pmp->pfs_hmps[i] != hmp)
 631                                         continue;
 632                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 633                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 634                                         hammer2_thr_delete(
 635                                                 &pmp->xop_groups[j].thrs[i]);
 636                                 }
 637                                 rchain = iroot->cluster.array[i].chain;
 638                                 iroot->cluster.array[i].chain = NULL;
 639                                 pmp->pfs_types[i] = 0;
 640                                 if (pmp->pfs_names[i]) {
 641                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 642                                         pmp->pfs_names[i] = NULL;
 643                                 }
 644                                 if (rchain) {
 645                                         hammer2_chain_drop(rchain);
 646                                         /* focus hint */
 647                                         if (iroot->cluster.focus == rchain)
 648                                                 iroot->cluster.focus = NULL;
 649                                 }
 650                                 pmp->pfs_hmps[i] = NULL;
 651                         }
 652                         hammer2_mtx_unlock(&iroot->lock);
 653                         didfreeze = 1;  /* remaster, unfreeze down below */
 654                 } else {
 655                         didfreeze = 0;
 656                 }
 657
 658                 /*
 659                  * Cleanup trailing chains.  Gaps may remain.
 660                  */
 661                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 662                         if (pmp->pfs_hmps[i])
 663                                 break;
 664                 }
 665                 iroot->cluster.nchains = i + 1;
 666
 667                 /*
 668                  * If the PMP has no elements remaining we can destroy it.
 669                  * (this will transition management threads from frozen->exit).
 670                  */
 671                 if (iroot->cluster.nchains == 0) {
 672                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 673                                 hmp, pmp);
 674                         hammer2_pfsfree(pmp);
 675                         goto again;
 676                 }
 677
 678                 /*
 679                  * If elements still remain we need to set the REMASTER
 680                  * flag and unfreeze it.
 681                  */
 682                 if (didfreeze) {
 683                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 684                                 if (pmp->pfs_hmps[i] == NULL)
 685                                         continue;
 686                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 687                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 688                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 689                                         hammer2_thr_remaster(
 690                                                 &pmp->xop_groups[j].thrs[i]);
 691                                         hammer2_thr_unfreeze(
 692                                                 &pmp->xop_groups[j].thrs[i]);
 693                                 }
 694                         }
 695                 }
 696         }
 697 }
 698
 699 /*
 700  * Mount or remount HAMMER2 fileystem from physical media
 701  *
 702  *      mountroot
 703  *              mp              mount point structure
 704  *              path            NULL
 705  *              data            <unused>
 706  *              cred            <unused>
 707  *
 708  *      mount
 709  *              mp              mount point structure
 710  *              path            path to mount point
 711  *              data            pointer to argument structure in user space
 712  *                      volume  volume path (device@LABEL form)
 713  *                      hflags  user mount flags
 714  *              cred            user credentials
 715  *
 716  * RETURNS:     0       Success
 717  *              !0      error number
 718  */
 719 static
 720 int
 721 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 722                   struct ucred *cred)
 723 {
 724         struct hammer2_mount_info info;
 725         hammer2_pfs_t *pmp;
 726         hammer2_pfs_t *spmp;
 727         hammer2_dev_t *hmp;
 728         hammer2_key_t key_next;
 729         hammer2_key_t key_dummy;
 730         hammer2_key_t lhc;
 731         struct vnode *devvp;
 732         struct nlookupdata nd;
 733         hammer2_chain_t *parent;
 734         hammer2_chain_t *chain;
 735         hammer2_cluster_t *cluster;
 736         const hammer2_inode_data_t *ripdata;
 737         hammer2_blockref_t bref;
 738         struct file *fp;
 739         char devstr[MNAMELEN];
 740         size_t size;
 741         size_t done;
 742         char *dev;
 743         char *label;
 744         int ronly = 1;
 745         int error;
 746         int cache_index;
 747         int i;
 748
 749         hmp = NULL;
 750         pmp = NULL;
 751         dev = NULL;
 752         label = NULL;
 753         devvp = NULL;
 754         cache_index = -1;
 755
 756         kprintf("hammer2_mount\n");
 757
 758         if (path == NULL) {
 759                 /*
 760                  * Root mount
 761                  */
 762                 bzero(&info, sizeof(info));
 763                 info.cluster_fd = -1;
 764                 return (EOPNOTSUPP);
 765         } else {
 766                 /*
 767                  * Non-root mount or updating a mount
 768                  */
 769                 error = copyin(data, &info, sizeof(info));
 770                 if (error)
 771                         return (error);
 772
 773                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 774                 if (error)
 775                         return (error);
 776
 777                 /* Extract device and label */
 778                 dev = devstr;
 779                 label = strchr(devstr, '@');
 780                 if (label == NULL ||
 781                     ((label + 1) - dev) > done) {
 782                         return (EINVAL);
 783                 }
 784                 *label = '\0';
 785                 label++;
 786                 if (*label == '\0')
 787                         return (EINVAL);
 788
 789                 if (mp->mnt_flag & MNT_UPDATE) {
 790                         /*
 791                          * Update mount.  Note that pmp->iroot->cluster is
 792                          * an inode-embedded cluster and thus cannot be
 793                          * directly locked.
 794                          *
 795                          * XXX HAMMER2 needs to implement NFS export via
 796                          *     mountctl.
 797                          */
 798                         pmp = MPTOPMP(mp);
 799                         cluster = &pmp->iroot->cluster;
 800                         for (i = 0; i < cluster->nchains; ++i) {
 801                                 if (cluster->array[i].chain == NULL)
 802                                         continue;
 803                                 hmp = cluster->array[i].chain->hmp;
 804                                 devvp = hmp->devvp;
 805                                 error = hammer2_remount(hmp, mp, path,
 806                                                         devvp, cred);
 807                                 if (error)
 808                                         break;
 809                         }
 810
 811                         return error;
 812                 }
 813         }
 814
 815         /*
 816          * HMP device mount
 817          *
 818          * Lookup name and verify it refers to a block device.
 819          */
 820         error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 821         if (error == 0)
 822                 error = nlookup(&nd);
 823         if (error == 0)
 824                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 825         nlookup_done(&nd);
 826
 827         if (error == 0) {
 828                 if (vn_isdisk(devvp, &error))
 829                         error = vfs_mountedon(devvp);
 830         }
 831
 832         /*
 833          * Determine if the device has already been mounted.  After this
 834          * check hmp will be non-NULL if we are doing the second or more
 835          * hammer2 mounts from the same device.
 836          */
 837         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 838         TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 839                 if (hmp->devvp == devvp)
 840                         break;
 841         }
 842
 843         /*
 844          * Open the device if this isn't a secondary mount and construct
 845          * the H2 device mount (hmp).
