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         hammer2_trans_t trans;
  71         int error;
  72         int waitfor;
  73 };
  74
  75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  76 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
  77 static struct hammer2_mntlist hammer2_mntlist;
  78 static struct hammer2_pfslist hammer2_pfslist;
  79 static struct lock hammer2_mntlk;
  80
  81 int hammer2_debug;
  82 int hammer2_cluster_enable = 1;
  83 int hammer2_hardlink_enable = 1;
  84 int hammer2_flush_pipe = 100;
  85 int hammer2_synchronous_flush = 1;
  86 int hammer2_dio_count;
  87 long hammer2_limit_dirty_chains;
  88 long hammer2_iod_file_read;
  89 long hammer2_iod_meta_read;
  90 long hammer2_iod_indr_read;
  91 long hammer2_iod_fmap_read;
  92 long hammer2_iod_volu_read;
  93 long hammer2_iod_file_write;
  94 long hammer2_iod_meta_write;
  95 long hammer2_iod_indr_write;
  96 long hammer2_iod_fmap_write;
  97 long hammer2_iod_volu_write;
  98 long hammer2_ioa_file_read;
  99 long hammer2_ioa_meta_read;
 100 long hammer2_ioa_indr_read;
 101 long hammer2_ioa_fmap_read;
 102 long hammer2_ioa_volu_read;
 103 long hammer2_ioa_fmap_write;
 104 long hammer2_ioa_file_write;
 105 long hammer2_ioa_meta_write;
 106 long hammer2_ioa_indr_write;
 107 long hammer2_ioa_volu_write;
 108
 109 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 110 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 111                 "Buffer used for compression.");
 112
 113 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 114 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 115                 "Buffer used for decompression.");
 116
 117 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 118
 119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 120            &hammer2_debug, 0, "");
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW,
 122            &hammer2_cluster_enable, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
 124            &hammer2_hardlink_enable, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 126            &hammer2_flush_pipe, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 128            &hammer2_synchronous_flush, 0, "");
 129 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 130            &hammer2_limit_dirty_chains, 0, "");
 131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 132            &hammer2_dio_count, 0, "");
 133
 134 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 135            &hammer2_iod_file_read, 0, "");
 136 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 137            &hammer2_iod_meta_read, 0, "");
 138 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 139            &hammer2_iod_indr_read, 0, "");
 140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 141            &hammer2_iod_fmap_read, 0, "");
 142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 143            &hammer2_iod_volu_read, 0, "");
 144
 145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 146            &hammer2_iod_file_write, 0, "");
 147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 148            &hammer2_iod_meta_write, 0, "");
 149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 150            &hammer2_iod_indr_write, 0, "");
 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 152            &hammer2_iod_fmap_write, 0, "");
 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 154            &hammer2_iod_volu_write, 0, "");
 155
 156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
 157            &hammer2_ioa_file_read, 0, "");
 158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
 159            &hammer2_ioa_meta_read, 0, "");
 160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
 161            &hammer2_ioa_indr_read, 0, "");
 162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
 163            &hammer2_ioa_fmap_read, 0, "");
 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
 165            &hammer2_ioa_volu_read, 0, "");
 166
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
 168            &hammer2_ioa_file_write, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
 170            &hammer2_ioa_meta_write, 0, "");
 171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
 172            &hammer2_ioa_indr_write, 0, "");
 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
 174            &hammer2_ioa_fmap_write, 0, "");
 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
 176            &hammer2_ioa_volu_write, 0, "");
 177
 178 static int hammer2_vfs_init(struct vfsconf *conf);
 179 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 180 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 181                                 struct ucred *cred);
 182 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 183                                 struct vnode *, struct ucred *);
 184 static int hammer2_recovery(hammer2_dev_t *hmp);
 185 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 186 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 187 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 188                                 struct ucred *cred);
 189 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 190                                 struct ucred *cred);
 191 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
 192                                 ino_t ino, struct vnode **vpp);
 193 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 194                                 struct fid *fhp, struct vnode **vpp);
 195 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 196 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 197                                 int *exflagsp, struct ucred **credanonp);
 198
 199 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 200 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 201
 202 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 203
 204 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 205 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 206                                 hammer2_dev_t *hmp);
 207
 208 /*
 209  * HAMMER2 vfs operations.
 210  */
 211 static struct vfsops hammer2_vfsops = {
 212         .vfs_init       = hammer2_vfs_init,
 213         .vfs_uninit     = hammer2_vfs_uninit,
 214         .vfs_sync       = hammer2_vfs_sync,
 215         .vfs_mount      = hammer2_vfs_mount,
 216         .vfs_unmount    = hammer2_vfs_unmount,
 217         .vfs_root       = hammer2_vfs_root,
 218         .vfs_statfs     = hammer2_vfs_statfs,
 219         .vfs_statvfs    = hammer2_vfs_statvfs,
 220         .vfs_vget       = hammer2_vfs_vget,
 221         .vfs_vptofh     = hammer2_vfs_vptofh,
 222         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 223         .vfs_checkexp   = hammer2_vfs_checkexp
 224 };
 225
 226 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 227
 228 VFS_SET(hammer2_vfsops, hammer2, 0);
 229 MODULE_VERSION(hammer2, 1);
 230
 231 static
 232 int
 233 hammer2_vfs_init(struct vfsconf *conf)
 234 {
 235         static struct objcache_malloc_args margs_read;
 236         static struct objcache_malloc_args margs_write;
 237         static struct objcache_malloc_args margs_vop;
 238
 239         int error;
 240
 241         error = 0;
 242
 243         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 244                 error = EINVAL;
 245         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 246                 error = EINVAL;
 247         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 248                 error = EINVAL;
 249
 250         if (error)
 251                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 252
 253         margs_read.objsize = 65536;
 254         margs_read.mtype = M_HAMMER2_DEBUFFER;
 255
 256         margs_write.objsize = 32768;
 257         margs_write.mtype = M_HAMMER2_CBUFFER;
 258
 259         margs_vop.objsize = sizeof(hammer2_xop_t);
 260         margs_vop.mtype = M_HAMMER2;
 261
 262         /*
 263          * Note thaht for the XOPS cache we want backing store allocations
 264          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 265          * confusion), so use the backing store function that does it.  This
 266          * means that initial XOPS objects are zerod but REUSED objects are
 267          * not.  So we are responsible for cleaning the object up sufficiently
 268          * for our needs before objcache_put()ing it back (typically just the
 269          * FIFO indices).
 270          */
 271         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 272                                 0, 1, NULL, NULL, NULL,
 273                                 objcache_malloc_alloc,
 274                                 objcache_malloc_free,
 275                                 &margs_read);
 276         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 277                                 0, 1, NULL, NULL, NULL,
 278                                 objcache_malloc_alloc,
 279                                 objcache_malloc_free,
 280                                 &margs_write);
 281         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 282                                 0, 1, NULL, NULL, NULL,
 283                                 objcache_malloc_alloc_zero,
 284                                 objcache_malloc_free,
 285                                 &margs_vop);
 286
 287
 288         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 289         TAILQ_INIT(&hammer2_mntlist);
 290         TAILQ_INIT(&hammer2_pfslist);
 291
 292         hammer2_limit_dirty_chains = desiredvnodes / 10;
 293
 294         return (error);
 295 }
 296
 297 static
 298 int
 299 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 300 {
 301         objcache_destroy(cache_buffer_read);
 302         objcache_destroy(cache_buffer_write);
 303         objcache_destroy(cache_xops);
 304         return 0;
 305 }
 306
 307 /*
 308  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
 309  * mounts and the spmp structure for media (hmp) structures.
 310  *
 311  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 312  * transactions.  Note that synchronization does not use this field.
 313  * (typically frontend operations and synchronization cannot run on the
 314  * same PFS node at the same time).
 315  *
 316  * XXX check locking
 317  */
 318 hammer2_pfs_t *
 319 hammer2_pfsalloc(hammer2_cluster_t *cluster,
 320                  const hammer2_inode_data_t *ripdata,
 321                  hammer2_tid_t modify_tid)
 322 {
 323         hammer2_chain_t *rchain;
 324         hammer2_inode_t *iroot;
 325         hammer2_pfs_t *pmp;
 326         int count;
 327         int i;
 328         int j;
 329
 330         /*
 331          * Locate or create the PFS based on the cluster id.  If ripdata
 332          * is NULL this is a spmp which is unique and is always allocated.
 333          */
 334         if (ripdata) {
 335                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 336                         if (bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 337                                  sizeof(pmp->pfs_clid)) == 0) {
 338                                         break;
 339                         }
 340                 }
 341         } else {
 342                 pmp = NULL;
 343         }
 344
 345         if (pmp == NULL) {
 346                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 347                 hammer2_trans_manage_init(&pmp->tmanage);
 348                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 349                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 350                 lockinit(&pmp->lock, "pfslk", 0, 0);
 351                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 352                 RB_INIT(&pmp->inum_tree);
 353                 TAILQ_INIT(&pmp->unlinkq);
 354                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 355
 356                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 357                         hammer2_xop_group_init(pmp, &pmp->xop_groups[j]);
 358
 359                 /*
 360                  * Save the last media transaction id for the flusher.  Set
 361                  * initial
 362                  */
 363                 if (ripdata)
 364                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 365                 hammer2_mtx_init(&pmp->wthread_mtx, "h2wthr");
 366                 bioq_init(&pmp->wthread_bioq);
 367                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 368
 369                 /*
 370                  * The synchronization thread may start too early, make
 371                  * sure it stays frozen until we are ready to let it go.
