sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2014 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
  62 #include "hammer2.h"
  63 #include "hammer2_lz4.h"
  64
  65 #include "zlib/hammer2_zlib.h"
  66
  67 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  68
  69 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  70
  71 struct hammer2_sync_info {
  72         hammer2_trans_t trans;
  73         int error;
  74         int waitfor;
  75 };
  76
  77 TAILQ_HEAD(hammer2_mntlist, hammer2_mount);
  78 TAILQ_HEAD(hammer2_pfslist, hammer2_pfsmount);
  79 static struct hammer2_mntlist hammer2_mntlist;
  80 static struct hammer2_pfslist hammer2_pfslist;
  81 static struct lock hammer2_mntlk;
  82
  83 int hammer2_debug;
  84 int hammer2_cluster_enable = 1;
  85 int hammer2_hardlink_enable = 1;
  86 int hammer2_flush_pipe = 100;
  87 int hammer2_synchronous_flush = 1;
  88 int hammer2_dio_count;
  89 long hammer2_limit_dirty_chains;
  90 long hammer2_iod_file_read;
  91 long hammer2_iod_meta_read;
  92 long hammer2_iod_indr_read;
  93 long hammer2_iod_fmap_read;
  94 long hammer2_iod_volu_read;
  95 long hammer2_iod_file_write;
  96 long hammer2_iod_meta_write;
  97 long hammer2_iod_indr_write;
  98 long hammer2_iod_fmap_write;
  99 long hammer2_iod_volu_write;
 100 long hammer2_ioa_file_read;
 101 long hammer2_ioa_meta_read;
 102 long hammer2_ioa_indr_read;
 103 long hammer2_ioa_fmap_read;
 104 long hammer2_ioa_volu_read;
 105 long hammer2_ioa_fmap_write;
 106 long hammer2_ioa_file_write;
 107 long hammer2_ioa_meta_write;
 108 long hammer2_ioa_indr_write;
 109 long hammer2_ioa_volu_write;
 110
 111 MALLOC_DECLARE(C_BUFFER);
 112 MALLOC_DEFINE(C_BUFFER, "compbuffer", "Buffer used for compression.");
 113
 114 MALLOC_DECLARE(D_BUFFER);
 115 MALLOC_DEFINE(D_BUFFER, "decompbuffer", "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_mount_t *, struct mount *, char *,
 183                                 struct vnode *, struct ucred *);
 184 static int hammer2_recovery(hammer2_mount_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_mount_t *hmp);
 200 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 201
 202 static void hammer2_write_thread(void *arg);
 203
 204 static void hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp);
 205 static void hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp);
 206
 207 /*
 208  * Functions for compression in threads,
 209  * from hammer2_vnops.c
 210  */
 211 static void hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
 212                                 hammer2_inode_t *ip,
 213                                 hammer2_inode_data_t *ipdata,
 214                                 hammer2_cluster_t *cparent,
 215                                 hammer2_key_t lbase, int ioflag, int pblksize,
 216                                 int *errorp);
 217 static void hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
 218                                 hammer2_inode_t *ip,
 219                                 const hammer2_inode_data_t *ipdata,
 220                                 hammer2_cluster_t *cparent,
 221                                 hammer2_key_t lbase, int ioflag,
 222                                 int pblksize, int *errorp, int comp_algo);
 223 static void hammer2_zero_check_and_write(struct buf *bp,
 224                                 hammer2_trans_t *trans, hammer2_inode_t *ip,
 225                                 const hammer2_inode_data_t *ipdata,
 226                                 hammer2_cluster_t *cparent,
 227                                 hammer2_key_t lbase,
 228                                 int ioflag, int pblksize, int *errorp);
 229 static int test_block_zeros(const char *buf, size_t bytes);
 230 static void zero_write(struct buf *bp, hammer2_trans_t *trans,
 231                                 hammer2_inode_t *ip,
 232                                 const hammer2_inode_data_t *ipdata,
 233                                 hammer2_cluster_t *cparent,
 234                                 hammer2_key_t lbase,
 235                                 int *errorp);
 236 static void hammer2_write_bp(hammer2_cluster_t *cluster, struct buf *bp,
 237                                 int ioflag, int pblksize, int *errorp);
 238
 239 static int hammer2_rcvdmsg(kdmsg_msg_t *msg);
 240 static void hammer2_autodmsg(kdmsg_msg_t *msg);
 241 static int hammer2_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg);
 242
 243
 244 /*
 245  * HAMMER2 vfs operations.
 246  */
 247 static struct vfsops hammer2_vfsops = {
 248         .vfs_init       = hammer2_vfs_init,
 249         .vfs_uninit     = hammer2_vfs_uninit,
 250         .vfs_sync       = hammer2_vfs_sync,
 251         .vfs_mount      = hammer2_vfs_mount,
 252         .vfs_unmount    = hammer2_vfs_unmount,
 253         .vfs_root       = hammer2_vfs_root,
 254         .vfs_statfs     = hammer2_vfs_statfs,
 255         .vfs_statvfs    = hammer2_vfs_statvfs,
 256         .vfs_vget       = hammer2_vfs_vget,
 257         .vfs_vptofh     = hammer2_vfs_vptofh,
 258         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 259         .vfs_checkexp   = hammer2_vfs_checkexp
 260 };
 261
 262 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 263
 264 VFS_SET(hammer2_vfsops, hammer2, 0);
 265 MODULE_VERSION(hammer2, 1);
 266
 267 static
 268 int
 269 hammer2_vfs_init(struct vfsconf *conf)
 270 {
 271         static struct objcache_malloc_args margs_read;
 272         static struct objcache_malloc_args margs_write;
 273
 274         int error;
 275
 276         error = 0;
 277
 278         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 279                 error = EINVAL;
 280         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 281                 error = EINVAL;
 282         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 283                 error = EINVAL;
 284
 285         if (error)
 286                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 287
 288         margs_read.objsize = 65536;
 289         margs_read.mtype = D_BUFFER;
 290
 291         margs_write.objsize = 32768;
 292         margs_write.mtype = C_BUFFER;
 293
 294         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 295                                 0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
 296                                 objcache_malloc_free, &margs_read);
 297         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 298                                 0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
 299                                 objcache_malloc_free, &margs_write);
 300
 301         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 302         TAILQ_INIT(&hammer2_mntlist);
 303         TAILQ_INIT(&hammer2_pfslist);
 304
 305         hammer2_limit_dirty_chains = desiredvnodes / 10;
 306
 307         hammer2_trans_manage_init();
 308
 309         return (error);
 310 }
 311
 312 static
 313 int
 314 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 315 {
 316         objcache_destroy(cache_buffer_read);
 317         objcache_destroy(cache_buffer_write);
 318         return 0;
 319 }
 320
 321 /*
 322  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
 323  * mounts and the spmp structure for media (hmp) structures.
 324  */
 325 static hammer2_pfsmount_t *
 326 hammer2_pfsalloc(const hammer2_inode_data_t *ipdata, hammer2_tid_t alloc_tid)
 327 {
 328         hammer2_pfsmount_t *pmp;
 329
 330         pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 331         kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 332         kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 333         lockinit(&pmp->lock, "pfslk", 0, 0);
 334         spin_init(&pmp->inum_spin);
 335         RB_INIT(&pmp->inum_tree);
 336         TAILQ_INIT(&pmp->unlinkq);
 337         spin_init(&pmp->list_spin);
 338
 339         pmp->alloc_tid = alloc_tid + 1;   /* our first media transaction id */
 340         pmp->flush_tid = pmp->alloc_tid;
 341         if (ipdata) {
 342                 pmp->inode_tid = ipdata->pfs_inum + 1;
 343                 pmp->pfs_clid = ipdata->pfs_clid;
 344         }
 345         mtx_init(&pmp->wthread_mtx);
 346         bioq_init(&pmp->wthread_bioq);
 347
 348         return pmp;
 349 }
 350
 351 /*
 352  * Mount or remount HAMMER2 fileystem from physical media
 353  *
 354  *      mountroot
 355  *              mp              mount point structure
 356  *              path            NULL
 357  *              data            <unused>
 358  *              cred            <unused>
 359  *
 360  *      mount
 361  *              mp              mount point structure
 362  *              path            path to mount point
 363  *              data            pointer to argument structure in user space
 364  *                      volume  volume path (device@LABEL form)
 365  *                      hflags  user mount flags
 366  *              cred            user credentials
 367  *
 368  * RETURNS:     0       Success
 369  *              !0      error number
 370  */
 371 static
 372 int
 373 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 374                   struct ucred *cred)
 375 {
 376         struct hammer2_mount_info info;
 377         hammer2_pfsmount_t *pmp;
 378         hammer2_pfsmount_t *spmp;
 379         hammer2_mount_t *hmp;
 380         hammer2_key_t key_next;
 381         hammer2_key_t key_dummy;
 382         hammer2_key_t lhc;
 383         struct vnode *devvp;
 384         struct nlookupdata nd;
 385         hammer2_chain_t *parent;
 386         hammer2_chain_t *rchain;
 387         hammer2_cluster_t *cluster;
 388         hammer2_cluster_t *cparent;
 389         const hammer2_inode_data_t *ipdata;
 390         hammer2_blockref_t bref;
 391         struct file *fp;
 392         char devstr[MNAMELEN];
 393         size_t size;
 394         size_t done;
 395         char *dev;
 396         char *label;
 397         int ronly = 1;
 398         int error;
 399         int cache_index;
 400         int ddflag;
 401         int i;
 402
 403         hmp = NULL;
 404         pmp = NULL;
 405         dev = NULL;
 406         label = NULL;
 407         devvp = NULL;
 408         cache_index = -1;
 409
 410         kprintf("hammer2_mount\n");
 411
 412         if (path == NULL) {
 413                 /*
 414                  * Root mount
 415                  */
 416                 bzero(&info, sizeof(info));
 417                 info.cluster_fd = -1;
 418                 return (EOPNOTSUPP);
 419         } else {
 420                 /*
 421                  * Non-root mount or updating a mount
 422                  */
 423                 error = copyin(data, &info, sizeof(info));
 424                 if (error)
 425                         return (error);
 426
 427                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 428                 if (error)
 429                         return (error);
 430
 431                 /* Extract device and label */
 432                 dev = devstr;
 433                 label = strchr(devstr, '@');
 434                 if (label == NULL ||
 435                     ((label + 1) - dev) > done) {
 436                         return (EINVAL);
 437                 }
 438                 *label = '\0';
 439                 label++;
 440                 if (*label == '\0')
 441                         return (EINVAL);
 442
 443                 if (mp->mnt_flag & MNT_UPDATE) {
 444                         /* Update mount */
 445                         /* HAMMER2 implements NFS export via mountctl */
 446                         pmp = MPTOPMP(mp);
 447                         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
 448                                 hmp = pmp->iroot->cluster.array[i]->hmp;
 449                                 devvp = hmp->devvp;
 450                                 error = hammer2_remount(hmp, mp, path,
 451                                                         devvp, cred);
 452                                 if (error)
 453                                         break;
 454                         }
 455                         /*hammer2_inode_install_hidden(pmp);*/
 456
 457                         return error;
 458                 }
 459         }
 460
 461         /*
 462          * HMP device mount
 463          *
 464          * Lookup name and verify it refers to a block device.
 465          */
 466         error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 467         if (error == 0)
 468                 error = nlookup(&nd);
 469         if (error == 0)
 470                 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 471         nlookup_done(&nd);
 472
 473         if (error == 0) {
 474                 if (vn_isdisk(devvp, &error))
 475                         error = vfs_mountedon(devvp);
 476         }
 477
 478         /*
 479          * Determine if the device has already been mounted.  After this
 480          * check hmp will be non-NULL if we are doing the second or more
 481          * hammer2 mounts from the same device.
 482          */
 483         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 484         TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 485                 if (hmp->devvp == devvp)
 486                         break;
 487         }
 488
 489         /*
 490          * Open the device if this isn't a secondary mount and construct
 491          * the H2 device mount (hmp).