 846          */
 847         if (hmp == NULL) {
 848                 hammer2_chain_t *schain;
 849                 hammer2_xid_t xid;
 850
 851                 if (error == 0 && vcount(devvp) > 0)
 852                         error = EBUSY;
 853
 854                 /*
 855                  * Now open the device
 856                  */
 857                 if (error == 0) {
 858                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
 859                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
 860                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
 861                         if (error == 0) {
 862                                 error = VOP_OPEN(devvp,
 863                                                  ronly ? FREAD : FREAD | FWRITE,
 864                                                  FSCRED, NULL);
 865                         }
 866                         vn_unlock(devvp);
 867                 }
 868                 if (error && devvp) {
 869                         vrele(devvp);
 870                         devvp = NULL;
 871                 }
 872                 if (error) {
 873                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 874                         return error;
 875                 }
 876                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
 877                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
 878                 hmp->ronly = ronly;
 879                 hmp->devvp = devvp;
 880                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
 881                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
 882                 RB_INIT(&hmp->iotree);
 883                 spin_init(&hmp->io_spin, "hm2mount_io");
 884                 spin_init(&hmp->list_spin, "hm2mount_list");
 885                 TAILQ_INIT(&hmp->flushq);
 886
 887                 lockinit(&hmp->vollk, "h2vol", 0, 0);
 888
 889                 /*
 890                  * vchain setup. vchain.data is embedded.
 891                  * vchain.refs is initialized and will never drop to 0.
 892                  *
 893                  * NOTE! voldata is not yet loaded.
 894                  */
 895                 hmp->vchain.hmp = hmp;
 896                 hmp->vchain.refs = 1;
 897                 hmp->vchain.data = (void *)&hmp->voldata;
 898                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
 899                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 900                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 901
 902                 hammer2_chain_core_init(&hmp->vchain);
 903                 /* hmp->vchain.u.xxx is left NULL */
 904
 905                 /*
 906                  * fchain setup.  fchain.data is embedded.
 907                  * fchain.refs is initialized and will never drop to 0.
 908                  *
 909                  * The data is not used but needs to be initialized to
 910                  * pass assertion muster.  We use this chain primarily
 911                  * as a placeholder for the freemap's top-level RBTREE
 912                  * so it does not interfere with the volume's topology
 913                  * RBTREE.
 914                  */
 915                 hmp->fchain.hmp = hmp;
 916                 hmp->fchain.refs = 1;
 917                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
 918                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
 919                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 920                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 921                 hmp->fchain.bref.methods =
 922                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
 923                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
 924
 925                 hammer2_chain_core_init(&hmp->fchain);
 926                 /* hmp->fchain.u.xxx is left NULL */
 927
 928                 /*
 929                  * Install the volume header and initialize fields from
 930                  * voldata.
 931                  */
 932                 error = hammer2_install_volume_header(hmp);
 933                 if (error) {
 934                         hammer2_unmount_helper(mp, NULL, hmp);
 935                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 936                         hammer2_vfs_unmount(mp, MNT_FORCE);
 937                         return error;
 938                 }
 939
 940                 /*
 941                  * Really important to get these right or flush will get
 942                  * confused.
 943                  */
 944                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0);
 945                 kprintf("alloc spmp %p tid %016jx\n",
 946                         hmp->spmp, hmp->voldata.mirror_tid);
 947                 spmp = hmp->spmp;
 948
 949                 /*
 950                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
 951                  * is inherited from the volume header.
 952                  */
 953                 xid = 0;
 954                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 955                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
 956                 hmp->vchain.pmp = spmp;
 957                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 958                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
 959                 hmp->fchain.pmp = spmp;
 960
 961                 /*
 962                  * First locate the super-root inode, which is key 0
 963                  * relative to the volume header's blockset.
 964                  *
 965                  * Then locate the root inode by scanning the directory keyspace
 966                  * represented by the label.
 967                  */
 968                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
 969                 schain = hammer2_chain_lookup(&parent, &key_dummy,
 970                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
 971                                       &cache_index, 0);
 972                 hammer2_chain_lookup_done(parent);
 973                 if (schain == NULL) {
 974                         kprintf("hammer2_mount: invalid super-root\n");
 975                         hammer2_unmount_helper(mp, NULL, hmp);
 976                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 977                         hammer2_vfs_unmount(mp, MNT_FORCE);
 978                         return EINVAL;
 979                 }
 980                 if (schain->error) {
 981                         kprintf("hammer2_mount: error %s reading super-root\n",
 982                                 hammer2_error_str(schain->error));
 983                         hammer2_chain_unlock(schain);
 984                         hammer2_chain_drop(schain);
 985                         schain = NULL;
 986                         hammer2_unmount_helper(mp, NULL, hmp);
 987                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 988                         hammer2_vfs_unmount(mp, MNT_FORCE);
 989                         return EINVAL;
 990                 }
 991
 992                 /*
 993                  * The super-root always uses an inode_tid of 1 when
 994                  * creating PFSs.
 995                  */
 996                 spmp->inode_tid = 1;
 997                 spmp->modify_tid = schain->bref.modify_tid + 1;
 998
 999                 /*
1000                  * Sanity-check schain's pmp and finish initialization.
1001                  * Any chain belonging to the super-root topology should
1002                  * have a NULL pmp (not even set to spmp).
1003                  */
1004                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1005                 KKASSERT(schain->pmp == NULL);
1006                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1007
1008                 /*
1009                  * Replace the dummy spmp->iroot with a real one.  It's
1010                  * easier to just do a wholesale replacement than to try
1011                  * to update the chain and fixup the iroot fields.
1012                  *
1013                  * The returned inode is locked with the supplied cluster.
1014                  */
1015                 cluster = hammer2_cluster_from_chain(schain);
1016                 hammer2_inode_drop(spmp->iroot);
1017                 spmp->iroot = NULL;
1018                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1019                 spmp->spmp_hmp = hmp;
1020                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1021                 spmp->pfs_hmps[0] = hmp;
1022                 hammer2_inode_ref(spmp->iroot);
1023                 hammer2_inode_unlock(spmp->iroot);
1024                 hammer2_cluster_unlock(cluster);
1025                 hammer2_cluster_drop(cluster);
1026                 schain = NULL;
1027                 /* leave spmp->iroot with one ref */
1028
1029                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1030                         error = hammer2_recovery(hmp);
1031                         /* XXX do something with error */
1032                 }
1033                 hammer2_update_pmps(hmp);
1034                 hammer2_iocom_init(hmp);
1035
1036                 /*
1037                  * Ref the cluster management messaging descriptor.  The mount
1038                  * program deals with the other end of the communications pipe.
