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_cluster_t *cluster,
 319                  const hammer2_inode_data_t *ripdata,
 320                  hammer2_tid_t modify_tid)
 321 {
 322         hammer2_chain_t *rchain;
 323         hammer2_inode_t *iroot;
 324         hammer2_pfs_t *pmp;
 325         int count;
 326         int i;
 327         int j;
 328
 329         /*
 330          * Locate or create the PFS based on the cluster id.  If ripdata
 331          * is NULL this is a spmp which is unique and is always allocated.
 332          */
 333         if (ripdata) {
 334                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 335                         if (bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 336                                  sizeof(pmp->pfs_clid)) == 0) {
 337                                         break;
 338                         }
 339                 }
 340         } else {
 341                 pmp = NULL;
 342         }
 343
 344         if (pmp == NULL) {
 345                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 346                 hammer2_trans_manage_init(pmp);
 347                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 348                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 349                 lockinit(&pmp->lock, "pfslk", 0, 0);
 350                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 351                 RB_INIT(&pmp->inum_tree);
 352                 TAILQ_INIT(&pmp->unlinkq);
 353                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 354
 355                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 356                         hammer2_xop_group_init(pmp, &pmp->xop_groups[j]);
 357
 358                 /*
 359                  * Save the last media transaction id for the flusher.  Set
 360                  * initial
 361                  */
 362                 if (ripdata)
 363                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 364                 hammer2_mtx_init(&pmp->wthread_mtx, "h2wthr");
 365                 bioq_init(&pmp->wthread_bioq);
 366                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 367
 368                 /*
 369                  * The synchronization thread may start too early, make
 370                  * sure it stays frozen until we are ready to let it go.
 371                  * XXX
 372                  */
 373                 /*
 374                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 375                                          HAMMER2_THREAD_REMASTER;
 376                 */
 377         }
 378
 379         /*
 380          * Create the PFS's root inode.
 381          */
 382         if ((iroot = pmp->iroot) == NULL) {
 383                 iroot = hammer2_inode_get(pmp, NULL, NULL);
 384                 pmp->iroot = iroot;
 385                 hammer2_inode_ref(iroot);
 386                 hammer2_inode_unlock(iroot, NULL);
 387         }
 388
 389         /*
 390          * Stop here if no cluster is passed in.
 391          */
 392         if (cluster == NULL)
 393                 goto done;
 394
 395         /*
 396          * When a cluster is passed in we must add the cluster's chains
 397          * to the PFS's root inode, update pmp->pfs_types[], and update
 398          * the syncronization threads.
 399          *
 400          * At the moment empty spots can develop due to removals or failures.
 401          * Ultimately we want to re-fill these spots but doing so might
 402          * confused running code. XXX
 403          */
 404         hammer2_inode_ref(iroot);
 405         hammer2_mtx_ex(&iroot->lock);
 406         j = iroot->cluster.nchains;
 407
 408         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 409
 410         for (i = 0; i < cluster->nchains; ++i) {
 411                 if (j == HAMMER2_MAXCLUSTER)
 412                         break;
 413                 rchain = cluster->array[i].chain;
 414                 KKASSERT(rchain->pmp == NULL);
 415                 rchain->pmp = pmp;
 416                 hammer2_chain_ref(rchain);
 417                 iroot->cluster.array[j].chain = rchain;
 418                 pmp->pfs_types[j] = ripdata->meta.pfs_type;
 419                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 420
 421                 /*
 422                  * If the PFS is already mounted we must account
 423                  * for the mount_count here.
 424                  */
 425                 if (pmp->mp)
 426                         ++rchain->hmp->mount_count;
 427
 428                 /*
 429                  * May have to fixup dirty chain tracking.  Previous
 430                  * pmp was NULL so nothing to undo.
 431                  */
 432                 if (rchain->flags & HAMMER2_CHAIN_MODIFIED)
 433                         hammer2_pfs_memory_inc(pmp);
 434                 ++j;
 435         }
 436         iroot->cluster.nchains = j;
 437
 438         if (i != cluster->nchains) {
 439                 kprintf("hammer2_mount: cluster full!\n");
 440                 /* XXX fatal error? */
 441         }
 442
 443         /*
 444          * Update nmasters from any PFS inode which is part of the cluster.
 445          * It is possible that this will result in a value which is too
 446          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 447          * override this value later on.
 448          *
 449          * (This informs us of masters that might not currently be
 450          *  discoverable by this mount).
 451          */
 452         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 453                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 454         }
 455
 456         /*
 457          * Count visible masters.  Masters are usually added with
 458          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 459          * are more (XXX and must update the master inodes).
 460          */
 461         count = 0;
 462         for (i = 0; i < iroot->cluster.nchains; ++i) {
 463                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 464                         ++count;
 465         }
 466         if (pmp->pfs_nmasters < count)
 467                 pmp->pfs_nmasters = count;
 468
 469         /*
 470          * Create missing synchronization and support threads.
 471          *
 472          * Single-node masters (including snapshots) have nothing to
 473          * synchronize and do not require this thread.
 474          *
 475          * Multi-node masters or any number of soft masters, slaves, copy,
 476          * or other PFS types need the thread.
 477          *
 478          * Each thread is responsible for its particular cluster index.
 479          * We use independent threads so stalls or mismatches related to
 480          * any given target do not affect other targets.
 481          */
 482         for (i = 0; i < iroot->cluster.nchains; ++i) {
 483                 /*
 484                  * Single-node masters (including snapshots) have nothing
 485                  * to synchronize and will make direct xops support calls,
 486                  * thus they do not require this thread.
 487                  *
 488                  * Note that there can be thousands of snapshots.  We do not
 489                  * want to create thousands of threads.
 490                  */
 491                 if (pmp->pfs_nmasters <= 1 &&
 492                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 493                         continue;
 494                 }
 495
 496                 /*
 497                  * Sync support thread
 498                  */
 499                 if (pmp->sync_thrs[i].td == NULL) {
 500                         hammer2_thr_create(&pmp->sync_thrs[i], pmp,
 501                                            "h2nod", i, -1,
 502                                            hammer2_primary_sync_thread);
 503                 }
 504         }
 505
 506         /*
 507          * Create missing Xop threads
 508          */
 509         if (pmp->mp)
 510                 hammer2_xop_helper_create(pmp);
 511
 512         hammer2_mtx_unlock(&iroot->lock);
 513         hammer2_inode_drop(iroot);
 514 done:
 515         return pmp;
 516 }
 517
 518 /*
 519  * Destroy a PFS, typically only occurs after the last mount on a device
 520  * has gone away.
 521  */
 522 static void
 523 hammer2_pfsfree(hammer2_pfs_t *pmp)
 524 {
 525         hammer2_inode_t *iroot;
 526         int i;
 527         int j;
 528
 529         /*
 530          * Cleanup our reference on iroot.  iroot is (should) not be needed
 531          * by the flush code.
 532          */
 533         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 534
 535         iroot = pmp->iroot;
 536         if (iroot) {
 537                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 538                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 539                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 540                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 541                 }
 542 #if REPORT_REFS_ERRORS
 543                 if (pmp->iroot->refs != 1)
 544                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 545                                 pmp->iroot, pmp->iroot->refs);
 546 #else
 547                 KKASSERT(pmp->iroot->refs == 1);
 548 #endif
 549                 /* ref for pmp->iroot */
 550                 hammer2_inode_drop(pmp->iroot);
 551                 pmp->iroot = NULL;
 552         }
 553
 554         kmalloc_destroy(&pmp->mmsg);
 555         kmalloc_destroy(&pmp->minode);
 556
 557         kfree(pmp, M_HAMMER2);
 558 }
 559
 560 /*
 561  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 562  * empty is terminated and freed.
 563  *
 564  * XXX inefficient.
 565  */
 566 static void
 567 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 568 {
 569         hammer2_pfs_t *pmp;
 570         hammer2_inode_t *iroot;
 571         hammer2_cluster_t *cluster;
 572         hammer2_chain_t *rchain;
 573         int didfreeze;
 574         int i;
 575         int j;
 576
 577 again:
 578         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 579                 if ((iroot = pmp->iroot) == NULL)
 580                         continue;
 581                 if (hmp->spmp == pmp) {
 582                         kprintf("unmount hmp %p remove spmp %p\n",
 583                                 hmp, pmp);
 584                         hmp->spmp = NULL;
 585                 }
 586
 587                 /*
 588                  * Determine if this PFS is affected.  If it is we must
 589                  * freeze all management threads and lock its iroot.