 492          */
 493         if (hmp == NULL) {
 494                 hammer2_chain_t *schain;
 495                 hammer2_xid_t xid;
 496
 497                 if (error == 0 && vcount(devvp) > 0)
 498                         error = EBUSY;
 499
 500                 /*
 501                  * Now open the device
 502                  */
 503                 if (error == 0) {
 504                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
 505                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
 506                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
 507                         if (error == 0) {
 508                                 error = VOP_OPEN(devvp,
 509                                                  ronly ? FREAD : FREAD | FWRITE,
 510                                                  FSCRED, NULL);
 511                         }
 512                         vn_unlock(devvp);
 513                 }
 514                 if (error && devvp) {
 515                         vrele(devvp);
 516                         devvp = NULL;
 517                 }
 518                 if (error) {
 519                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 520                         return error;
 521                 }
 522                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
 523                 hmp->ronly = ronly;
 524                 hmp->devvp = devvp;
 525                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
 526                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
 527                 RB_INIT(&hmp->iotree);
 528                 spin_init(&hmp->io_spin);
 529                 spin_init(&hmp->list_spin);
 530                 TAILQ_INIT(&hmp->flushq);
 531
 532                 lockinit(&hmp->vollk, "h2vol", 0, 0);
 533
 534                 /*
 535                  * vchain setup. vchain.data is embedded.
 536                  * vchain.refs is initialized and will never drop to 0.
 537                  *
 538                  * NOTE! voldata is not yet loaded.
 539                  */
 540                 hmp->vchain.hmp = hmp;
 541                 hmp->vchain.refs = 1;
 542                 hmp->vchain.data = (void *)&hmp->voldata;
 543                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
 544                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 545                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 546
 547                 hammer2_chain_core_alloc(NULL, &hmp->vchain);
 548                 /* hmp->vchain.u.xxx is left NULL */
 549
 550                 /*
 551                  * fchain setup.  fchain.data is embedded.
 552                  * fchain.refs is initialized and will never drop to 0.
 553                  *
 554                  * The data is not used but needs to be initialized to
 555                  * pass assertion muster.  We use this chain primarily
 556                  * as a placeholder for the freemap's top-level RBTREE
 557                  * so it does not interfere with the volume's topology
 558                  * RBTREE.
 559                  */
 560                 hmp->fchain.hmp = hmp;
 561                 hmp->fchain.refs = 1;
 562                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
 563                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
 564                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
 565                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 566                 hmp->fchain.bref.methods =
 567                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
 568                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
 569
 570                 hammer2_chain_core_alloc(NULL, &hmp->fchain);
 571                 /* hmp->fchain.u.xxx is left NULL */
 572
 573                 /*
 574                  * Install the volume header and initialize fields from
 575                  * voldata.
 576                  */
 577                 error = hammer2_install_volume_header(hmp);
 578                 if (error) {
 579                         ++hmp->pmp_count;
 580                         hammer2_vfs_unmount_hmp1(mp, hmp);
 581                         hammer2_vfs_unmount_hmp2(mp, hmp);
 582                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 583                         hammer2_vfs_unmount(mp, MNT_FORCE);
 584                         return error;
 585                 }
 586
 587                 /*
 588                  * Really important to get these right or flush will get
 589                  * confused.
 590                  */
 591                 hmp->spmp = hammer2_pfsalloc(NULL, hmp->voldata.mirror_tid);
 592                 kprintf("alloc spmp %p tid %016jx\n",
 593                         hmp->spmp, hmp->voldata.mirror_tid);
 594                 spmp = hmp->spmp;
 595                 spmp->inode_tid = 1;
 596
 597                 xid = 0;
 598                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
 599                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
 600                 hmp->vchain.pmp = spmp;
 601                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
 602                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
 603                 hmp->fchain.pmp = spmp;
 604
 605                 /*
 606                  * First locate the super-root inode, which is key 0
 607                  * relative to the volume header's blockset.
 608                  *
 609                  * Then locate the root inode by scanning the directory keyspace
 610                  * represented by the label.
 611                  */
 612                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
 613                 schain = hammer2_chain_lookup(&parent, &key_dummy,
 614                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
 615                                       &cache_index, 0, &ddflag);
 616                 hammer2_chain_lookup_done(parent);
 617                 if (schain == NULL) {
 618                         kprintf("hammer2_mount: invalid super-root\n");
 619                         ++hmp->pmp_count;
 620                         hammer2_vfs_unmount_hmp1(mp, hmp);
 621                         hammer2_vfs_unmount_hmp2(mp, hmp);
 622                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 623                         hammer2_vfs_unmount(mp, MNT_FORCE);
 624                         return EINVAL;
 625                 }
 626
 627                 /*
 628                  * Sanity-check schain's pmp, finish initializing spmp.
 629                  */
 630                 KKASSERT(schain->pmp == spmp);
 631                 spmp->pfs_clid = schain->data->ipdata.pfs_clid;
 632
 633                 /*
 634                  * NOTE: The CHAIN_PFSROOT is not set on the super-root inode.
 635                  * NOTE: inode_get sucks up schain's lock.
 636                  */
 637                 cluster = hammer2_cluster_from_chain(schain);
 638                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster);
 639                 spmp->spmp_hmp = hmp;
 640                 hammer2_inode_ref(spmp->iroot);
 641                 hammer2_inode_unlock_ex(spmp->iroot, cluster);
 642                 schain = NULL;
 643                 /* leave spmp->iroot with one ref */
 644
 645                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
 646                         error = hammer2_recovery(hmp);
 647                         /* XXX do something with error */
 648                 }
 649                 ++hmp->pmp_count;
 650
 651                 /*
 652                  * XXX RDONLY stuff is totally broken FIXME XXX
 653                  *
 654                  * Automatic LNK_CONN
 655                  * Automatic handling of received LNK_SPAN
 656                  * Automatic handling of received LNK_CIRC
 657                  * No automatic LNK_SPAN generation - we do this ourselves
 658                  * No automatic LNK_CIRC generation - we do this ourselves
 659                  */
 660                 kdmsg_iocom_init(&hmp->iocom, hmp,
 661                                  KDMSG_IOCOMF_AUTOCONN |
 662                                  KDMSG_IOCOMF_AUTORXSPAN,
 663                                  hmp->mchain, hammer2_rcvdmsg);
 664
 665                 /*
 666                  * Ref the cluster management messaging descriptor.  The mount
 667                  * program deals with the other end of the communications pipe.
 668                  */
 669                 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
 670                 if (fp) {
 671                         hammer2_cluster_reconnect(hmp, fp);
 672                 } else {
 673                         kprintf("hammer2_mount: bad cluster_fd!\n");
 674                 }
 675         } else {
 676                 spmp = hmp->spmp;
 677                 ++hmp->pmp_count;
 678         }
 679
 680         /*
 681          * Lookup mount point under the media-localized super-root.
 682          *
 683          * cluster->pmp will incorrectly point to spmp and must be fixed
 684          * up later on.
 685          */
 686         cparent = hammer2_inode_lock_ex(spmp->iroot);
 687         lhc = hammer2_dirhash(label, strlen(label));
 688         cluster = hammer2_cluster_lookup(cparent, &key_next,
 689                                       lhc, lhc + HAMMER2_DIRHASH_LOMASK,
 690                                       0, &ddflag);
 691         while (cluster) {
 692                 if (hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE &&
 693                     strcmp(label,
 694                        hammer2_cluster_data(cluster)->ipdata.filename) == 0) {
 695                         break;
 696                 }
 697                 cluster = hammer2_cluster_next(cparent, cluster, &key_next,
 698                                             key_next,
 699                                             lhc + HAMMER2_DIRHASH_LOMASK, 0);
 700         }
 701         hammer2_inode_unlock_ex(spmp->iroot, cparent);
 702
 703         if (cluster == NULL) {
 704                 kprintf("hammer2_mount: PFS label not found\n");
 705                 hammer2_vfs_unmount_hmp1(mp, hmp);
 706                 hammer2_vfs_unmount_hmp2(mp, hmp);
 707                 lockmgr(&hammer2_mntlk, LK_RELEASE);
 708                 hammer2_vfs_unmount(mp, MNT_FORCE);
 709                 return EINVAL;
 710         }
 711
 712         for (i = 0; i < cluster->nchains; ++i) {
 713                 rchain = cluster->array[i];
 714                 if (rchain->flags & HAMMER2_CHAIN_MOUNTED) {
 715                         kprintf("hammer2_mount: PFS label already mounted!\n");
 716                         hammer2_cluster_unlock(cluster);
 717                         hammer2_vfs_unmount_hmp1(mp, hmp);
 718                         hammer2_vfs_unmount_hmp2(mp, hmp);
 719                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 720                         hammer2_vfs_unmount(mp, MNT_FORCE);
 721                         return EBUSY;
 722                 }
 723                 KKASSERT(rchain->pmp == NULL);
 724 #if 0
 725                 if (rchain->flags & HAMMER2_CHAIN_RECYCLE) {
 726                         kprintf("hammer2_mount: PFS label is recycling\n");
 727                         hammer2_cluster_unlock(cluster);
 728                         hammer2_vfs_unmount_hmp1(mp, hmp);
 729                         hammer2_vfs_unmount_hmp2(mp, hmp);
 730                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 731                         hammer2_vfs_unmount(mp, MNT_FORCE);
 732                         return EBUSY;
 733                 }
 734 #endif
 735         }
 736
 737         /*
 738          * Check to see if the cluster id is already mounted at the mount
 739          * point.  If it is, add us to the cluster.
 740          */
 741         ipdata = &hammer2_cluster_data(cluster)->ipdata;
 742         hammer2_cluster_bref(cluster, &bref);
 743         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 744                 if (pmp->spmp_hmp == NULL &&
 745                     bcmp(&pmp->pfs_clid, &ipdata->pfs_clid,
 746                          sizeof(pmp->pfs_clid)) == 0) {
 747                         break;
 748                 }
 749         }
 750
 751         if (pmp) {
 752                 int i;
 753                 int j;
 754
 755                 hammer2_inode_ref(pmp->iroot);
 756                 ccms_thread_lock(&pmp->iroot->topo_cst, CCMS_STATE_EXCLUSIVE);
 757
 758                 if (pmp->iroot->cluster.nchains + cluster->nchains >
 759                     HAMMER2_MAXCLUSTER) {
 760                         kprintf("hammer2_mount: cluster full!\n");
 761
 762                         ccms_thread_unlock(&pmp->iroot->topo_cst);
 763                         hammer2_inode_drop(pmp->iroot);
 764
 765                         hammer2_cluster_unlock(cluster);
 766                         hammer2_vfs_unmount_hmp1(mp, hmp);
 767                         hammer2_vfs_unmount_hmp2(mp, hmp);
 768                         lockmgr(&hammer2_mntlk, LK_RELEASE);
 769                         hammer2_vfs_unmount(mp, MNT_FORCE);
 770                         return EBUSY;
 771                 }
 772                 kprintf("hammer2_vfs_mount: Adding pfs to existing cluster\n");
 773                 j = pmp->iroot->cluster.nchains;
 774                 for (i = 0; i < cluster->nchains; ++i) {
 775                         rchain = cluster->array[i];
 776                         KKASSERT(rchain->pmp == NULL);
 777                         rchain->pmp = pmp;
 778                         hammer2_chain_ref(cluster->array[i]);
 779                         pmp->iroot->cluster.array[j] = cluster->array[i];
 780                         ++j;
 781                 }
 782                 pmp->iroot->cluster.nchains = j;
 783                 ccms_thread_unlock(&pmp->iroot->topo_cst);
 784                 hammer2_inode_drop(pmp->iroot);
 785                 hammer2_cluster_unlock(cluster);
 786                 lockmgr(&hammer2_mntlk, LK_RELEASE);
 787
 788                 kprintf("ok\n");
 789                 hammer2_inode_install_hidden(pmp);
 790
 791                 return ERANGE;
 792         }
 793
 794         /*
 795          * Block device opened successfully, finish initializing the
 796          * mount structure.
 797          *
 798          * From this point on we have to call hammer2_unmount() on failure.