 372                  * XXX
 373                  */
 374                 /*
 375                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 376                                          HAMMER2_THREAD_REMASTER;
 377                 */
 378         }
 379
 380         /*
 381          * Create the PFS's root inode.
 382          */
 383         if ((iroot = pmp->iroot) == NULL) {
 384                 iroot = hammer2_inode_get(pmp, NULL, NULL);
 385                 pmp->iroot = iroot;
 386                 hammer2_inode_ref(iroot);
 387                 hammer2_inode_unlock(iroot, NULL);
 388         }
 389
 390         /*
 391          * Stop here if no cluster is passed in.
 392          */
 393         if (cluster == NULL)
 394                 goto done;
 395
 396         /*
 397          * When a cluster is passed in we must add the cluster's chains
 398          * to the PFS's root inode, update pmp->pfs_types[], and update
 399          * the syncronization threads.
 400          *
 401          * At the moment empty spots can develop due to removals or failures.
 402          * Ultimately we want to re-fill these spots but doing so might
 403          * confused running code. XXX
 404          */
 405         hammer2_inode_ref(iroot);
 406         hammer2_mtx_ex(&iroot->lock);
 407         j = iroot->cluster.nchains;
 408
 409         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 410
 411         for (i = 0; i < cluster->nchains; ++i) {
 412                 if (j == HAMMER2_MAXCLUSTER)
 413                         break;
 414                 rchain = cluster->array[i].chain;
 415                 KKASSERT(rchain->pmp == NULL);
 416                 rchain->pmp = pmp;
 417                 hammer2_chain_ref(rchain);
 418                 iroot->cluster.array[j].chain = rchain;
 419                 pmp->pfs_types[j] = ripdata->meta.pfs_type;
 420                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 421
 422                 /*
 423                  * If the PFS is already mounted we must account
 424                  * for the mount_count here.
 425                  */
 426                 if (pmp->mp)
 427                         ++rchain->hmp->mount_count;
 428
 429                 /*
 430                  * May have to fixup dirty chain tracking.  Previous
 431                  * pmp was NULL so nothing to undo.
 432                  */
 433                 if (rchain->flags & HAMMER2_CHAIN_MODIFIED)
 434                         hammer2_pfs_memory_inc(pmp);
 435                 ++j;
 436         }
 437         iroot->cluster.nchains = j;
 438
 439         if (i != cluster->nchains) {
 440                 kprintf("hammer2_mount: cluster full!\n");
 441                 /* XXX fatal error? */
 442         }
 443
 444         /*
 445          * Update nmasters from any PFS inode which is part of the cluster.
 446          * It is possible that this will result in a value which is too
 447          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 448          * override this value later on.
 449          *
 450          * (This informs us of masters that might not currently be
 451          *  discoverable by this mount).
 452          */
 453         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 454                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 455         }
 456
 457         /*
 458          * Count visible masters.  Masters are usually added with
 459          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 460          * are more (XXX and must update the master inodes).
 461          */
 462         count = 0;
 463         for (i = 0; i < iroot->cluster.nchains; ++i) {
 464                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 465                         ++count;
 466         }
 467         if (pmp->pfs_nmasters < count)
 468                 pmp->pfs_nmasters = count;
 469
 470         /*
 471          * Create missing synchronization and support threads.
 472          *
 473          * Single-node masters (including snapshots) have nothing to
 474          * synchronize and do not require this thread.
 475          *
 476          * Multi-node masters or any number of soft masters, slaves, copy,
 477          * or other PFS types need the thread.
 478          *
 479          * Each thread is responsible for its particular cluster index.
 480          * We use independent threads so stalls or mismatches related to
 481          * any given target do not affect other targets.
 482          */
 483         for (i = 0; i < iroot->cluster.nchains; ++i) {
 484                 /*
 485                  * Single-node masters (including snapshots) have nothing
 486                  * to synchronize and will make direct xops support calls,
 487                  * thus they do not require this thread.
 488                  *
 489                  * Note that there can be thousands of snapshots.  We do not
 490                  * want to create thousands of threads.
 491                  */
 492                 if (pmp->pfs_nmasters <= 1 &&
 493                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 494                         continue;
 495                 }
 496
 497                 /*
 498                  * Sync support thread
 499                  */
 500                 if (pmp->sync_thrs[i].td == NULL) {
 501                         hammer2_thr_create(&pmp->sync_thrs[i], pmp,
 502                                            "h2nod", i, -1,
 503                                            hammer2_primary_sync_thread);
 504                 }
 505         }
 506
 507         /*
 508          * Create missing Xop threads
 509          */
 510         if (pmp->mp)
 511                 hammer2_xop_helper_create(pmp);
 512
 513         hammer2_mtx_unlock(&iroot->lock);
 514         hammer2_inode_drop(iroot);
 515 done:
 516         return pmp;
 517 }
 518
 519 /*
 520  * Destroy a PFS, typically only occurs after the last mount on a device
 521  * has gone away.
 522  */
 523 static void
 524 hammer2_pfsfree(hammer2_pfs_t *pmp)
 525 {
 526         hammer2_inode_t *iroot;
 527         int i;
 528         int j;
 529
 530         /*
 531          * Cleanup our reference on iroot.  iroot is (should) not be needed
 532          * by the flush code.
 533          */
 534         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 535
 536         iroot = pmp->iroot;
 537         if (iroot) {
 538                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 539                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 540                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 541                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 542                 }
 543 #if REPORT_REFS_ERRORS
 544                 if (pmp->iroot->refs != 1)
 545                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 546                                 pmp->iroot, pmp->iroot->refs);
 547 #else
 548                 KKASSERT(pmp->iroot->refs == 1);
 549 #endif
 550                 /* ref for pmp->iroot */
 551                 hammer2_inode_drop(pmp->iroot);
 552                 pmp->iroot = NULL;
 553         }
 554
 555         kmalloc_destroy(&pmp->mmsg);
 556         kmalloc_destroy(&pmp->minode);
 557
 558         kfree(pmp, M_HAMMER2);
 559 }
 560
 561 /*
 562  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 563  * empty is terminated and freed.
 564  *
 565  * XXX inefficient.
 566  */
 567 static void
 568 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 569 {
 570         hammer2_pfs_t *pmp;
 571         hammer2_inode_t *iroot;
 572         hammer2_cluster_t *cluster;
 573         hammer2_chain_t *rchain;
 574         int didfreeze;
 575         int i;
 576         int j;
 577
 578 again:
 579         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 580                 if ((iroot = pmp->iroot) == NULL)
 581                         continue;
 582                 if (hmp->spmp == pmp) {
 583                         kprintf("unmount hmp %p remove spmp %p\n",
 584                                 hmp, pmp);
 585                         hmp->spmp = NULL;
 586                 }
 587
 588                 /*
 589                  * Determine if this PFS is affected.  If it is we must
 590                  * freeze all management threads and lock its iroot.
 591                  *
 592                  * Freezing a management thread forces it idle, operations
 593                  * in-progress will be aborted and it will have to start
 594                  * over again when unfrozen, or exit if told to exit.
 595                  */
 596                 cluster = &iroot->cluster;
 597                 for (i = 0; i < cluster->nchains; ++i) {
 598                         rchain = cluster->array[i].chain;
 599                         if (rchain == NULL || rchain->hmp != hmp)
 600                                 continue;
 601                         break;
 602                 }
 603                 if (i != cluster->nchains) {
 604                         /*
 605                          * Make sure all synchronization threads are locked
 606                          * down.
 607                          */
 608                         for (i = 0; i < iroot->cluster.nchains; ++i) {
 609                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 610                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 611                                         hammer2_thr_freeze_async(
 612                                                 &pmp->xop_groups[j].thrs[i]);
 613                                 }
 614                         }
 615                         for (i = 0; i < iroot->cluster.nchains; ++i) {
 616                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 617                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 618                                         hammer2_thr_freeze(
 619                                                 &pmp->xop_groups[j].thrs[i]);
 620                                 }
 621                         }
 622
 623                         /*
 624                          * Lock the inode and clean out matching chains.
 625                          * Note that we cannot use hammer2_inode_lock_*()
 626                          * here because that would attempt to validate the
 627                          * cluster that we are in the middle of ripping
 628                          * apart.
 629                          *
 630                          * WARNING! We are working directly on the inodes
 631                          *          embedded cluster.
 632                          */
 633                         hammer2_mtx_ex(&iroot->lock);
 634
 635                         /*
 636                          * Remove the chain from matching elements of the PFS.
 637                          */
 638                         for (i = 0; i < cluster->nchains; ++i) {
 639                                 rchain = cluster->array[i].chain;
 640                                 if (rchain == NULL || rchain->hmp != hmp)
 641                                         continue;
 642                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 643                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 644                                         hammer2_thr_delete(
 645                                                 &pmp->xop_groups[j].thrs[i]);
 646                                 }
 647                                 rchain = cluster->array[i].chain;
 648                                 cluster->array[i].chain = NULL;
 649                                 pmp->pfs_types[i] = 0;
 650                                 if (pmp->pfs_names[i]) {
 651                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 652                                         pmp->pfs_names[i] = NULL;
 653                                 }
 654                                 hammer2_chain_drop(rchain);
 655
 656                                 /* focus hint */
 657                                 if (cluster->focus == rchain)
 658                                         cluster->focus = NULL;
 659                         }
 660                         hammer2_mtx_unlock(&iroot->lock);
 661                         didfreeze = 1;  /* remaster, unfreeze down below */
 662                 } else {
 663                         didfreeze = 0;
 664                 }
 665
 666                 /*
 667                  * Cleanup trailing chains.  Do not reorder chains (for now).