 590                  *
 591                  * Freezing a management thread forces it idle, operations
 592                  * in-progress will be aborted and it will have to start
 593                  * over again when unfrozen, or exit if told to exit.
 594                  */
 595                 cluster = &iroot->cluster;
 596                 for (i = 0; i < cluster->nchains; ++i) {
 597                         rchain = cluster->array[i].chain;
 598                         if (rchain == NULL || rchain->hmp != hmp)
 599                                 continue;
 600                         break;
 601                 }
 602                 if (i != cluster->nchains) {
 603                         /*
 604                          * Make sure all synchronization threads are locked
 605                          * down.
 606                          */
 607                         for (i = 0; i < iroot->cluster.nchains; ++i) {
 608                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 609                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 610                                         hammer2_thr_freeze_async(
 611                                                 &pmp->xop_groups[j].thrs[i]);
 612                                 }
 613                         }
 614                         for (i = 0; i < iroot->cluster.nchains; ++i) {
 615                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 616                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 617                                         hammer2_thr_freeze(
 618                                                 &pmp->xop_groups[j].thrs[i]);
 619                                 }
 620                         }
 621
 622                         /*
 623                          * Lock the inode and clean out matching chains.
 624                          * Note that we cannot use hammer2_inode_lock_*()
 625                          * here because that would attempt to validate the
 626                          * cluster that we are in the middle of ripping
 627                          * apart.
 628                          *
 629                          * WARNING! We are working directly on the inodes
 630                          *          embedded cluster.
 631                          */
 632                         hammer2_mtx_ex(&iroot->lock);
 633
 634                         /*
 635                          * Remove the chain from matching elements of the PFS.
 636                          */
 637                         for (i = 0; i < cluster->nchains; ++i) {
 638                                 rchain = cluster->array[i].chain;
 639                                 if (rchain == NULL || rchain->hmp != hmp)
 640                                         continue;
 641                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 642                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 643                                         hammer2_thr_delete(
 644                                                 &pmp->xop_groups[j].thrs[i]);
 645                                 }
 646                                 rchain = cluster->array[i].chain;
 647                                 cluster->array[i].chain = NULL;
 648                                 pmp->pfs_types[i] = 0;
 649                                 if (pmp->pfs_names[i]) {
 650                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 651                                         pmp->pfs_names[i] = NULL;
 652                                 }
 653                                 hammer2_chain_drop(rchain);
 654
 655                                 /* focus hint */
 656                                 if (cluster->focus == rchain)
 657                                         cluster->focus = NULL;
 658                         }
 659                         hammer2_mtx_unlock(&iroot->lock);
 660                         didfreeze = 1;  /* remaster, unfreeze down below */
 661                 } else {
 662                         didfreeze = 0;
 663                 }
 664
 665                 /*
 666                  * Cleanup trailing chains.  Do not reorder chains (for now).
 667                  * XXX might remove more than we intended.
 668                  */
 669                 while (i > 0) {
 670                         if (cluster->array[i - 1].chain)
 671                                 break;
 672                         --i;
 673                 }
 674                 cluster->nchains = i;
 675
 676                 /*
 677                  * If the PMP has no elements remaining we can destroy it.
 678                  * (this will transition management threads from frozen->exit).
 679                  */
 680                 if (cluster->nchains == 0) {
 681                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 682                                 hmp, pmp);
 683                         hammer2_pfsfree(pmp);
 684                         goto again;
 685                 }
 686
 687                 /*
 688                  * If elements still remain we need to set the REMASTER
 689                  * flag and unfreeze it.
 690                  */
 691                 if (didfreeze) {
 692                         for (i = 0; i < iroot->cluster.nchains; ++i) {
 693                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 694                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 695                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 696                                         hammer2_thr_remaster(
 697                                                 &pmp->xop_groups[j].thrs[i]);
 698                                         hammer2_thr_unfreeze(
 699                                                 &pmp->xop_groups[j].thrs[i]);
 700                                 }
 701                         }
 702                 }
 703         }
 704 }
 705
 706 /*
 707  * Mount or remount HAMMER2 fileystem from physical media
 708  *
 709  *      mountroot
 710  *              mp              mount point structure
 711  *              path            NULL
 712  *              data            <unused>
 713  *              cred            <unused>
 714  *
 715  *      mount
 716  *              mp              mount point structure
 717  *              path            path to mount point
 718  *              data            pointer to argument structure in user space
 719  *                      volume  volume path (device@LABEL form)
 720  *                      hflags  user mount flags
 721  *              cred            user credentials
 722  *
 723  * RETURNS:     0       Success
 724  *              !0      error number
 725  */
 726 static
 727 int
 728 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 729                   struct ucred *cred)
 730 {
 731         struct hammer2_mount_info info;
 732         hammer2_pfs_t *pmp;
 733         hammer2_pfs_t *spmp;
 734         hammer2_dev_t *hmp;
 735         hammer2_key_t key_next;
 736         hammer2_key_t key_dummy;
 737         hammer2_key_t lhc;
 738         struct vnode *devvp;
 739         struct nlookupdata nd;
 740         hammer2_chain_t *parent;
 741         hammer2_cluster_t *cluster;
 742         hammer2_cluster_t *cparent;
 743         const hammer2_inode_data_t *ripdata;
 744         hammer2_blockref_t bref;
 745         struct file *fp;
 746         char devstr[MNAMELEN];
 747         size_t size;
 748         size_t done;
 749         char *dev;
 750         char *label;
 751         int ronly = 1;
 752         int error;
 753         int cache_index;
 754         int i;
 755
 756         hmp = NULL;
 757         pmp = NULL;
 758         dev = NULL;
 759         label = NULL;
 760         devvp = NULL;
 761         cache_index = -1;
 762
 763         kprintf("hammer2_mount\n");
 764
 765         if (path == NULL) {
 766                 /*
 767                  * Root mount
 768                  */
 769                 bzero(&info, sizeof(info));
 770                 info.cluster_fd = -1;
 771                 return (EOPNOTSUPP);
 772         } else {
 773                 /*
 774                  * Non-root mount or updating a mount
 775                  */
 776                 error = copyin(data, &info, sizeof(info));
 777                 if (error)
 778                         return (error);
 779
 780                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 781                 if (error)
 782                         return (error);
 783
 784                 /* Extract device and label */
 785                 dev = devstr;
 786                 label = strchr(devstr, '@');
 787                 if (label == NULL ||
 788                     ((label + 1) - dev) > done) {
 789                         return (EINVAL);
 790                 }
 791                 *label = '\0';
 792                 label++;
 793                 if (*label == '\0')
 794                         return (EINVAL);
 795
 796                 if (mp->mnt_flag & MNT_UPDATE) {
 797                         /*
 798                          * Update mount.  Note that pmp->iroot->cluster is
 799                          * an inode-embedded cluster and thus cannot be
 800                          * directly locked.
 801                          *
 802                          * XXX HAMMER2 needs to implement NFS export via
 803                          *     mountctl.
 804                          */
 805                         pmp = MPTOPMP(mp);
 806                         cluster = &pmp->iroot->cluster;
 807                         for (i = 0; i < cluster->nchains; ++i) {
 808                                 if (cluster->array[i].chain == NULL)
 809                                         continue;
 810                                 hmp = cluster->array[i].chain->hmp;
 811                                 devvp = hmp->devvp;
 812                                 error = hammer2_remount(hmp, mp, path,
 813                                                         devvp, cred);
 814                                 if (error)
 815                                         break;
 816                         }
 817                         /*hammer2_inode_install_hidden(pmp);*/
 818
 819                         return error;
 820                 }
 821         }
 822
 823         /*
 824          * HMP device mount
 825          *
 826          * Lookup name and verify it refers to a block device.
 827          */
 828         error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 829         if (error == 0)
 830                 error = nlookup(&nd);
 831         if (error == 0)
 832                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 833         nlookup_done(&nd);
 834
 835         if (error == 0) {
 836                 if (vn_isdisk(devvp, &error))
 837                         error = vfs_mountedon(devvp);
 838         }
 839
 840         /*
 841          * Determine if the device has already been mounted.  After this
 842          * check hmp will be non-NULL if we are doing the second or more
 843          * hammer2 mounts from the same device.
 844          */
 845         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 846         TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 847                 if (hmp->devvp == devvp)
 848                         break;
 849         }
 850
 851         /*
 852          * Open the device if this isn't a secondary mount and construct
 853          * the H2 device mount (hmp).