1039                  */
1040                 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1041                 if (fp) {
1042                         hammer2_cluster_reconnect(hmp, fp);
1043                 } else {
1044                         kprintf("hammer2_mount: bad cluster_fd!\n");
1045                 }
1046         } else {
1047                 spmp = hmp->spmp;
1048         }
1049
1050         /*
1051          * Lookup the mount point under the media-localized super-root.
1052          * Scanning hammer2_pfslist doesn't help us because it represents
1053          * PFS cluster ids which can aggregate several named PFSs together.
1054          *
1055          * cluster->pmp will incorrectly point to spmp and must be fixed
1056          * up later on.
1057          */
1058         hammer2_inode_lock(spmp->iroot, 0);
1059         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1060         lhc = hammer2_dirhash(label, strlen(label));
1061         chain = hammer2_chain_lookup(&parent, &key_next,
1062                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1063                                      &cache_index, 0);
1064         while (chain) {
1065                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1066                     strcmp(label, chain->data->ipdata.filename) == 0) {
1067                         break;
1068                 }
1069                 chain = hammer2_chain_next(&parent, chain, &key_next,
1070                                             key_next,
1071                                             lhc + HAMMER2_DIRHASH_LOMASK,
1072                                             &cache_index, 0);
1073         }
1074         if (parent) {
1075                 hammer2_chain_unlock(parent);
1076                 hammer2_chain_drop(parent);
1077         }
1078         hammer2_inode_unlock(spmp->iroot);
1079
1080         /*
1081          * PFS could not be found?
1082          */
1083         if (chain == NULL) {
1084                 kprintf("hammer2_mount: PFS label not found\n");
1085                 hammer2_unmount_helper(mp, NULL, hmp);
1086                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1087                 hammer2_vfs_unmount(mp, MNT_FORCE);
1088
1089                 return EINVAL;
1090         }
1091
1092         /*
1093          * Acquire the pmp structure (it should have already been allocated
1094          * via hammer2_update_pmps() so do not pass cluster in to add to
1095          * available chains).
1096          *
1097          * Check if the cluster has already been mounted.  A cluster can
1098          * only be mounted once, use null mounts to mount additional copies.
1099          */
1100         ripdata = &chain->data->ipdata;
1101         bref = chain->bref;
1102         pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid);
1103         hammer2_chain_unlock(chain);
1104         hammer2_chain_drop(chain);
1105
1106         if (pmp->mp) {
1107                 kprintf("hammer2_mount: PFS already mounted!\n");
1108                 hammer2_unmount_helper(mp, NULL, hmp);
1109                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1110                 hammer2_vfs_unmount(mp, MNT_FORCE);
1111
1112                 return EBUSY;
1113         }
1114
1115         /*
1116          * Finish the mount
1117          */
1118         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1119
1120         mp->mnt_flag = MNT_LOCAL;
1121         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1122         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1123
1124         /*
1125          * required mount structure initializations
1126          */
1127         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1128         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1129
1130         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1131         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1132
1133         /*
1134          * Optional fields
1135          */
1136         mp->mnt_iosize_max = MAXPHYS;
1137
1138         /*
1139          * Connect up mount pointers.
1140          */
1141         hammer2_mount_helper(mp, pmp);
1142
1143         lockmgr(&hammer2_mntlk, LK_RELEASE);
1144
1145         /*
1146          * Finish setup
1147          */
1148         vfs_getnewfsid(mp);
1149         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1150         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1151         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1152
1153         copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
1154         bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1155         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1156         copyinstr(path, mp->mnt_stat.f_mntonname,
1157                   sizeof(mp->mnt_stat.f_mntonname) - 1,
1158                   &size);
1159
1160         /*
1161          * Initial statfs to prime mnt_stat.
1162          */
1163         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1164
1165         return 0;
1166 }
1167
1168 /*
1169  * Scan PFSs under the super-root and create hammer2_pfs structures.
1170  */
1171 static
1172 void
1173 hammer2_update_pmps(hammer2_dev_t *hmp)
1174 {
1175         const hammer2_inode_data_t *ripdata;
1176         hammer2_chain_t *parent;
1177         hammer2_chain_t *chain;
1178         hammer2_blockref_t bref;
1179         hammer2_pfs_t *spmp;
1180         hammer2_pfs_t *pmp;
1181         hammer2_key_t key_next;
1182         int cache_index = -1;
1183
1184         /*
1185          * Lookup mount point under the media-localized super-root.
1186          *
1187          * cluster->pmp will incorrectly point to spmp and must be fixed
1188          * up later on.
1189          */
1190         spmp = hmp->spmp;
1191         hammer2_inode_lock(spmp->iroot, 0);
1192         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1193         chain = hammer2_chain_lookup(&parent, &key_next,
1194                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1195                                          &cache_index, 0);
1196         while (chain) {
1197                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1198                         continue;
1199                 ripdata = &chain->data->ipdata;
1200                 bref = chain->bref;
1201                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1202
1203                 pmp = hammer2_pfsalloc(chain, ripdata, bref.modify_tid);
1204                 chain = hammer2_chain_next(&parent, chain, &key_next,
1205                                            key_next, HAMMER2_KEY_MAX,
1206                                            &cache_index, 0);
1207         }
1208         if (parent) {
1209                 hammer2_chain_unlock(parent);
1210                 hammer2_chain_drop(parent);
1211         }
1212         hammer2_inode_unlock(spmp->iroot);
1213 }
1214
1215 static
1216 int
1217 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path,
1218                 struct vnode *devvp, struct ucred *cred)
1219 {
1220         int error;
1221
1222         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1223                 error = hammer2_recovery(hmp);
1224         } else {
1225                 error = 0;
1226         }
1227         return error;
1228 }
1229
1230 static
1231 int
1232 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1233 {
1234         hammer2_pfs_t *pmp;
1235         int flags;
1236         int error = 0;
1237
1238         pmp = MPTOPMP(mp);
1239
1240         if (pmp == NULL)
1241                 return(0);
1242
1243         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1244
1245         /*
1246          * If mount initialization proceeded far enough we must flush
1247          * its vnodes and sync the underlying mount points.  Three syncs
1248          * are required to fully flush the filesystem (freemap updates lag
1249          * by one flush, and one extra for safety).
1250          */
1251         if (mntflags & MNT_FORCE)
1252                 flags = FORCECLOSE;
1253         else
1254                 flags = 0;
1255         if (pmp->iroot) {
1256                 error = vflush(mp, 0, flags);
1257                 if (error)
1258                         goto failed;
1259                 hammer2_vfs_sync(mp, MNT_WAIT);
1260                 hammer2_vfs_sync(mp, MNT_WAIT);
1261                 hammer2_vfs_sync(mp, MNT_WAIT);
1262         }
1263
1264         /*
1265          * Cleanup the frontend support XOPS threads
1266          */
1267         hammer2_xop_helper_cleanup(pmp);
1268
1269         /*
1270          * Cleanup our reference on ihidden.