 668                  * XXX might remove more than we intended.
 669                  */
 670                 while (i > 0) {
 671                         if (cluster->array[i - 1].chain)
 672                                 break;
 673                         --i;
 674                 }
 675                 cluster->nchains = i;
 676
 677                 /*
 678                  * If the PMP has no elements remaining we can destroy it.
 679                  * (this will transition management threads from frozen->exit).
 680                  */
 681                 if (cluster->nchains == 0) {
 682                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 683                                 hmp, pmp);
 684                         hammer2_pfsfree(pmp);
 685                         goto again;
 686                 }
 687
 688                 /*
 689                  * If elements still remain we need to set the REMASTER
 690                  * flag and unfreeze it.
 691                  */
 692                 if (didfreeze) {
 693                         for (i = 0; i < iroot->cluster.nchains; ++i) {
 694                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 695                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 696                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 697                                         hammer2_thr_remaster(
 698                                                 &pmp->xop_groups[j].thrs[i]);
 699                                         hammer2_thr_unfreeze(
 700                                                 &pmp->xop_groups[j].thrs[i]);
 701                                 }
 702                         }
 703                 }
 704         }
 705 }
 706
 707 /*
 708  * Mount or remount HAMMER2 fileystem from physical media
 709  *
 710  *      mountroot
 711  *              mp              mount point structure
 712  *              path            NULL
 713  *              data            <unused>
 714  *              cred            <unused>
 715  *
 716  *      mount
 717  *              mp              mount point structure
 718  *              path            path to mount point
 719  *              data            pointer to argument structure in user space
 720  *                      volume  volume path (device@LABEL form)
 721  *                      hflags  user mount flags
 722  *              cred            user credentials
 723  *
 724  * RETURNS:     0       Success
 725  *              !0      error number
 726  */
 727 static
 728 int
 729 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 730                   struct ucred *cred)
 731 {
 732         struct hammer2_mount_info info;
 733         hammer2_pfs_t *pmp;
 734         hammer2_pfs_t *spmp;
 735         hammer2_dev_t *hmp;
 736         hammer2_key_t key_next;
 737         hammer2_key_t key_dummy;
 738         hammer2_key_t lhc;
 739         struct vnode *devvp;
 740         struct nlookupdata nd;
 741         hammer2_chain_t *parent;
 742         hammer2_cluster_t *cluster;
 743         hammer2_cluster_t *cparent;
 744         const hammer2_inode_data_t *ripdata;
 745         hammer2_blockref_t bref;
 746         struct file *fp;
 747         char devstr[MNAMELEN];
 748         size_t size;
 749         size_t done;
 750         char *dev;
 751         char *label;
 752         int ronly = 1;
 753         int error;
 754         int cache_index;
 755         int i;
 756
 757         hmp = NULL;
 758         pmp = NULL;
 759         dev = NULL;
 760         label = NULL;
 761         devvp = NULL;
 762         cache_index = -1;
 763
 764         kprintf("hammer2_mount\n");
 765
 766         if (path == NULL) {
 767                 /*
 768                  * Root mount
 769                  */
 770                 bzero(&info, sizeof(info));
 771                 info.cluster_fd = -1;
 772                 return (EOPNOTSUPP);
 773         } else {
 774                 /*
 775                  * Non-root mount or updating a mount
 776                  */
 777                 error = copyin(data, &info, sizeof(info));
 778                 if (error)
 779                         return (error);
 780
 781                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 782                 if (error)
 783                         return (error);
 784
 785                 /* Extract device and label */
 786                 dev = devstr;
 787                 label = strchr(devstr, '@');
 788                 if (label == NULL ||
 789                     ((label + 1) - dev) > done) {
 790                         return (EINVAL);
 791                 }
 792                 *label = '\0';
 793                 label++;
 794                 if (*label == '\0')
 795                         return (EINVAL);
 796
 797                 if (mp->mnt_flag & MNT_UPDATE) {
 798                         /*
 799                          * Update mount.  Note that pmp->iroot->cluster is
 800                          * an inode-embedded cluster and thus cannot be
 801                          * directly locked.
 802                          *
 803                          * XXX HAMMER2 needs to implement NFS export via
 804                          *     mountctl.
 805                          */
 806                         pmp = MPTOPMP(mp);
 807                         cluster = &pmp->iroot->cluster;
 808                         for (i = 0; i < cluster->nchains; ++i) {
 809                                 if (cluster->array[i].chain == NULL)
 810                                         continue;
 811                                 hmp = cluster->array[i].chain->hmp;
 812                                 devvp = hmp->devvp;
 813                                 error = hammer2_remount(hmp, mp, path,
 814                                                         devvp, cred);
 815                                 if (error)
 816                                         break;
 817                         }
 818                         /*hammer2_inode_install_hidden(pmp);*/
 819
 820                         return error;
 821                 }
 822         }
 823
 824         /*
 825          * HMP device mount
 826          *
 827          * Lookup name and verify it refers to a block device.
 828          */
 829         error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 830         if (error == 0)
 831                 error = nlookup(&nd);
 832         if (error == 0)
 833                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 834         nlookup_done(&nd);
 835
 836         if (error == 0) {
 837                 if (vn_isdisk(devvp, &error))
 838                         error = vfs_mountedon(devvp);
 839         }
 840
 841         /*
 842          * Determine if the device has already been mounted.  After this
 843          * check hmp will be non-NULL if we are doing the second or more
 844          * hammer2 mounts from the same device.
 845          */
 846         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 847         TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 848                 if (hmp->devvp == devvp)
 849                         break;
 850         }
 851
 852         /*
 853          * Open the device if this isn't a secondary mount and construct
 854          * the H2 device mount (hmp).
 855          */
 856         if (hmp == NULL) {
 857                 hammer2_chain_t *schain;
 858                 hammer2_xid_t xid;
 859
 860                 if (error == 0 && vcount(devvp) > 0)
 861                         error = EBUSY;
 862
 863                 /*
 864                  * Now open the device
 865                  */
 866                 if (error == 0) {
 867                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
 868                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
 869                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
 870                         if (error == 0) {
 871                                 error = VOP_OPEN(devvp,
 872                                                  ronly ? FREAD : FREAD | FWRITE,
 873                                                  FSCRED, NULL);
 874                         }
 875                         vn_unlock(devvp);
 876                 }
 877                 if (error && devvp) {
 878                         vrele(devvp);
 879                         devvp = NULL;
 880                 }
 881                 if (error) {
 882                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 883                         return error;
 884                 }
 885                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
 886                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
 887                 hmp->ronly = ronly;
 888                 hmp->devvp = devvp;
 889                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
 890                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
 891                 RB_INIT(&hmp->iotree);
 892                 spin_init(&hmp->io_spin, "hm2mount_io");
 893                 spin_init(&hmp->list_spin, "hm2mount_list");
 894                 TAILQ_INIT(&hmp->flushq);
 895
 896                 lockinit(&hmp->vollk, "h2vol", 0, 0);
 897
 898                 /*
 899                  * vchain setup. vchain.data is embedded.
 900                  * vchain.refs is initialized and will never drop to 0.
 901                  *
 902                  * NOTE! voldata is not yet loaded.
 903                  */
 904                 hmp->vchain.hmp = hmp;
 905                 hmp->vchain.refs = 1;
 906                 hmp->vchain.data = (void *)&hmp->voldata;
 907                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
 908                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 909                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 910
 911                 hammer2_chain_core_init(&hmp->vchain);
 912                 /* hmp->vchain.u.xxx is left NULL */
 913
 914                 /*
 915                  * fchain setup.  fchain.data is embedded.
 916                  * fchain.refs is initialized and will never drop to 0.
 917                  *
 918                  * The data is not used but needs to be initialized to
 919                  * pass assertion muster.  We use this chain primarily
 920                  * as a placeholder for the freemap's top-level RBTREE
 921                  * so it does not interfere with the volume's topology
 922                  * RBTREE.
 923                  */
 924                 hmp->fchain.hmp = hmp;
 925                 hmp->fchain.refs = 1;
 926                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
 927                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
 928                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 929                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 930                 hmp->fchain.bref.methods =
 931                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
 932                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
 933
 934                 hammer2_chain_core_init(&hmp->fchain);
 935                 /* hmp->fchain.u.xxx is left NULL */
 936
 937                 /*
 938                  * Install the volume header and initialize fields from
 939                  * voldata.
 940                  */
 941                 error = hammer2_install_volume_header(hmp);
 942                 if (error) {
 943                         hammer2_unmount_helper(mp, NULL, hmp);
 944                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 945                         hammer2_vfs_unmount(mp, MNT_FORCE);
 946                         return error;
 947                 }
 948
 949                 /*
 950                  * Really important to get these right or flush will get
 951                  * confused.
 952                  */
 953                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0);
 954                 kprintf("alloc spmp %p tid %016jx\n",
 955                         hmp->spmp, hmp->voldata.mirror_tid);
 956                 spmp = hmp->spmp;
 957
 958                 /*
 959                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
 960                  * is inherited from the volume header.