 799          */
 800         pmp = hammer2_pfsalloc(ipdata, bref.mirror_tid);
 801         kprintf("PMP mirror_tid is %016jx\n", bref.mirror_tid);
 802         for (i = 0; i < cluster->nchains; ++i) {
 803                 rchain = cluster->array[i];
 804                 KKASSERT(rchain->pmp == NULL);
 805                 rchain->pmp = pmp;
 806                 atomic_set_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
 807         }
 808         cluster->pmp = pmp;
 809
 810         ccms_domain_init(&pmp->ccms_dom);
 811         TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 812         lockmgr(&hammer2_mntlk, LK_RELEASE);
 813
 814         kprintf("hammer2_mount hmp=%p pmp=%p pmpcnt=%d\n",
 815                 hmp, pmp, hmp->pmp_count);
 816
 817         mp->mnt_flag = MNT_LOCAL;
 818         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
 819         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
 820
 821         /*
 822          * required mount structure initializations
 823          */
 824         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
 825         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
 826
 827         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
 828         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
 829
 830         /*
 831          * Optional fields
 832          */
 833         mp->mnt_iosize_max = MAXPHYS;
 834         mp->mnt_data = (qaddr_t)pmp;
 835         pmp->mp = mp;
 836
 837         /*
 838          * After this point hammer2_vfs_unmount() has visibility on hmp
 839          * and manual hmp1/hmp2 calls are not needed on fatal errors.
 840          */
 841         pmp->iroot = hammer2_inode_get(pmp, NULL, cluster);
 842         hammer2_inode_ref(pmp->iroot);          /* ref for pmp->iroot */
 843         hammer2_inode_unlock_ex(pmp->iroot, cluster);
 844
 845         /*
 846          * The logical file buffer bio write thread handles things
 847          * like physical block assignment and compression.
 848          *
 849          * (only applicable to pfs mounts, not applicable to spmp)
 850          */
 851         pmp->wthread_destroy = 0;
 852         lwkt_create(hammer2_write_thread, pmp,
 853                     &pmp->wthread_td, NULL, 0, -1, "hwrite-%s", label);
 854
 855         /*
 856          * With the cluster operational install ihidden.
 857          * (only applicable to pfs mounts, not applicable to spmp)
 858          */
 859         hammer2_inode_install_hidden(pmp);
 860
 861         /*
 862          * Finish setup
 863          */
 864         vfs_getnewfsid(mp);
 865         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
 866         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
 867         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
 868
 869         copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
 870         bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
 871         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
 872         copyinstr(path, mp->mnt_stat.f_mntonname,
 873                   sizeof(mp->mnt_stat.f_mntonname) - 1,
 874                   &size);
 875
 876         /*
 877          * Initial statfs to prime mnt_stat.
 878          */
 879         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
 880
 881         return 0;
 882 }
 883
 884 /*
 885  * Handle bioq for strategy write
 886  */
 887 static
 888 void
 889 hammer2_write_thread(void *arg)
 890 {
 891         hammer2_pfsmount_t *pmp;
 892         struct bio *bio;
 893         struct buf *bp;
 894         hammer2_trans_t trans;
 895         struct vnode *vp;
 896         hammer2_inode_t *ip;
 897         hammer2_cluster_t *cparent;
 898         hammer2_inode_data_t *wipdata;
 899         hammer2_key_t lbase;
 900         int lblksize;
 901         int pblksize;
 902         int error;
 903
 904         pmp = arg;
 905
 906         mtx_lock(&pmp->wthread_mtx);
 907         while (pmp->wthread_destroy == 0) {
 908                 if (bioq_first(&pmp->wthread_bioq) == NULL) {
 909                         mtxsleep(&pmp->wthread_bioq, &pmp->wthread_mtx,
 910                                  0, "h2bioqw", 0);
 911                 }
 912                 cparent = NULL;
 913
 914                 hammer2_trans_init(&trans, pmp, HAMMER2_TRANS_BUFCACHE);
 915
 916                 while ((bio = bioq_takefirst(&pmp->wthread_bioq)) != NULL) {
 917                         /*
 918                          * dummy bio for synchronization.  The transaction
 919                          * must be reinitialized.
 920                          */
 921                         if (bio->bio_buf == NULL) {
 922                                 bio->bio_flags |= BIO_DONE;
 923                                 wakeup(bio);
 924                                 hammer2_trans_done(&trans);
 925                                 hammer2_trans_init(&trans, pmp,
 926                                                    HAMMER2_TRANS_BUFCACHE);
 927                                 continue;
 928                         }
 929
 930                         /*
 931                          * else normal bio processing
 932                          */
 933                         mtx_unlock(&pmp->wthread_mtx);
 934
 935                         hammer2_lwinprog_drop(pmp);
 936
 937                         error = 0;
 938                         bp = bio->bio_buf;
 939                         vp = bp->b_vp;
 940                         ip = VTOI(vp);
 941
 942                         /*
 943                          * Inode is modified, flush size and mtime changes
 944                          * to ensure that the file size remains consistent
 945                          * with the buffers being flushed.
 946                          *
 947                          * NOTE: The inode_fsync() call only flushes the
 948                          *       inode's meta-data state, it doesn't try
 949                          *       to flush underlying buffers or chains.
 950                          */
 951                         cparent = hammer2_inode_lock_ex(ip);
 952                         if (ip->flags & (HAMMER2_INODE_RESIZED |
 953                                          HAMMER2_INODE_MTIME)) {
 954                                 hammer2_inode_fsync(&trans, ip, cparent);
 955                         }
 956                         wipdata = hammer2_cluster_modify_ip(&trans, ip,
 957                                                          cparent, 0);
 958                         lblksize = hammer2_calc_logical(ip, bio->bio_offset,
 959                                                         &lbase, NULL);
 960                         pblksize = hammer2_calc_physical(ip, wipdata, lbase);
 961                         hammer2_write_file_core(bp, &trans, ip, wipdata,
 962                                                 cparent,
 963                                                 lbase, IO_ASYNC,
 964                                                 pblksize, &error);
 965                         hammer2_cluster_modsync(cparent);
 966                         hammer2_inode_unlock_ex(ip, cparent);
 967                         if (error) {
 968                                 kprintf("hammer2: error in buffer write\n");
 969                                 bp->b_flags |= B_ERROR;
 970                                 bp->b_error = EIO;
 971                         }
 972                         biodone(bio);
 973                         mtx_lock(&pmp->wthread_mtx);
 974                 }
 975                 hammer2_trans_done(&trans);
 976         }
 977         pmp->wthread_destroy = -1;
 978         wakeup(&pmp->wthread_destroy);
 979
 980         mtx_unlock(&pmp->wthread_mtx);
 981 }
 982
 983 void
 984 hammer2_bioq_sync(hammer2_pfsmount_t *pmp)
 985 {
 986         struct bio sync_bio;
 987
 988         bzero(&sync_bio, sizeof(sync_bio));     /* dummy with no bio_buf */
 989         mtx_lock(&pmp->wthread_mtx);
 990         if (pmp->wthread_destroy == 0 &&
 991             TAILQ_FIRST(&pmp->wthread_bioq.queue)) {
 992                 bioq_insert_tail(&pmp->wthread_bioq, &sync_bio);
 993                 while ((sync_bio.bio_flags & BIO_DONE) == 0)
 994                         mtxsleep(&sync_bio, &pmp->wthread_mtx, 0, "h2bioq", 0);
 995         }
 996         mtx_unlock(&pmp->wthread_mtx);
 997 }
 998
 999 /*
1000  * Return a chain suitable for I/O, creating the chain if necessary
1001  * and assigning its physical block.
1002  */
1003 static
1004 hammer2_cluster_t *
1005 hammer2_assign_physical(hammer2_trans_t *trans,
1006                         hammer2_inode_t *ip, hammer2_cluster_t *cparent,
1007                         hammer2_key_t lbase, int pblksize, int *errorp)
1008 {
1009         hammer2_cluster_t *cluster;
1010         hammer2_cluster_t *dparent;
1011         hammer2_key_t key_dummy;
1012         int pradix = hammer2_getradix(pblksize);
1013         int ddflag;
1014
1015         /*
1016          * Locate the chain associated with lbase, return a locked chain.
1017          * However, do not instantiate any data reference (which utilizes a
1018          * device buffer) because we will be using direct IO via the
1019          * logical buffer cache buffer.
1020          */
1021         *errorp = 0;
1022         KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
1023 retry:
1024         dparent = hammer2_cluster_lookup_init(cparent, 0);
1025         cluster = hammer2_cluster_lookup(dparent, &key_dummy,
1026                                      lbase, lbase,
1027                                      HAMMER2_LOOKUP_NODATA, &ddflag);
1028
1029         if (cluster == NULL) {
1030                 /*
1031                  * We found a hole, create a new chain entry.
1032                  *
1033                  * NOTE: DATA chains are created without device backing
1034                  *       store (nor do we want any).
1035                  */
1036                 *errorp = hammer2_cluster_create(trans, dparent, &cluster,
1037                                                lbase, HAMMER2_PBUFRADIX,
1038                                                HAMMER2_BREF_TYPE_DATA,
1039                                                pblksize);
1040                 if (cluster == NULL) {
1041                         hammer2_cluster_lookup_done(dparent);
1042                         panic("hammer2_cluster_create: par=%p error=%d\n",
1043                                 dparent->focus, *errorp);
1044                         goto retry;
1045                 }
1046                 /*ip->delta_dcount += pblksize;*/
1047         } else {
1048                 switch (hammer2_cluster_type(cluster)) {
1049                 case HAMMER2_BREF_TYPE_INODE:
1050                         /*
1051                          * The data is embedded in the inode.  The
1052                          * caller is responsible for marking the inode
1053                          * modified and copying the data to the embedded
1054                          * area.
1055                          */
1056                         break;
1057                 case HAMMER2_BREF_TYPE_DATA:
1058                         if (hammer2_cluster_bytes(cluster) != pblksize) {
1059                                 hammer2_cluster_resize(trans, ip,
1060                                                      dparent, cluster,
1061                                                      pradix,
1062                                                      HAMMER2_MODIFY_OPTDATA);
1063                         }
1064                         hammer2_cluster_modify(trans, cluster,
1065                                              HAMMER2_MODIFY_OPTDATA);
1066                         break;
1067                 default:
1068                         panic("hammer2_assign_physical: bad type");
1069                         /* NOT REACHED */
1070                         break;
1071                 }
1072         }
1073
1074         /*
1075          * Cleanup.  If cluster wound up being the inode itself, i.e.
1076          * the DIRECTDATA case for offset 0, then we need to update cparent.
1077          * The caller expects cparent to not become stale.
1078          */
1079         hammer2_cluster_lookup_done(dparent);
1080         /* dparent = NULL; safety */
1081         if (cluster && ddflag)
1082                 hammer2_cluster_replace_locked(cparent, cluster);
1083         return (cluster);
1084 }
1085
1086 /*
1087  * From hammer2_vnops.c.
1088  * The core write function which determines which path to take
1089  * depending on compression settings.
1090  */
1091 static
1092 void
1093 hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
1094                         hammer2_inode_t *ip, hammer2_inode_data_t *ipdata,
1095                         hammer2_cluster_t *cparent,
1096                         hammer2_key_t lbase, int ioflag, int pblksize,
1097                         int *errorp)
1098 {
1099         hammer2_cluster_t *cluster;
1100
1101         switch(HAMMER2_DEC_COMP(ipdata->comp_algo)) {
1102         case HAMMER2_COMP_NONE:
1103                 /*
1104                  * We have to assign physical storage to the buffer
1105                  * we intend to dirty or write now to avoid deadlocks
1106                  * in the strategy code later.
1107                  *
1108                  * This can return NOOFFSET for inode-embedded data.
1109                  * The strategy code will take care of it in that case.
1110                  */
1111                 cluster = hammer2_assign_physical(trans, ip, cparent,
1112                                                 lbase, pblksize,
1113                                                 errorp);
1114                 hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp);
1115                 if (cluster)
1116                         hammer2_cluster_unlock(cluster);
1117                 break;
1118         case HAMMER2_COMP_AUTOZERO:
1119                 /*
1120                  * Check for zero-fill only
1121                  */
1122                 hammer2_zero_check_and_write(bp, trans, ip,
1123                                     ipdata, cparent, lbase,
1124                                     ioflag, pblksize, errorp);
1125                 break;
1126         case HAMMER2_COMP_LZ4:
1127         case HAMMER2_COMP_ZLIB:
1128         default:
1129                 /*
1130                  * Check for zero-fill and attempt compression.
1131                  */
1132                 hammer2_compress_and_write(bp, trans, ip,
1133                                            ipdata, cparent,
1134                                            lbase, ioflag,
1135                                            pblksize, errorp,
1136                                            ipdata->comp_algo);
1137                 break;
1138         }
1139 }
1140
1141 /*
1142  * Generic function that will perform the compression in compression
1143  * write path. The compression algorithm is determined by the settings
1144  * obtained from inode.