1271          */
1272         if (pmp->ihidden) {
1273                 hammer2_inode_drop(pmp->ihidden);
1274                 pmp->ihidden = NULL;
1275         }
1276         if (pmp->mp)
1277                 hammer2_unmount_helper(mp, pmp, NULL);
1278
1279         error = 0;
1280 failed:
1281         lockmgr(&hammer2_mntlk, LK_RELEASE);
1282
1283         return (error);
1284 }
1285
1286 /*
1287  * Mount helper, hook the system mount into our PFS.
1288  * The mount lock is held.
1289  *
1290  * We must bump the mount_count on related devices for any
1291  * mounted PFSs.
1292  */
1293 static
1294 void
1295 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1296 {
1297         hammer2_cluster_t *cluster;
1298         hammer2_chain_t *rchain;
1299         int i;
1300
1301         mp->mnt_data = (qaddr_t)pmp;
1302         pmp->mp = mp;
1303
1304         /*
1305          * After pmp->mp is set we have to adjust hmp->mount_count.
1306          */
1307         cluster = &pmp->iroot->cluster;
1308         for (i = 0; i < cluster->nchains; ++i) {
1309                 rchain = cluster->array[i].chain;
1310                 if (rchain == NULL)
1311                         continue;
1312                 ++rchain->hmp->mount_count;
1313                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1314                         rchain->hmp, rchain->hmp->mount_count);
1315         }
1316
1317         /*
1318          * Create missing Xop threads
1319          */
1320         hammer2_xop_helper_create(pmp);
1321 }
1322
1323 /*
1324  * Mount helper, unhook the system mount from our PFS.
1325  * The mount lock is held.
1326  *
1327  * If hmp is supplied a mount responsible for being the first to open
1328  * the block device failed and the block device and all PFSs using the
1329  * block device must be cleaned up.
1330  *
1331  * If pmp is supplied multiple devices might be backing the PFS and each
1332  * must be disconnect.  This might not be the last PFS using some of the
1333  * underlying devices.  Also, we have to adjust our hmp->mount_count
1334  * accounting for the devices backing the pmp which is now undergoing an
1335  * unmount.
1336  */
1337 static
1338 void
1339 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1340 {
1341         hammer2_cluster_t *cluster;
1342         hammer2_chain_t *rchain;
1343         struct vnode *devvp;
1344         int dumpcnt;
1345         int ronly = 0;
1346         int i;
1347
1348         /*
1349          * If no device supplied this is a high-level unmount and we have to
1350          * to disconnect the mount, adjust mount_count, and locate devices
1351          * that might now have no mounts.
1352          */
1353         if (pmp) {
1354                 KKASSERT(hmp == NULL);
1355                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1356                 pmp->mp = NULL;
1357                 mp->mnt_data = NULL;
1358
1359                 /*
1360                  * After pmp->mp is cleared we have to account for
1361                  * mount_count.
1362                  */
1363                 cluster = &pmp->iroot->cluster;
1364                 for (i = 0; i < cluster->nchains; ++i) {
1365                         rchain = cluster->array[i].chain;
1366                         if (rchain == NULL)
1367                                 continue;
1368                         --rchain->hmp->mount_count;
1369                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1370                                 rchain->hmp, rchain->hmp->mount_count);
1371                         /* scrapping hmp now may invalidate the pmp */
1372                 }
1373 again:
1374                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1375                         if (hmp->mount_count == 0) {
1376                                 hammer2_unmount_helper(NULL, NULL, hmp);
1377                                 goto again;
1378                         }
1379                 }
1380                 return;
1381         }
1382
1383         /*
1384          * Try to terminate the block device.  We can't terminate it if
1385          * there are still PFSs referencing it.
1386          */
1387         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1388                 hmp, hmp->mount_count);
1389         if (hmp->mount_count)
1390                 return;
1391
1392         hammer2_pfsfree_scan(hmp);
1393         hammer2_dev_exlock(hmp);        /* XXX order */
1394
1395         /*
1396          * Cycle the volume data lock as a safety (probably not needed any
1397          * more).  To ensure everything is out we need to flush at least
1398          * three times.  (1) The running of the unlinkq can dirty the
1399          * filesystem, (2) A normal flush can dirty the freemap, and
1400          * (3) ensure that the freemap is fully synchronized.
1401          *
1402          * The next mount's recovery scan can clean everything up but we want
1403          * to leave the filesystem in a 100% clean state on a normal unmount.
1404          */
1405 #if 0
1406         hammer2_voldata_lock(hmp);
1407         hammer2_voldata_unlock(hmp);
1408 #endif
1409         hammer2_iocom_uninit(hmp);
1410
1411         if ((hmp->vchain.flags | hmp->fchain.flags) &
1412             HAMMER2_CHAIN_FLUSH_MASK) {
1413                 kprintf("hammer2_unmount: chains left over "
1414                         "after final sync\n");
1415                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1416                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1417
1418                 if (hammer2_debug & 0x0010)
1419                         Debugger("entered debugger");
1420         }
1421
1422         KKASSERT(hmp->spmp == NULL);
1423
1424         /*
1425          * Finish up with the device vnode
1426          */
1427         if ((devvp = hmp->devvp) != NULL) {
1428                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1429                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1430                 hmp->devvp = NULL;
1431                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1432                 vn_unlock(devvp);
1433                 vrele(devvp);
1434                 devvp = NULL;
1435         }
1436
1437         /*
1438          * Clear vchain/fchain flags that might prevent final cleanup
1439          * of these chains.
1440          */
1441         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1442                 atomic_clear_int(&hmp->vchain.flags,
1443                                  HAMMER2_CHAIN_MODIFIED);
1444                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1445                 hammer2_chain_drop(&hmp->vchain);
1446         }
1447         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1448                 atomic_clear_int(&hmp->vchain.flags,
1449                                  HAMMER2_CHAIN_UPDATE);
1450                 hammer2_chain_drop(&hmp->vchain);
1451         }
1452
1453         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1454                 atomic_clear_int(&hmp->fchain.flags,
1455                                  HAMMER2_CHAIN_MODIFIED);
1456                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1457                 hammer2_chain_drop(&hmp->fchain);
1458         }
1459         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1460                 atomic_clear_int(&hmp->fchain.flags,
1461                                  HAMMER2_CHAIN_UPDATE);
1462                 hammer2_chain_drop(&hmp->fchain);
1463         }
1464
1465         /*
1466          * Final drop of embedded freemap root chain to
1467          * clean up fchain.core (fchain structure is not
1468          * flagged ALLOCATED so it is cleaned out and then
1469          * left to rot).
