1 /* 2 * Copyright (c) 2011-2018 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression) 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in 16 * the documentation and/or other materials provided with the 17 * distribution. 18 * 3. Neither the name of The DragonFly Project nor the names of its 19 * contributors may be used to endorse or promote products derived 20 * from this software without specific, prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 #include <sys/param.h> 36 #include <sys/systm.h> 37 #include <sys/kernel.h> 38 #include <sys/nlookup.h> 39 #include <sys/vnode.h> 40 #include <sys/mount.h> 41 #include <sys/fcntl.h> 42 #include <sys/buf.h> 43 #include <sys/uuid.h> 44 #include <sys/vfsops.h> 45 #include <sys/sysctl.h> 46 #include <sys/socket.h> 47 #include <sys/objcache.h> 48 49 #include <sys/proc.h> 50 #include <sys/mountctl.h> 51 #include <sys/dirent.h> 52 #include <sys/uio.h> 53 54 #include "hammer2.h" 55 #include "hammer2_disk.h" 56 #include "hammer2_mount.h" 57 #include "hammer2_lz4.h" 58 59 #include "zlib/hammer2_zlib.h" 60 61 #define REPORT_REFS_ERRORS 1 /* XXX remove me */ 62 63 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache"); 64 65 struct hammer2_sync_info { 66 int error; 67 int waitfor; 68 int pass; 69 }; 70 71 TAILQ_HEAD(hammer2_mntlist, hammer2_dev); 72 static struct hammer2_mntlist hammer2_mntlist; 73 74 struct hammer2_pfslist hammer2_pfslist; 75 struct hammer2_pfslist hammer2_spmplist; 76 struct lock hammer2_mntlk; 77 78 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT; 79 int hammer2_debug; 80 int hammer2_xopgroups; 81 long hammer2_debug_inode; 82 int hammer2_cluster_meta_read = 1; /* physical read-ahead */ 83 int hammer2_cluster_data_read = 4; /* physical read-ahead */ 84 int hammer2_cluster_write = 0; /* physical write clustering */ 85 int hammer2_dedup_enable = 1; 86 int hammer2_always_compress = 0; /* always try to compress */ 87 int hammer2_flush_pipe = 100; 88 int hammer2_dio_count; 89 int hammer2_dio_limit = 256; 90 int hammer2_bulkfree_tps = 5000; 91 int hammer2_worker_rmask = 3; 92 long hammer2_chain_allocs; 93 long hammer2_limit_dirty_chains; 94 long hammer2_limit_dirty_inodes; 95 long hammer2_count_modified_chains; 96 long hammer2_iod_file_read; 97 long hammer2_iod_meta_read; 98 long hammer2_iod_indr_read; 99 long hammer2_iod_fmap_read; 100 long hammer2_iod_volu_read; 101 long hammer2_iod_file_write; 102 long hammer2_iod_file_wembed; 103 long hammer2_iod_file_wzero; 104 long hammer2_iod_file_wdedup; 105 long hammer2_iod_meta_write; 106 long hammer2_iod_indr_write; 107 long hammer2_iod_fmap_write; 108 long hammer2_iod_volu_write; 109 static long hammer2_iod_inode_creates; 110 static long hammer2_iod_inode_deletes; 111 112 long hammer2_process_icrc32; 113 long hammer2_process_xxhash64; 114 115 MALLOC_DECLARE(M_HAMMER2_CBUFFER); 116 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer", 117 "Buffer used for compression."); 118 119 MALLOC_DECLARE(M_HAMMER2_DEBUFFER); 120 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer", 121 "Buffer used for decompression."); 122 123 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem"); 124 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD, 126 &hammer2_supported_version, 0, ""); 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW, 128 &hammer2_debug, 0, ""); 129 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW, 130 &hammer2_debug_inode, 0, ""); 131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW, 132 &hammer2_cluster_meta_read, 0, ""); 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW, 134 &hammer2_cluster_data_read, 0, ""); 135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW, 136 &hammer2_cluster_write, 0, ""); 137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW, 138 &hammer2_dedup_enable, 0, ""); 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW, 140 &hammer2_always_compress, 0, ""); 141 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW, 142 &hammer2_flush_pipe, 0, ""); 143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, worker_rmask, CTLFLAG_RW, 144 &hammer2_worker_rmask, 0, ""); 145 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW, 146 &hammer2_bulkfree_tps, 0, ""); 147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW, 148 &hammer2_chain_allocs, 0, ""); 149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW, 150 &hammer2_limit_dirty_chains, 0, ""); 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW, 152 &hammer2_limit_dirty_inodes, 0, ""); 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW, 154 &hammer2_count_modified_chains, 0, ""); 155 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD, 156 &hammer2_dio_count, 0, ""); 157 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW, 158 &hammer2_dio_limit, 0, ""); 159 160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW, 161 &hammer2_iod_file_read, 0, ""); 162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW, 163 &hammer2_iod_meta_read, 0, ""); 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW, 165 &hammer2_iod_indr_read, 0, ""); 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW, 167 &hammer2_iod_fmap_read, 0, ""); 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW, 169 &hammer2_iod_volu_read, 0, ""); 170 171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW, 172 &hammer2_iod_file_write, 0, ""); 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW, 174 &hammer2_iod_file_wembed, 0, ""); 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW, 176 &hammer2_iod_file_wzero, 0, ""); 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW, 178 &hammer2_iod_file_wdedup, 0, ""); 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW, 180 &hammer2_iod_meta_write, 0, ""); 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW, 182 &hammer2_iod_indr_write, 0, ""); 183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW, 