 854          */
 855         if (hmp == NULL) {
 856                 hammer2_chain_t *schain;
 857                 hammer2_xid_t xid;
 858
 859                 if (error == 0 && vcount(devvp) > 0)
 860                         error = EBUSY;
 861
 862                 /*
 863                  * Now open the device
 864                  */
 865                 if (error == 0) {
 866                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
 867                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
 868                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
 869                         if (error == 0) {
 870                                 error = VOP_OPEN(devvp,
 871                                                  ronly ? FREAD : FREAD | FWRITE,
 872                                                  FSCRED, NULL);
 873                         }
 874                         vn_unlock(devvp);
 875                 }
 876                 if (error && devvp) {
 877                         vrele(devvp);
 878                         devvp = NULL;
 879                 }
 880                 if (error) {
 881                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 882                         return error;
 883                 }
 884                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
 885                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
 886                 hmp->ronly = ronly;
 887                 hmp->devvp = devvp;
 888                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
 889                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
 890                 RB_INIT(&hmp->iotree);
 891                 spin_init(&hmp->io_spin, "hm2mount_io");
 892                 spin_init(&hmp->list_spin, "hm2mount_list");
 893                 TAILQ_INIT(&hmp->flushq);
 894
 895                 lockinit(&hmp->vollk, "h2vol", 0, 0);
 896
 897                 /*
 898                  * vchain setup. vchain.data is embedded.
 899                  * vchain.refs is initialized and will never drop to 0.
 900                  *
 901                  * NOTE! voldata is not yet loaded.
 902                  */
 903                 hmp->vchain.hmp = hmp;
 904                 hmp->vchain.refs = 1;
 905                 hmp->vchain.data = (void *)&hmp->voldata;
 906                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
 907                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 908                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 909
 910                 hammer2_chain_core_init(&hmp->vchain);
 911                 /* hmp->vchain.u.xxx is left NULL */
 912
 913                 /*
 914                  * fchain setup.  fchain.data is embedded.
 915                  * fchain.refs is initialized and will never drop to 0.
 916                  *
 917                  * The data is not used but needs to be initialized to
 918                  * pass assertion muster.  We use this chain primarily
 919                  * as a placeholder for the freemap's top-level RBTREE
 920                  * so it does not interfere with the volume's topology
 921                  * RBTREE.
 922                  */
 923                 hmp->fchain.hmp = hmp;
 924                 hmp->fchain.refs = 1;
 925                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
 926                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
 927                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 928                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 929                 hmp->fchain.bref.methods =
 930                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
 931                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
 932
 933                 hammer2_chain_core_init(&hmp->fchain);
 934                 /* hmp->fchain.u.xxx is left NULL */
 935
 936                 /*
 937                  * Install the volume header and initialize fields from
 938                  * voldata.
 939                  */
 940                 error = hammer2_install_volume_header(hmp);
 941                 if (error) {
 942                         hammer2_unmount_helper(mp, NULL, hmp);
 943                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 944                         hammer2_vfs_unmount(mp, MNT_FORCE);
 945                         return error;
 946                 }
 947
 948                 /*
 949                  * Really important to get these right or flush will get
 950                  * confused.
 951                  */
 952                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0);
 953                 kprintf("alloc spmp %p tid %016jx\n",
 954                         hmp->spmp, hmp->voldata.mirror_tid);
 955                 spmp = hmp->spmp;
 956
 957                 /*
 958                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
 959                  * is inherited from the volume header.
 960                  */
 961                 xid = 0;
 962                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 963                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
 964                 hmp->vchain.pmp = spmp;
 965                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 966                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
 967                 hmp->fchain.pmp = spmp;
 968
 969                 /*
 970                  * First locate the super-root inode, which is key 0
 971                  * relative to the volume header's blockset.
 972                  *
 973                  * Then locate the root inode by scanning the directory keyspace
 974                  * represented by the label.
 975                  */
 976                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
 977                 schain = hammer2_chain_lookup(&parent, &key_dummy,
 978                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
 979                                       &cache_index, 0);
 980                 hammer2_chain_lookup_done(parent);
 981                 if (schain == NULL) {
 982                         kprintf("hammer2_mount: invalid super-root\n");
 983                         hammer2_unmount_helper(mp, NULL, hmp);
 984                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 985                         hammer2_vfs_unmount(mp, MNT_FORCE);
 986                         return EINVAL;
 987                 }
 988                 if (schain->error) {
 989                         kprintf("hammer2_mount: error %s reading super-root\n",
 990                                 hammer2_error_str(schain->error));
 991                         hammer2_chain_unlock(schain);
 992                         hammer2_chain_drop(schain);
 993                         schain = NULL;
 994                         hammer2_unmount_helper(mp, NULL, hmp);
 995                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 996                         hammer2_vfs_unmount(mp, MNT_FORCE);
 997                         return EINVAL;
 998                 }
 999
1000                 /*
1001                  * The super-root always uses an inode_tid of 1 when
1002                  * creating PFSs.
1003                  */
1004                 spmp->inode_tid = 1;
1005                 spmp->modify_tid = schain->bref.modify_tid + 1;
1006
1007                 /*
1008                  * Sanity-check schain's pmp and finish initialization.
1009                  * Any chain belonging to the super-root topology should
1010                  * have a NULL pmp (not even set to spmp).
1011                  */
1012                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1013                 KKASSERT(schain->pmp == NULL);
1014                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1015
1016                 /*
1017                  * Replace the dummy spmp->iroot with a real one.  It's
1018                  * easier to just do a wholesale replacement than to try
1019                  * to update the chain and fixup the iroot fields.
1020                  *
1021                  * The returned inode is locked with the supplied cluster.
1022                  */
1023                 cluster = hammer2_cluster_from_chain(schain);
1024                 hammer2_inode_drop(spmp->iroot);
1025                 spmp->iroot = NULL;
1026                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster);
1027                 spmp->spmp_hmp = hmp;
1028                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1029                 hammer2_inode_ref(spmp->iroot);
1030                 hammer2_inode_unlock(spmp->iroot, cluster);
1031                 schain = NULL;
1032                 /* leave spmp->iroot with one ref */
1033
1034                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1035                         error = hammer2_recovery(hmp);
1036                         /* XXX do something with error */
1037                 }
1038                 hammer2_update_pmps(hmp);
1039                 hammer2_iocom_init(hmp);
1040
1041                 /*
1042                  * Ref the cluster management messaging descriptor.  The mount
1043                  * program deals with the other end of the communications pipe.
1044                  */
1045                 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1046                 if (fp) {
1047                         hammer2_cluster_reconnect(hmp, fp);
1048                 } else {
1049                         kprintf("hammer2_mount: bad cluster_fd!\n");
1050                 }
1051         } else {
1052                 spmp = hmp->spmp;
1053         }
1054
1055         /*
1056          * Lookup the mount point under the media-localized super-root.
1057          * Scanning hammer2_pfslist doesn't help us because it represents
1058          * PFS cluster ids which can aggregate several named PFSs together.
1059          *
1060          * cluster->pmp will incorrectly point to spmp and must be fixed
1061          * up later on.
1062          */
1063         hammer2_inode_lock(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1064         cparent = hammer2_inode_cluster(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1065         lhc = hammer2_dirhash(label, strlen(label));
1066         cluster = hammer2_cluster_lookup(cparent, &key_next,
1067                                       lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1068                                       0);
1069         while (cluster) {
1070                 if (hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE &&
1071                     strcmp(label,
1072                        hammer2_cluster_rdata(cluster)->ipdata.filename) == 0) {
1073                         break;
1074                 }
1075                 cluster = hammer2_cluster_next(cparent, cluster, &key_next,
1076                                             key_next,
1077                                             lhc + HAMMER2_DIRHASH_LOMASK, 0);
1078         }
1079         hammer2_inode_unlock(spmp->iroot, cparent);
1080
1081         /*
1082          * PFS could not be found?
1083          */
1084         if (cluster == NULL) {
1085                 kprintf("hammer2_mount: PFS label not found\n");
1086                 hammer2_unmount_helper(mp, NULL, hmp);
1087                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1088                 hammer2_vfs_unmount(mp, MNT_FORCE);
1089
1090                 return EINVAL;
1091         }
1092
1093         /*
1094          * Acquire the pmp structure (it should have already been allocated
1095          * via hammer2_update_pmps() so do not pass cluster in to add to
1096          * available chains).
1097          *
1098          * Check if the cluster has already been mounted.  A cluster can
1099          * only be mounted once, use null mounts to mount additional copies.