1470          */
1471         hammer2_chain_drop(&hmp->fchain);
1472
1473         /*
1474          * Final drop of embedded volume root chain to clean
1475          * up vchain.core (vchain structure is not flagged
1476          * ALLOCATED so it is cleaned out and then left to
1477          * rot).
1478          */
1479         dumpcnt = 50;
1480         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1481         dumpcnt = 50;
1482         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1483         hammer2_dev_unlock(hmp);
1484         hammer2_chain_drop(&hmp->vchain);
1485
1486         hammer2_io_cleanup(hmp, &hmp->iotree);
1487         if (hmp->iofree_count) {
1488                 kprintf("io_cleanup: %d I/O's left hanging\n",
1489                         hmp->iofree_count);
1490         }
1491
1492         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1493         kmalloc_destroy(&hmp->mchain);
1494         kfree(hmp, M_HAMMER2);
1495 }
1496
1497 static
1498 int
1499 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1500              ino_t ino, struct vnode **vpp)
1501 {
1502         kprintf("hammer2_vget\n");
1503         return (EOPNOTSUPP);
1504 }
1505
1506 static
1507 int
1508 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1509 {
1510         hammer2_pfs_t *pmp;
1511         int error;
1512         struct vnode *vp;
1513
1514         pmp = MPTOPMP(mp);
1515         if (pmp->iroot == NULL) {
1516                 *vpp = NULL;
1517                 return EINVAL;
1518         }
1519
1520         error = 0;
1521         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1522
1523         while (pmp->inode_tid == 0) {
1524                 hammer2_xop_ipcluster_t *xop;
1525                 hammer2_inode_meta_t *meta;
1526
1527                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1528                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1529                 error = hammer2_xop_collect(&xop->head, 0);
1530
1531                 if (error == 0) {
1532                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1533                         pmp->iroot->meta = *meta;
1534                         pmp->inode_tid = meta->pfs_inum + 1;
1535                         if (pmp->inode_tid < HAMMER2_INODE_START)
1536                                 pmp->inode_tid = HAMMER2_INODE_START;
1537                         pmp->modify_tid =
1538                                 xop->head.cluster.focus->bref.modify_tid + 1;
1539                         kprintf("PFS: Starting inode %jd\n",
1540                                 (intmax_t)pmp->inode_tid);
1541                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1542                                 pmp->inode_tid, pmp->modify_tid);
1543                         wakeup(&pmp->iroot);
1544
1545                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1546
1547                         /*
1548                          * Prime the mount info.
1549                          */
1550                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1551
1552                         /*
1553                          * With the cluster operational, check for and
1554                          * install ihidden if needed.  The install_hidden
1555                          * code needs to get a transaction so we must unlock
1556                          * iroot around it.
1557                          *
1558                          * This is only applicable PFS mounts, there is no
1559                          * hidden directory in the spmp.
1560                          */
1561                         hammer2_inode_unlock(pmp->iroot);
1562                         hammer2_inode_install_hidden(pmp);
1563                         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1564
1565                         break;
1566                 }
1567
1568                 /*
1569                  * Loop, try again
1570                  */
1571                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1572                 hammer2_inode_unlock(pmp->iroot);
1573                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1574                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1575                 if (error == EINTR)
1576                         break;
1577         }
1578
1579         if (error) {
1580                 hammer2_inode_unlock(pmp->iroot);
1581                 *vpp = NULL;
1582         } else {
1583                 vp = hammer2_igetv(pmp->iroot, &error);
1584                 hammer2_inode_unlock(pmp->iroot);
1585                 *vpp = vp;
1586         }
1587
1588         return (error);
1589 }
1590
1591 /*
1592  * Filesystem status
1593  *
1594  * XXX incorporate ipdata->meta.inode_quota and data_quota
1595  */
1596 static
1597 int
1598 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1599 {
1600         hammer2_pfs_t *pmp;
1601         hammer2_dev_t *hmp;
1602         hammer2_blockref_t bref;
1603         int i;
1604
1605         /*
1606          * NOTE: iroot might not have validated the cluster yet.
1607          */
1608         pmp = MPTOPMP(mp);
1609
1610         mp->mnt_stat.f_files = 0;
1611         mp->mnt_stat.f_ffree = 0;
1612         mp->mnt_stat.f_blocks = 0;
1613         mp->mnt_stat.f_bfree = 0;
1614         mp->mnt_stat.f_bavail = 0;
1615
1616         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1617                 hmp = pmp->pfs_hmps[i];
1618                 if (hmp == NULL)
1619                         continue;
1620                 if (pmp->iroot->cluster.array[i].chain)
1621                         bref = pmp->iroot->cluster.array[i].chain->bref;
1622                 else
1623                         bzero(&bref, sizeof(bref));
1624
1625                 mp->mnt_stat.f_files = bref.inode_count;
1626                 mp->mnt_stat.f_ffree = 0;
1627                 mp->mnt_stat.f_blocks = (bref.data_count +
1628                                          hmp->voldata.allocator_free) /
1629                                         mp->mnt_vstat.f_bsize;
1630                 mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free /
1631                                         mp->mnt_vstat.f_bsize;
1632                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1633
1634                 *sbp = mp->mnt_stat;
1635         }
1636         return (0);
1637 }
1638
1639 static
1640 int
1641 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1642 {
1643         hammer2_pfs_t *pmp;
1644         hammer2_dev_t *hmp;
1645         hammer2_blockref_t bref;
1646         int i;
1647
1648         /*
1649          * NOTE: iroot might not have validated the cluster yet.
1650          */
1651         pmp = MPTOPMP(mp);
1652
1653         mp->mnt_vstat.f_bsize = 0;
1654         mp->mnt_vstat.f_files = 0;
1655         mp->mnt_vstat.f_ffree = 0;
1656         mp->mnt_vstat.f_blocks = 0;
1657         mp->mnt_vstat.f_bfree = 0;
1658         mp->mnt_vstat.f_bavail = 0;
1659
1660         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1661                 hmp = pmp->pfs_hmps[i];
1662                 if (hmp == NULL)
1663                         continue;
1664                 if (pmp->iroot->cluster.array[i].chain)
1665                         bref = pmp->iroot->cluster.array[i].chain->bref;
1666                 else
1667                         bzero(&bref, sizeof(bref));
1668
1669                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1670                 mp->mnt_vstat.f_files = bref.inode_count;
1671                 mp->mnt_vstat.f_ffree = 0;
1672                 mp->mnt_vstat.f_blocks = (bref.data_count +
1673                                          hmp->voldata.allocator_free) /
1674                                         mp->mnt_vstat.f_bsize;
1675                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1676                                         mp->mnt_vstat.f_bsize;
1677                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1678
1679                 *sbp = mp->mnt_vstat;
1680         }
1681         return (0);
1682 }
1683
1684 /*
1685  * Mount-time recovery (RW mounts)
1686  *
1687  * Updates to the free block table are allowed to lag flushes by one
1688  * transaction.  In case of a crash, then on a fresh mount we must do an
1689  * incremental scan of the last committed transaction id and make sure that
1690  * all related blocks have been marked allocated.