1145  */
1146 static
1147 void
1148 hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
1149         hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1150         hammer2_cluster_t *cparent,
1151         hammer2_key_t lbase, int ioflag, int pblksize,
1152         int *errorp, int comp_algo)
1153 {
1154         hammer2_cluster_t *cluster;
1155         hammer2_chain_t *chain;
1156         int comp_size;
1157         int comp_block_size;
1158         int i;
1159         char *comp_buffer;
1160
1161         if (test_block_zeros(bp->b_data, pblksize)) {
1162                 zero_write(bp, trans, ip, ipdata, cparent, lbase, errorp);
1163                 return;
1164         }
1165
1166         comp_size = 0;
1167         comp_buffer = NULL;
1168
1169         KKASSERT(pblksize / 2 <= 32768);
1170
1171         if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
1172                 z_stream strm_compress;
1173                 int comp_level;
1174                 int ret;
1175
1176                 switch(HAMMER2_DEC_COMP(comp_algo)) {
1177                 case HAMMER2_COMP_LZ4:
1178                         comp_buffer = objcache_get(cache_buffer_write,
1179                                                    M_INTWAIT);
1180                         comp_size = LZ4_compress_limitedOutput(
1181                                         bp->b_data,
1182                                         &comp_buffer[sizeof(int)],
1183                                         pblksize,
1184                                         pblksize / 2 - sizeof(int));
1185                         /*
1186                          * We need to prefix with the size, LZ4
1187                          * doesn't do it for us.  Add the related
1188                          * overhead.
1189                          */
1190                         *(int *)comp_buffer = comp_size;
1191                         if (comp_size)
1192                                 comp_size += sizeof(int);
1193                         break;
1194                 case HAMMER2_COMP_ZLIB:
1195                         comp_level = HAMMER2_DEC_LEVEL(comp_algo);
1196                         if (comp_level == 0)
1197                                 comp_level = 6; /* default zlib compression */
1198                         else if (comp_level < 6)
1199                                 comp_level = 6;
1200                         else if (comp_level > 9)
1201                                 comp_level = 9;
1202                         ret = deflateInit(&strm_compress, comp_level);
1203                         if (ret != Z_OK) {
1204                                 kprintf("HAMMER2 ZLIB: fatal error "
1205                                         "on deflateInit.\n");
1206                         }
1207
1208                         comp_buffer = objcache_get(cache_buffer_write,
1209                                                    M_INTWAIT);
1210                         strm_compress.next_in = bp->b_data;
1211                         strm_compress.avail_in = pblksize;
1212                         strm_compress.next_out = comp_buffer;
1213                         strm_compress.avail_out = pblksize / 2;
1214                         ret = deflate(&strm_compress, Z_FINISH);
1215                         if (ret == Z_STREAM_END) {
1216                                 comp_size = pblksize / 2 -
1217                                             strm_compress.avail_out;
1218                         } else {
1219                                 comp_size = 0;
1220                         }
1221                         ret = deflateEnd(&strm_compress);
1222                         break;
1223                 default:
1224                         kprintf("Error: Unknown compression method.\n");
1225                         kprintf("Comp_method = %d.\n", comp_algo);
1226                         break;
1227                 }
1228         }
1229
1230         if (comp_size == 0) {
1231                 /*
1232                  * compression failed or turned off
1233                  */
1234                 comp_block_size = pblksize;     /* safety */
1235                 if (++ip->comp_heuristic > 128)
1236                         ip->comp_heuristic = 8;
1237         } else {
1238                 /*
1239                  * compression succeeded
1240                  */
1241                 ip->comp_heuristic = 0;
1242                 if (comp_size <= 1024) {
1243                         comp_block_size = 1024;
1244                 } else if (comp_size <= 2048) {
1245                         comp_block_size = 2048;
1246                 } else if (comp_size <= 4096) {
1247                         comp_block_size = 4096;
1248                 } else if (comp_size <= 8192) {
1249                         comp_block_size = 8192;
1250                 } else if (comp_size <= 16384) {
1251                         comp_block_size = 16384;
1252                 } else if (comp_size <= 32768) {
1253                         comp_block_size = 32768;
1254                 } else {
1255                         panic("hammer2: WRITE PATH: "
1256                               "Weird comp_size value.");
1257                         /* NOT REACHED */
1258                         comp_block_size = pblksize;
1259                 }
1260         }
1261
1262         cluster = hammer2_assign_physical(trans, ip, cparent,
1263                                           lbase, comp_block_size,
1264                                           errorp);
1265         ipdata = &hammer2_cluster_data(cparent)->ipdata;
1266
1267         if (*errorp) {
1268                 kprintf("WRITE PATH: An error occurred while "
1269                         "assigning physical space.\n");
1270                 KKASSERT(cluster == NULL);
1271                 goto done;
1272         }
1273
1274         for (i = 0; i < cluster->nchains; ++i) {
1275                 hammer2_io_t *dio;
1276                 char *bdata;
1277                 int temp_check;
1278
1279                 chain = cluster->array[i];
1280                 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1281
1282                 switch(chain->bref.type) {
1283                 case HAMMER2_BREF_TYPE_INODE:
1284                         KKASSERT(chain->data->ipdata.op_flags &
1285                                  HAMMER2_OPFLAG_DIRECTDATA);
1286                         KKASSERT(bp->b_loffset == 0);
1287                         bcopy(bp->b_data, chain->data->ipdata.u.data,
1288                               HAMMER2_EMBEDDED_BYTES);
1289                         break;
1290                 case HAMMER2_BREF_TYPE_DATA:
1291                         temp_check = HAMMER2_DEC_CHECK(chain->bref.methods);
1292
1293                         /*
1294                          * Optimize out the read-before-write
1295                          * if possible.
1296                          */
1297                         *errorp = hammer2_io_newnz(chain->hmp,
1298                                                    chain->bref.data_off,
1299                                                    chain->bytes,
1300                                                    &dio);
1301                         if (*errorp) {
1302                                 hammer2_io_brelse(&dio);
1303                                 kprintf("hammer2: WRITE PATH: "
1304                                         "dbp bread error\n");
1305                                 break;
1306                         }
1307                         bdata = hammer2_io_data(dio, chain->bref.data_off);
1308
1309                         /*
1310                          * When loading the block make sure we don't
1311                          * leave garbage after the compressed data.
1312                          */
1313                         if (comp_size) {
1314                                 chain->bref.methods =
1315                                         HAMMER2_ENC_COMP(comp_algo) +
1316                                         HAMMER2_ENC_CHECK(temp_check);
1317                                 bcopy(comp_buffer, bdata, comp_size);
1318                                 if (comp_size != comp_block_size) {
1319                                         bzero(bdata + comp_size,
1320                                               comp_block_size - comp_size);
1321                                 }
1322                         } else {
1323                                 chain->bref.methods =
1324                                         HAMMER2_ENC_COMP(
1325                                                 HAMMER2_COMP_NONE) +
1326                                         HAMMER2_ENC_CHECK(temp_check);
1327                                 bcopy(bp->b_data, bdata, pblksize);
1328                         }
1329
1330                         /*
1331                          * Device buffer is now valid, chain is no longer in
1332                          * the initial state.
1333                          *
1334                          * (No blockref table worries with file data)
1335                          */
1336                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1337
1338                         /* Now write the related bdp. */
1339                         if (ioflag & IO_SYNC) {
1340                                 /*
1341                                  * Synchronous I/O requested.
1342                                  */
1343                                 hammer2_io_bwrite(&dio);
1344                         /*
1345                         } else if ((ioflag & IO_DIRECT) &&
1346                                    loff + n == pblksize) {
1347                                 hammer2_io_bdwrite(&dio);
1348                         */
1349                         } else if (ioflag & IO_ASYNC) {
1350                                 hammer2_io_bawrite(&dio);
1351                         } else {
1352                                 hammer2_io_bdwrite(&dio);
1353                         }
1354                         break;
1355                 default:
1356                         panic("hammer2_write_bp: bad chain type %d\n",
1357                                 chain->bref.type);
1358                         /* NOT REACHED */
1359                         break;
1360                 }
1361         }
1362 done:
1363         if (cluster)
1364                 hammer2_cluster_unlock(cluster);
1365         if (comp_buffer)
1366                 objcache_put(cache_buffer_write, comp_buffer);
1367 }
1368
1369 /*
1370  * Function that performs zero-checking and writing without compression,
1371  * it corresponds to default zero-checking path.
1372  */
1373 static
1374 void
1375 hammer2_zero_check_and_write(struct buf *bp, hammer2_trans_t *trans,
1376         hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1377         hammer2_cluster_t *cparent,
1378         hammer2_key_t lbase, int ioflag, int pblksize, int *errorp)
1379 {
1380         hammer2_cluster_t *cluster;
1381
1382         if (test_block_zeros(bp->b_data, pblksize)) {
1383                 zero_write(bp, trans, ip, ipdata, cparent, lbase, errorp);
1384         } else {
1385                 cluster = hammer2_assign_physical(trans, ip, cparent,
1386                                                   lbase, pblksize, errorp);
1387                 hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp);
1388                 if (cluster)
1389                         hammer2_cluster_unlock(cluster);
1390         }
1391 }
1392
1393 /*
1394  * A function to test whether a block of data contains only zeros,
1395  * returns TRUE (non-zero) if the block is all zeros.
1396  */
1397 static
1398 int
1399 test_block_zeros(const char *buf, size_t bytes)
1400 {
1401         size_t i;
1402
1403         for (i = 0; i < bytes; i += sizeof(long)) {
1404                 if (*(const long *)(buf + i) != 0)
1405                         return (0);
1406         }
1407         return (1);
1408 }
1409
1410 /*
1411  * Function to "write" a block that contains only zeros.
1412  */
1413 static
1414 void
1415 zero_write(struct buf *bp, hammer2_trans_t *trans,
1416            hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1417            hammer2_cluster_t *cparent,
1418            hammer2_key_t lbase, int *errorp __unused)
1419 {
1420         hammer2_cluster_t *cluster;
1421         hammer2_media_data_t *data;
1422         hammer2_key_t key_dummy;
1423         int ddflag;
1424
1425         cparent = hammer2_cluster_lookup_init(cparent, 0);
1426         cluster = hammer2_cluster_lookup(cparent, &key_dummy, lbase, lbase,
1427                                      HAMMER2_LOOKUP_NODATA, &ddflag);
1428         if (cluster) {
1429                 data = hammer2_cluster_wdata(cluster);
1430
1431                 if (ddflag) {
1432                         KKASSERT(cluster->focus->flags &
1433                                  HAMMER2_CHAIN_MODIFIED);
1434                         bzero(data->ipdata.u.data, HAMMER2_EMBEDDED_BYTES);
1435                         hammer2_cluster_modsync(cluster);
1436                 } else {
1437                         hammer2_cluster_delete(trans, cparent, cluster,
1438                                                HAMMER2_DELETE_PERMANENT);
1439                 }
1440                 hammer2_cluster_unlock(cluster);
1441         }
1442         hammer2_cluster_lookup_done(cparent);
1443 }
1444
1445 /*
1446  * Function to write the data as it is, without performing any sort of
1447  * compression. This function is used in path without compression and
1448  * default zero-checking path.
1449  */
1450 static
1451 void
1452 hammer2_write_bp(hammer2_cluster_t *cluster, struct buf *bp, int ioflag,
1453                                 int pblksize, int *errorp)
1454 {
1455         hammer2_chain_t *chain;
1456         hammer2_io_t *dio;
1457         char *bdata;
1458         int error;
1459         int i;
1460         int temp_check;
1461
1462         error = 0;      /* XXX TODO below */
1463
1464         for (i = 0; i < cluster->nchains; ++i) {
1465                 chain = cluster->array[i];
1466
1467                 temp_check = HAMMER2_DEC_CHECK(chain->bref.methods);
1468
1469                 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1470
1471                 switch(chain->bref.type) {
1472                 case HAMMER2_BREF_TYPE_INODE:
1473                         KKASSERT(chain->data->ipdata.op_flags &
1474                                  HAMMER2_OPFLAG_DIRECTDATA);
1475                         KKASSERT(bp->b_loffset == 0);
1476                         bcopy(bp->b_data, chain->data->ipdata.u.data,
1477                               HAMMER2_EMBEDDED_BYTES);
1478                         error = 0;
1479                         break;
1480                 case HAMMER2_BREF_TYPE_DATA:
1481                         error = hammer2_io_newnz(chain->hmp,
1482                                                  chain->bref.data_off,
1483                                                  chain->bytes, &dio);
1484                         if (error) {
1485                                 hammer2_io_bqrelse(&dio);
1486                                 kprintf("hammer2: WRITE PATH: "
1487                                         "dbp bread error\n");
1488                                 break;
1489                         }
1490                         bdata = hammer2_io_data(dio, chain->bref.data_off);
1491
1492                         chain->bref.methods = HAMMER2_ENC_COMP(
1493                                                         HAMMER2_COMP_NONE) +
1494                                               HAMMER2_ENC_CHECK(temp_check);
1495                         bcopy(bp->b_data, bdata, chain->bytes);
1496
1497                         /*
1498                          * Device buffer is now valid, chain is no longer in
1499                          * the initial state.