1100          */
1101         ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1102         hammer2_cluster_bref(cluster, &bref);
1103         pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid);
1104         hammer2_cluster_unlock(cluster);
1105         hammer2_cluster_drop(cluster);
1106
1107         if (pmp->mp) {
1108                 kprintf("hammer2_mount: PFS already mounted!\n");
1109                 hammer2_unmount_helper(mp, NULL, hmp);
1110                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1111                 hammer2_vfs_unmount(mp, MNT_FORCE);
1112
1113                 return EBUSY;
1114         }
1115
1116         /*
1117          * Finish the mount
1118          */
1119         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1120
1121         mp->mnt_flag = MNT_LOCAL;
1122         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1123         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1124
1125         /*
1126          * required mount structure initializations
1127          */
1128         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1129         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1130
1131         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1132         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1133
1134         /*
1135          * Optional fields
1136          */
1137         mp->mnt_iosize_max = MAXPHYS;
1138
1139         /*
1140          * Connect up mount pointers.
1141          */
1142         hammer2_mount_helper(mp, pmp);
1143
1144         lockmgr(&hammer2_mntlk, LK_RELEASE);
1145
1146         /*
1147          * A mounted PFS needs a write thread for logical buffers and
1148          * a hidden directory for deletions of open files.  These features
1149          * are not used by unmounted PFSs.
1150          *
1151          * The logical file buffer bio write thread handles things like
1152          * physical block assignment and compression.
1153          */
1154         pmp->wthread_destroy = 0;
1155         lwkt_create(hammer2_write_thread, pmp,
1156                     &pmp->wthread_td, NULL, 0, -1, "h2pfs-%s", label);
1157
1158         /*
1159          * With the cluster operational install ihidden.
1160          * (only applicable to pfs mounts, not applicable to spmp)
1161          */
1162         hammer2_inode_install_hidden(pmp);
1163
1164         /*
1165          * Finish setup
1166          */
1167         vfs_getnewfsid(mp);
1168         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1169         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1170         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1171
1172         copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
1173         bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1174         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1175         copyinstr(path, mp->mnt_stat.f_mntonname,
1176                   sizeof(mp->mnt_stat.f_mntonname) - 1,
1177                   &size);
1178
1179         /*
1180          * Initial statfs to prime mnt_stat.
1181          */
1182         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1183
1184         return 0;
1185 }
1186
1187 /*
1188  * Scan PFSs under the super-root and create hammer2_pfs structures.
1189  */
1190 static
1191 void
1192 hammer2_update_pmps(hammer2_dev_t *hmp)
1193 {
1194         const hammer2_inode_data_t *ripdata;
1195         hammer2_cluster_t *cparent;
1196         hammer2_cluster_t *cluster;
1197         hammer2_blockref_t bref;
1198         hammer2_pfs_t *spmp;
1199         hammer2_pfs_t *pmp;
1200         hammer2_key_t key_next;
1201
1202         /*
1203          * Lookup mount point under the media-localized super-root.
1204          *
1205          * cluster->pmp will incorrectly point to spmp and must be fixed
1206          * up later on.
1207          */
1208         spmp = hmp->spmp;
1209         hammer2_inode_lock(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1210         cparent = hammer2_inode_cluster(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1211         cluster = hammer2_cluster_lookup(cparent, &key_next,
1212                                          HAMMER2_KEY_MIN,
1213                                          HAMMER2_KEY_MAX,
1214                                          0);
1215         while (cluster) {
1216                 if (hammer2_cluster_type(cluster) != HAMMER2_BREF_TYPE_INODE)
1217                         continue;
1218                 ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1219                 hammer2_cluster_bref(cluster, &bref);
1220                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1221
1222                 pmp = hammer2_pfsalloc(cluster, ripdata, bref.modify_tid);
1223                 cluster = hammer2_cluster_next(cparent, cluster,
1224                                                &key_next,
1225                                                key_next,
1226                                                HAMMER2_KEY_MAX,
1227                                                0);
1228         }
1229         hammer2_inode_unlock(spmp->iroot, cparent);
1230 }
1231
1232 static
1233 int
1234 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path,
1235                 struct vnode *devvp, struct ucred *cred)
1236 {
1237         int error;
1238
1239         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1240                 error = hammer2_recovery(hmp);
1241         } else {
1242                 error = 0;
1243         }
1244         return error;
1245 }
1246
1247 static
1248 int
1249 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1250 {
1251         hammer2_pfs_t *pmp;
1252         int flags;
1253         int error = 0;
1254
1255         pmp = MPTOPMP(mp);
1256
1257         if (pmp == NULL)
1258                 return(0);
1259
1260         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1261
1262         /*
1263          * If mount initialization proceeded far enough we must flush
1264          * its vnodes and sync the underlying mount points.  Three syncs
1265          * are required to fully flush the filesystem (freemap updates lag
1266          * by one flush, and one extra for safety).
1267          */
1268         if (mntflags & MNT_FORCE)
1269                 flags = FORCECLOSE;
1270         else
1271                 flags = 0;
1272         if (pmp->iroot) {
1273                 error = vflush(mp, 0, flags);
1274                 if (error)
1275                         goto failed;
1276                 hammer2_vfs_sync(mp, MNT_WAIT);
1277                 hammer2_vfs_sync(mp, MNT_WAIT);
1278                 hammer2_vfs_sync(mp, MNT_WAIT);
1279         }
1280
1281         if (pmp->wthread_td) {
1282                 hammer2_mtx_ex(&pmp->wthread_mtx);
1283                 pmp->wthread_destroy = 1;
1284                 wakeup(&pmp->wthread_bioq);
1285                 while (pmp->wthread_destroy != -1) {
1286                         mtxsleep(&pmp->wthread_destroy,
1287                                 &pmp->wthread_mtx, 0,
1288                                 "umount-sleep", 0);
1289                 }
1290                 hammer2_mtx_unlock(&pmp->wthread_mtx);
1291                 pmp->wthread_td = NULL;
1292         }
1293
1294         /*
1295          * Cleanup the frontend support XOPS threads
1296          */
1297         hammer2_xop_helper_cleanup(pmp);
1298
1299         /*
1300          * Cleanup our reference on ihidden.
1301          */
1302         if (pmp->ihidden) {
1303                 hammer2_inode_drop(pmp->ihidden);
1304                 pmp->ihidden = NULL;
1305         }
1306         if (pmp->mp)
1307                 hammer2_unmount_helper(mp, pmp, NULL);
1308
1309         error = 0;
1310 failed:
1311         lockmgr(&hammer2_mntlk, LK_RELEASE);
1312
1313         return (error);
1314 }
1315
1316 /*
1317  * Mount helper, hook the system mount into our PFS.
1318  * The mount lock is held.
1319  *
1320  * We must bump the mount_count on related devices for any
1321  * mounted PFSs.
1322  */
1323 static
1324 void
1325 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1326 {
1327         hammer2_cluster_t *cluster;
1328         hammer2_chain_t *rchain;
1329         int i;
1330
1331         mp->mnt_data = (qaddr_t)pmp;
1332         pmp->mp = mp;
1333
1334         /*
1335          * After pmp->mp is set we have to adjust hmp->mount_count.
1336          */
1337         cluster = &pmp->iroot->cluster;
1338         for (i = 0; i < cluster->nchains; ++i) {
1339                 rchain = cluster->array[i].chain;
1340                 if (rchain == NULL)
1341                         continue;
1342                 ++rchain->hmp->mount_count;
1343                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1344                         rchain->hmp, rchain->hmp->mount_count);
1345         }
1346
1347         /*
1348          * Create missing Xop threads
1349          */
1350         hammer2_xop_helper_create(pmp);
1351 }
1352
1353 /*
1354  * Mount helper, unhook the system mount from our PFS.
1355  * The mount lock is held.
1356  *
1357  * If hmp is supplied a mount responsible for being the first to open
1358  * the block device failed and the block device and all PFSs using the
1359  * block device must be cleaned up.
1360  *
1361  * If pmp is supplied multiple devices might be backing the PFS and each
1362  * must be disconnect.  This might not be the last PFS using some of the
1363  * underlying devices.  Also, we have to adjust our hmp->mount_count
1364  * accounting for the devices backing the pmp which is now undergoing an
1365  * unmount.
1366  */
1367 static
1368 void
1369 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1370 {
1371         hammer2_cluster_t *cluster;
1372         hammer2_chain_t *rchain;
1373         struct vnode *devvp;
1374         int dumpcnt;
1375         int ronly = 0;
1376         int i;
1377
1378         /*
1379          * If no device supplied this is a high-level unmount and we have to
1380          * to disconnect the mount, adjust mount_count, and locate devices
1381          * that might now have no mounts.