184 &hammer2_iod_fmap_write, 0, ""); 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW, 186 &hammer2_iod_volu_write, 0, ""); 187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW, 188 &hammer2_iod_inode_creates, 0, ""); 189 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW, 190 &hammer2_iod_inode_deletes, 0, ""); 191 192 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW, 193 &hammer2_process_icrc32, 0, ""); 194 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW, 195 &hammer2_process_xxhash64, 0, ""); 196 197 static int hammer2_vfs_init(struct vfsconf *conf); 198 static int hammer2_vfs_uninit(struct vfsconf *vfsp); 199 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data, 200 struct ucred *cred); 201 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *, 202 struct vnode *, struct ucred *); 203 static int hammer2_recovery(hammer2_dev_t *hmp); 204 static int hammer2_vfs_unmount(struct mount *mp, int mntflags); 205 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp); 206 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, 207 struct ucred *cred); 208 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, 209 struct ucred *cred); 210 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp, 211 struct fid *fhp, struct vnode **vpp); 212 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp); 213 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 214 int *exflagsp, struct ucred **credanonp); 215 static int hammer2_vfs_modifying(struct mount *mp); 216 217 static int hammer2_install_volume_header(hammer2_dev_t *hmp); 218 219 static void hammer2_update_pmps(hammer2_dev_t *hmp); 220 221 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp); 222 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, 223 hammer2_dev_t *hmp); 224 static int hammer2_fixup_pfses(hammer2_dev_t *hmp); 225 226 /* 227 * HAMMER2 vfs operations. 228 */ 229 static struct vfsops hammer2_vfsops = { 230 .vfs_flags = 0, 231 .vfs_init = hammer2_vfs_init, 232 .vfs_uninit = hammer2_vfs_uninit, 233 .vfs_sync = hammer2_vfs_sync, 234 .vfs_mount = hammer2_vfs_mount, 235 .vfs_unmount = hammer2_vfs_unmount, 236 .vfs_root = hammer2_vfs_root, 237 .vfs_statfs = hammer2_vfs_statfs, 238 .vfs_statvfs = hammer2_vfs_statvfs, 239 .vfs_vget = hammer2_vfs_vget, 240 .vfs_vptofh = hammer2_vfs_vptofh, 241 .vfs_fhtovp = hammer2_vfs_fhtovp, 242 .vfs_checkexp = hammer2_vfs_checkexp, 243 .vfs_modifying = hammer2_vfs_modifying 244 }; 245 246 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", ""); 247 248 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE); 249 MODULE_VERSION(hammer2, 1); 250 251 static 252 int 253 hammer2_vfs_init(struct vfsconf *conf) 254 { 255 static struct objcache_malloc_args margs_read; 256 static struct objcache_malloc_args margs_write; 257 static struct objcache_malloc_args margs_vop; 258 259 int error; 260 261 error = 0; 262 kmalloc_raise_limit(M_HAMMER2, 0); /* unlimited */ 263 264 /* 265 * hammer2_xopgroups must be even and is most optimal if 266 * 2 x ncpus so strategy functions can be queued to the same 267 * cpu. 268 */ 269 hammer2_xopgroups = HAMMER2_XOPGROUPS_MIN; 270 if (hammer2_xopgroups < ncpus * 2) 271 hammer2_xopgroups = ncpus * 2; 272 273 /* 274 * A large DIO cache is needed to retain dedup enablement masks. 275 * The bulkfree code clears related masks as part of the disk block 276 * recycling algorithm, preventing it from being used for a later 277 * dedup. 278 * 279 * NOTE: A large buffer cache can actually interfere with dedup 280 * operation because we dedup based on media physical buffers 281 * and not logical buffers. Try to make the DIO case large 282 * enough to avoid this problem, but also cap it. 283 */ 284 hammer2_dio_limit = nbuf * 2; 285 if (hammer2_dio_limit > 100000) 286 hammer2_dio_limit = 100000; 287 288 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref)) 289 error = EINVAL; 290 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data)) 291 error = EINVAL; 292 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data)) 293 error = EINVAL; 294 295 if (error) 296 kprintf("HAMMER2 structure size mismatch; cannot continue.\n"); 297 298 margs_read.objsize = 65536; 299 margs_read.mtype = M_HAMMER2_DEBUFFER; 300 301 margs_write.objsize = 32768; 302 margs_write.mtype = M_HAMMER2_CBUFFER; 303 304 margs_vop.objsize = sizeof(hammer2_xop_t); 305 margs_vop.mtype = M_HAMMER2; 306 307 /* 308 * Note thaht for the XOPS cache we want backing store allocations 309 * to use M_ZERO. This is not allowed in objcache_get() (to avoid 310 * confusion), so use the backing store function that does it. This 311 * means that initial XOPS objects are zerod but REUSED objects are 312 * not. So we are responsible for cleaning the object up sufficiently 313 * for our needs before objcache_put()ing it back (typically just the 314 * FIFO indices). 315 */ 316 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc, 317 0, 1, NULL, NULL, NULL, 318 objcache_malloc_alloc, 319 objcache_malloc_free, 320 &margs_read); 321 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc, 322 0, 1, NULL, NULL, NULL, 323 objcache_malloc_alloc, 324 objcache_malloc_free, 325 &margs_write); 326 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc, 327 0, 1, NULL, NULL, NULL, 328 objcache_malloc_alloc_zero, 329 objcache_malloc_free, 330 &margs_vop); 331 332 333 lockinit(&hammer2_mntlk, "mntlk", 0, 0); 334 TAILQ_INIT(&hammer2_mntlist); 335 TAILQ_INIT(&hammer2_pfslist); 336 TAILQ_INIT(&hammer2_spmplist); 337 338 hammer2_limit_dirty_chains = maxvnodes / 10; 339 if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS) 340 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS; 341 if (hammer2_limit_dirty_chains < 1000) 342 hammer2_limit_dirty_chains = 1000; 343 344 hammer2_limit_dirty_inodes = maxvnodes / 25; 345 if (hammer2_limit_dirty_inodes < 100) 346 hammer2_limit_dirty_inodes = 100; 347 if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES) 348 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES; 349 350 return (error); 351 } 352 353 static 354 int 355 hammer2_vfs_uninit(struct vfsconf *vfsp __unused) 356 { 357 objcache_destroy(cache_buffer_read); 358 objcache_destroy(cache_buffer_write); 359 objcache_destroy(cache_xops); 360 return 0; 361 } 362 363 /* 364 * Core PFS allocator. Used to allocate or reference the pmp structure 365 * for PFS cluster mounts and the spmp structure for media (hmp) structures. 