 961                  */
 962                 xid = 0;
 963                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 964                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
 965                 hmp->vchain.pmp = spmp;
 966                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 967                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
 968                 hmp->fchain.pmp = spmp;
 969
 970                 /*
 971                  * First locate the super-root inode, which is key 0
 972                  * relative to the volume header's blockset.
 973                  *
 974                  * Then locate the root inode by scanning the directory keyspace
 975                  * represented by the label.
 976                  */
 977                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
 978                 schain = hammer2_chain_lookup(&parent, &key_dummy,
 979                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
 980                                       &cache_index, 0);
 981                 hammer2_chain_lookup_done(parent);
 982                 if (schain == NULL) {
 983                         kprintf("hammer2_mount: invalid super-root\n");
 984                         hammer2_unmount_helper(mp, NULL, hmp);
 985                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 986                         hammer2_vfs_unmount(mp, MNT_FORCE);
 987                         return EINVAL;
 988                 }
 989                 if (schain->error) {
 990                         kprintf("hammer2_mount: error %s reading super-root\n",
 991                                 hammer2_error_str(schain->error));
 992                         hammer2_chain_unlock(schain);
 993                         hammer2_chain_drop(schain);
 994                         schain = NULL;
 995                         hammer2_unmount_helper(mp, NULL, hmp);
 996                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 997                         hammer2_vfs_unmount(mp, MNT_FORCE);
 998                         return EINVAL;
 999                 }
1000
1001                 /*
1002                  * The super-root always uses an inode_tid of 1 when
1003                  * creating PFSs.
1004                  */
1005                 spmp->inode_tid = 1;
1006                 spmp->modify_tid = schain->bref.modify_tid;
1007
1008                 /*
1009                  * Sanity-check schain's pmp and finish initialization.
1010                  * Any chain belonging to the super-root topology should
1011                  * have a NULL pmp (not even set to spmp).
1012                  */
1013                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1014                 KKASSERT(schain->pmp == NULL);
1015                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1016
1017                 /*
1018                  * Replace the dummy spmp->iroot with a real one.  It's
1019                  * easier to just do a wholesale replacement than to try
1020                  * to update the chain and fixup the iroot fields.
1021                  *
1022                  * The returned inode is locked with the supplied cluster.
1023                  */
1024                 cluster = hammer2_cluster_from_chain(schain);
1025                 hammer2_inode_drop(spmp->iroot);
1026                 spmp->iroot = NULL;
1027                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster);
1028                 spmp->spmp_hmp = hmp;
1029                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1030                 hammer2_inode_ref(spmp->iroot);
1031                 hammer2_inode_unlock(spmp->iroot, cluster);
1032                 schain = NULL;
1033                 /* leave spmp->iroot with one ref */
1034
1035                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1036                         error = hammer2_recovery(hmp);
1037                         /* XXX do something with error */
1038                 }
1039                 hammer2_update_pmps(hmp);
1040                 hammer2_iocom_init(hmp);
1041
1042                 /*
1043                  * Ref the cluster management messaging descriptor.  The mount
1044                  * program deals with the other end of the communications pipe.
1045                  */
1046                 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1047                 if (fp) {
1048                         hammer2_cluster_reconnect(hmp, fp);
1049                 } else {
1050                         kprintf("hammer2_mount: bad cluster_fd!\n");
1051                 }
1052         } else {
1053                 spmp = hmp->spmp;
1054         }
1055
1056         /*
1057          * Lookup the mount point under the media-localized super-root.
1058          * Scanning hammer2_pfslist doesn't help us because it represents
1059          * PFS cluster ids which can aggregate several named PFSs together.
1060          *
1061          * cluster->pmp will incorrectly point to spmp and must be fixed
1062          * up later on.
1063          */
1064         hammer2_inode_lock(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1065         cparent = hammer2_inode_cluster(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1066         lhc = hammer2_dirhash(label, strlen(label));
1067         cluster = hammer2_cluster_lookup(cparent, &key_next,
1068                                       lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1069                                       0);
1070         while (cluster) {
1071                 if (hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE &&
1072                     strcmp(label,
1073                        hammer2_cluster_rdata(cluster)->ipdata.filename) == 0) {
1074                         break;
1075                 }
1076                 cluster = hammer2_cluster_next(cparent, cluster, &key_next,
1077                                             key_next,
1078                                             lhc + HAMMER2_DIRHASH_LOMASK, 0);
1079         }
1080         hammer2_inode_unlock(spmp->iroot, cparent);
1081
1082         /*
1083          * PFS could not be found?
1084          */
1085         if (cluster == NULL) {
1086                 kprintf("hammer2_mount: PFS label not found\n");
1087                 hammer2_unmount_helper(mp, NULL, hmp);
1088                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1089                 hammer2_vfs_unmount(mp, MNT_FORCE);
1090
1091                 return EINVAL;
1092         }
1093
1094         /*
1095          * Acquire the pmp structure (it should have already been allocated
1096          * via hammer2_update_pmps() so do not pass cluster in to add to
1097          * available chains).
1098          *
1099          * Check if the cluster has already been mounted.  A cluster can
1100          * only be mounted once, use null mounts to mount additional copies.
1101          */
1102         ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1103         hammer2_cluster_bref(cluster, &bref);
1104         pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid);
1105         hammer2_cluster_unlock(cluster);
1106         hammer2_cluster_drop(cluster);
1107
1108         if (pmp->mp) {
1109                 kprintf("hammer2_mount: PFS already mounted!\n");
1110                 hammer2_unmount_helper(mp, NULL, hmp);
1111                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1112                 hammer2_vfs_unmount(mp, MNT_FORCE);
1113
1114                 return EBUSY;
1115         }
1116
1117         /*
1118          * Finish the mount
1119          */
1120         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1121
1122         mp->mnt_flag = MNT_LOCAL;
1123         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1124         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1125
1126         /*
1127          * required mount structure initializations
1128          */
1129         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1130         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1131
1132         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1133         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1134
1135         /*
1136          * Optional fields
1137          */
1138         mp->mnt_iosize_max = MAXPHYS;
1139
1140         /*
1141          * Connect up mount pointers.
1142          */
1143         hammer2_mount_helper(mp, pmp);
1144
1145         lockmgr(&hammer2_mntlk, LK_RELEASE);
1146
1147         /*
1148          * A mounted PFS needs a write thread for logical buffers and
1149          * a hidden directory for deletions of open files.  These features
1150          * are not used by unmounted PFSs.
1151          *
1152          * The logical file buffer bio write thread handles things like
1153          * physical block assignment and compression.
1154          */
1155         pmp->wthread_destroy = 0;
1156         lwkt_create(hammer2_write_thread, pmp,
1157                     &pmp->wthread_td, NULL, 0, -1, "h2pfs-%s", label);
1158
1159         /*
1160          * With the cluster operational install ihidden.
1161          * (only applicable to pfs mounts, not applicable to spmp)
1162          */
1163         hammer2_inode_install_hidden(pmp);
1164
1165         /*
1166          * Finish setup
1167          */
1168         vfs_getnewfsid(mp);
1169         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1170         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1171         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1172
1173         copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
1174         bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1175         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1176         copyinstr(path, mp->mnt_stat.f_mntonname,
1177                   sizeof(mp->mnt_stat.f_mntonname) - 1,
1178                   &size);
1179
1180         /*
1181          * Initial statfs to prime mnt_stat.
1182          */
1183         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1184
1185         return 0;
1186 }
1187
1188 /*
1189  * Scan PFSs under the super-root and create hammer2_pfs structures.
1190  */
1191 static
1192 void
1193 hammer2_update_pmps(hammer2_dev_t *hmp)
1194 {
1195         const hammer2_inode_data_t *ripdata;
1196         hammer2_cluster_t *cparent;
1197         hammer2_cluster_t *cluster;
1198         hammer2_blockref_t bref;
1199         hammer2_pfs_t *spmp;
1200         hammer2_pfs_t *pmp;
1201         hammer2_key_t key_next;
1202
1203         /*
1204          * Lookup mount point under the media-localized super-root.
1205          *
1206          * cluster->pmp will incorrectly point to spmp and must be fixed
1207          * up later on.
1208          */
1209         spmp = hmp->spmp;
1210         hammer2_inode_lock(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1211         cparent = hammer2_inode_cluster(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1212         cluster = hammer2_cluster_lookup(cparent, &key_next,
1213                                          HAMMER2_KEY_MIN,
1214                                          HAMMER2_KEY_MAX,
1215                                          0);
1216         while (cluster) {
1217                 if (hammer2_cluster_type(cluster) != HAMMER2_BREF_TYPE_INODE)
1218                         continue;
1219                 ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1220                 hammer2_cluster_bref(cluster, &bref);
1221                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1222
1223                 pmp = hammer2_pfsalloc(cluster, ripdata, bref.modify_tid);
1224                 cluster = hammer2_cluster_next(cparent, cluster,
1225                                                &key_next,
1226                                                key_next,
1227                                                HAMMER2_KEY_MAX,
1228                                                0);
1229         }
1230         hammer2_inode_unlock(spmp->iroot, cparent);
1231 }
1232
1233 static
1234 int
1235 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path,
1236                 struct vnode *devvp, struct ucred *cred)
1237 {
1238         int error;
1239
1240         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1241                 error = hammer2_recovery(hmp);
1242         } else {
1243                 error = 0;
1244         }
1245         return error;
1246 }
1247
1248 static
1249 int
1250 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1251 {
1252         hammer2_pfs_t *pmp;
1253         int flags;
1254         int error = 0;
1255
1256         pmp = MPTOPMP(mp);
1257
1258         if (pmp == NULL)
1259                 return(0);
1260
1261         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1262
1263         /*
1264          * If mount initialization proceeded far enough we must flush
1265          * its vnodes and sync the underlying mount points.  Three syncs
1266          * are required to fully flush the filesystem (freemap updates lag
1267          * by one flush, and one extra for safety).