1500                          *
1501                          * (No blockref table worries with file data)
1502                          */
1503                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1504
1505                         if (ioflag & IO_SYNC) {
1506                                 /*
1507                                  * Synchronous I/O requested.
1508                                  */
1509                                 hammer2_io_bwrite(&dio);
1510                         /*
1511                         } else if ((ioflag & IO_DIRECT) &&
1512                                    loff + n == pblksize) {
1513                                 hammer2_io_bdwrite(&dio);
1514                         */
1515                         } else if (ioflag & IO_ASYNC) {
1516                                 hammer2_io_bawrite(&dio);
1517                         } else {
1518                                 hammer2_io_bdwrite(&dio);
1519                         }
1520                         break;
1521                 default:
1522                         panic("hammer2_write_bp: bad chain type %d\n",
1523                               chain->bref.type);
1524                         /* NOT REACHED */
1525                         error = 0;
1526                         break;
1527                 }
1528                 KKASSERT(error == 0);   /* XXX TODO */
1529         }
1530         *errorp = error;
1531 }
1532
1533 static
1534 int
1535 hammer2_remount(hammer2_mount_t *hmp, struct mount *mp, char *path,
1536                 struct vnode *devvp, struct ucred *cred)
1537 {
1538         int error;
1539
1540         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1541                 error = hammer2_recovery(hmp);
1542         } else {
1543                 error = 0;
1544         }
1545         return error;
1546 }
1547
1548 static
1549 int
1550 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1551 {
1552         hammer2_pfsmount_t *pmp;
1553         hammer2_mount_t *hmp;
1554         hammer2_chain_t *rchain;
1555         hammer2_cluster_t *cluster;
1556         int flags;
1557         int error = 0;
1558         int i;
1559
1560         pmp = MPTOPMP(mp);
1561
1562         if (pmp == NULL)
1563                 return(0);
1564
1565         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1566         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
1567
1568         /*
1569          * If mount initialization proceeded far enough we must flush
1570          * its vnodes.
1571          */
1572         if (mntflags & MNT_FORCE)
1573                 flags = FORCECLOSE;
1574         else
1575                 flags = 0;
1576         if (pmp->iroot) {
1577                 error = vflush(mp, 0, flags);
1578                 if (error)
1579                         goto failed;
1580         }
1581
1582         ccms_domain_uninit(&pmp->ccms_dom);
1583
1584         if (pmp->wthread_td) {
1585                 mtx_lock(&pmp->wthread_mtx);
1586                 pmp->wthread_destroy = 1;
1587                 wakeup(&pmp->wthread_bioq);
1588                 while (pmp->wthread_destroy != -1) {
1589                         mtxsleep(&pmp->wthread_destroy,
1590                                 &pmp->wthread_mtx, 0,
1591                                 "umount-sleep", 0);
1592                 }
1593                 mtx_unlock(&pmp->wthread_mtx);
1594                 pmp->wthread_td = NULL;
1595         }
1596
1597         /*
1598          * Cleanup our reference on ihidden.
1599          */
1600         if (pmp->ihidden) {
1601                 hammer2_inode_drop(pmp->ihidden);
1602                 pmp->ihidden = NULL;
1603         }
1604
1605         /*
1606          * Cleanup our reference on iroot.  iroot is (should) not be needed
1607          * by the flush code.
1608          */
1609         if (pmp->iroot) {
1610                 cluster = &pmp->iroot->cluster;
1611                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1612                         rchain = pmp->iroot->cluster.array[i];
1613                         if (rchain == NULL)
1614                                 continue;
1615                         hmp = rchain->hmp;
1616                         hammer2_vfs_unmount_hmp1(mp, hmp);
1617
1618                         atomic_clear_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
1619 #if REPORT_REFS_ERRORS
1620                         if (rchain->refs != 1)
1621                                 kprintf("PMP->RCHAIN %p REFS WRONG %d\n",
1622                                         rchain, rchain->refs);
1623 #else
1624                         KKASSERT(rchain->refs == 1);
1625 #endif
1626                         hammer2_chain_drop(rchain);
1627                         cluster->array[i] = NULL;
1628                         hammer2_vfs_unmount_hmp2(mp, hmp);
1629                 }
1630                 cluster->focus = NULL;
1631
1632 #if REPORT_REFS_ERRORS
1633                 if (pmp->iroot->refs != 1)
1634                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
1635                                 pmp->iroot, pmp->iroot->refs);
1636 #else
1637                 KKASSERT(pmp->iroot->refs == 1);
1638 #endif
1639                 /* ref for pmp->iroot */
1640                 hammer2_inode_drop(pmp->iroot);
1641                 pmp->iroot = NULL;
1642         }
1643
1644         pmp->mp = NULL;
1645         mp->mnt_data = NULL;
1646
1647         kmalloc_destroy(&pmp->mmsg);
1648         kmalloc_destroy(&pmp->minode);
1649
1650         kfree(pmp, M_HAMMER2);
1651         error = 0;
1652
1653 failed:
1654         lockmgr(&hammer2_mntlk, LK_RELEASE);
1655
1656         return (error);
1657 }
1658
1659 static
1660 void
1661 hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp)
1662 {
1663         hammer2_mount_exlock(hmp);
1664         --hmp->pmp_count;
1665
1666         kprintf("hammer2_unmount hmp=%p pmpcnt=%d\n", hmp, hmp->pmp_count);
1667
1668         kdmsg_iocom_uninit(&hmp->iocom);        /* XXX chain depend deadlck? */
1669
1670         /*
1671          * Cycle the volume data lock as a safety (probably not needed any
1672          * more).  To ensure everything is out we need to flush at least
1673          * three times.  (1) The running of the unlinkq can dirty the
1674          * filesystem, (2) A normal flush can dirty the freemap, and
1675          * (3) ensure that the freemap is fully synchronized.
1676          *
1677          * The next mount's recovery scan can clean everything up but we want
1678          * to leave the filesystem in a 100% clean state on a normal unmount.
1679          */
1680         hammer2_voldata_lock(hmp);
1681         hammer2_voldata_unlock(hmp);
1682         hammer2_vfs_sync(mp, MNT_WAIT);
1683         hammer2_vfs_sync(mp, MNT_WAIT);
1684         hammer2_vfs_sync(mp, MNT_WAIT);
1685
1686         if (hmp->pmp_count == 0) {
1687                 if ((hmp->vchain.flags | hmp->fchain.flags) &
1688                     HAMMER2_CHAIN_FLUSH_MASK) {
1689                         kprintf("hammer2_unmount: chains left over "
1690                                 "after final sync\n");
1691                         kprintf("    vchain %08x\n", hmp->vchain.flags);
1692                         kprintf("    fchain %08x\n", hmp->fchain.flags);
1693
1694                         if (hammer2_debug & 0x0010)
1695                                 Debugger("entered debugger");
1696                 }
1697         }
1698 }
1699
1700 static
1701 void
1702 hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp)
1703 {
1704         hammer2_pfsmount_t *spmp;
1705         struct vnode *devvp;
1706         int dumpcnt;
1707         int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1708
1709         /*
1710          * If no PFS's left drop the master hammer2_mount for the
1711          * device.
1712          */
1713         if (hmp->pmp_count == 0) {
1714                 /*
1715                  * Clean up SPMP and the super-root inode
1716                  */
1717                 spmp = hmp->spmp;
1718                 if (spmp) {
1719                         if (spmp->iroot) {
1720                                 hammer2_inode_drop(spmp->iroot);
1721                                 spmp->iroot = NULL;
1722                         }
1723                         hmp->spmp = NULL;
1724                         kmalloc_destroy(&spmp->mmsg);
1725                         kmalloc_destroy(&spmp->minode);
1726                         kfree(spmp, M_HAMMER2);
1727                 }
1728
1729                 /*
1730                  * Finish up with the device vnode
1731                  */
1732                 if ((devvp = hmp->devvp) != NULL) {
1733                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1734                         vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1735                         hmp->devvp = NULL;
1736                         VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1737                         vn_unlock(devvp);
1738                         vrele(devvp);
1739                         devvp = NULL;
1740                 }
1741
1742                 /*
1743                  * Clear vchain/fchain flags that might prevent final cleanup
1744                  * of these chains.
1745                  */
1746                 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1747                         atomic_clear_int(&hmp->vchain.flags,
1748                                          HAMMER2_CHAIN_MODIFIED);
1749                         hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1750                         hammer2_chain_drop(&hmp->vchain);
1751                 }
1752                 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1753                         atomic_clear_int(&hmp->vchain.flags,
1754                                          HAMMER2_CHAIN_UPDATE);
1755                         hammer2_chain_drop(&hmp->vchain);
1756                 }
1757
1758                 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1759                         atomic_clear_int(&hmp->fchain.flags,
1760                                          HAMMER2_CHAIN_MODIFIED);
1761                         hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1762                         hammer2_chain_drop(&hmp->fchain);
1763                 }
1764                 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1765                         atomic_clear_int(&hmp->fchain.flags,
1766                                          HAMMER2_CHAIN_UPDATE);
1767                         hammer2_chain_drop(&hmp->fchain);
1768                 }
1769
1770                 /*
1771                  * Final drop of embedded freemap root chain to
1772                  * clean up fchain.core (fchain structure is not
1773                  * flagged ALLOCATED so it is cleaned out and then
1774                  * left to rot).
1775                  */
1776                 hammer2_chain_drop(&hmp->fchain);
1777
1778                 /*
1779                  * Final drop of embedded volume root chain to clean
1780                  * up vchain.core (vchain structure is not flagged
1781                  * ALLOCATED so it is cleaned out and then left to
1782                  * rot).
1783                  */
1784                 dumpcnt = 50;
1785                 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1786                 dumpcnt = 50;
1787                 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1788                 hammer2_mount_unlock(hmp);
1789                 hammer2_chain_drop(&hmp->vchain);
1790
1791                 hammer2_io_cleanup(hmp, &hmp->iotree);
1792                 if (hmp->iofree_count) {
1793                         kprintf("io_cleanup: %d I/O's left hanging\n",
1794                                 hmp->iofree_count);
1795                 }
1796
1797                 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1798                 kmalloc_destroy(&hmp->mchain);
1799                 kfree(hmp, M_HAMMER2);
1800         } else {
1801                 hammer2_mount_unlock(hmp);
1802         }
1803 }
1804
1805 static
1806 int
1807 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1808              ino_t ino, struct vnode **vpp)
1809 {
1810         kprintf("hammer2_vget\n");
1811         return (EOPNOTSUPP);
1812 }
1813
1814 static
1815 int
1816 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1817 {
1818         hammer2_pfsmount_t *pmp;
1819         hammer2_cluster_t *cparent;
1820         int error;
1821         struct vnode *vp;
1822
1823         pmp = MPTOPMP(mp);
1824         if (pmp->iroot == NULL) {
1825                 *vpp = NULL;
1826                 error = EINVAL;
1827         } else {
1828                 cparent = hammer2_inode_lock_sh(pmp->iroot);
1829                 vp = hammer2_igetv(pmp->iroot, cparent, &error);
1830                 hammer2_inode_unlock_sh(pmp->iroot, cparent);
1831                 *vpp = vp;
1832                 if (vp == NULL)
1833                         kprintf("vnodefail\n");
1834         }
1835
1836         return (error);
1837 }
1838
1839 /*
1840  * Filesystem status
1841  *
1842  * XXX incorporate ipdata->inode_quota and data_quota
1843  */
1844 static
1845 int
1846 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1847 {
1848         hammer2_pfsmount_t *pmp;
1849         hammer2_mount_t *hmp;
1850
1851         pmp = MPTOPMP(mp);
1852         KKASSERT(pmp->iroot->cluster.nchains >= 1);
1853         hmp = pmp->iroot->cluster.focus->hmp;   /* XXX */
1854
1855         mp->mnt_stat.f_files = pmp->inode_count;
1856         mp->mnt_stat.f_ffree = 0;
1857         mp->mnt_stat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1858         mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1859         mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1860
1861         *sbp = mp->mnt_stat;
1862         return (0);
1863 }
1864
1865 static
1866 int
1867 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1868 {
1869         hammer2_pfsmount_t *pmp;
1870         hammer2_mount_t *hmp;
1871
1872         pmp = MPTOPMP(mp);
1873         KKASSERT(pmp->iroot->cluster.nchains >= 1);
1874         hmp = pmp->iroot->cluster.focus->hmp;   /* XXX */
1875
1876         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1877         mp->mnt_vstat.f_files = pmp->inode_count;
1878         mp->mnt_vstat.f_ffree = 0;
1879         mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1880         mp->mnt_vstat.f_bfree =  hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1881         mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1882
1883         *sbp = mp->mnt_vstat;
1884         return (0);
1885 }
1886
1887 /*
1888  * Mount-time recovery (RW mounts)
1889  *
1890  * Updates to the free block table are allowed to lag flushes by one
1891  * transaction.  In case of a crash, then on a fresh mount we must do an
1892  * incremental scan of the last committed transaction id and make sure that
1893  * all related blocks have been marked allocated.