366 * The pmp can be passed in or loaded by this function using the chain and 367 * inode data. 368 * 369 * pmp->modify_tid tracks new modify_tid transaction ids for front-end 370 * transactions. Note that synchronization does not use this field. 371 * (typically frontend operations and synchronization cannot run on the 372 * same PFS node at the same time). 373 * 374 * XXX check locking 375 */ 376 hammer2_pfs_t * 377 hammer2_pfsalloc(hammer2_chain_t *chain, 378 const hammer2_inode_data_t *ripdata, 379 hammer2_tid_t modify_tid, hammer2_dev_t *force_local) 380 { 381 hammer2_pfs_t *pmp; 382 hammer2_inode_t *iroot; 383 int count; 384 int i; 385 int j; 386 387 pmp = NULL; 388 389 /* 390 * Locate or create the PFS based on the cluster id. If ripdata 391 * is NULL this is a spmp which is unique and is always allocated. 392 * 393 * If the device is mounted in local mode all PFSs are considered 394 * independent and not part of any cluster (for debugging only). 395 */ 396 if (ripdata) { 397 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 398 if (force_local != pmp->force_local) 399 continue; 400 if (force_local == NULL && 401 bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid, 402 sizeof(pmp->pfs_clid)) == 0) { 403 break; 404 } else if (force_local && pmp->pfs_names[0] && 405 strcmp(pmp->pfs_names[0], ripdata->filename) == 0) { 406 break; 407 } 408 } 409 } 410 411 if (pmp == NULL) { 412 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO); 413 pmp->force_local = force_local; 414 hammer2_trans_manage_init(pmp); 415 kmalloc_create(&pmp->minode, "HAMMER2-inodes"); 416 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg"); 417 lockinit(&pmp->lock, "pfslk", 0, 0); 418 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0); 419 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum"); 420 spin_init(&pmp->xop_spin, "h2xop"); 421 spin_init(&pmp->lru_spin, "h2lru"); 422 RB_INIT(&pmp->inum_tree); 423 TAILQ_INIT(&pmp->syncq); 424 TAILQ_INIT(&pmp->depq); 425 TAILQ_INIT(&pmp->lru_list); 426 spin_init(&pmp->list_spin, "h2pfsalloc_list"); 427 428 /* 429 * Save the last media transaction id for the flusher. Set 430 * initial 431 */ 432 if (ripdata) { 433 pmp->pfs_clid = ripdata->meta.pfs_clid; 434 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry); 435 } else { 436 pmp->flags |= HAMMER2_PMPF_SPMP; 437 TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry); 438 } 439 440 /* 441 * The synchronization thread may start too early, make 442 * sure it stays frozen until we are ready to let it go. 443 * XXX 444 */ 445 /* 446 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN | 447 HAMMER2_THREAD_REMASTER; 448 */ 449 } 450 451 /* 452 * Create the PFS's root inode and any missing XOP helper threads. 453 */ 454 if ((iroot = pmp->iroot) == NULL) { 455 iroot = hammer2_inode_get(pmp, NULL, 1, -1); 456 if (ripdata) 457 iroot->meta = ripdata->meta; 458 pmp->iroot = iroot; 459 hammer2_inode_ref(iroot); 460 hammer2_inode_unlock(iroot); 461 } 462 463 /* 464 * Stop here if no chain is passed in. 465 */ 466 if (chain == NULL) 467 goto done; 468 469 /* 470 * When a chain is passed in we must add it to the PFS's root 471 * inode, update pmp->pfs_types[], and update the syncronization 472 * threads. 473 * 474 * When forcing local mode, mark the PFS as a MASTER regardless. 475 * 476 * At the moment empty spots can develop due to removals or failures. 477 * Ultimately we want to re-fill these spots but doing so might 478 * confused running code. XXX 479 */ 480 hammer2_inode_ref(iroot); 481 hammer2_mtx_ex(&iroot->lock); 482 j = iroot->cluster.nchains; 483 484 if (j == HAMMER2_MAXCLUSTER) { 485 kprintf("hammer2_pfsalloc: cluster full!\n"); 486 /* XXX fatal error? */ 487 } else { 488 KKASSERT(chain->pmp == NULL); 489 chain->pmp = pmp; 490 hammer2_chain_ref(chain); 491 iroot->cluster.array[j].chain = chain; 492 if (force_local) 493 pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER; 494 else 495 pmp->pfs_types[j] = ripdata->meta.pfs_type; 496 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2); 497 pmp->pfs_hmps[j] = chain->hmp; 498 hammer2_spin_ex(&pmp->inum_spin); 499 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset; 500 hammer2_spin_unex(&pmp->inum_spin); 501 502 /* 503 * If the PFS is already mounted we must account 504 * for the mount_count here. 505 */ 506 if (pmp->mp) 507 ++chain->hmp->mount_count; 508 509 /* 510 * May have to fixup dirty chain tracking. Previous 511 * pmp was NULL so nothing to undo. 512 */ 513 if (chain->flags & HAMMER2_CHAIN_MODIFIED) 514 hammer2_pfs_memory_inc(pmp); 515 ++j; 516 } 517 iroot->cluster.nchains = j; 518 519 /* 520 * Update nmasters from any PFS inode which is part of the cluster. 521 * It is possible that this will result in a value which is too 522 * high. MASTER PFSs are authoritative for pfs_nmasters and will 523 * override this value later on. 524 * 525 * (This informs us of masters that might not currently be 526 * discoverable by this mount). 527 */ 528 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) { 529 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters; 530 } 531 532 /* 533 * Count visible masters. Masters are usually added with 534 * ripdata->meta.pfs_nmasters set to 1. This detects when there 535 * are more (XXX and must update the master inodes). 