1268          */
1269         if (mntflags & MNT_FORCE)
1270                 flags = FORCECLOSE;
1271         else
1272                 flags = 0;
1273         if (pmp->iroot) {
1274                 error = vflush(mp, 0, flags);
1275                 if (error)
1276                         goto failed;
1277                 hammer2_vfs_sync(mp, MNT_WAIT);
1278                 hammer2_vfs_sync(mp, MNT_WAIT);
1279                 hammer2_vfs_sync(mp, MNT_WAIT);
1280         }
1281
1282         if (pmp->wthread_td) {
1283                 hammer2_mtx_ex(&pmp->wthread_mtx);
1284                 pmp->wthread_destroy = 1;
1285                 wakeup(&pmp->wthread_bioq);
1286                 while (pmp->wthread_destroy != -1) {
1287                         mtxsleep(&pmp->wthread_destroy,
1288                                 &pmp->wthread_mtx, 0,
1289                                 "umount-sleep", 0);
1290                 }
1291                 hammer2_mtx_unlock(&pmp->wthread_mtx);
1292                 pmp->wthread_td = NULL;
1293         }
1294
1295         /*
1296          * Cleanup the frontend support XOPS threads
1297          */
1298         hammer2_xop_helper_cleanup(pmp);
1299
1300         /*
1301          * Cleanup our reference on ihidden.
1302          */
1303         if (pmp->ihidden) {
1304                 hammer2_inode_drop(pmp->ihidden);
1305                 pmp->ihidden = NULL;
1306         }
1307         if (pmp->mp)
1308                 hammer2_unmount_helper(mp, pmp, NULL);
1309
1310         error = 0;
1311 failed:
1312         lockmgr(&hammer2_mntlk, LK_RELEASE);
1313
1314         return (error);
1315 }
1316
1317 /*
1318  * Mount helper, hook the system mount into our PFS.
1319  * The mount lock is held.
1320  *
1321  * We must bump the mount_count on related devices for any
1322  * mounted PFSs.
1323  */
1324 static
1325 void
1326 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1327 {
1328         hammer2_cluster_t *cluster;
1329         hammer2_chain_t *rchain;
1330         int i;
1331
1332         mp->mnt_data = (qaddr_t)pmp;
1333         pmp->mp = mp;
1334
1335         /*
1336          * After pmp->mp is set we have to adjust hmp->mount_count.
1337          */
1338         cluster = &pmp->iroot->cluster;
1339         for (i = 0; i < cluster->nchains; ++i) {
1340                 rchain = cluster->array[i].chain;
1341                 if (rchain == NULL)
1342                         continue;
1343                 ++rchain->hmp->mount_count;
1344                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1345                         rchain->hmp, rchain->hmp->mount_count);
1346         }
1347
1348         /*
1349          * Create missing Xop threads
1350          */
1351         hammer2_xop_helper_create(pmp);
1352 }
1353
1354 /*
1355  * Mount helper, unhook the system mount from our PFS.
1356  * The mount lock is held.
1357  *
1358  * If hmp is supplied a mount responsible for being the first to open
1359  * the block device failed and the block device and all PFSs using the
1360  * block device must be cleaned up.
1361  *
1362  * If pmp is supplied multiple devices might be backing the PFS and each
1363  * must be disconnect.  This might not be the last PFS using some of the
1364  * underlying devices.  Also, we have to adjust our hmp->mount_count
1365  * accounting for the devices backing the pmp which is now undergoing an
1366  * unmount.
1367  */
1368 static
1369 void
1370 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1371 {
1372         hammer2_cluster_t *cluster;
1373         hammer2_chain_t *rchain;
1374         struct vnode *devvp;
1375         int dumpcnt;
1376         int ronly = 0;
1377         int i;
1378
1379         /*
1380          * If no device supplied this is a high-level unmount and we have to
1381          * to disconnect the mount, adjust mount_count, and locate devices
1382          * that might now have no mounts.
1383          */
1384         if (pmp) {
1385                 KKASSERT(hmp == NULL);
1386                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1387                 pmp->mp = NULL;
1388                 mp->mnt_data = NULL;
1389
1390                 /*
1391                  * After pmp->mp is cleared we have to account for
1392                  * mount_count.
1393                  */
1394                 cluster = &pmp->iroot->cluster;
1395                 for (i = 0; i < cluster->nchains; ++i) {
1396                         rchain = cluster->array[i].chain;
1397                         if (rchain == NULL)
1398                                 continue;
1399                         --rchain->hmp->mount_count;
1400                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1401                                 rchain->hmp, rchain->hmp->mount_count);
1402                         /* scrapping hmp now may invalidate the pmp */
1403                 }
1404 again:
1405                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1406                         if (hmp->mount_count == 0) {
1407                                 hammer2_unmount_helper(NULL, NULL, hmp);
1408                                 goto again;
1409                         }
1410                 }
1411                 return;
1412         }
1413
1414         /*
1415          * Try to terminate the block device.  We can't terminate it if
1416          * there are still PFSs referencing it.
1417          */
1418         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1419                 hmp, hmp->mount_count);
1420         if (hmp->mount_count)
1421                 return;
1422
1423         hammer2_pfsfree_scan(hmp);
1424         hammer2_dev_exlock(hmp);        /* XXX order */
1425
1426         /*
1427          * Cycle the volume data lock as a safety (probably not needed any
1428          * more).  To ensure everything is out we need to flush at least
1429          * three times.  (1) The running of the unlinkq can dirty the
1430          * filesystem, (2) A normal flush can dirty the freemap, and
1431          * (3) ensure that the freemap is fully synchronized.
1432          *
1433          * The next mount's recovery scan can clean everything up but we want
1434          * to leave the filesystem in a 100% clean state on a normal unmount.
1435          */
1436 #if 0
1437         hammer2_voldata_lock(hmp);
1438         hammer2_voldata_unlock(hmp);
1439 #endif
1440         hammer2_iocom_uninit(hmp);
1441
1442         if ((hmp->vchain.flags | hmp->fchain.flags) &
1443             HAMMER2_CHAIN_FLUSH_MASK) {
1444                 kprintf("hammer2_unmount: chains left over "
1445                         "after final sync\n");
1446                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1447                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1448
1449                 if (hammer2_debug & 0x0010)
1450                         Debugger("entered debugger");
1451         }
1452
1453         KKASSERT(hmp->spmp == NULL);
1454
1455         /*
1456          * Finish up with the device vnode
1457          */
1458         if ((devvp = hmp->devvp) != NULL) {
1459                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1460                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1461                 hmp->devvp = NULL;
1462                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1463                 vn_unlock(devvp);
1464                 vrele(devvp);
1465                 devvp = NULL;
1466         }
1467
1468         /*
1469          * Clear vchain/fchain flags that might prevent final cleanup
1470          * of these chains.
1471          */
1472         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1473                 atomic_clear_int(&hmp->vchain.flags,
1474                                  HAMMER2_CHAIN_MODIFIED);
1475                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1476                 hammer2_chain_drop(&hmp->vchain);
1477         }
1478         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1479                 atomic_clear_int(&hmp->vchain.flags,
1480                                  HAMMER2_CHAIN_UPDATE);
1481                 hammer2_chain_drop(&hmp->vchain);
1482         }
1483
1484         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1485                 atomic_clear_int(&hmp->fchain.flags,
1486                                  HAMMER2_CHAIN_MODIFIED);
1487                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1488                 hammer2_chain_drop(&hmp->fchain);
1489         }
1490         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1491                 atomic_clear_int(&hmp->fchain.flags,
1492                                  HAMMER2_CHAIN_UPDATE);
1493                 hammer2_chain_drop(&hmp->fchain);
1494         }
1495
1496         /*
1497          * Final drop of embedded freemap root chain to
1498          * clean up fchain.core (fchain structure is not
1499          * flagged ALLOCATED so it is cleaned out and then
1500          * left to rot).
1501          */
1502         hammer2_chain_drop(&hmp->fchain);
1503
1504         /*
1505          * Final drop of embedded volume root chain to clean
1506          * up vchain.core (vchain structure is not flagged
1507          * ALLOCATED so it is cleaned out and then left to
1508          * rot).