1382          */
1383         if (pmp) {
1384                 KKASSERT(hmp == NULL);
1385                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1386                 pmp->mp = NULL;
1387                 mp->mnt_data = NULL;
1388
1389                 /*
1390                  * After pmp->mp is cleared we have to account for
1391                  * mount_count.
1392                  */
1393                 cluster = &pmp->iroot->cluster;
1394                 for (i = 0; i < cluster->nchains; ++i) {
1395                         rchain = cluster->array[i].chain;
1396                         if (rchain == NULL)
1397                                 continue;
1398                         --rchain->hmp->mount_count;
1399                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1400                                 rchain->hmp, rchain->hmp->mount_count);
1401                         /* scrapping hmp now may invalidate the pmp */
1402                 }
1403 again:
1404                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1405                         if (hmp->mount_count == 0) {
1406                                 hammer2_unmount_helper(NULL, NULL, hmp);
1407                                 goto again;
1408                         }
1409                 }
1410                 return;
1411         }
1412
1413         /*
1414          * Try to terminate the block device.  We can't terminate it if
1415          * there are still PFSs referencing it.
1416          */
1417         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1418                 hmp, hmp->mount_count);
1419         if (hmp->mount_count)
1420                 return;
1421
1422         hammer2_pfsfree_scan(hmp);
1423         hammer2_dev_exlock(hmp);        /* XXX order */
1424
1425         /*
1426          * Cycle the volume data lock as a safety (probably not needed any
1427          * more).  To ensure everything is out we need to flush at least
1428          * three times.  (1) The running of the unlinkq can dirty the
1429          * filesystem, (2) A normal flush can dirty the freemap, and
1430          * (3) ensure that the freemap is fully synchronized.
1431          *
1432          * The next mount's recovery scan can clean everything up but we want
1433          * to leave the filesystem in a 100% clean state on a normal unmount.
1434          */
1435 #if 0
1436         hammer2_voldata_lock(hmp);
1437         hammer2_voldata_unlock(hmp);
1438 #endif
1439         hammer2_iocom_uninit(hmp);
1440
1441         if ((hmp->vchain.flags | hmp->fchain.flags) &
1442             HAMMER2_CHAIN_FLUSH_MASK) {
1443                 kprintf("hammer2_unmount: chains left over "
1444                         "after final sync\n");
1445                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1446                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1447
1448                 if (hammer2_debug & 0x0010)
1449                         Debugger("entered debugger");
1450         }
1451
1452         KKASSERT(hmp->spmp == NULL);
1453
1454         /*
1455          * Finish up with the device vnode
1456          */
1457         if ((devvp = hmp->devvp) != NULL) {
1458                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1459                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1460                 hmp->devvp = NULL;
1461                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1462                 vn_unlock(devvp);
1463                 vrele(devvp);
1464                 devvp = NULL;
1465         }
1466
1467         /*
1468          * Clear vchain/fchain flags that might prevent final cleanup
1469          * of these chains.
1470          */
1471         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1472                 atomic_clear_int(&hmp->vchain.flags,
1473                                  HAMMER2_CHAIN_MODIFIED);
1474                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1475                 hammer2_chain_drop(&hmp->vchain);
1476         }
1477         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1478                 atomic_clear_int(&hmp->vchain.flags,
1479                                  HAMMER2_CHAIN_UPDATE);
1480                 hammer2_chain_drop(&hmp->vchain);
1481         }
1482
1483         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1484                 atomic_clear_int(&hmp->fchain.flags,
1485                                  HAMMER2_CHAIN_MODIFIED);
1486                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1487                 hammer2_chain_drop(&hmp->fchain);
1488         }
1489         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1490                 atomic_clear_int(&hmp->fchain.flags,
1491                                  HAMMER2_CHAIN_UPDATE);
1492                 hammer2_chain_drop(&hmp->fchain);
1493         }
1494
1495         /*
1496          * Final drop of embedded freemap root chain to
1497          * clean up fchain.core (fchain structure is not
1498          * flagged ALLOCATED so it is cleaned out and then
1499          * left to rot).
1500          */
1501         hammer2_chain_drop(&hmp->fchain);
1502
1503         /*
1504          * Final drop of embedded volume root chain to clean
1505          * up vchain.core (vchain structure is not flagged
1506          * ALLOCATED so it is cleaned out and then left to
1507          * rot).
1508          */
1509         dumpcnt = 50;
1510         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1511         dumpcnt = 50;
1512         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1513         hammer2_dev_unlock(hmp);
1514         hammer2_chain_drop(&hmp->vchain);
1515
1516         hammer2_io_cleanup(hmp, &hmp->iotree);
1517         if (hmp->iofree_count) {
1518                 kprintf("io_cleanup: %d I/O's left hanging\n",
1519                         hmp->iofree_count);
1520         }
1521
1522         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1523         kmalloc_destroy(&hmp->mchain);
1524         kfree(hmp, M_HAMMER2);
1525 }
1526
1527 static
1528 int
1529 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1530              ino_t ino, struct vnode **vpp)
1531 {
1532         kprintf("hammer2_vget\n");
1533         return (EOPNOTSUPP);
1534 }
1535
1536 static
1537 int
1538 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1539 {
1540         hammer2_pfs_t *pmp;
1541         hammer2_cluster_t *cparent;
1542         int error;
1543         struct vnode *vp;
1544
1545         pmp = MPTOPMP(mp);
1546         if (pmp->iroot == NULL) {
1547                 *vpp = NULL;
1548                 error = EINVAL;
1549         } else {
1550                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_ALWAYS |
1551                                                HAMMER2_RESOLVE_SHARED);
1552                 cparent = hammer2_inode_cluster(pmp->iroot,
1553                                                 HAMMER2_RESOLVE_ALWAYS |
1554                                                 HAMMER2_RESOLVE_SHARED);
1555
1556                 /*
1557                  * Initialize pmp->inode_tid and pmp->modify_tid on first access
1558                  * to the root of mount that resolves good.
1559                  * XXX probably not the best place for this.
1560                  */
1561                 if (pmp->inode_tid == 0 &&
1562                     cparent->error == 0 && cparent->focus) {
1563                         const hammer2_inode_data_t *ripdata;
1564                         hammer2_blockref_t bref;
1565
1566                         ripdata = &hammer2_cluster_rdata(cparent)->ipdata;
1567                         hammer2_cluster_bref(cparent, &bref);
1568                         pmp->inode_tid = ripdata->meta.pfs_inum + 1;
1569                         if (pmp->inode_tid < HAMMER2_INODE_START)
1570                                 pmp->inode_tid = HAMMER2_INODE_START;
1571                         pmp->modify_tid = bref.modify_tid + 1;
1572                         pmp->iroot->meta = ripdata->meta;
1573                         hammer2_cluster_bref(cparent, &pmp->iroot->bref);
1574                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1575                                 pmp->inode_tid, pmp->modify_tid);
1576                 }
1577
1578                 vp = hammer2_igetv(pmp->iroot, &error);
1579                 hammer2_inode_unlock(pmp->iroot, cparent);
1580                 *vpp = vp;
1581                 if (vp == NULL)
1582                         kprintf("vnodefail\n");
1583         }
1584
1585         return (error);
1586 }
1587
1588 /*
1589  * Filesystem status
1590  *
1591  * XXX incorporate ipdata->meta.inode_quota and data_quota
1592  */
1593 static
1594 int
1595 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1596 {
1597         hammer2_pfs_t *pmp;
1598         hammer2_dev_t *hmp;
1599         hammer2_blockref_t bref;
1600
1601         pmp = MPTOPMP(mp);
1602         KKASSERT(pmp->iroot->cluster.nchains >= 1);
1603         hmp = pmp->iroot->cluster.focus->hmp;   /* iroot retains focus */
1604         bref = pmp->iroot->cluster.focus->bref; /* no lock */
1605
1606         mp->mnt_stat.f_files = bref.inode_count;
1607         mp->mnt_stat.f_ffree = 0;
1608         mp->mnt_stat.f_blocks = (bref.data_count +
1609                                  hmp->voldata.allocator_free) /
1610                                 mp->mnt_vstat.f_bsize;
1611         mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free /
1612                                 mp->mnt_vstat.f_bsize;
1613         mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1614
1615         *sbp = mp->mnt_stat;
1616         return (0);
1617 }
1618
1619 static
1620 int
1621 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1622 {
1623         hammer2_pfs_t *pmp;
1624         hammer2_dev_t *hmp;
1625         hammer2_blockref_t bref;
1626
1627         pmp = MPTOPMP(mp);
1628         KKASSERT(pmp->iroot->cluster.nchains >= 1);
1629         hmp = pmp->iroot->cluster.focus->hmp;   /* iroot retains focus */
1630         bref = pmp->iroot->cluster.focus->bref; /* no lock */
1631
1632         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1633         mp->mnt_vstat.f_files = bref.inode_count;
1634         mp->mnt_vstat.f_ffree = 0;
1635         mp->mnt_vstat.f_blocks = (bref.data_count +
1636                                  hmp->voldata.allocator_free) /
1637                                 mp->mnt_vstat.f_bsize;
1638         mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1639                                 mp->mnt_vstat.f_bsize;
1640         mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1641
1642         *sbp = mp->mnt_vstat;
1643         return (0);
1644 }
1645
1646 /*
1647  * Mount-time recovery (RW mounts)
1648  *
1649  * Updates to the free block table are allowed to lag flushes by one
1650  * transaction.  In case of a crash, then on a fresh mount we must do an
1651  * incremental scan of the last committed transaction id and make sure that
1652  * all related blocks have been marked allocated.