536 */ 537 count = 0; 538 for (i = 0; i < iroot->cluster.nchains; ++i) { 539 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) 540 ++count; 541 } 542 if (pmp->pfs_nmasters < count) 543 pmp->pfs_nmasters = count; 544 545 /* 546 * Create missing synchronization and support threads. 547 * 548 * Single-node masters (including snapshots) have nothing to 549 * synchronize and do not require this thread. 550 * 551 * Multi-node masters or any number of soft masters, slaves, copy, 552 * or other PFS types need the thread. 553 * 554 * Each thread is responsible for its particular cluster index. 555 * We use independent threads so stalls or mismatches related to 556 * any given target do not affect other targets. 557 */ 558 for (i = 0; i < iroot->cluster.nchains; ++i) { 559 /* 560 * Single-node masters (including snapshots) have nothing 561 * to synchronize and will make direct xops support calls, 562 * thus they do not require this thread. 563 * 564 * Note that there can be thousands of snapshots. We do not 565 * want to create thousands of threads. 566 */ 567 if (pmp->pfs_nmasters <= 1 && 568 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) { 569 continue; 570 } 571 572 /* 573 * Sync support thread 574 */ 575 if (pmp->sync_thrs[i].td == NULL) { 576 hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL, 577 "h2nod", i, -1, 578 hammer2_primary_sync_thread); 579 } 580 } 581 582 /* 583 * Create missing Xop threads 584 * 585 * NOTE: We create helper threads for all mounted PFSs or any 586 * PFSs with 2+ nodes (so the sync thread can update them, 587 * even if not mounted). 588 */ 589 if (pmp->mp || iroot->cluster.nchains >= 2) 590 hammer2_xop_helper_create(pmp); 591 592 hammer2_mtx_unlock(&iroot->lock); 593 hammer2_inode_drop(iroot); 594 done: 595 return pmp; 596 } 597 598 /* 599 * Deallocate an element of a probed PFS. If destroying and this is a 600 * MASTER, adjust nmasters. 601 * 602 * This function does not physically destroy the PFS element in its device 603 * under the super-root (see hammer2_ioctl_pfs_delete()). 604 */ 605 void 606 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying) 607 { 608 hammer2_inode_t *iroot; 609 hammer2_chain_t *chain; 610 int j; 611 612 /* 613 * Cleanup our reference on iroot. iroot is (should) not be needed 614 * by the flush code. 615 */ 616 iroot = pmp->iroot; 617 if (iroot) { 618 /* 619 * Stop synchronizing 620 * 621 * XXX flush after acquiring the iroot lock. 622 * XXX clean out the cluster index from all inode structures. 623 */ 624 hammer2_thr_delete(&pmp->sync_thrs[clindex]); 625 626 /* 627 * Remove the cluster index from the group. If destroying 628 * the PFS and this is a master, adjust pfs_nmasters. 629 */ 630 hammer2_mtx_ex(&iroot->lock); 631 chain = iroot->cluster.array[clindex].chain; 632 iroot->cluster.array[clindex].chain = NULL; 633 634 switch(pmp->pfs_types[clindex]) { 635 case HAMMER2_PFSTYPE_MASTER: 636 if (destroying && pmp->pfs_nmasters > 0) 637 --pmp->pfs_nmasters; 638 /* XXX adjust ripdata->meta.pfs_nmasters */ 639 break; 640 default: 641 break; 642 } 643 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE; 644 645 hammer2_mtx_unlock(&iroot->lock); 646 647 /* 648 * Release the chain. 649 */ 650 if (chain) { 651 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 652 hammer2_chain_drop(chain); 653 } 654 655 /* 656 * Terminate all XOP threads for the cluster index. 657 */ 658 if (pmp->xop_groups) { 659 for (j = 0; j < hammer2_xopgroups; ++j) { 660 hammer2_thr_delete( 661 &pmp->xop_groups[j].thrs[clindex]); 662 } 663 } 664 } 665 } 666 667 /* 668 * Destroy a PFS, typically only occurs after the last mount on a device 669 * has gone away. 670 */ 671 static void 672 hammer2_pfsfree(hammer2_pfs_t *pmp) 673 { 674 hammer2_inode_t *iroot; 675 hammer2_chain_t *chain; 676 int chains_still_present = 0; 677 int i; 678 int j; 679 680 /* 681 * Cleanup our reference on iroot. iroot is (should) not be needed 682 * by the flush code. 683 */ 684 if (pmp->flags & HAMMER2_PMPF_SPMP) 685 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry); 686 else 687 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry); 688 689 /* 690 * Cleanup chains remaining on LRU list. 691 */ 692 hammer2_spin_ex(&pmp->lru_spin); 693 while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) { 694 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU); 695 atomic_add_int(&pmp->lru_count, -1); 696 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU); 697 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node); 698 hammer2_chain_ref(chain); 699 hammer2_spin_unex(&pmp->lru_spin); 700 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 701 hammer2_chain_drop(chain); 702 hammer2_spin_ex(&pmp->lru_spin); 703 } 704 hammer2_spin_unex(&pmp->lru_spin); 705 706 /* 707 * Clean up iroot 708 */ 709 iroot = pmp->iroot; 710 if (iroot) { 711 for (i = 0; i < iroot->cluster.nchains; ++i) { 712 hammer2_thr_delete(&pmp->sync_thrs[i]); 713 if (pmp->xop_groups) { 714 for (j = 0; j < hammer2_xopgroups; ++j) 715 hammer2_thr_delete( 716 &pmp->xop_groups[j].thrs[i]); 717 } 718 chain = iroot->cluster.array[i].chain; 719 if (chain && !RB_EMPTY(&chain->core.rbtree)) { 720 kprintf("hammer2: Warning pmp %p still " 721 "has active chains\n", pmp); 722 chains_still_present = 1; 723 } 724 } 725 #if REPORT_REFS_ERRORS 726 if (iroot->refs != 1) 727 kprintf("PMP->IROOT %p REFS WRONG %d\n", 728 iroot, iroot->refs); 729 #else 730 KKASSERT(iroot->refs == 1); 731 #endif 732 /* ref for iroot */ 733 hammer2_inode_drop(iroot); 734 pmp->iroot = NULL; 735 } 736 737 /* 738 * Free remaining pmp resources 739 */ 740 if (chains_still_present) { 741 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp); 742 } else { 743 kmalloc_destroy(&pmp->mmsg); 744 kmalloc_destroy(&pmp->minode); 745 kfree(pmp, M_HAMMER2); 746 } 747 } 748 749 /* 750 * Remove all references to hmp from the pfs list. Any PFS which becomes 751 * empty is terminated and freed. 