1509          */
1510         dumpcnt = 50;
1511         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1512         dumpcnt = 50;
1513         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1514         hammer2_dev_unlock(hmp);
1515         hammer2_chain_drop(&hmp->vchain);
1516
1517         hammer2_io_cleanup(hmp, &hmp->iotree);
1518         if (hmp->iofree_count) {
1519                 kprintf("io_cleanup: %d I/O's left hanging\n",
1520                         hmp->iofree_count);
1521         }
1522
1523         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1524         kmalloc_destroy(&hmp->mchain);
1525         kfree(hmp, M_HAMMER2);
1526 }
1527
1528 static
1529 int
1530 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1531              ino_t ino, struct vnode **vpp)
1532 {
1533         kprintf("hammer2_vget\n");
1534         return (EOPNOTSUPP);
1535 }
1536
1537 static
1538 int
1539 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1540 {
1541         hammer2_pfs_t *pmp;
1542         hammer2_cluster_t *cparent;
1543         int error;
1544         struct vnode *vp;
1545
1546         pmp = MPTOPMP(mp);
1547         if (pmp->iroot == NULL) {
1548                 *vpp = NULL;
1549                 error = EINVAL;
1550         } else {
1551                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_ALWAYS |
1552                                                HAMMER2_RESOLVE_SHARED);
1553                 cparent = hammer2_inode_cluster(pmp->iroot,
1554                                                 HAMMER2_RESOLVE_ALWAYS |
1555                                                 HAMMER2_RESOLVE_SHARED);
1556
1557                 /*
1558                  * Initialize pmp->inode_tid and pmp->modify_tid on first access
1559                  * to the root of mount that resolves good.
1560                  * XXX probably not the best place for this.
1561                  */
1562                 if (pmp->inode_tid == 0 &&
1563                     cparent->error == 0 && cparent->focus) {
1564                         const hammer2_inode_data_t *ripdata;
1565                         hammer2_blockref_t bref;
1566
1567                         ripdata = &hammer2_cluster_rdata(cparent)->ipdata;
1568                         hammer2_cluster_bref(cparent, &bref);
1569                         pmp->inode_tid = ripdata->meta.pfs_inum + 1;
1570                         pmp->modify_tid = bref.modify_tid;
1571                         pmp->iroot->meta = ripdata->meta;
1572                         hammer2_cluster_bref(cparent, &pmp->iroot->bref);
1573                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1574                                 pmp->inode_tid, pmp->modify_tid);
1575                 }
1576
1577                 vp = hammer2_igetv(pmp->iroot, cparent, &error);
1578                 hammer2_inode_unlock(pmp->iroot, cparent);
1579                 *vpp = vp;
1580                 if (vp == NULL)
1581                         kprintf("vnodefail\n");
1582         }
1583
1584         return (error);
1585 }
1586
1587 /*
1588  * Filesystem status
1589  *
1590  * XXX incorporate ipdata->meta.inode_quota and data_quota
1591  */
1592 static
1593 int
1594 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1595 {
1596         hammer2_pfs_t *pmp;
1597         hammer2_dev_t *hmp;
1598         hammer2_blockref_t bref;
1599
1600         pmp = MPTOPMP(mp);
1601         KKASSERT(pmp->iroot->cluster.nchains >= 1);
1602         hmp = pmp->iroot->cluster.focus->hmp;   /* iroot retains focus */
1603         bref = pmp->iroot->cluster.focus->bref; /* no lock */
1604
1605         mp->mnt_stat.f_files = bref.inode_count;
1606         mp->mnt_stat.f_ffree = 0;
1607         mp->mnt_stat.f_blocks = (bref.data_count +
1608                                  hmp->voldata.allocator_free) /
1609                                 mp->mnt_vstat.f_bsize;
1610         mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free /
1611                                 mp->mnt_vstat.f_bsize;
1612         mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1613
1614         *sbp = mp->mnt_stat;
1615         return (0);
1616 }
1617
1618 static
1619 int
1620 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1621 {
1622         hammer2_pfs_t *pmp;
1623         hammer2_dev_t *hmp;
1624         hammer2_blockref_t bref;
1625
1626         pmp = MPTOPMP(mp);
1627         KKASSERT(pmp->iroot->cluster.nchains >= 1);
1628         hmp = pmp->iroot->cluster.focus->hmp;   /* iroot retains focus */
1629         bref = pmp->iroot->cluster.focus->bref; /* no lock */
1630
1631         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1632         mp->mnt_vstat.f_files = bref.inode_count;
1633         mp->mnt_vstat.f_ffree = 0;
1634         mp->mnt_vstat.f_blocks = (bref.data_count +
1635                                  hmp->voldata.allocator_free) /
1636                                 mp->mnt_vstat.f_bsize;
1637         mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1638                                 mp->mnt_vstat.f_bsize;
1639         mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1640
1641         *sbp = mp->mnt_vstat;
1642         return (0);
1643 }
1644
1645 /*
1646  * Mount-time recovery (RW mounts)
1647  *
1648  * Updates to the free block table are allowed to lag flushes by one
1649  * transaction.  In case of a crash, then on a fresh mount we must do an
1650  * incremental scan of the last committed transaction id and make sure that
1651  * all related blocks have been marked allocated.
1652  *
1653  * The super-root topology and each PFS has its own transaction id domain,
1654  * so we must track PFS boundary transitions.
1655  */
1656 struct hammer2_recovery_elm {
1657         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1658         hammer2_chain_t *chain;
1659         hammer2_tid_t sync_tid;
1660 };
1661
1662 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1663
1664 struct hammer2_recovery_info {
1665         struct hammer2_recovery_list list;
1666         int     depth;
1667 };
1668
1669 static int hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_dev_t *hmp,
1670                         hammer2_chain_t *parent,
1671                         struct hammer2_recovery_info *info,
1672                         hammer2_tid_t sync_tid);
1673
1674 #define HAMMER2_RECOVERY_MAXDEPTH       10
1675
1676 static
1677 int
1678 hammer2_recovery(hammer2_dev_t *hmp)
1679 {
1680         hammer2_trans_t trans;
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(&trans, 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(&trans, 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(&trans, hmp, parent,
1717                                               &info,
1718                                               hmp->voldata.freemap_tid);
1719                 hammer2_chain_unlock(parent);
1720                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
1721                 if (error)
1722                         cumulative_error = error;
1723         }
1724         hammer2_trans_done(&trans);
1725
1726         return cumulative_error;
1727 }
1728
1729 static
1730 int
1731 hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_dev_t *hmp,
1732                       hammer2_chain_t *parent,
1733                       struct hammer2_recovery_info *info,
1734                       hammer2_tid_t sync_tid)
1735 {
1736         const hammer2_inode_data_t *ripdata;
1737         hammer2_chain_t *chain;
1738         int cache_index;
1739         int cumulative_error = 0;
1740         int error;
1741
1742         /*
1743          * Adjust freemap to ensure that the block(s) are marked allocated.
1744          */
1745         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1746                 hammer2_freemap_adjust(trans, hmp, &parent->bref,
1747                                        HAMMER2_FREEMAP_DORECOVER);
1748         }
1749
1750         /*
1751          * Check type for recursive scan
1752          */
1753         switch(parent->bref.type) {
1754         case HAMMER2_BREF_TYPE_VOLUME:
1755                 /* data already instantiated */
1756                 break;
1757         case HAMMER2_BREF_TYPE_INODE:
1758                 /*
1759                  * Must instantiate data for DIRECTDATA test and also
1760                  * for recursion.
1761                  */
1762                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1763                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
1764                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1765                         /* not applicable to recovery scan */
1766                         hammer2_chain_unlock(parent);
1767                         return 0;
1768                 }
1769                 hammer2_chain_unlock(parent);
1770                 break;
1771         case HAMMER2_BREF_TYPE_INDIRECT:
1772                 /*
1773                  * Must instantiate data for recursion
1774                  */
1775                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1776                 hammer2_chain_unlock(parent);
1777                 break;
1778         case HAMMER2_BREF_TYPE_DATA:
1779         case HAMMER2_BREF_TYPE_FREEMAP:
1780         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1781         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1782                 /* not applicable to recovery scan */
1783                 return 0;
1784                 break;
1785         default:
1786                 return EDOM;
1787         }
1788
1789         /*
1790          * Defer operation if depth limit reached or if we are crossing a
1791          * PFS boundary.
1792          */
1793         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1794                 struct hammer2_recovery_elm *elm;
1795
1796                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1797                 elm->chain = parent;
1798                 elm->sync_tid = sync_tid;
1799                 hammer2_chain_ref(parent);
1800                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
1801                 /* unlocked by caller */
1802
1803                 return(0);
1804         }
1805
1806
1807         /*
1808          * Recursive scan of the last flushed transaction only.  We are
1809          * doing this without pmp assignments so don't leave the chains
1810          * hanging around after we are done with them.
1811          */
1812         cache_index = 0;
1813         chain = hammer2_chain_scan(parent, NULL, &cache_index,
1814                                    HAMMER2_LOOKUP_NODATA);
1815         while (chain) {
1816                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
1817                 if (chain->bref.mirror_tid > sync_tid) {
1818                         ++info->depth;
1819                         error = hammer2_recovery_scan(trans, hmp, chain,
1820                                                       info, sync_tid);
1821                         --info->depth;
1822                         if (error)
1823                                 cumulative_error = error;
1824                 }
1825
1826                 /*
1827                  * Flush the recovery at the PFS boundary to stage it for
1828                  * the final flush of the super-root topology.
1829                  */
1830                 if ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
1831                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
1832                         hammer2_flush(trans, chain, 1);
1833                 }
1834                 chain = hammer2_chain_scan(parent, chain, &cache_index,
1835                                            HAMMER2_LOOKUP_NODATA);
1836         }
1837
1838         return cumulative_error;
1839 }
1840
1841 /*
1842  * Sync a mount point; this is called on a per-mount basis from the
1843  * filesystem syncer process periodically and whenever a user issues
1844  * a sync.