1653  *
1654  * The super-root topology and each PFS has its own transaction id domain,
1655  * so we must track PFS boundary transitions.
1656  */
1657 struct hammer2_recovery_elm {
1658         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1659         hammer2_chain_t *chain;
1660         hammer2_tid_t sync_tid;
1661 };
1662
1663 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1664
1665 struct hammer2_recovery_info {
1666         struct hammer2_recovery_list list;
1667         int     depth;
1668 };
1669
1670 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1671                         hammer2_chain_t *parent,
1672                         struct hammer2_recovery_info *info,
1673                         hammer2_tid_t sync_tid);
1674
1675 #define HAMMER2_RECOVERY_MAXDEPTH       10
1676
1677 static
1678 int
1679 hammer2_recovery(hammer2_dev_t *hmp)
1680 {
1681         struct hammer2_recovery_info info;
1682         struct hammer2_recovery_elm *elm;
1683         hammer2_chain_t *parent;
1684         hammer2_tid_t sync_tid;
1685         hammer2_tid_t mirror_tid;
1686         int error;
1687         int cumulative_error = 0;
1688
1689         hammer2_trans_init(hmp->spmp, 0);
1690
1691         sync_tid = hmp->voldata.freemap_tid;
1692         mirror_tid = hmp->voldata.mirror_tid;
1693
1694         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1695         if (sync_tid >= mirror_tid) {
1696                 kprintf(" no recovery needed\n");
1697         } else {
1698                 kprintf(" freemap recovery %016jx-%016jx\n",
1699                         sync_tid + 1, mirror_tid);
1700         }
1701
1702         TAILQ_INIT(&info.list);
1703         info.depth = 0;
1704         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1705         cumulative_error = hammer2_recovery_scan(hmp, parent,
1706                                                  &info, sync_tid);
1707         hammer2_chain_lookup_done(parent);
1708
1709         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1710                 TAILQ_REMOVE(&info.list, elm, entry);
1711                 parent = elm->chain;
1712                 sync_tid = elm->sync_tid;
1713                 kfree(elm, M_HAMMER2);
1714
1715                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1716                 error = hammer2_recovery_scan(hmp, parent,
1717                                               &info, hmp->voldata.freemap_tid);
1718                 hammer2_chain_unlock(parent);
1719                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
1720                 if (error)
1721                         cumulative_error = error;
1722         }
1723         hammer2_trans_done(hmp->spmp);
1724
1725         return cumulative_error;
1726 }
1727
1728 static
1729 int
1730 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1731                       struct hammer2_recovery_info *info,
1732                       hammer2_tid_t sync_tid)
1733 {
1734         const hammer2_inode_data_t *ripdata;
1735         hammer2_chain_t *chain;
1736         int cache_index;
1737         int cumulative_error = 0;
1738         int error;
1739
1740         /*
1741          * Adjust freemap to ensure that the block(s) are marked allocated.
1742          */
1743         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1744                 hammer2_freemap_adjust(hmp, &parent->bref,
1745                                        HAMMER2_FREEMAP_DORECOVER);
1746         }
1747
1748         /*
1749          * Check type for recursive scan
1750          */
1751         switch(parent->bref.type) {
1752         case HAMMER2_BREF_TYPE_VOLUME:
1753                 /* data already instantiated */
1754                 break;
1755         case HAMMER2_BREF_TYPE_INODE:
1756                 /*
1757                  * Must instantiate data for DIRECTDATA test and also
1758                  * for recursion.
1759                  */
1760                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1761                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
1762                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1763                         /* not applicable to recovery scan */
1764                         hammer2_chain_unlock(parent);
1765                         return 0;
1766                 }
1767                 hammer2_chain_unlock(parent);
1768                 break;
1769         case HAMMER2_BREF_TYPE_INDIRECT:
1770                 /*
1771                  * Must instantiate data for recursion
1772                  */
1773                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1774                 hammer2_chain_unlock(parent);
1775                 break;
1776         case HAMMER2_BREF_TYPE_DATA:
1777         case HAMMER2_BREF_TYPE_FREEMAP:
1778         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1779         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1780                 /* not applicable to recovery scan */
1781                 return 0;
1782                 break;
1783         default:
1784                 return EDOM;
1785         }
1786
1787         /*
1788          * Defer operation if depth limit reached or if we are crossing a
1789          * PFS boundary.
1790          */
1791         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1792                 struct hammer2_recovery_elm *elm;
1793
1794                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1795                 elm->chain = parent;
1796                 elm->sync_tid = sync_tid;
1797                 hammer2_chain_ref(parent);
1798                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
1799                 /* unlocked by caller */
1800
1801                 return(0);
1802         }
1803
1804
1805         /*
1806          * Recursive scan of the last flushed transaction only.  We are
1807          * doing this without pmp assignments so don't leave the chains
1808          * hanging around after we are done with them.
1809          */
1810         cache_index = 0;
1811         chain = hammer2_chain_scan(parent, NULL, &cache_index,
1812                                    HAMMER2_LOOKUP_NODATA);
1813         while (chain) {
1814                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
1815                 if (chain->bref.mirror_tid > sync_tid) {
1816                         ++info->depth;
1817                         error = hammer2_recovery_scan(hmp, chain,
1818                                                       info, sync_tid);
1819                         --info->depth;
1820                         if (error)
1821                                 cumulative_error = error;
1822                 }
1823
1824                 /*
1825                  * Flush the recovery at the PFS boundary to stage it for
1826                  * the final flush of the super-root topology.
1827                  */
1828                 if ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
1829                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
1830                         hammer2_flush(chain, 1);
1831                 }
1832                 chain = hammer2_chain_scan(parent, chain, &cache_index,
1833                                            HAMMER2_LOOKUP_NODATA);
1834         }
1835
1836         return cumulative_error;
1837 }
1838
1839 /*
1840  * Sync a mount point; this is called on a per-mount basis from the
1841  * filesystem syncer process periodically and whenever a user issues
1842  * a sync.
1843  */
1844 int
1845 hammer2_vfs_sync(struct mount *mp, int waitfor)
1846 {
1847         struct hammer2_sync_info info;
1848         hammer2_inode_t *iroot;
1849         hammer2_chain_t *chain;
1850         hammer2_chain_t *parent;
1851         hammer2_pfs_t *pmp;
1852         hammer2_dev_t *hmp;
1853         int flags;
1854         int error;
1855         int total_error;
1856         int i;
1857         int j;
1858
1859         pmp = MPTOPMP(mp);
1860         iroot = pmp->iroot;
1861         KKASSERT(iroot);
1862         KKASSERT(iroot->pmp == pmp);
1863
1864         /*
1865          * We can't acquire locks on existing vnodes while in a transaction
1866          * without risking a deadlock.  This assumes that vfsync() can be
1867          * called without the vnode locked (which it can in DragonFly).
1868          * Otherwise we'd have to implement a multi-pass or flag the lock
1869          * failures and retry.
1870          *
1871          * The reclamation code interlocks with the sync list's token
1872          * (by removing the vnode from the scan list) before unlocking
1873          * the inode, giving us time to ref the inode.
1874          */
1875         /*flags = VMSC_GETVP;*/
1876         flags = 0;
1877         if (waitfor & MNT_LAZY)
1878                 flags |= VMSC_ONEPASS;
1879
1880 #if 0
1881         /*
1882          * Preflush the vnodes using a normal transaction before interlocking
1883          * with a flush transaction.