752 * 753 * XXX inefficient. 754 */ 755 static void 756 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which) 757 { 758 hammer2_pfs_t *pmp; 759 hammer2_inode_t *iroot; 760 hammer2_chain_t *rchain; 761 int i; 762 int j; 763 struct hammer2_pfslist *wlist; 764 765 if (which == 0) 766 wlist = &hammer2_pfslist; 767 else 768 wlist = &hammer2_spmplist; 769 again: 770 TAILQ_FOREACH(pmp, wlist, mntentry) { 771 if ((iroot = pmp->iroot) == NULL) 772 continue; 773 774 /* 775 * Determine if this PFS is affected. If it is we must 776 * freeze all management threads and lock its iroot. 777 * 778 * Freezing a management thread forces it idle, operations 779 * in-progress will be aborted and it will have to start 780 * over again when unfrozen, or exit if told to exit. 781 */ 782 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 783 if (pmp->pfs_hmps[i] == hmp) 784 break; 785 } 786 if (i == HAMMER2_MAXCLUSTER) 787 continue; 788 789 hammer2_vfs_sync_pmp(pmp, MNT_WAIT); 790 791 /* 792 * Make sure all synchronization threads are locked 793 * down. 794 */ 795 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 796 if (pmp->pfs_hmps[i] == NULL) 797 continue; 798 hammer2_thr_freeze_async(&pmp->sync_thrs[i]); 799 if (pmp->xop_groups) { 800 for (j = 0; j < hammer2_xopgroups; ++j) { 801 hammer2_thr_freeze_async( 802 &pmp->xop_groups[j].thrs[i]); 803 } 804 } 805 } 806 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 807 if (pmp->pfs_hmps[i] == NULL) 808 continue; 809 hammer2_thr_freeze(&pmp->sync_thrs[i]); 810 if (pmp->xop_groups) { 811 for (j = 0; j < hammer2_xopgroups; ++j) { 812 hammer2_thr_freeze( 813 &pmp->xop_groups[j].thrs[i]); 814 } 815 } 816 } 817 818 /* 819 * Lock the inode and clean out matching chains. 820 * Note that we cannot use hammer2_inode_lock_*() 821 * here because that would attempt to validate the 822 * cluster that we are in the middle of ripping 823 * apart. 824 * 825 * WARNING! We are working directly on the inodes 826 * embedded cluster. 827 */ 828 hammer2_mtx_ex(&iroot->lock); 829 830 /* 831 * Remove the chain from matching elements of the PFS. 832 */ 833 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 834 if (pmp->pfs_hmps[i] != hmp) 835 continue; 836 hammer2_thr_delete(&pmp->sync_thrs[i]); 837 if (pmp->xop_groups) { 838 for (j = 0; j < hammer2_xopgroups; ++j) { 839 hammer2_thr_delete( 840 &pmp->xop_groups[j].thrs[i]); 841 } 842 } 843 rchain = iroot->cluster.array[i].chain; 844 iroot->cluster.array[i].chain = NULL; 845 pmp->pfs_types[i] = 0; 846 if (pmp->pfs_names[i]) { 847 kfree(pmp->pfs_names[i], M_HAMMER2); 848 pmp->pfs_names[i] = NULL; 849 } 850 if (rchain) { 851 hammer2_chain_drop(rchain); 852 /* focus hint */ 853 if (iroot->cluster.focus == rchain) 854 iroot->cluster.focus = NULL; 855 } 856 pmp->pfs_hmps[i] = NULL; 857 } 858 hammer2_mtx_unlock(&iroot->lock); 859 860 /* 861 * Cleanup trailing chains. Gaps may remain. 862 */ 863 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) { 864 if (pmp->pfs_hmps[i]) 865 break; 866 } 867 iroot->cluster.nchains = i + 1; 868 869 /* 870 * If the PMP has no elements remaining we can destroy it. 871 * (this will transition management threads from frozen->exit). 872 */ 873 if (iroot->cluster.nchains == 0) { 874 /* 875 * If this was the hmp's spmp, we need to clean 876 * a little more stuff out. 877 */ 878 if (hmp->spmp == pmp) { 879 hmp->spmp = NULL; 880 hmp->vchain.pmp = NULL; 881 hmp->fchain.pmp = NULL; 882 } 883 884 /* 885 * Free the pmp and restart the loop 886 */ 887 KKASSERT(TAILQ_EMPTY(&pmp->syncq)); 888 KKASSERT(TAILQ_EMPTY(&pmp->depq)); 889 hammer2_pfsfree(pmp); 890 goto again; 891 } 892 893 /* 894 * If elements still remain we need to set the REMASTER 895 * flag and unfreeze it. 896 */ 897 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 898 if (pmp->pfs_hmps[i] == NULL) 899 continue; 900 hammer2_thr_remaster(&pmp->sync_thrs[i]); 901 hammer2_thr_unfreeze(&pmp->sync_thrs[i]); 902 if (pmp->xop_groups) { 903 for (j = 0; j < hammer2_xopgroups; ++j) { 904 hammer2_thr_remaster( 905 &pmp->xop_groups[j].thrs[i]); 906 hammer2_thr_unfreeze( 907 &pmp->xop_groups[j].thrs[i]); 908 } 909 } 910 } 911 } 912 } 913 914 /* 915 * Mount or remount HAMMER2 fileystem from physical media 916 * 917 * mountroot 918 * mp mount point structure 919 * path NULL 920 * data <unused> 921 * cred <unused> 922 * 923 * mount 924 * mp mount point structure 925 * path path to mount point 926 * data pointer to argument structure in user space 927 * volume volume path (device@LABEL form) 928 * hflags user mount flags 929 * cred user credentials 930 * 931 * RETURNS: 0 Success 932 * !0 error number 933 */ 934 static 935 int 936 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data, 937 struct ucred *cred) 938 { 939 struct hammer2_mount_info info; 940 hammer2_pfs_t *pmp; 941 hammer2_pfs_t *spmp; 942 hammer2_dev_t *hmp; 943 hammer2_dev_t *force_local; 944 hammer2_key_t key_next; 945 hammer2_key_t key_dummy; 946 hammer2_key_t lhc; 947 struct vnode *devvp; 948 struct nlookupdata nd; 949 hammer2_chain_t *parent; 950 hammer2_chain_t *chain; 951 const hammer2_inode_data_t *ripdata; 952 hammer2_blockref_t bref; 953 struct file *fp; 954 char devstr[MNAMELEN]; 955 size_t size; 956 size_t done; 957 char *dev; 958 char *label; 959 int ronly = 1; 960 int error; 961 int i; 962 963 hmp = NULL; 964 pmp = NULL; 965 dev = NULL; 966 label = NULL; 967 devvp = NULL; 968 bzero(&info, sizeof(info)); 969 970 if (path) { 971 /* 972 * Non-root mount or updating a mount 973 */ 974 error = copyin(data, &info, sizeof(info)); 975 if (error) 976 return (error); 977 } 978 979 if (mp->mnt_flag & MNT_UPDATE) { 980 /* 981 * Update mount. Note that pmp->iroot->cluster is 982 * an inode-embedded cluster and thus cannot be 983 * directly locked. 984 * 985 * XXX HAMMER2 needs to implement NFS export via 986 * mountctl. 