1845  */
1846 int
1847 hammer2_vfs_sync(struct mount *mp, int waitfor)
1848 {
1849         struct hammer2_sync_info info;
1850         hammer2_inode_t *iroot;
1851         hammer2_chain_t *chain;
1852         hammer2_chain_t *parent;
1853         hammer2_pfs_t *pmp;
1854         hammer2_dev_t *hmp;
1855         int flags;
1856         int error;
1857         int total_error;
1858         int i;
1859         int j;
1860
1861         pmp = MPTOPMP(mp);
1862         iroot = pmp->iroot;
1863         KKASSERT(iroot);
1864         KKASSERT(iroot->pmp == pmp);
1865
1866         /*
1867          * We can't acquire locks on existing vnodes while in a transaction
1868          * without risking a deadlock.  This assumes that vfsync() can be
1869          * called without the vnode locked (which it can in DragonFly).
1870          * Otherwise we'd have to implement a multi-pass or flag the lock
1871          * failures and retry.
1872          *
1873          * The reclamation code interlocks with the sync list's token
1874          * (by removing the vnode from the scan list) before unlocking
1875          * the inode, giving us time to ref the inode.
1876          */
1877         /*flags = VMSC_GETVP;*/
1878         flags = 0;
1879         if (waitfor & MNT_LAZY)
1880                 flags |= VMSC_ONEPASS;
1881
1882 #if 0
1883         /*
1884          * Preflush the vnodes using a normal transaction before interlocking
1885          * with a flush transaction.
1886          */
1887         hammer2_trans_init(&info.trans, pmp, 0);
1888         info.error = 0;
1889         info.waitfor = MNT_NOWAIT;
1890         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1891         hammer2_trans_done(&info.trans);
1892 #endif
1893
1894         /*
1895          * Start our flush transaction.  This does not return until all
1896          * concurrent transactions have completed and will prevent any
1897          * new transactions from running concurrently, except for the
1898          * buffer cache transactions.
1899          *
1900          * For efficiency do an async pass before making sure with a
1901          * synchronous pass on all related buffer cache buffers.  It
1902          * should theoretically not be possible for any new file buffers
1903          * to be instantiated during this sequence.
1904          */
1905         hammer2_trans_init(&info.trans, pmp, HAMMER2_TRANS_ISFLUSH |
1906                                              HAMMER2_TRANS_PREFLUSH);
1907         hammer2_run_unlinkq(&info.trans, pmp);
1908
1909         info.error = 0;
1910         info.waitfor = MNT_NOWAIT;
1911         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1912         info.waitfor = MNT_WAIT;
1913         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
1914
1915         /*
1916          * Clear PREFLUSH.  This prevents (or asserts on) any new logical
1917          * buffer cache flushes which occur during the flush.  Device buffers
1918          * are not affected.
1919          */
1920         hammer2_bioq_sync(info.trans.pmp);
1921         atomic_clear_int(&info.trans.flags, HAMMER2_TRANS_PREFLUSH);
1922
1923         total_error = 0;
1924
1925         /*
1926          * Flush all nodes to synchronize the PFSROOT subtopology to the media.
1927          *
1928          * Note that this flush will not be visible on crash recovery until
1929          * we flush the super-root topology in the next loop.
1930          */
1931         for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
1932                 chain = iroot->cluster.array[i].chain;
1933                 if (chain == NULL)
1934                         continue;
1935
1936                 hammer2_chain_ref(chain);
1937                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
1938                 if (chain->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1939                         hammer2_flush(&info.trans, chain, 1);
1940                         parent = chain->parent;
1941                         KKASSERT(chain->pmp != parent->pmp);
1942                         hammer2_chain_setflush(&info.trans, parent);
1943                 }
1944                 hammer2_chain_unlock(chain);
1945                 hammer2_chain_drop(chain);
1946         }
1947         hammer2_trans_done(&info.trans);
1948
1949         /*
1950          * Flush all volume roots to synchronize PFS flushes with the
1951          * storage media volume header.  This will flush the freemap and
1952          * the superroot topology but stops when it reaches a PFSROOT
1953          * (which we already flushed above).
1954          *
1955          * This is the last step which connects the volume root to the
1956          * PFSROOT dirs flushed above.
1957          *
1958          * Each spmp (representing the hmp's super-root) requires its own
1959          * transaction.
1960          */
1961         for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
1962                 hammer2_chain_t *tmp;
1963
1964                 chain = iroot->cluster.array[i].chain;
1965                 if (chain == NULL)
1966                         continue;
1967
1968                 hmp = chain->hmp;
1969
1970                 /*
1971                  * We only have to flush each hmp once
1972                  */
1973                 for (j = i - 1; j >= 0; --j) {
1974                         if ((tmp = iroot->cluster.array[j].chain) != NULL) {
1975                                 if (tmp->hmp == hmp)
1976                                         break;
1977                         }
1978                 }
1979                 if (j >= 0)
1980                         continue;
1981
1982                 /*
1983                  * spmp transaction.  The super-root is never directly
1984                  * mounted so there shouldn't be any vnodes, let alone any
1985                  * dirty vnodes associated with it.
1986                  */
1987                 hammer2_trans_init(&info.trans, hmp->spmp,
1988                                    HAMMER2_TRANS_ISFLUSH);
1989
1990                 /*
1991                  * Media mounts have two 'roots', vchain for the topology
1992                  * and fchain for the free block table.  Flush both.
1993                  *
1994                  * Note that the topology and free block table are handled
1995                  * independently, so the free block table can wind up being
1996                  * ahead of the topology.  We depend on the bulk free scan
1997                  * code to deal with any loose ends.
1998                  */
1999                 hammer2_chain_ref(&hmp->vchain);
2000                 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2001                 hammer2_chain_ref(&hmp->fchain);
2002                 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2003                 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2004                         /*
2005                          * This will also modify vchain as a side effect,
2006                          * mark vchain as modified now.
2007                          */
2008                         hammer2_voldata_modify(hmp);
2009                         chain = &hmp->fchain;
2010                         hammer2_flush(&info.trans, chain, 1);
2011                         KKASSERT(chain == &hmp->fchain);
2012                 }
2013                 hammer2_chain_unlock(&hmp->fchain);
2014                 hammer2_chain_unlock(&hmp->vchain);
2015                 hammer2_chain_drop(&hmp->fchain);
2016                 /* vchain dropped down below */
2017
2018                 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2019                 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2020                         chain = &hmp->vchain;
2021                         hammer2_flush(&info.trans, chain, 1);
2022                         KKASSERT(chain == &hmp->vchain);
2023                 }
2024                 hammer2_chain_unlock(&hmp->vchain);
2025                 hammer2_chain_drop(&hmp->vchain);
2026
2027                 error = 0;
2028
2029                 /*
2030                  * We can't safely flush the volume header until we have
2031                  * flushed any device buffers which have built up.
2032                  *
2033                  * XXX this isn't being incremental
2034                  */
2035                 vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
2036                 error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
2037                 vn_unlock(hmp->devvp);
2038
2039                 /*
2040                  * The flush code sets CHAIN_VOLUMESYNC to indicate that the
2041                  * volume header needs synchronization via hmp->volsync.
2042                  *
2043                  * XXX synchronize the flag & data with only this flush XXX
2044                  */
2045                 if (error == 0 &&
2046                     (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
2047                         struct buf *bp;
2048
2049                         /*
2050                          * Synchronize the disk before flushing the volume
2051                          * header.
2052                          */
2053                         bp = getpbuf(NULL);
2054                         bp->b_bio1.bio_offset = 0;
2055                         bp->b_bufsize = 0;
2056                         bp->b_bcount = 0;
2057                         bp->b_cmd = BUF_CMD_FLUSH;
2058                         bp->b_bio1.bio_done = biodone_sync;
2059                         bp->b_bio1.bio_flags |= BIO_SYNC;
2060                         vn_strategy(hmp->devvp, &bp->b_bio1);
2061                         biowait(&bp->b_bio1, "h2vol");
2062                         relpbuf(bp, NULL);
2063
2064                         /*
2065                          * Then we can safely flush the version of the
2066                          * volume header synchronized by the flush code.
2067                          */
2068                         i = hmp->volhdrno + 1;
2069                         if (i >= HAMMER2_NUM_VOLHDRS)
2070                                 i = 0;
2071                         if (i * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
2072                             hmp->volsync.volu_size) {
2073                                 i = 0;
2074                         }
2075                         kprintf("sync volhdr %d %jd\n",
2076                                 i, (intmax_t)hmp->volsync.volu_size);
2077                         bp = getblk(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2078                                     HAMMER2_PBUFSIZE, 0, 0);
2079                         atomic_clear_int(&hmp->vchain.flags,
2080                                          HAMMER2_CHAIN_VOLUMESYNC);
2081                         bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
2082                         bawrite(bp);
2083                         hmp->volhdrno = i;
2084                 }
2085                 if (error)
2086                         total_error = error;
2087
2088                 hammer2_trans_done(&info.trans);        /* spmp trans */
2089         }
2090         return (total_error);
2091 }
2092
2093 /*
2094  * Sync passes.
2095  */
2096 static int
2097 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2098 {
2099         struct hammer2_sync_info *info = data;
2100         hammer2_inode_t *ip;
2101         int error;
2102
2103         /*
2104          * Degenerate cases.  Note that ip == NULL typically means the
2105          * syncer vnode itself and we don't want to vclrisdirty() in that
2106          * situation.