1884          */
1885         hammer2_trans_init(pmp, 0);
1886         info.error = 0;
1887         info.waitfor = MNT_NOWAIT;
1888         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1889         hammer2_trans_done(pmp);
1890 #endif
1891
1892         /*
1893          * Start our flush transaction.  This does not return until all
1894          * concurrent transactions have completed and will prevent any
1895          * new transactions from running concurrently, except for the
1896          * buffer cache transactions.
1897          *
1898          * For efficiency do an async pass before making sure with a
1899          * synchronous pass on all related buffer cache buffers.  It
1900          * should theoretically not be possible for any new file buffers
1901          * to be instantiated during this sequence.
1902          */
1903         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH |
1904                                 HAMMER2_TRANS_PREFLUSH);
1905         hammer2_run_unlinkq(pmp);
1906
1907         info.error = 0;
1908         info.waitfor = MNT_NOWAIT;
1909         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1910         info.waitfor = MNT_WAIT;
1911         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
1912
1913         /*
1914          * Clear PREFLUSH.  This prevents (or asserts on) any new logical
1915          * buffer cache flushes which occur during the flush.  Device buffers
1916          * are not affected.
1917          */
1918         hammer2_bioq_sync(pmp);
1919         atomic_clear_int(&pmp->trans.flags, HAMMER2_TRANS_PREFLUSH);
1920
1921         total_error = 0;
1922
1923         /*
1924          * Flush all nodes to synchronize the PFSROOT subtopology to the media.
1925          *
1926          * Note that this flush will not be visible on crash recovery until
1927          * we flush the super-root topology in the next loop.
1928          */
1929         for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
1930                 chain = iroot->cluster.array[i].chain;
1931                 if (chain == NULL)
1932                         continue;
1933
1934                 hammer2_chain_ref(chain);
1935                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
1936                 if (chain->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1937                         hammer2_flush(chain, 1);
1938                         parent = chain->parent;
1939                         KKASSERT(chain->pmp != parent->pmp);
1940                         hammer2_chain_setflush(parent);
1941                 }
1942                 hammer2_chain_unlock(chain);
1943                 hammer2_chain_drop(chain);
1944         }
1945         hammer2_trans_done(pmp);
1946
1947         /*
1948          * Flush all volume roots to synchronize PFS flushes with the
1949          * storage media volume header.  This will flush the freemap and
1950          * the superroot topology but stops when it reaches a PFSROOT
1951          * (which we already flushed above).
1952          *
1953          * This is the last step which connects the volume root to the
1954          * PFSROOT dirs flushed above.
1955          *
1956          * Each spmp (representing the hmp's super-root) requires its own
1957          * transaction.
1958          */
1959         for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
1960                 hammer2_chain_t *tmp;
1961
1962                 chain = iroot->cluster.array[i].chain;
1963                 if (chain == NULL)
1964                         continue;
1965
1966                 hmp = chain->hmp;
1967
1968                 /*
1969                  * We only have to flush each hmp once
1970                  */
1971                 for (j = i - 1; j >= 0; --j) {
1972                         if ((tmp = iroot->cluster.array[j].chain) != NULL) {
1973                                 if (tmp->hmp == hmp)
1974                                         break;
1975                         }
1976                 }
1977                 if (j >= 0)
1978                         continue;
1979
1980                 /*
1981                  * spmp transaction.  The super-root is never directly
1982                  * mounted so there shouldn't be any vnodes, let alone any
1983                  * dirty vnodes associated with it.
1984                  */
1985                 hammer2_trans_init(hmp->spmp, HAMMER2_TRANS_ISFLUSH);
1986
1987                 /*
1988                  * Media mounts have two 'roots', vchain for the topology
1989                  * and fchain for the free block table.  Flush both.
1990                  *
1991                  * Note that the topology and free block table are handled
1992                  * independently, so the free block table can wind up being
1993                  * ahead of the topology.  We depend on the bulk free scan
1994                  * code to deal with any loose ends.
1995                  */
1996                 hammer2_chain_ref(&hmp->vchain);
1997                 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1998                 hammer2_chain_ref(&hmp->fchain);
1999                 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2000                 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2001                         /*
2002                          * This will also modify vchain as a side effect,
2003                          * mark vchain as modified now.
2004                          */
2005                         hammer2_voldata_modify(hmp);
2006                         chain = &hmp->fchain;
2007                         hammer2_flush(chain, 1);
2008                         KKASSERT(chain == &hmp->fchain);
2009                 }
2010                 hammer2_chain_unlock(&hmp->fchain);
2011                 hammer2_chain_unlock(&hmp->vchain);
2012                 hammer2_chain_drop(&hmp->fchain);
2013                 /* vchain dropped down below */
2014
2015                 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2016                 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2017                         chain = &hmp->vchain;
2018                         hammer2_flush(chain, 1);
2019                         KKASSERT(chain == &hmp->vchain);
2020                 }
2021                 hammer2_chain_unlock(&hmp->vchain);
2022                 hammer2_chain_drop(&hmp->vchain);
2023
2024                 error = 0;
2025
2026                 /*
2027                  * We can't safely flush the volume header until we have
2028                  * flushed any device buffers which have built up.
2029                  *
2030                  * XXX this isn't being incremental
2031                  */
2032                 vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
2033                 error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
2034                 vn_unlock(hmp->devvp);
2035
2036                 /*
2037                  * The flush code sets CHAIN_VOLUMESYNC to indicate that the
2038                  * volume header needs synchronization via hmp->volsync.
2039                  *
2040                  * XXX synchronize the flag & data with only this flush XXX
2041                  */
2042                 if (error == 0 &&
2043                     (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
2044                         struct buf *bp;
2045
2046                         /*
2047                          * Synchronize the disk before flushing the volume
2048                          * header.
2049                          */
2050                         bp = getpbuf(NULL);
2051                         bp->b_bio1.bio_offset = 0;
2052                         bp->b_bufsize = 0;
2053                         bp->b_bcount = 0;
2054                         bp->b_cmd = BUF_CMD_FLUSH;
2055                         bp->b_bio1.bio_done = biodone_sync;
2056                         bp->b_bio1.bio_flags |= BIO_SYNC;
2057                         vn_strategy(hmp->devvp, &bp->b_bio1);
2058                         biowait(&bp->b_bio1, "h2vol");
2059                         relpbuf(bp, NULL);
2060
2061                         /*
2062                          * Then we can safely flush the version of the
2063                          * volume header synchronized by the flush code.
2064                          */
2065                         i = hmp->volhdrno + 1;
2066                         if (i >= HAMMER2_NUM_VOLHDRS)
2067                                 i = 0;
2068                         if (i * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
2069                             hmp->volsync.volu_size) {
2070                                 i = 0;
2071                         }
2072                         kprintf("sync volhdr %d %jd\n",
2073                                 i, (intmax_t)hmp->volsync.volu_size);
2074                         bp = getblk(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2075                                     HAMMER2_PBUFSIZE, 0, 0);
2076                         atomic_clear_int(&hmp->vchain.flags,
2077                                          HAMMER2_CHAIN_VOLUMESYNC);
2078                         bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
2079                         bawrite(bp);
2080                         hmp->volhdrno = i;
2081                 }
2082                 if (error)
2083                         total_error = error;
2084
2085                 hammer2_trans_done(hmp->spmp);  /* spmp trans */
2086         }
2087         return (total_error);
2088 }
2089
2090 /*
2091  * Sync passes.
2092  */
2093 static int
2094 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2095 {
2096         struct hammer2_sync_info *info = data;
2097         hammer2_inode_t *ip;
2098         int error;
2099
2100         /*
2101          * Degenerate cases.  Note that ip == NULL typically means the
2102          * syncer vnode itself and we don't want to vclrisdirty() in that
2103          * situation.
2104          */
2105         ip = VTOI(vp);
2106         if (ip == NULL) {
2107                 return(0);
2108         }
2109         if (vp->v_type == VNON || vp->v_type == VBAD) {
2110                 vclrisdirty(vp);
2111                 return(0);
2112         }
2113
2114         /*
2115          * VOP_FSYNC will start a new transaction so replicate some code
2116          * here to do it inline (see hammer2_vop_fsync()).
2117          *
2118          * WARNING: The vfsync interacts with the buffer cache and might
2119          *          block, we can't hold the inode lock at that time.
2120          *          However, we MUST ref ip before blocking to ensure that
2121          *          it isn't ripped out from under us (since we do not
2122          *          hold a lock on the vnode).