987 */ 988 hammer2_cluster_t *cluster; 989 990 pmp = MPTOPMP(mp); 991 pmp->hflags = info.hflags; 992 cluster = &pmp->iroot->cluster; 993 for (i = 0; i < cluster->nchains; ++i) { 994 if (cluster->array[i].chain == NULL) 995 continue; 996 hmp = cluster->array[i].chain->hmp; 997 devvp = hmp->devvp; 998 error = hammer2_remount(hmp, mp, path, 999 devvp, cred); 1000 if (error) 1001 break; 1002 } 1003 1004 return error; 1005 } 1006 1007 if (path == NULL) { 1008 /* 1009 * Root mount 1010 */ 1011 info.cluster_fd = -1; 1012 ksnprintf(devstr, sizeof(devstr), "%s", 1013 mp->mnt_stat.f_mntfromname); 1014 done = strlen(devstr) + 1; 1015 kprintf("hammer2_mount: root devstr=\"%s\"\n", devstr); 1016 } else { 1017 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done); 1018 if (error) 1019 return (error); 1020 kprintf("hammer2_mount: devstr=\"%s\"\n", devstr); 1021 } 1022 1023 /* 1024 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA 1025 * if no label specified, based on the partition id. Error out if no 1026 * label or device (with partition id) is specified. This is strictly 1027 * a convenience to match the default label created by newfs_hammer2, 1028 * our preference is that a label always be specified. 1029 * 1030 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command 1031 * that does not specify a device, as long as some H2 label 1032 * has already been mounted from that device. This makes 1033 * mounting snapshots a lot easier. 1034 */ 1035 dev = devstr; 1036 label = strchr(devstr, '@'); 1037 if (label && ((label + 1) - dev) > done) { 1038 kprintf("hammer2: mount: bad label %s/%zd\n", devstr, done); 1039 return (EINVAL); 1040 } 1041 if (label == NULL || label[1] == 0) { 1042 char slice; 1043 1044 if (label == NULL) 1045 label = devstr + strlen(devstr); 1046 else 1047 *label = '\0'; /* clean up trailing @ */ 1048 1049 slice = label[-1]; 1050 switch(slice) { 1051 case 'a': 1052 label = "BOOT"; 1053 break; 1054 case 'd': 1055 label = "ROOT"; 1056 break; 1057 default: 1058 label = "DATA"; 1059 break; 1060 } 1061 } else { 1062 *label = '\0'; 1063 label++; 1064 } 1065 1066 kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n", 1067 dev, label, (mp->mnt_flag & MNT_RDONLY)); 1068 1069 /* 1070 * HMP device mount 1071 * 1072 * If a path is specified and dev is not an empty string, lookup the 1073 * name and verify that it referes to a block device. 1074 * 1075 * If a path is specified and dev is an empty string we fall through 1076 * and locate the label in the hmp search. 1077 */ 1078 if (path && *dev != 0) { 1079 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW); 1080 if (error == 0) 1081 error = nlookup(&nd); 1082 if (error == 0) 1083 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp); 1084 nlookup_done(&nd); 1085 } else if (path == NULL) { 1086 /* root mount */ 1087 cdev_t cdev = kgetdiskbyname(dev); 1088 error = bdevvp(cdev, &devvp); 1089 if (error) 1090 kprintf("hammer2: cannot find '%s'\n", dev); 1091 } else { 1092 /* 1093 * We will locate the hmp using the label in the hmp loop. 1094 */ 1095 error = 0; 1096 } 1097 1098 /* 1099 * Make sure its a block device. Do not check to see if it is 1100 * already mounted until we determine that its a fresh H2 device. 1101 */ 1102 if (error == 0 && devvp) { 1103 vn_isdisk(devvp, &error); 1104 } 1105 1106 /* 1107 * Determine if the device has already been mounted. After this 1108 * check hmp will be non-NULL if we are doing the second or more 1109 * hammer2 mounts from the same device. 1110 */ 1111 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1112 if (devvp) { 1113 /* 1114 * Match the device. Due to the way devfs works, 1115 * we may not be able to directly match the vnode pointer, 1116 * so also check to see if the underlying device matches. 1117 */ 1118 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1119 if (hmp->devvp == devvp) 1120 break; 1121 if (devvp->v_rdev && 1122 hmp->devvp->v_rdev == devvp->v_rdev) { 1123 break; 1124 } 1125 } 1126 1127 /* 1128 * If no match this may be a fresh H2 mount, make sure 1129 * the device is not mounted on anything else. 1130 */ 1131 if (hmp == NULL) 1132 error = vfs_mountedon(devvp); 1133 } else if (error == 0) { 1134 /* 1135 * Match the label to a pmp already probed. 1136 */ 1137 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 1138 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 1139 if (pmp->pfs_names[i] && 1140 strcmp(pmp->pfs_names[i], label) == 0) { 1141 hmp = pmp->pfs_hmps[i]; 1142 break; 1143 } 1144 } 1145 if (hmp) 1146 break; 1147 } 1148 if (hmp == NULL) { 1149 lockmgr(&hammer2_mntlk, LK_RELEASE); 1150 kprintf("hammer2_mount: PFS label \"%s\" not found\n", 1151 label); 1152 return ENOENT; 1153 } 1154 } 1155 1156 /* 1157 * Open the device if this isn't a secondary mount and construct 1158 * the H2 device mount (hmp). 1159 */ 1160 if (hmp == NULL) { 1161 hammer2_chain_t *schain; 1162 hammer2_xid_t xid; 1163 hammer2_xop_head_t xop; 1164 1165 if (error == 0 && vcount(devvp) > 0) { 1166 kprintf("Primary device already has references\n"); 1167 error = EBUSY; 1168 } 1169 1170 /* 1171 * Now open the device 1172 */ 1173 if (error == 0) { 1174 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0); 1175 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1176 error = vinvalbuf(devvp, V_SAVE, 0, 0); 1177 if (error == 0) { 1178 error = VOP_OPEN(devvp, 1179 (ronly ? FREAD : FREAD | FWRITE), 1180 FSCRED, NULL); 1181 } 1182 vn_unlock(devvp); 1183 } 1184 if (error && devvp) { 1185 vrele(devvp); 1186 devvp = NULL; 1187 } 1188 if (error) { 1189 lockmgr(&hammer2_mntlk, LK_RELEASE); 1190 return error; 1191 } 1192 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO); 1193 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev); 1194 hmp->ronly = ronly; 1195 hmp->devvp = devvp; 1196 hmp->hflags = info.hflags & HMNT2_DEVFLAGS; 1197 kmalloc_create(&hmp->mchain, "HAMMER2-chains"); 1198 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry); 1199 RB_INIT(&hmp->iotree); 1200 spin_init(&hmp->io_spin, "h2mount_io"); 1201 spin_init(&hmp->list_spin, "h2mount_list"); 1202 1203 lockinit(&hmp->vollk, "h2vol", 0, 0); 1204 lockinit(&hmp->bulklk, "h2bulk", 0, 0); 1205 lockinit(&hmp->bflock, "h2bflk", 0, 0); 1206 1207 /* 1208 * vchain setup. vchain.data is embedded. 