2107          */
2108         ip = VTOI(vp);
2109         if (ip == NULL) {
2110                 return(0);
2111         }
2112         if (vp->v_type == VNON || vp->v_type == VBAD) {
2113                 vclrisdirty(vp);
2114                 return(0);
2115         }
2116
2117         /*
2118          * VOP_FSYNC will start a new transaction so replicate some code
2119          * here to do it inline (see hammer2_vop_fsync()).
2120          *
2121          * WARNING: The vfsync interacts with the buffer cache and might
2122          *          block, we can't hold the inode lock at that time.
2123          *          However, we MUST ref ip before blocking to ensure that
2124          *          it isn't ripped out from under us (since we do not
2125          *          hold a lock on the vnode).
2126          */
2127         hammer2_inode_ref(ip);
2128         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2129             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2130                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2131                 hammer2_inode_fsync(&info->trans, ip, NULL);
2132         }
2133         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2134             RB_EMPTY(&vp->v_rbdirty_tree)) {
2135                 vclrisdirty(vp);
2136         }
2137
2138         hammer2_inode_drop(ip);
2139 #if 1
2140         error = 0;
2141         if (error)
2142                 info->error = error;
2143 #endif
2144         return(0);
2145 }
2146
2147 static
2148 int
2149 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2150 {
2151         return (0);
2152 }
2153
2154 static
2155 int
2156 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2157                struct fid *fhp, struct vnode **vpp)
2158 {
2159         return (0);
2160 }
2161
2162 static
2163 int
2164 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2165                  int *exflagsp, struct ucred **credanonp)
2166 {
2167         return (0);
2168 }
2169
2170 /*
2171  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2172  * header into the HMP
2173  *
2174  * XXX read four volhdrs and use the one with the highest TID whos CRC
2175  *     matches.
2176  *
2177  * XXX check iCRCs.
2178  *
2179  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2180  *     nonexistant locations.
2181  *
2182  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2183  */
2184 static
2185 int
2186 hammer2_install_volume_header(hammer2_dev_t *hmp)
2187 {
2188         hammer2_volume_data_t *vd;
2189         struct buf *bp;
2190         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2191         int error_reported;
2192         int error;
2193         int valid;
2194         int i;
2195
2196         error_reported = 0;
2197         error = 0;
2198         valid = 0;
2199         bp = NULL;
2200
2201         /*
2202          * There are up to 4 copies of the volume header (syncs iterate
2203          * between them so there is no single master).  We don't trust the
2204          * volu_size field so we don't know precisely how large the filesystem
2205          * is, so depend on the OS to return an error if we go beyond the
2206          * block device's EOF.
2207          */
2208         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2209                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2210                               HAMMER2_VOLUME_BYTES, &bp);
2211                 if (error) {
2212                         brelse(bp);
2213                         bp = NULL;
2214                         continue;
2215                 }
2216
2217                 vd = (struct hammer2_volume_data *) bp->b_data;
2218                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2219                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2220                         brelse(bp);
2221                         bp = NULL;
2222                         continue;
2223                 }
2224
2225                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2226                         /* XXX: Reversed-endianness filesystem */
2227                         kprintf("hammer2: reverse-endian filesystem detected");
2228                         brelse(bp);
2229                         bp = NULL;
2230                         continue;
2231                 }
2232
2233                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2234                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2235                                       HAMMER2_VOLUME_ICRC0_SIZE);
2236                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2237                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2238                                        HAMMER2_VOLUME_ICRC1_SIZE);
2239                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2240                         kprintf("hammer2 volume header crc "
2241                                 "mismatch copy #%d %08x/%08x\n",
2242                                 i, crc0, crc);
2243                         error_reported = 1;
2244                         brelse(bp);
2245                         bp = NULL;
2246                         continue;
2247                 }
2248                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2249                         valid = 1;
2250                         hmp->voldata = *vd;
2251                         hmp->volhdrno = i;
2252                 }
2253                 brelse(bp);
2254                 bp = NULL;
2255         }
2256         if (valid) {
2257                 hmp->volsync = hmp->voldata;
2258                 error = 0;
2259                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2260                         kprintf("hammer2: using volume header #%d\n",
2261                                 hmp->volhdrno);
2262                 }
2263         } else {
2264                 error = EINVAL;
2265                 kprintf("hammer2: no valid volume headers found!\n");
2266         }
2267         return (error);
2268 }
2269
2270 /*
2271  * This handles hysteresis on regular file flushes.  Because the BIOs are
2272  * routed to a thread it is possible for an excessive number to build up
2273  * and cause long front-end stalls long before the runningbuffspace limit
2274  * is hit, so we implement hammer2_flush_pipe to control the
2275  * hysteresis.
2276  *
2277  * This is a particular problem when compression is used.
2278  */
2279 void
2280 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2281 {
2282         atomic_add_int(&pmp->count_lwinprog, 1);
2283 }
2284
2285 void
2286 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2287 {
2288         int lwinprog;
2289
2290         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2291         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2292             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2293                 atomic_clear_int(&pmp->count_lwinprog,
2294                                  HAMMER2_LWINPROG_WAITING);
2295                 wakeup(&pmp->count_lwinprog);
2296         }
2297 }
2298
2299 void
2300 hammer2_lwinprog_wait(hammer2_pfs_t *pmp)
2301 {
2302         int lwinprog;
2303
2304         for (;;) {
2305                 lwinprog = pmp->count_lwinprog;
2306                 cpu_ccfence();
2307                 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2308                         break;
2309                 tsleep_interlock(&pmp->count_lwinprog, 0);
2310                 atomic_set_int(&pmp->count_lwinprog, HAMMER2_LWINPROG_WAITING);
2311                 lwinprog = pmp->count_lwinprog;
2312                 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2313                         break;
2314                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2315         }
2316 }
2317
2318 /*
2319  * Manage excessive memory resource use for chain and related
2320  * structures.
2321  */
2322 void
2323 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2324 {
2325         uint32_t waiting;
2326         uint32_t count;
2327         uint32_t limit;
2328 #if 0
2329         static int zzticks;
2330 #endif
2331
2332         /*
2333          * Atomic check condition and wait.  Also do an early speedup of
2334          * the syncer to try to avoid hitting the wait.
2335          */
2336         for (;;) {
2337                 waiting = pmp->inmem_dirty_chains;
2338                 cpu_ccfence();
2339                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2340
2341                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2342                 if (limit < hammer2_limit_dirty_chains)
2343                         limit = hammer2_limit_dirty_chains;
2344                 if (limit < 1000)
2345                         limit = 1000;
2346
2347 #if 0
2348                 if ((int)(ticks - zzticks) > hz) {
2349                         zzticks = ticks;
2350                         kprintf("count %ld %ld\n", count, limit);
2351                 }
2352 #endif
2353
2354                 /*
2355                  * Block if there are too many dirty chains present, wait
2356                  * for the flush to clean some out.
2357                  */
2358                 if (count > limit) {
2359                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2360                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2361                                                waiting,
2362                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2363                                 speedup_syncer(pmp->mp);
2364                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2365                                        "chnmem", hz);
2366                         }
2367                         continue;       /* loop on success or fail */
2368                 }
2369
2370                 /*
2371                  * Try to start an early flush before we are forced to block.
2372                  */
2373                 if (count > limit * 7 / 10)
2374                         speedup_syncer(pmp->mp);
2375                 break;
2376         }
2377 }
2378
2379 void
2380 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2381 {
2382         if (pmp) {
2383                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2384         }
2385 }
2386
2387 void
2388 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2389 {
2390         uint32_t waiting;
2391
2392         if (pmp == NULL)
2393                 return;
2394
2395         for (;;) {
2396                 waiting = pmp->inmem_dirty_chains;
2397                 cpu_ccfence();
2398                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2399                                        waiting,
2400                                        (waiting - 1) &
2401                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2402                         break;
2403                 }
2404         }
2405
2406         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2407                 wakeup(&pmp->inmem_dirty_chains);
2408 }
2409
2410 /*
2411  * Debugging
2412  */
2413 void
2414 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2415 {
2416         hammer2_chain_t *scan;
2417         hammer2_chain_t *parent;
2418
2419         --*countp;
2420         if (*countp == 0) {
2421                 kprintf("%*.*s...\n", tab, tab, "");
2422                 return;
2423         }
2424         if (*countp < 0)
2425                 return;
2426         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2427                 tab, tab, "", pfx,
2428                 chain, chain->bref.type,
2429                 chain->bref.key, chain->bref.keybits,
2430                 chain->bref.mirror_tid);
2431
2432         kprintf("%*.*s      [%08x] (%s) refs=%d",
2433                 tab, tab, "",
2434                 chain->flags,
2435                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2436                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2437                 chain->refs);
2438
2439         parent = chain->parent;
2440         if (parent)
2441                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2442                         tab, tab, "",
2443                         parent, parent->flags, parent->refs);
2444         if (RB_EMPTY(&chain->core.rbtree)) {
2445                 kprintf("\n");
2446         } else {
2447                 kprintf(" {\n");
2448                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2449                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2450                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2451                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2452                                 chain->data->ipdata.filename);
2453                 else
2454                         kprintf("%*.*s}\n", tab, tab, "");
2455         }
2456 }