1894  *
1895  * The super-root topology and each PFS has its own transaction id domain,
1896  * so we must track PFS boundary transitions.
1897  */
1898 struct hammer2_recovery_elm {
1899         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1900         hammer2_chain_t *chain;
1901         hammer2_tid_t sync_tid;
1902 };
1903
1904 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1905
1906 struct hammer2_recovery_info {
1907         struct hammer2_recovery_list list;
1908         int     depth;
1909 };
1910
1911 static int hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1912                         hammer2_chain_t *parent,
1913                         struct hammer2_recovery_info *info,
1914                         hammer2_tid_t sync_tid);
1915
1916 #define HAMMER2_RECOVERY_MAXDEPTH       10
1917
1918 static
1919 int
1920 hammer2_recovery(hammer2_mount_t *hmp)
1921 {
1922         hammer2_trans_t trans;
1923         struct hammer2_recovery_info info;
1924         struct hammer2_recovery_elm *elm;
1925         hammer2_chain_t *parent;
1926         hammer2_tid_t sync_tid;
1927         int error;
1928         int cumulative_error = 0;
1929
1930         hammer2_trans_init(&trans, hmp->spmp, 0);
1931
1932         sync_tid = 0;
1933         TAILQ_INIT(&info.list);
1934         info.depth = 0;
1935         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1936         cumulative_error = hammer2_recovery_scan(&trans, hmp, parent,
1937                                                  &info, sync_tid);
1938         hammer2_chain_lookup_done(parent);
1939
1940         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1941                 TAILQ_REMOVE(&info.list, elm, entry);
1942                 parent = elm->chain;
1943                 sync_tid = elm->sync_tid;
1944                 kfree(elm, M_HAMMER2);
1945
1946                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
1947                                            HAMMER2_RESOLVE_NOREF);
1948                 error = hammer2_recovery_scan(&trans, hmp, parent,
1949                                               &info, sync_tid);
1950                 hammer2_chain_unlock(parent);
1951                 if (error)
1952                         cumulative_error = error;
1953         }
1954         hammer2_trans_done(&trans);
1955
1956         return cumulative_error;
1957 }
1958
1959 static
1960 int
1961 hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1962                       hammer2_chain_t *parent,
1963                       struct hammer2_recovery_info *info,
1964                       hammer2_tid_t sync_tid)
1965 {
1966         hammer2_chain_t *chain;
1967         int cache_index;
1968         int cumulative_error = 0;
1969         int pfs_boundary = 0;
1970         int error;
1971
1972         /*
1973          * Adjust freemap to ensure that the block(s) are marked allocated.
1974          */
1975         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1976                 hammer2_freemap_adjust(trans, hmp, &parent->bref,
1977                                        HAMMER2_FREEMAP_DORECOVER);
1978         }
1979
1980         /*
1981          * Check type for recursive scan
1982          */
1983         switch(parent->bref.type) {
1984         case HAMMER2_BREF_TYPE_VOLUME:
1985                 /* data already instantiated */
1986                 break;
1987         case HAMMER2_BREF_TYPE_INODE:
1988                 /*
1989                  * Must instantiate data for DIRECTDATA test and also
1990                  * for recursion.
1991                  */
1992                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1993                 if (parent->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1994                         /* not applicable to recovery scan */
1995                         hammer2_chain_unlock(parent);
1996                         return 0;
1997                 }
1998                 if ((parent->data->ipdata.op_flags & HAMMER2_OPFLAG_PFSROOT) &&
1999                     info->depth != 0) {
2000                         pfs_boundary = 1;
2001                         sync_tid = parent->bref.mirror_tid - 1;
2002                 }
2003                 hammer2_chain_unlock(parent);
2004                 break;
2005         case HAMMER2_BREF_TYPE_INDIRECT:
2006                 /*
2007                  * Must instantiate data for recursion
2008                  */
2009                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2010                 hammer2_chain_unlock(parent);
2011                 break;
2012         case HAMMER2_BREF_TYPE_DATA:
2013         case HAMMER2_BREF_TYPE_FREEMAP:
2014         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2015         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2016                 /* not applicable to recovery scan */
2017                 return 0;
2018                 break;
2019         default:
2020                 return EDOM;
2021         }
2022
2023         /*
2024          * Defer operation if depth limit reached or if we are crossing a
2025          * PFS boundary.
2026          */
2027         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH || pfs_boundary) {
2028                 struct hammer2_recovery_elm *elm;
2029
2030                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2031                 elm->chain = parent;
2032                 elm->sync_tid = sync_tid;
2033                 hammer2_chain_ref(parent);
2034                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2035                 /* unlocked by caller */
2036
2037                 return(0);
2038         }
2039
2040
2041         /*
2042          * Recursive scan of the last flushed transaction only.  We are
2043          * doing this without pmp assignments so don't leave the chains
2044          * hanging around after we are done with them.
2045          */
2046         cache_index = 0;
2047         chain = hammer2_chain_scan(parent, NULL, &cache_index,
2048                                    HAMMER2_LOOKUP_NODATA);
2049         while (chain) {
2050                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2051                 if (chain->bref.mirror_tid >= sync_tid) {
2052                         ++info->depth;
2053                         error = hammer2_recovery_scan(trans, hmp, chain,
2054                                                       info, sync_tid);
2055                         --info->depth;
2056                         if (error)
2057                                 cumulative_error = error;
2058                 }
2059                 chain = hammer2_chain_scan(parent, chain, &cache_index,
2060                                            HAMMER2_LOOKUP_NODATA);
2061         }
2062
2063         return cumulative_error;
2064 }
2065
2066 /*
2067  * Sync the entire filesystem; this is called from the filesystem syncer
2068  * process periodically and whenever a user calls sync(1) on the hammer
2069  * mountpoint.
2070  *
2071  * Currently is actually called from the syncer! \o/
2072  *
2073  * This task will have to snapshot the state of the dirty inode chain.
2074  * From that, it will have to make sure all of the inodes on the dirty
2075  * chain have IO initiated. We make sure that io is initiated for the root
2076  * block.
2077  *
2078  * If waitfor is set, we wait for media to acknowledge the new rootblock.
2079  *
2080  * THINKS: side A vs side B, to have sync not stall all I/O?
2081  */
2082 int
2083 hammer2_vfs_sync(struct mount *mp, int waitfor)
2084 {
2085         struct hammer2_sync_info info;
2086         hammer2_inode_t *iroot;
2087         hammer2_chain_t *chain;
2088         hammer2_chain_t *parent;
2089         hammer2_pfsmount_t *pmp;
2090         hammer2_mount_t *hmp;
2091         int flags;
2092         int error;
2093         int total_error;
2094         int force_fchain;
2095         int i;
2096         int j;
2097
2098         pmp = MPTOPMP(mp);
2099         iroot = pmp->iroot;
2100         KKASSERT(iroot);
2101         KKASSERT(iroot->pmp == pmp);
2102
2103         /*
2104          * We can't acquire locks on existing vnodes while in a transaction
2105          * without risking a deadlock.  This assumes that vfsync() can be
2106          * called without the vnode locked (which it can in DragonFly).
2107          * Otherwise we'd have to implement a multi-pass or flag the lock
2108          * failures and retry.
2109          *
2110          * The reclamation code interlocks with the sync list's token
2111          * (by removing the vnode from the scan list) before unlocking
2112          * the inode, giving us time to ref the inode.
2113          */
2114         /*flags = VMSC_GETVP;*/
2115         flags = 0;
2116         if (waitfor & MNT_LAZY)
2117                 flags |= VMSC_ONEPASS;
2118
2119         /*
2120          * Start our flush transaction.  This does not return until all
2121          * concurrent transactions have completed and will prevent any
2122          * new transactions from running concurrently, except for the
2123          * buffer cache transactions.
2124          *
2125          * For efficiency do an async pass before making sure with a
2126          * synchronous pass on all related buffer cache buffers.  It
2127          * should theoretically not be possible for any new file buffers
2128          * to be instantiated during this sequence.
2129          */
2130         hammer2_trans_init(&info.trans, pmp, HAMMER2_TRANS_ISFLUSH |
2131                                              HAMMER2_TRANS_PREFLUSH);
2132         hammer2_run_unlinkq(&info.trans, pmp);
2133
2134         info.error = 0;
2135         info.waitfor = MNT_NOWAIT;
2136         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2137         info.waitfor = MNT_WAIT;
2138         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2139
2140         /*
2141          * Clear PREFLUSH.  This prevents (or asserts on) any new logical
2142          * buffer cache flushes which occur during the flush.  Device buffers
2143          * are not affected.
2144          */
2145
2146 #if 0
2147         if (info.error == 0 && (waitfor & MNT_WAIT)) {
2148                 info.waitfor = waitfor;
2149                     vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2150
2151         }
2152 #endif
2153         hammer2_bioq_sync(info.trans.pmp);
2154         atomic_clear_int(&info.trans.flags, HAMMER2_TRANS_PREFLUSH);
2155
2156         total_error = 0;
2157
2158         /*
2159          * Flush all storage elements making up the cluster
2160          *
2161          * We must also flush any deleted siblings because the super-root
2162          * flush won't do it for us.  They all must be staged or the
2163          * super-root flush will not be able to update its block table
2164          * properly.
2165          *
2166          * XXX currently done serially instead of concurrently
2167          */
2168         for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2169                 chain = iroot->cluster.array[i];
2170                 if (chain) {
2171                         hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
2172                         hammer2_flush(&info.trans, chain);
2173                         hammer2_chain_unlock(chain);
2174                 }
2175         }
2176 #if 0
2177         hammer2_trans_done(&info.trans);
2178 #endif
2179
2180         /*
2181          * Flush all volume roots to synchronize PFS flushes with the
2182          * storage media.  Use a super-root transaction for each one.
2183          *
2184          * The flush code will detect super-root -> pfs-root chain
2185          * transitions using the last pfs-root flush.
2186          */
2187         for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2188                 chain = iroot->cluster.array[i];
2189                 if (chain == NULL)
2190                         continue;
2191
2192                 hmp = chain->hmp;
2193
2194                 /*
2195                  * We only have to flush each hmp once
2196                  */
2197                 for (j = i - 1; j >= 0; --j) {
2198                         if (iroot->cluster.array[j] &&
2199                             iroot->cluster.array[j]->hmp == hmp)
2200                                 break;
2201                 }
2202                 if (j >= 0)
2203                         continue;
2204                 hammer2_trans_spmp(&info.trans, hmp->spmp);
2205
2206                 /*
2207                  * Force an update of the XID from the PFS root to the
2208                  * topology root.  We couldn't do this from the PFS
2209                  * transaction because a SPMP transaction is needed.
2210                  * This does not modify blocks, instead what it does is
2211                  * allow the flush code to find the transition point and
2212                  * then update on the way back up.