1691  *
1692  * The super-root topology and each PFS has its own transaction id domain,
1693  * so we must track PFS boundary transitions.
1694  */
1695 struct hammer2_recovery_elm {
1696         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1697         hammer2_chain_t *chain;
1698         hammer2_tid_t sync_tid;
1699 };
1700
1701 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1702
1703 struct hammer2_recovery_info {
1704         struct hammer2_recovery_list list;
1705         hammer2_tid_t   mtid;
1706         int     depth;
1707 };
1708
1709 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1710                         hammer2_chain_t *parent,
1711                         struct hammer2_recovery_info *info,
1712                         hammer2_tid_t sync_tid);
1713
1714 #define HAMMER2_RECOVERY_MAXDEPTH       10
1715
1716 static
1717 int
1718 hammer2_recovery(hammer2_dev_t *hmp)
1719 {
1720         struct hammer2_recovery_info info;
1721         struct hammer2_recovery_elm *elm;
1722         hammer2_chain_t *parent;
1723         hammer2_tid_t sync_tid;
1724         hammer2_tid_t mirror_tid;
1725         int error;
1726         int cumulative_error = 0;
1727
1728         hammer2_trans_init(hmp->spmp, 0);
1729
1730         sync_tid = hmp->voldata.freemap_tid;
1731         mirror_tid = hmp->voldata.mirror_tid;
1732
1733         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1734         if (sync_tid >= mirror_tid) {
1735                 kprintf(" no recovery needed\n");
1736         } else {
1737                 kprintf(" freemap recovery %016jx-%016jx\n",
1738                         sync_tid + 1, mirror_tid);
1739         }
1740
1741         TAILQ_INIT(&info.list);
1742         info.depth = 0;
1743         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1744         cumulative_error = hammer2_recovery_scan(hmp, parent,
1745                                                  &info, sync_tid);
1746         hammer2_chain_lookup_done(parent);
1747
1748         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1749                 TAILQ_REMOVE(&info.list, elm, entry);
1750                 parent = elm->chain;
1751                 sync_tid = elm->sync_tid;
1752                 kfree(elm, M_HAMMER2);
1753
1754                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1755                 error = hammer2_recovery_scan(hmp, parent,
1756                                               &info, hmp->voldata.freemap_tid);
1757                 hammer2_chain_unlock(parent);
1758                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
1759                 if (error)
1760                         cumulative_error = error;
1761         }
1762         hammer2_trans_done(hmp->spmp);
1763
1764         return cumulative_error;
1765 }
1766
1767 static
1768 int
1769 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1770                       struct hammer2_recovery_info *info,
1771                       hammer2_tid_t sync_tid)
1772 {
1773         const hammer2_inode_data_t *ripdata;
1774         hammer2_chain_t *chain;
1775         int cache_index;
1776         int cumulative_error = 0;
1777         int error;
1778
1779         /*
1780          * Adjust freemap to ensure that the block(s) are marked allocated.
1781          */
1782         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1783                 hammer2_freemap_adjust(hmp, &parent->bref,
1784                                        HAMMER2_FREEMAP_DORECOVER);
1785         }
1786
1787         /*
1788          * Check type for recursive scan
1789          */
1790         switch(parent->bref.type) {
1791         case HAMMER2_BREF_TYPE_VOLUME:
1792                 /* data already instantiated */
1793                 break;
1794         case HAMMER2_BREF_TYPE_INODE:
1795                 /*
1796                  * Must instantiate data for DIRECTDATA test and also
1797                  * for recursion.
1798                  */
1799                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1800                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
1801                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1802                         /* not applicable to recovery scan */
1803                         hammer2_chain_unlock(parent);
1804                         return 0;
1805                 }
1806                 hammer2_chain_unlock(parent);
1807                 break;
1808         case HAMMER2_BREF_TYPE_INDIRECT:
1809                 /*
1810                  * Must instantiate data for recursion
1811                  */
1812                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1813                 hammer2_chain_unlock(parent);
1814                 break;
1815         case HAMMER2_BREF_TYPE_DATA:
1816         case HAMMER2_BREF_TYPE_FREEMAP:
1817         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1818         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1819                 /* not applicable to recovery scan */
1820                 return 0;
1821                 break;
1822         default:
1823                 return EDOM;
1824         }
1825
1826         /*
1827          * Defer operation if depth limit reached or if we are crossing a
1828          * PFS boundary.
1829          */
1830         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1831                 struct hammer2_recovery_elm *elm;
1832
1833                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1834                 elm->chain = parent;
1835                 elm->sync_tid = sync_tid;
1836                 hammer2_chain_ref(parent);
1837                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
1838                 /* unlocked by caller */
1839
1840                 return(0);
1841         }
1842
1843
1844         /*
1845          * Recursive scan of the last flushed transaction only.  We are
1846          * doing this without pmp assignments so don't leave the chains
1847          * hanging around after we are done with them.
1848          */
1849         cache_index = 0;
1850         chain = hammer2_chain_scan(parent, NULL, &cache_index,
1851                                    HAMMER2_LOOKUP_NODATA);
1852         while (chain) {
1853                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
1854                 if (chain->bref.mirror_tid > sync_tid) {
1855                         ++info->depth;
1856                         error = hammer2_recovery_scan(hmp, chain,
1857                                                       info, sync_tid);
1858                         --info->depth;
1859                         if (error)
1860                                 cumulative_error = error;
1861                 }
1862
1863                 /*
1864                  * Flush the recovery at the PFS boundary to stage it for
1865                  * the final flush of the super-root topology.
1866                  */
1867                 if ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
1868                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
1869                         hammer2_flush(chain, info->mtid, 1);
1870                 }
1871                 chain = hammer2_chain_scan(parent, chain, &cache_index,
1872                                            HAMMER2_LOOKUP_NODATA);
1873         }
1874
1875         return cumulative_error;
1876 }
1877
1878 /*
1879  * Sync a mount point; this is called on a per-mount basis from the
1880  * filesystem syncer process periodically and whenever a user issues
1881  * a sync.