1209 * vchain.refs is initialized and will never drop to 0. 1210 * 1211 * NOTE! voldata is not yet loaded. 1212 */ 1213 hmp->vchain.hmp = hmp; 1214 hmp->vchain.refs = 1; 1215 hmp->vchain.data = (void *)&hmp->voldata; 1216 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME; 1217 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1218 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1219 hammer2_chain_core_init(&hmp->vchain); 1220 1221 /* 1222 * fchain setup. fchain.data is embedded. 1223 * fchain.refs is initialized and will never drop to 0. 1224 * 1225 * The data is not used but needs to be initialized to 1226 * pass assertion muster. We use this chain primarily 1227 * as a placeholder for the freemap's top-level radix tree 1228 * so it does not interfere with the volume's topology 1229 * radix tree. 1230 */ 1231 hmp->fchain.hmp = hmp; 1232 hmp->fchain.refs = 1; 1233 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset; 1234 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP; 1235 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1236 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1237 hmp->fchain.bref.methods = 1238 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) | 1239 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE); 1240 hammer2_chain_core_init(&hmp->fchain); 1241 1242 /* 1243 * Install the volume header and initialize fields from 1244 * voldata. 1245 */ 1246 error = hammer2_install_volume_header(hmp); 1247 if (error) { 1248 hammer2_unmount_helper(mp, NULL, hmp); 1249 lockmgr(&hammer2_mntlk, LK_RELEASE); 1250 hammer2_vfs_unmount(mp, MNT_FORCE); 1251 return error; 1252 } 1253 1254 /* 1255 * Really important to get these right or the flush and 1256 * teardown code will get confused. 1257 */ 1258 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL); 1259 spmp = hmp->spmp; 1260 spmp->pfs_hmps[0] = hmp; 1261 1262 /* 1263 * Dummy-up vchain and fchain's modify_tid. mirror_tid 1264 * is inherited from the volume header. 1265 */ 1266 xid = 0; 1267 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1268 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid; 1269 hmp->vchain.pmp = spmp; 1270 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1271 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid; 1272 hmp->fchain.pmp = spmp; 1273 1274 /* 1275 * First locate the super-root inode, which is key 0 1276 * relative to the volume header's blockset. 1277 * 1278 * Then locate the root inode by scanning the directory keyspace 1279 * represented by the label. 1280 */ 1281 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 1282 schain = hammer2_chain_lookup(&parent, &key_dummy, 1283 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY, 1284 &error, 0); 1285 hammer2_chain_lookup_done(parent); 1286 if (schain == NULL) { 1287 kprintf("hammer2_mount: invalid super-root\n"); 1288 hammer2_unmount_helper(mp, NULL, hmp); 1289 lockmgr(&hammer2_mntlk, LK_RELEASE); 1290 hammer2_vfs_unmount(mp, MNT_FORCE); 1291 return EINVAL; 1292 } 1293 if (schain->error) { 1294 kprintf("hammer2_mount: error %s reading super-root\n", 1295 hammer2_error_str(schain->error)); 1296 hammer2_chain_unlock(schain); 1297 hammer2_chain_drop(schain); 1298 schain = NULL; 1299 hammer2_unmount_helper(mp, NULL, hmp); 1300 lockmgr(&hammer2_mntlk, LK_RELEASE); 1301 hammer2_vfs_unmount(mp, MNT_FORCE); 1302 return EINVAL; 1303 } 1304 1305 /* 1306 * The super-root always uses an inode_tid of 1 when 1307 * creating PFSs. 1308 */ 1309 spmp->inode_tid = 1; 1310 spmp->modify_tid = schain->bref.modify_tid + 1; 1311 1312 /* 1313 * Sanity-check schain's pmp and finish initialization. 1314 * Any chain belonging to the super-root topology should 1315 * have a NULL pmp (not even set to spmp). 1316 */ 1317 ripdata = &schain->data->ipdata; 1318 KKASSERT(schain->pmp == NULL); 1319 spmp->pfs_clid = ripdata->meta.pfs_clid; 1320 1321 /* 1322 * Replace the dummy spmp->iroot with a real one. It's 1323 * easier to just do a wholesale replacement than to try 1324 * to update the chain and fixup the iroot fields. 1325 * 1326 * The returned inode is locked with the supplied cluster. 1327 */ 1328 hammer2_dummy_xop_from_chain(&xop, schain); 1329 hammer2_inode_drop(spmp->iroot); 1330 spmp->iroot = NULL; 1331 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1); 1332 spmp->spmp_hmp = hmp; 1333 spmp->pfs_types[0] = ripdata->meta.pfs_type; 1334 spmp->pfs_hmps[0] = hmp; 1335 hammer2_inode_ref(spmp->iroot); 1336 hammer2_inode_unlock(spmp->iroot); 1337 hammer2_cluster_unlock(&xop.cluster); 1338 hammer2_chain_drop(schain); 1339 /* do not call hammer2_cluster_drop() on an embedded cluster */ 1340 schain = NULL; /* now invalid */ 1341 /* leave spmp->iroot with one ref */ 1342 1343 if ((mp->mnt_flag & MNT_RDONLY) == 0) { 1344 error = hammer2_recovery(hmp); 1345 if (error == 0) 1346 error |= hammer2_fixup_pfses(hmp); 1347 /* XXX do something with error */ 1348 } 1349 hammer2_update_pmps(hmp); 1350 hammer2_iocom_init(hmp); 1351 hammer2_bulkfree_init(hmp); 1352 1353 /* 1354 * Ref the cluster management messaging descriptor. The mount 1355 * program deals with the other end of the communications pipe. 1356 * 1357 * Root mounts typically do not supply one. 1358 */ 1359 if (info.cluster_fd >= 0) { 1360 fp = holdfp(curthread, info.cluster_fd, -1); 1361 if (fp) { 1362 hammer2_cluster_reconnect(hmp, fp); 1363 } else { 1364 kprintf("hammer2_mount: bad cluster_fd!