2123          */
2124         hammer2_inode_ref(ip);
2125         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2126             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2127                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2128                 hammer2_inode_fsync(ip, NULL);
2129         }
2130         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2131             RB_EMPTY(&vp->v_rbdirty_tree)) {
2132                 vclrisdirty(vp);
2133         }
2134
2135         hammer2_inode_drop(ip);
2136 #if 1
2137         error = 0;
2138         if (error)
2139                 info->error = error;
2140 #endif
2141         return(0);
2142 }
2143
2144 static
2145 int
2146 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2147 {
2148         return (0);
2149 }
2150
2151 static
2152 int
2153 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2154                struct fid *fhp, struct vnode **vpp)
2155 {
2156         return (0);
2157 }
2158
2159 static
2160 int
2161 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2162                  int *exflagsp, struct ucred **credanonp)
2163 {
2164         return (0);
2165 }
2166
2167 /*
2168  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2169  * header into the HMP
2170  *
2171  * XXX read four volhdrs and use the one with the highest TID whos CRC
2172  *     matches.
2173  *
2174  * XXX check iCRCs.
2175  *
2176  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2177  *     nonexistant locations.
2178  *
2179  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2180  */
2181 static
2182 int
2183 hammer2_install_volume_header(hammer2_dev_t *hmp)
2184 {
2185         hammer2_volume_data_t *vd;
2186         struct buf *bp;
2187         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2188         int error_reported;
2189         int error;
2190         int valid;
2191         int i;
2192
2193         error_reported = 0;
2194         error = 0;
2195         valid = 0;
2196         bp = NULL;
2197
2198         /*
2199          * There are up to 4 copies of the volume header (syncs iterate
2200          * between them so there is no single master).  We don't trust the
2201          * volu_size field so we don't know precisely how large the filesystem
2202          * is, so depend on the OS to return an error if we go beyond the
2203          * block device's EOF.
2204          */
2205         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2206                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2207                               HAMMER2_VOLUME_BYTES, &bp);
2208                 if (error) {
2209                         brelse(bp);
2210                         bp = NULL;
2211                         continue;
2212                 }
2213
2214                 vd = (struct hammer2_volume_data *) bp->b_data;
2215                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2216                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2217                         brelse(bp);
2218                         bp = NULL;
2219                         continue;
2220                 }
2221
2222                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2223                         /* XXX: Reversed-endianness filesystem */
2224                         kprintf("hammer2: reverse-endian filesystem detected");
2225                         brelse(bp);
2226                         bp = NULL;
2227                         continue;
2228                 }
2229
2230                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2231                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2232                                       HAMMER2_VOLUME_ICRC0_SIZE);
2233                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2234                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2235                                        HAMMER2_VOLUME_ICRC1_SIZE);
2236                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2237                         kprintf("hammer2 volume header crc "
2238                                 "mismatch copy #%d %08x/%08x\n",
2239                                 i, crc0, crc);
2240                         error_reported = 1;
2241                         brelse(bp);
2242                         bp = NULL;
2243                         continue;
2244                 }
2245                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2246                         valid = 1;
2247                         hmp->voldata = *vd;
2248                         hmp->volhdrno = i;
2249                 }
2250                 brelse(bp);
2251                 bp = NULL;
2252         }
2253         if (valid) {
2254                 hmp->volsync = hmp->voldata;
2255                 error = 0;
2256                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2257                         kprintf("hammer2: using volume header #%d\n",
2258                                 hmp->volhdrno);
2259                 }
2260         } else {
2261                 error = EINVAL;
2262                 kprintf("hammer2: no valid volume headers found!\n");
2263         }
2264         return (error);
2265 }
2266
2267 /*
2268  * This handles hysteresis on regular file flushes.  Because the BIOs are
2269  * routed to a thread it is possible for an excessive number to build up
2270  * and cause long front-end stalls long before the runningbuffspace limit
2271  * is hit, so we implement hammer2_flush_pipe to control the
2272  * hysteresis.
2273  *
2274  * This is a particular problem when compression is used.
2275  */
2276 void
2277 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2278 {
2279         atomic_add_int(&pmp->count_lwinprog, 1);
2280 }
2281
2282 void
2283 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2284 {
2285         int lwinprog;
2286
2287         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2288         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2289             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2290                 atomic_clear_int(&pmp->count_lwinprog,
2291                                  HAMMER2_LWINPROG_WAITING);
2292                 wakeup(&pmp->count_lwinprog);
2293         }
2294 }
2295
2296 void
2297 hammer2_lwinprog_wait(hammer2_pfs_t *pmp)
2298 {
2299         int lwinprog;
2300
2301         for (;;) {
2302                 lwinprog = pmp->count_lwinprog;
2303                 cpu_ccfence();
2304                 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2305                         break;
2306                 tsleep_interlock(&pmp->count_lwinprog, 0);
2307                 atomic_set_int(&pmp->count_lwinprog, HAMMER2_LWINPROG_WAITING);
2308                 lwinprog = pmp->count_lwinprog;
2309                 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2310                         break;
2311                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2312         }
2313 }
2314
2315 /*
2316  * Manage excessive memory resource use for chain and related
2317  * structures.
2318  */
2319 void
2320 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2321 {
2322         uint32_t waiting;
2323         uint32_t count;
2324         uint32_t limit;
2325 #if 0
2326         static int zzticks;
2327 #endif
2328
2329         /*
2330          * Atomic check condition and wait.  Also do an early speedup of
2331          * the syncer to try to avoid hitting the wait.
2332          */
2333         for (;;) {
2334                 waiting = pmp->inmem_dirty_chains;
2335                 cpu_ccfence();
2336                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2337
2338                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2339                 if (limit < hammer2_limit_dirty_chains)
2340                         limit = hammer2_limit_dirty_chains;
2341                 if (limit < 1000)
2342                         limit = 1000;
2343
2344 #if 0
2345                 if ((int)(ticks - zzticks) > hz) {
2346                         zzticks = ticks;
2347                         kprintf("count %ld %ld\n", count, limit);
2348                 }
2349 #endif
2350
2351                 /*
2352                  * Block if there are too many dirty chains present, wait
2353                  * for the flush to clean some out.
2354                  */
2355                 if (count > limit) {
2356                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2357                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2358                                                waiting,
2359                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2360                                 speedup_syncer(pmp->mp);
2361                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2362                                        "chnmem", hz);
2363                         }
2364                         continue;       /* loop on success or fail */
2365                 }
2366
2367                 /*
2368                  * Try to start an early flush before we are forced to block.
2369                  */
2370                 if (count > limit * 7 / 10)
2371                         speedup_syncer(pmp->mp);
2372                 break;
2373         }
2374 }
2375
2376 void
2377 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2378 {
2379         if (pmp) {
2380                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2381         }
2382 }
2383
2384 void
2385 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2386 {
2387         uint32_t waiting;
2388
2389         if (pmp == NULL)
2390                 return;
2391
2392         for (;;) {
2393                 waiting = pmp->inmem_dirty_chains;
2394                 cpu_ccfence();
2395                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2396                                        waiting,
2397                                        (waiting - 1) &
2398                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2399                         break;
2400                 }
2401         }
2402
2403         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2404                 wakeup(&pmp->inmem_dirty_chains);
2405 }
2406
2407 /*
2408  * Debugging
2409  */
2410 void
2411 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2412 {
2413         hammer2_chain_t *scan;
2414         hammer2_chain_t *parent;
2415
2416         --*countp;
2417         if (*countp == 0) {
2418                 kprintf("%*.*s...\n", tab, tab, "");
2419                 return;
2420         }
2421         if (*countp < 0)
2422                 return;
2423         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2424                 tab, tab, "", pfx,
2425                 chain, chain->bref.type,
2426                 chain->bref.key, chain->bref.keybits,
2427                 chain->bref.mirror_tid);
2428
2429         kprintf("%*.*s      [%08x] (%s) refs=%d",
2430                 tab, tab, "",
2431                 chain->flags,
2432                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2433                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2434                 chain->refs);
2435
2436         parent = chain->parent;
2437         if (parent)
2438                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2439                         tab, tab, "",
2440                         parent, parent->flags, parent->refs);
2441         if (RB_EMPTY(&chain->core.rbtree)) {
2442                 kprintf("\n");
2443         } else {
2444                 kprintf(" {\n");
2445                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2446                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2447                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2448                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2449                                 chain->data->ipdata.filename);
2450                 else
2451                         kprintf("%*.*s}\n", tab, tab, "");
2452         }
2453 }