1882  */
1883 int
1884 hammer2_vfs_sync(struct mount *mp, int waitfor)
1885 {
1886         hammer2_xop_flush_t *xop;
1887         struct hammer2_sync_info info;
1888         hammer2_inode_t *iroot;
1889         hammer2_pfs_t *pmp;
1890         int flags;
1891         int error;
1892
1893         pmp = MPTOPMP(mp);
1894         iroot = pmp->iroot;
1895         KKASSERT(iroot);
1896         KKASSERT(iroot->pmp == pmp);
1897
1898         /*
1899          * We can't acquire locks on existing vnodes while in a transaction
1900          * without risking a deadlock.  This assumes that vfsync() can be
1901          * called without the vnode locked (which it can in DragonFly).
1902          * Otherwise we'd have to implement a multi-pass or flag the lock
1903          * failures and retry.
1904          *
1905          * The reclamation code interlocks with the sync list's token
1906          * (by removing the vnode from the scan list) before unlocking
1907          * the inode, giving us time to ref the inode.
1908          */
1909         /*flags = VMSC_GETVP;*/
1910         flags = 0;
1911         if (waitfor & MNT_LAZY)
1912                 flags |= VMSC_ONEPASS;
1913
1914 #if 0
1915         /*
1916          * Preflush the vnodes using a normal transaction before interlocking
1917          * with a flush transaction.
1918          */
1919         hammer2_trans_init(pmp, 0);
1920         info.error = 0;
1921         info.waitfor = MNT_NOWAIT;
1922         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1923         hammer2_trans_done(pmp);
1924 #endif
1925
1926         /*
1927          * Start our flush transaction.  This does not return until all
1928          * concurrent transactions have completed and will prevent any
1929          * new transactions from running concurrently, except for the
1930          * buffer cache transactions.
1931          *
1932          * For efficiency do an async pass before making sure with a
1933          * synchronous pass on all related buffer cache buffers.  It
1934          * should theoretically not be possible for any new file buffers
1935          * to be instantiated during this sequence.
1936          */
1937         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH |
1938                                 HAMMER2_TRANS_PREFLUSH);
1939         hammer2_inode_run_unlinkq(pmp);
1940
1941         info.error = 0;
1942         info.waitfor = MNT_NOWAIT;
1943         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1944         info.waitfor = MNT_WAIT;
1945         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
1946
1947         /*
1948          * Clear PREFLUSH.  This prevents (or asserts on) any new logical
1949          * buffer cache flushes which occur during the flush.  Device buffers
1950          * are not affected.
1951          */
1952         hammer2_bioq_sync(pmp);
1953         hammer2_trans_clear_preflush(pmp);
1954
1955         /*
1956          * Use the XOP interface to concurrently flush all nodes to
1957          * synchronize the PFSROOT subtopology to the media.  A standard
1958          * end-of-scan ENOENT error indicates cluster sufficiency.
1959          *
1960          * Note that this flush will not be visible on crash recovery until
1961          * we flush the super-root topology in the next loop.
1962          *
1963          * XXX For now wait for all flushes to complete.
1964          */
1965         if (iroot) {
1966                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
1967                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
1968                 error = hammer2_xop_collect(&xop->head,
1969                                             HAMMER2_XOP_COLLECT_WAITALL);
1970                 if (error == ENOENT)
1971                         error = 0;
1972         } else {
1973                 error = 0;
1974         }
1975         hammer2_trans_done(pmp);
1976
1977         return (error);
1978 }
1979
1980 /*
1981  * Sync passes.
1982  *
1983  * Note that we ignore the tranasction mtid we got above.  Instead,
1984  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
1985  * transactions.
1986  */
1987 static int
1988 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
1989 {
1990         struct hammer2_sync_info *info = data;
1991         hammer2_inode_t *ip;
1992         int error;
1993
1994         /*
1995          * Degenerate cases.  Note that ip == NULL typically means the
1996          * syncer vnode itself and we don't want to vclrisdirty() in that
1997          * situation.
1998          */
1999         ip = VTOI(vp);
2000         if (ip == NULL) {
2001                 return(0);
2002         }
2003         if (vp->v_type == VNON || vp->v_type == VBAD) {
2004                 vclrisdirty(vp);
2005                 return(0);
2006         }
2007
2008         /*
2009          * VOP_FSYNC will start a new transaction so replicate some code
2010          * here to do it inline (see hammer2_vop_fsync()).
2011          *
2012          * WARNING: The vfsync interacts with the buffer cache and might
2013          *          block, we can't hold the inode lock at that time.
2014          *          However, we MUST ref ip before blocking to ensure that
2015          *          it isn't ripped out from under us (since we do not
2016          *          hold a lock on the vnode).
2017          */
2018         hammer2_inode_ref(ip);
2019         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2020             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2021                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2022                 hammer2_inode_fsync(ip);
2023         }
2024         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2025             RB_EMPTY(&vp->v_rbdirty_tree)) {
2026                 vclrisdirty(vp);
2027         }
2028
2029         hammer2_inode_drop(ip);
2030 #if 1
2031         error = 0;
2032         if (error)
2033                 info->error = error;
2034 #endif
2035         return(0);
2036 }
2037
2038 static
2039 int
2040 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2041 {
2042         return (0);
2043 }
2044
2045 static
2046 int
2047 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2048                struct fid *fhp, struct vnode **vpp)
2049 {
2050         return (0);
2051 }
2052
2053 static
2054 int
2055 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2056                  int *exflagsp, struct ucred **credanonp)
2057 {
2058         return (0);
2059 }
2060
2061 /*
2062  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2063  * header into the HMP
2064  *
2065  * XXX read four volhdrs and use the one with the highest TID whos CRC
2066  *     matches.
2067  *
2068  * XXX check iCRCs.
2069  *
2070  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2071  *     nonexistant locations.
2072  *
2073  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2074  */
2075 static
2076 int
2077 hammer2_install_volume_header(hammer2_dev_t *hmp)
2078 {
2079         hammer2_volume_data_t *vd;
2080         struct buf *bp;
2081         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2082         int error_reported;
2083         int error;
2084         int valid;
2085         int i;
2086
2087         error_reported = 0;
2088         error = 0;
2089         valid = 0;
2090         bp = NULL;
2091
2092         /*
2093          * There are up to 4 copies of the volume header (syncs iterate
2094          * between them so there is no single master).  We don't trust the
2095          * volu_size field so we don't know precisely how large the filesystem
2096          * is, so depend on the OS to return an error if we go beyond the
2097          * block device's EOF.