2213                  */
2214                 parent = chain->parent;
2215                 KKASSERT(chain->pmp != parent->pmp);
2216                 hammer2_chain_setflush(&info.trans, parent);
2217
2218                 /*
2219                  * Media mounts have two 'roots', vchain for the topology
2220                  * and fchain for the free block table.  Flush both.
2221                  *
2222                  * Note that the topology and free block table are handled
2223                  * independently, so the free block table can wind up being
2224                  * ahead of the topology.  We depend on the bulk free scan
2225                  * code to deal with any loose ends.
2226                  */
2227                 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2228                 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2229                 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2230                         /*
2231                          * This will also modify vchain as a side effect,
2232                          * mark vchain as modified now.
2233                          */
2234                         hammer2_voldata_modify(hmp);
2235                         chain = &hmp->fchain;
2236                         hammer2_flush(&info.trans, chain);
2237                         KKASSERT(chain == &hmp->fchain);
2238                 }
2239                 hammer2_chain_unlock(&hmp->fchain);
2240                 hammer2_chain_unlock(&hmp->vchain);
2241
2242                 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2243                 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2244                         chain = &hmp->vchain;
2245                         hammer2_flush(&info.trans, chain);
2246                         KKASSERT(chain == &hmp->vchain);
2247                         force_fchain = 1;
2248                 } else {
2249                         force_fchain = 0;
2250                 }
2251                 hammer2_chain_unlock(&hmp->vchain);
2252
2253 #if 0
2254                 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2255                 if ((hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) ||
2256                     force_fchain) {
2257                         /* this will also modify vchain as a side effect */
2258                         chain = &hmp->fchain;
2259                         hammer2_flush(&info.trans, chain);
2260                         KKASSERT(chain == &hmp->fchain);
2261                 }
2262                 hammer2_chain_unlock(&hmp->fchain);
2263 #endif
2264
2265                 error = 0;
2266
2267                 /*
2268                  * We can't safely flush the volume header until we have
2269                  * flushed any device buffers which have built up.
2270                  *
2271                  * XXX this isn't being incremental
2272                  */
2273                 vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
2274                 error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
2275                 vn_unlock(hmp->devvp);
2276
2277                 /*
2278                  * The flush code sets CHAIN_VOLUMESYNC to indicate that the
2279                  * volume header needs synchronization via hmp->volsync.
2280                  *
2281                  * XXX synchronize the flag & data with only this flush XXX
2282                  */
2283                 if (error == 0 &&
2284                     (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
2285                         struct buf *bp;
2286
2287                         /*
2288                          * Synchronize the disk before flushing the volume
2289                          * header.
2290                          */
2291                         bp = getpbuf(NULL);
2292                         bp->b_bio1.bio_offset = 0;
2293                         bp->b_bufsize = 0;
2294                         bp->b_bcount = 0;
2295                         bp->b_cmd = BUF_CMD_FLUSH;
2296                         bp->b_bio1.bio_done = biodone_sync;
2297                         bp->b_bio1.bio_flags |= BIO_SYNC;
2298                         vn_strategy(hmp->devvp, &bp->b_bio1);
2299                         biowait(&bp->b_bio1, "h2vol");
2300                         relpbuf(bp, NULL);
2301
2302                         /*
2303                          * Then we can safely flush the version of the
2304                          * volume header synchronized by the flush code.
2305                          */
2306                         i = hmp->volhdrno + 1;
2307                         if (i >= HAMMER2_NUM_VOLHDRS)
2308                                 i = 0;
2309                         if (i * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
2310                             hmp->volsync.volu_size) {
2311                                 i = 0;
2312                         }
2313                         kprintf("sync volhdr %d %jd\n",
2314                                 i, (intmax_t)hmp->volsync.volu_size);
2315                         bp = getblk(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2316                                     HAMMER2_PBUFSIZE, 0, 0);
2317                         atomic_clear_int(&hmp->vchain.flags,
2318                                          HAMMER2_CHAIN_VOLUMESYNC);
2319                         bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
2320                         bawrite(bp);
2321                         hmp->volhdrno = i;
2322                 }
2323                 if (error)
2324                         total_error = error;
2325
2326 #if 0
2327                 hammer2_trans_done(&info.trans);
2328 #endif
2329         }
2330         hammer2_trans_done(&info.trans);
2331
2332         return (total_error);
2333 }
2334
2335 /*
2336  * Sync passes.
2337  */
2338 static int
2339 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2340 {
2341         struct hammer2_sync_info *info = data;
2342         hammer2_inode_t *ip;
2343         int error;
2344
2345         /*
2346          *
2347          */
2348         ip = VTOI(vp);
2349         if (ip == NULL)
2350                 return(0);
2351         if (vp->v_type == VNON || vp->v_type == VBAD) {
2352                 vclrisdirty(vp);
2353                 return(0);
2354         }
2355         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2356             RB_EMPTY(&vp->v_rbdirty_tree)) {
2357                 vclrisdirty(vp);
2358                 return(0);
2359         }
2360
2361         /*
2362          * VOP_FSYNC will start a new transaction so replicate some code
2363          * here to do it inline (see hammer2_vop_fsync()).
2364          *
2365          * WARNING: The vfsync interacts with the buffer cache and might
2366          *          block, we can't hold the inode lock at that time.
2367          *          However, we MUST ref ip before blocking to ensure that
2368          *          it isn't ripped out from under us (since we do not
2369          *          hold a lock on the vnode).
2370          */
2371         hammer2_inode_ref(ip);
2372         atomic_clear_int(&ip->flags, HAMMER2_INODE_MODIFIED);
2373         if (vp)
2374                 vfsync(vp, MNT_NOWAIT, 1, NULL, NULL);
2375
2376         hammer2_inode_drop(ip);
2377 #if 1
2378         error = 0;
2379         if (error)
2380                 info->error = error;
2381 #endif
2382         return(0);
2383 }
2384
2385 static
2386 int
2387 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2388 {
2389         return (0);
2390 }
2391
2392 static
2393 int
2394 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2395                struct fid *fhp, struct vnode **vpp)
2396 {
2397         return (0);
2398 }
2399
2400 static
2401 int
2402 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2403                  int *exflagsp, struct ucred **credanonp)
2404 {
2405         return (0);
2406 }
2407
2408 /*
2409  * Support code for hammer2_mount().  Read, verify, and install the volume
2410  * header into the HMP
2411  *
2412  * XXX read four volhdrs and use the one with the highest TID whos CRC
2413  *     matches.
2414  *
2415  * XXX check iCRCs.
2416  *
2417  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2418  *     nonexistant locations.
2419  *
2420  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2421  */
2422 static
2423 int
2424 hammer2_install_volume_header(hammer2_mount_t *hmp)
2425 {
2426         hammer2_volume_data_t *vd;
2427         struct buf *bp;
2428         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2429         int error_reported;
2430         int error;
2431         int valid;
2432         int i;
2433
2434         error_reported = 0;
2435         error = 0;
2436         valid = 0;
2437         bp = NULL;
2438
2439         /*
2440          * There are up to 4 copies of the volume header (syncs iterate
2441          * between them so there is no single master).  We don't trust the
2442          * volu_size field so we don't know precisely how large the filesystem
2443          * is, so depend on the OS to return an error if we go beyond the
2444          * block device's EOF.
2445          */
2446         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2447                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2448                               HAMMER2_VOLUME_BYTES, &bp);
2449                 if (error) {
2450                         brelse(bp);
2451                         bp = NULL;
2452                         continue;
2453                 }
2454
2455                 vd = (struct hammer2_volume_data *) bp->b_data;
2456                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2457                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2458                         brelse(bp);
2459                         bp = NULL;
2460                         continue;
2461                 }
2462
2463                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2464                         /* XXX: Reversed-endianness filesystem */
2465                         kprintf("hammer2: reverse-endian filesystem detected");
2466                         brelse(bp);
2467                         bp = NULL;
2468                         continue;
2469                 }
2470
2471                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2472                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2473                                       HAMMER2_VOLUME_ICRC0_SIZE);
2474                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2475                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2476                                        HAMMER2_VOLUME_ICRC1_SIZE);
2477                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2478                         kprintf("hammer2 volume header crc "
2479                                 "mismatch copy #%d %08x/%08x\n",
2480                                 i, crc0, crc);
2481                         error_reported = 1;
2482                         brelse(bp);
2483                         bp = NULL;
2484                         continue;
2485                 }
2486                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2487                         valid = 1;
2488                         hmp->voldata = *vd;
2489                         hmp->volhdrno = i;
2490                 }
2491                 brelse(bp);
2492                 bp = NULL;
2493         }
2494         if (valid) {
2495                 hmp->volsync = hmp->voldata;
2496                 error = 0;
2497                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2498                         kprintf("hammer2: using volume header #%d\n",
2499                                 hmp->volhdrno);
2500                 }
2501         } else {
2502                 error = EINVAL;
2503                 kprintf("hammer2: no valid volume headers found!\n");
2504         }
2505         return (error);
2506 }
2507
2508 /*
2509  * Reconnect using the passed file pointer.  The caller must ref the
2510  * fp for us.
2511  */
2512 void
2513 hammer2_cluster_reconnect(hammer2_mount_t *hmp, struct file *fp)
2514 {
2515         size_t name_len;
2516         const char *name = "disk-volume";
2517
2518         /*
2519          * Closes old comm descriptor, kills threads, cleans up
2520          * states, then installs the new descriptor and creates
2521          * new threads.
2522          */
2523         kdmsg_iocom_reconnect(&hmp->iocom, fp, "hammer2");
2524
2525         /*
2526          * Setup LNK_CONN fields for autoinitiated state machine.  We
2527          * will use SPANs to advertise multiple PFSs so only pass the
2528          * fsid and HAMMER2_PFSTYPE_SUPROOT for the AUTOCONN.
2529          *
2530          * We are not initiating a LNK_SPAN so we do not have to set-up
2531          * iocom.auto_lnk_span.
2532          */
2533         bzero(&hmp->iocom.auto_lnk_conn.pfs_clid,
2534               sizeof(hmp->iocom.auto_lnk_conn.pfs_clid));
2535         hmp->iocom.auto_lnk_conn.pfs_fsid = hmp->voldata.fsid;
2536         hmp->iocom.auto_lnk_conn.pfs_type = HAMMER2_PFSTYPE_SUPROOT;
2537         hmp->iocom.auto_lnk_conn.proto_version = DMSG_SPAN_PROTO_1;
2538 #if 0
2539         hmp->iocom.auto_lnk_conn.peer_type = hmp->voldata.peer_type;
2540 #endif
2541         hmp->iocom.auto_lnk_conn.peer_type = DMSG_PEER_HAMMER2;
2542
2543         /*
2544          * Filter adjustment.  Clients do not need visibility into other
2545          * clients (otherwise millions of clients would present a serious
2546          * problem).  The fs_label also serves to restrict the namespace.
2547          */
2548         hmp->iocom.auto_lnk_conn.peer_mask = 1LLU << DMSG_PEER_HAMMER2;
2549         hmp->iocom.auto_lnk_conn.pfs_mask = (uint64_t)-1;
2550
2551 #if 0
2552         switch (ipdata->pfs_type) {
2553         case DMSG_PFSTYPE_CLIENT:
2554                 hmp->iocom.auto_lnk_conn.peer_mask &=
2555                                 ~(1LLU << DMSG_PFSTYPE_CLIENT);
2556                 break;
2557         default:
2558                 break;
2559         }
2560 #endif
2561
2562         name_len = strlen(name);
2563         if (name_len >= sizeof(hmp->iocom.auto_lnk_conn.fs_label))
2564                 name_len = sizeof(hmp->iocom.auto_lnk_conn.fs_label) - 1;
2565         bcopy(name, hmp->iocom.auto_lnk_conn.fs_label, name_len);
2566         hmp->iocom.auto_lnk_conn.fs_label[name_len] = 0;
2567
2568         kdmsg_iocom_autoinitiate(&hmp->iocom, hammer2_autodmsg);
2569 }
2570
2571 static int
2572 hammer2_rcvdmsg(kdmsg_msg_t *msg)
2573 {
2574         kprintf("RCVMSG %08x\n", msg->tcmd);
2575
2576         switch(msg->tcmd) {
2577         case DMSG_DBG_SHELL:
2578                 /*
2579                  * (non-transaction)
2580                  * Execute shell command (not supported atm)
2581                  */
2582                 kdmsg_msg_result(msg, DMSG_ERR_NOSUPP);
2583                 break;
2584         case DMSG_DBG_SHELL | DMSGF_REPLY:
2585                 /*
2586                  * (non-transaction)
2587                  */
2588                 if (msg->aux_data) {
2589                         msg->aux_data[msg->aux_size - 1] = 0;
2590                         kprintf("HAMMER2 DBG: %s\n", msg->aux_data);
2591                 }
2592                 break;
2593         default:
2594                 /*
2595                  * Unsupported message received.  We only need to
2596                  * reply if it's a transaction in order to close our end.