2098          */
2099         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2100                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2101                               HAMMER2_VOLUME_BYTES, &bp);
2102                 if (error) {
2103                         brelse(bp);
2104                         bp = NULL;
2105                         continue;
2106                 }
2107
2108                 vd = (struct hammer2_volume_data *) bp->b_data;
2109                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2110                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2111                         brelse(bp);
2112                         bp = NULL;
2113                         continue;
2114                 }
2115
2116                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2117                         /* XXX: Reversed-endianness filesystem */
2118                         kprintf("hammer2: reverse-endian filesystem detected");
2119                         brelse(bp);
2120                         bp = NULL;
2121                         continue;
2122                 }
2123
2124                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2125                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2126                                       HAMMER2_VOLUME_ICRC0_SIZE);
2127                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2128                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2129                                        HAMMER2_VOLUME_ICRC1_SIZE);
2130                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2131                         kprintf("hammer2 volume header crc "
2132                                 "mismatch copy #%d %08x/%08x\n",
2133                                 i, crc0, crc);
2134                         error_reported = 1;
2135                         brelse(bp);
2136                         bp = NULL;
2137                         continue;
2138                 }
2139                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2140                         valid = 1;
2141                         hmp->voldata = *vd;
2142                         hmp->volhdrno = i;
2143                 }
2144                 brelse(bp);
2145                 bp = NULL;
2146         }
2147         if (valid) {
2148                 hmp->volsync = hmp->voldata;
2149                 error = 0;
2150                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2151                         kprintf("hammer2: using volume header #%d\n",
2152                                 hmp->volhdrno);
2153                 }
2154         } else {
2155                 error = EINVAL;
2156                 kprintf("hammer2: no valid volume headers found!\n");
2157         }
2158         return (error);
2159 }
2160
2161 /*
2162  * This handles hysteresis on regular file flushes.  Because the BIOs are
2163  * routed to a thread it is possible for an excessive number to build up
2164  * and cause long front-end stalls long before the runningbuffspace limit
2165  * is hit, so we implement hammer2_flush_pipe to control the
2166  * hysteresis.
2167  *
2168  * This is a particular problem when compression is used.
2169  */
2170 void
2171 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2172 {
2173         atomic_add_int(&pmp->count_lwinprog, 1);
2174 }
2175
2176 void
2177 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2178 {
2179         int lwinprog;
2180
2181         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2182         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2183             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2184                 atomic_clear_int(&pmp->count_lwinprog,
2185                                  HAMMER2_LWINPROG_WAITING);
2186                 wakeup(&pmp->count_lwinprog);
2187         }
2188         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2189             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2190                 atomic_clear_int(&pmp->count_lwinprog,
2191                                  HAMMER2_LWINPROG_WAITING0);
2192                 wakeup(&pmp->count_lwinprog);
2193         }
2194 }
2195
2196 void
2197 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2198 {
2199         int lwinprog;
2200         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2201                                     HAMMER2_LWINPROG_WAITING0;
2202
2203         for (;;) {
2204                 lwinprog = pmp->count_lwinprog;
2205                 cpu_ccfence();
2206                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2207                         break;
2208                 tsleep_interlock(&pmp->count_lwinprog, 0);
2209                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2210                 lwinprog = pmp->count_lwinprog;
2211                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2212                         break;
2213                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2214         }
2215 }
2216
2217 /*
2218  * Manage excessive memory resource use for chain and related
2219  * structures.
2220  */
2221 void
2222 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2223 {
2224         uint32_t waiting;
2225         uint32_t count;
2226         uint32_t limit;
2227 #if 0
2228         static int zzticks;
2229 #endif
2230
2231         /*
2232          * Atomic check condition and wait.  Also do an early speedup of
2233          * the syncer to try to avoid hitting the wait.
2234          */
2235         for (;;) {
2236                 waiting = pmp->inmem_dirty_chains;
2237                 cpu_ccfence();
2238                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2239
2240                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2241                 if (limit < hammer2_limit_dirty_chains)
2242                         limit = hammer2_limit_dirty_chains;
2243                 if (limit < 1000)
2244                         limit = 1000;
2245
2246 #if 0
2247                 if ((int)(ticks - zzticks) > hz) {
2248                         zzticks = ticks;
2249                         kprintf("count %ld %ld\n", count, limit);
2250                 }
2251 #endif
2252
2253                 /*
2254                  * Block if there are too many dirty chains present, wait
2255                  * for the flush to clean some out.
2256                  */
2257                 if (count > limit) {
2258                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2259                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2260                                                waiting,
2261                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2262                                 speedup_syncer(pmp->mp);
2263                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2264                                        "chnmem", hz);
2265                         }
2266                         continue;       /* loop on success or fail */
2267                 }
2268
2269                 /*
2270                  * Try to start an early flush before we are forced to block.
2271                  */
2272                 if (count > limit * 7 / 10)
2273                         speedup_syncer(pmp->mp);
2274                 break;
2275         }
2276 }
2277
2278 void
2279 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2280 {
2281         if (pmp) {
2282                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2283         }
2284 }
2285
2286 void
2287 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2288 {
2289         uint32_t waiting;
2290
2291         if (pmp == NULL)
2292                 return;
2293
2294         for (;;) {
2295                 waiting = pmp->inmem_dirty_chains;
2296                 cpu_ccfence();
2297                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2298                                        waiting,
2299                                        (waiting - 1) &
2300                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2301                         break;
2302                 }
2303         }
2304
2305         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2306                 wakeup(&pmp->inmem_dirty_chains);
2307 }
2308
2309 /*
2310  * Debugging
2311  */
2312 void
2313 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2314 {
2315         hammer2_chain_t *scan;
2316         hammer2_chain_t *parent;
2317
2318         --*countp;
2319         if (*countp == 0) {
2320                 kprintf("%*.*s...\n", tab, tab, "");
2321                 return;
2322         }
2323         if (*countp < 0)
2324                 return;
2325         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2326                 tab, tab, "", pfx,
2327                 chain, chain->bref.type,
2328                 chain->bref.key, chain->bref.keybits,
2329                 chain->bref.mirror_tid);
2330
2331         kprintf("%*.*s      [%08x] (%s) refs=%d",
2332                 tab, tab, "",
2333                 chain->flags,
2334                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2335                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2336                 chain->refs);
2337
2338         parent = chain->parent;
2339         if (parent)
2340                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2341                         tab, tab, "",
2342                         parent, parent->flags, parent->refs);
2343         if (RB_EMPTY(&chain->core.rbtree)) {
2344                 kprintf("\n");
2345         } else {
2346                 kprintf(" {\n");
2347                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2348                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2349                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2350                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2351                                 chain->data->ipdata.filename);
2352                 else
2353                         kprintf("%*.*s}\n", tab, tab, "");
2354         }
2355 }