2597                  * Ignore any one-way messages or any further messages
2598                  * associated with the transaction.
2599                  *
2600                  * NOTE: This case also includes DMSG_LNK_ERROR messages
2601                  *       which might be one-way, replying to those would
2602                  *       cause an infinite ping-pong.
2603                  */
2604                 if (msg->any.head.cmd & DMSGF_CREATE)
2605                         kdmsg_msg_reply(msg, DMSG_ERR_NOSUPP);
2606                 break;
2607         }
2608         return(0);
2609 }
2610
2611 /*
2612  * This function is called after KDMSG has automatically handled processing
2613  * of a LNK layer message (typically CONN, SPAN, or CIRC).
2614  *
2615  * We tag off the LNK_CONN to trigger our LNK_VOLCONF messages which
2616  * advertises all available hammer2 super-root volumes.
2617  */
2618 static void hammer2_update_spans(hammer2_mount_t *hmp, kdmsg_state_t *state);
2619
2620 static void
2621 hammer2_autodmsg(kdmsg_msg_t *msg)
2622 {
2623         hammer2_mount_t *hmp = msg->state->iocom->handle;
2624         int copyid;
2625
2626         kprintf("RCAMSG %08x\n", msg->tcmd);
2627
2628         switch(msg->tcmd) {
2629         case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_REPLY:
2630         case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_DELETE | DMSGF_REPLY:
2631                 if (msg->any.head.cmd & DMSGF_CREATE) {
2632                         kprintf("HAMMER2: VOLDATA DUMP\n");
2633
2634                         /*
2635                          * Dump the configuration stored in the volume header.
2636                          * This will typically be import/export access rights,
2637                          * master encryption keys (encrypted), etc.
2638                          */
2639                         hammer2_voldata_lock(hmp);
2640                         copyid = 0;
2641                         while (copyid < HAMMER2_COPYID_COUNT) {
2642                                 if (hmp->voldata.copyinfo[copyid].copyid)
2643                                         hammer2_volconf_update(hmp, copyid);
2644                                 ++copyid;
2645                         }
2646                         hammer2_voldata_unlock(hmp);
2647
2648                         kprintf("HAMMER2: INITIATE SPANs\n");
2649                         hammer2_update_spans(hmp, msg->state);
2650                 }
2651                 if ((msg->any.head.cmd & DMSGF_DELETE) &&
2652                     msg->state && (msg->state->txcmd & DMSGF_DELETE) == 0) {
2653                         kprintf("HAMMER2: CONN WAS TERMINATED\n");
2654                 }
2655                 break;
2656         default:
2657                 break;
2658         }
2659 }
2660
2661 /*
2662  * Update LNK_SPAN state
2663  */
2664 static void
2665 hammer2_update_spans(hammer2_mount_t *hmp, kdmsg_state_t *state)
2666 {
2667         const hammer2_inode_data_t *ipdata;
2668         hammer2_cluster_t *cparent;
2669         hammer2_cluster_t *cluster;
2670         hammer2_pfsmount_t *spmp;
2671         hammer2_key_t key_next;
2672         kdmsg_msg_t *rmsg;
2673         size_t name_len;
2674         int ddflag;
2675
2676         /*
2677          * Lookup mount point under the media-localized super-root.
2678          *
2679          * cluster->pmp will incorrectly point to spmp and must be fixed
2680          * up later on.
2681          */
2682         spmp = hmp->spmp;
2683         cparent = hammer2_inode_lock_ex(spmp->iroot);
2684         cluster = hammer2_cluster_lookup(cparent, &key_next,
2685                                          HAMMER2_KEY_MIN,
2686                                          HAMMER2_KEY_MAX,
2687                                          0, &ddflag);
2688         while (cluster) {
2689                 if (hammer2_cluster_type(cluster) != HAMMER2_BREF_TYPE_INODE)
2690                         continue;
2691                 ipdata = &hammer2_cluster_data(cluster)->ipdata;
2692                 kprintf("UPDATE SPANS: %s\n", ipdata->filename);
2693
2694                 rmsg = kdmsg_msg_alloc(state, DMSG_LNK_SPAN | DMSGF_CREATE,
2695                                        hammer2_lnk_span_reply, NULL);
2696                 rmsg->any.lnk_span.pfs_clid = ipdata->pfs_clid;
2697                 rmsg->any.lnk_span.pfs_fsid = ipdata->pfs_fsid;
2698                 rmsg->any.lnk_span.pfs_type = ipdata->pfs_type;
2699                 rmsg->any.lnk_span.peer_type = DMSG_PEER_HAMMER2;
2700                 rmsg->any.lnk_span.proto_version = DMSG_SPAN_PROTO_1;
2701                 name_len = ipdata->name_len;
2702                 if (name_len >= sizeof(rmsg->any.lnk_span.fs_label))
2703                         name_len = sizeof(rmsg->any.lnk_span.fs_label) - 1;
2704                 bcopy(ipdata->filename, rmsg->any.lnk_span.fs_label, name_len);
2705
2706                 kdmsg_msg_write(rmsg);
2707
2708                 cluster = hammer2_cluster_next(cparent, cluster,
2709                                                &key_next,
2710                                                key_next,
2711                                                HAMMER2_KEY_MAX,
2712                                                0);
2713         }
2714         hammer2_inode_unlock_ex(spmp->iroot, cparent);
2715 }
2716
2717 static
2718 int
2719 hammer2_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg)
2720 {
2721         if ((state->txcmd & DMSGF_DELETE) == 0 &&
2722             (msg->any.head.cmd & DMSGF_DELETE)) {
2723                 kdmsg_msg_reply(msg, 0);
2724         }
2725         return 0;
2726 }
2727
2728 /*
2729  * Volume configuration updates are passed onto the userland service
2730  * daemon via the open LNK_CONN transaction.
2731  */
2732 void
2733 hammer2_volconf_update(hammer2_mount_t *hmp, int index)
2734 {
2735         kdmsg_msg_t *msg;
2736
2737         /* XXX interlock against connection state termination */
2738         kprintf("volconf update %p\n", hmp->iocom.conn_state);
2739         if (hmp->iocom.conn_state) {
2740                 kprintf("TRANSMIT VOLCONF VIA OPEN CONN TRANSACTION\n");
2741                 msg = kdmsg_msg_alloc(hmp->iocom.conn_state,
2742                                       DMSG_LNK_HAMMER2_VOLCONF,
2743                                       NULL, NULL);
2744                 H2_LNK_VOLCONF(msg)->copy = hmp->voldata.copyinfo[index];
2745                 H2_LNK_VOLCONF(msg)->mediaid = hmp->voldata.fsid;
2746                 H2_LNK_VOLCONF(msg)->index = index;
2747                 kdmsg_msg_write(msg);
2748         }
2749 }
2750
2751 /*
2752  * This handles hysteresis on regular file flushes.  Because the BIOs are
2753  * routed to a thread it is possible for an excessive number to build up
2754  * and cause long front-end stalls long before the runningbuffspace limit
2755  * is hit, so we implement hammer2_flush_pipe to control the
2756  * hysteresis.
2757  *
2758  * This is a particular problem when compression is used.
2759  */
2760 void
2761 hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp)
2762 {
2763         atomic_add_int(&pmp->count_lwinprog, 1);
2764 }
2765
2766 void
2767 hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp)
2768 {
2769         int lwinprog;
2770
2771         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2772         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2773             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2774                 atomic_clear_int(&pmp->count_lwinprog,
2775                                  HAMMER2_LWINPROG_WAITING);
2776                 wakeup(&pmp->count_lwinprog);
2777         }
2778 }
2779
2780 void
2781 hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp)
2782 {
2783         int lwinprog;
2784
2785         for (;;) {
2786                 lwinprog = pmp->count_lwinprog;
2787                 cpu_ccfence();
2788                 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2789                         break;
2790                 tsleep_interlock(&pmp->count_lwinprog, 0);
2791                 atomic_set_int(&pmp->count_lwinprog, HAMMER2_LWINPROG_WAITING);
2792                 lwinprog = pmp->count_lwinprog;
2793                 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2794                         break;
2795                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2796         }
2797 }
2798
2799 /*
2800  * Manage excessive memory resource use for chain and related
2801  * structures.
2802  */
2803 void
2804 hammer2_pfs_memory_wait(hammer2_pfsmount_t *pmp)
2805 {
2806         uint32_t waiting;
2807         uint32_t count;
2808         uint32_t limit;
2809 #if 0
2810         static int zzticks;
2811 #endif
2812
2813         /*
2814          * Atomic check condition and wait.  Also do an early speedup of
2815          * the syncer to try to avoid hitting the wait.
2816          */
2817         for (;;) {
2818                 waiting = pmp->inmem_dirty_chains;
2819                 cpu_ccfence();
2820                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2821
2822                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2823                 if (limit < hammer2_limit_dirty_chains)
2824                         limit = hammer2_limit_dirty_chains;
2825                 if (limit < 1000)
2826                         limit = 1000;
2827
2828 #if 0
2829                 if ((int)(ticks - zzticks) > hz) {
2830                         zzticks = ticks;
2831                         kprintf("count %ld %ld\n", count, limit);
2832                 }
2833 #endif
2834
2835                 /*
2836                  * Block if there are too many dirty chains present, wait
2837                  * for the flush to clean some out.
2838                  */
2839                 if (count > limit) {
2840                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2841                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2842                                                waiting,
2843                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2844                                 speedup_syncer(pmp->mp);
2845                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2846                                        "chnmem", hz);
2847                         }
2848                         continue;       /* loop on success or fail */
2849                 }
2850
2851                 /*
2852                  * Try to start an early flush before we are forced to block.
2853                  */
2854                 if (count > limit * 7 / 10)
2855                         speedup_syncer(pmp->mp);
2856                 break;
2857         }
2858 }
2859
2860 void
2861 hammer2_pfs_memory_inc(hammer2_pfsmount_t *pmp)
2862 {
2863         if (pmp) {
2864                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2865         }
2866 }
2867
2868 void
2869 hammer2_pfs_memory_wakeup(hammer2_pfsmount_t *pmp)
2870 {
2871         uint32_t waiting;
2872
2873         if (pmp == NULL)
2874                 return;
2875
2876         for (;;) {
2877                 waiting = pmp->inmem_dirty_chains;
2878                 cpu_ccfence();
2879                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2880                                        waiting,
2881                                        (waiting - 1) &
2882                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2883                         break;
2884                 }
2885         }
2886
2887         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2888                 wakeup(&pmp->inmem_dirty_chains);
2889 }
2890
2891 /*
2892  * Debugging
2893  */
2894 void
2895 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2896 {
2897         hammer2_chain_t *scan;
2898         hammer2_chain_t *parent;
2899
2900         --*countp;
2901         if (*countp == 0) {
2902                 kprintf("%*.*s...\n", tab, tab, "");
2903                 return;
2904         }
2905         if (*countp < 0)
2906                 return;
2907         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2908                 tab, tab, "", pfx,
2909                 chain, chain->bref.type,
2910                 chain->bref.key, chain->bref.keybits,
2911                 chain->bref.mirror_tid);
2912
2913         kprintf("%*.*s      [%08x] (%s) refs=%d\n",
2914                 tab, tab, "",
2915                 chain->flags,
2916                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2917                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2918                 chain->refs);
2919
2920         kprintf("%*.*s      core [%08x]",
2921                 tab, tab, "",
2922                 chain->core.flags);
2923
2924         parent = chain->parent;
2925         if (parent)
2926                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2927                         tab, tab, "",
2928                         parent, parent->flags, parent->refs);
2929         if (RB_EMPTY(&chain->core.rbtree)) {
2930                 kprintf("\n");
2931         } else {
2932                 kprintf(" {\n");
2933                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2934                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2935                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2936                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2937                                 chain->data->ipdata.filename);
2938                 else
2939                         kprintf("%*.*s}\n", tab, tab, "");
2940         }
2941 }