\n"); 1365 } 1366 } 1367 } else { 1368 spmp = hmp->spmp; 1369 if (info.hflags & HMNT2_DEVFLAGS) { 1370 kprintf("hammer2: Warning: mount flags pertaining " 1371 "to the whole device may only be specified " 1372 "on the first mount of the device: %08x\n", 1373 info.hflags & HMNT2_DEVFLAGS); 1374 } 1375 } 1376 1377 /* 1378 * Force local mount (disassociate all PFSs from their clusters). 1379 * Used primarily for debugging. 1380 */ 1381 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1382 1383 /* 1384 * Lookup the mount point under the media-localized super-root. 1385 * Scanning hammer2_pfslist doesn't help us because it represents 1386 * PFS cluster ids which can aggregate several named PFSs together. 1387 * 1388 * cluster->pmp will incorrectly point to spmp and must be fixed 1389 * up later on. 1390 */ 1391 hammer2_inode_lock(spmp->iroot, 0); 1392 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1393 lhc = hammer2_dirhash(label, strlen(label)); 1394 chain = hammer2_chain_lookup(&parent, &key_next, 1395 lhc, lhc + HAMMER2_DIRHASH_LOMASK, 1396 &error, 0); 1397 while (chain) { 1398 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && 1399 strcmp(label, chain->data->ipdata.filename) == 0) { 1400 break; 1401 } 1402 chain = hammer2_chain_next(&parent, chain, &key_next, 1403 key_next, 1404 lhc + HAMMER2_DIRHASH_LOMASK, 1405 &error, 0); 1406 } 1407 if (parent) { 1408 hammer2_chain_unlock(parent); 1409 hammer2_chain_drop(parent); 1410 } 1411 hammer2_inode_unlock(spmp->iroot); 1412 1413 /* 1414 * PFS could not be found? 1415 */ 1416 if (chain == NULL) { 1417 hammer2_unmount_helper(mp, NULL, hmp); 1418 lockmgr(&hammer2_mntlk, LK_RELEASE); 1419 hammer2_vfs_unmount(mp, MNT_FORCE); 1420 1421 if (error) { 1422 kprintf("hammer2_mount: PFS label I/O error\n"); 1423 return EINVAL; 1424 } else { 1425 kprintf("hammer2_mount: PFS label \"%s\" not found\n", 1426 label); 1427 return ENOENT; 1428 } 1429 } 1430 1431 /* 1432 * Acquire the pmp structure (it should have already been allocated 1433 * via hammer2_update_pmps() so do not pass cluster in to add to 1434 * available chains). 1435 * 1436 * Check if the cluster has already been mounted. A cluster can 1437 * only be mounted once, use null mounts to mount additional copies. 1438 */ 1439 if (chain->error) { 1440 kprintf("hammer2_mount: PFS label I/O error\n"); 1441 } else { 1442 ripdata = &chain->data->ipdata; 1443 bref = chain->bref; 1444 pmp = hammer2_pfsalloc(NULL, ripdata, 1445 bref.modify_tid, force_local); 1446 } 1447 hammer2_chain_unlock(chain); 1448 hammer2_chain_drop(chain); 1449 1450 /* 1451 * Finish the mount 1452 */ 1453 kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp); 1454 1455 if (pmp->mp) { 1456 kprintf("hammer2_mount: PFS already mounted!\n"); 1457 hammer2_unmount_helper(mp, NULL, hmp); 1458 lockmgr(&hammer2_mntlk, LK_RELEASE); 1459 hammer2_vfs_unmount(mp, MNT_FORCE); 1460 1461 return EBUSY; 1462 } 1463 1464 pmp->hflags = info.hflags; 1465 mp->mnt_flag |= MNT_LOCAL; 1466 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */ 1467 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */ 1468 1469 /* 1470 * required mount structure initializations 1471 */ 1472 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE; 1473 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE; 1474 1475 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE; 1476 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE; 1477 1478 /* 1479 * Optional fields 1480 */ 1481 mp->mnt_iosize_max = MAXPHYS; 1482 1483 /* 1484 * Connect up mount pointers. 1485 */ 1486 hammer2_mount_helper(mp, pmp); 1487 hmp->devvp->v_rdev->si_mountpoint = mp; 1488 1489 lockmgr(&hammer2_mntlk, LK_RELEASE); 1490 1491 /* 1492 * Finish setup 1493 */ 1494 vfs_getnewfsid(mp); 1495 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops); 1496 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops); 1497 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops); 1498 1499 if (path) { 1500 copyinstr(info.volume, mp->mnt_stat.f_mntfromname, 1501 MNAMELEN - 1, &size); 1502 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size); 1503 } /* else root mount, already in there */ 1504 1505 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname)); 1506 if (path) { 1507 copyinstr(path, mp->mnt_stat.f_mntonname, 1508 sizeof(mp->mnt_stat.f_mntonname) - 1, 1509 &size); 1510 } else { 1511 /* root mount */ 1512 mp->mnt_stat.f_mntonname[0] = '/'; 1513 } 1514 1515 /* 1516 * Initial statfs to prime mnt_stat. 1517 */ 1518 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred); 1519 1520 return 0; 1521 } 1522 1523 /* 1524 * Scan PFSs under the super-root and create hammer2_pfs structures. 1525 */ 1526 static 1527 void 1528 hammer2_update_pmps(hammer2_dev_t *hmp) 1529 { 1530 const hammer2_inode_data_t *ripdata; 1531 hammer2_chain_t *parent; 1532 hammer2_chain_t *chain; 1533 hammer2_blockref_t bref; 1534 hammer2_dev_t *force_local; 1535 hammer2_pfs_t *spmp; 1536 hammer2_pfs_t *pmp; 1537 hammer2_key_t key_next; 1538 int error; 1539 1540 /* 1541 * Force local mount (disassociate all PFSs from their clusters). 1542 * Used primarily for debugging. 1543 */ 1544 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1545 1546 /* 1547 * Lookup mount point under the media-localized super-root. 1548 * 1549 * cluster->pmp will incorrectly point to spmp and must be fixed 1550 * up later on. 1551 */ 1552 spmp = hmp->spmp; 1553 hammer2_inode_lock(spmp->iroot, 0); 1554 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1555 chain = hammer2_chain_lookup(&parent, &key_next, 1556 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 1557 &error, 0); 1558 while (chain) { 1559 if (chain->error) { 1560 kprintf("I/O error scanning PFS labels\n"); 1561 } else if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) { 1562 kprintf("Non inode chain type %d under super-root\n", 1563 chain->bref.type); 1564 } else { 1565 ripdata = &chain->data->ipdata; 1566 bref = chain->bref; 1567 pmp = hammer2_pfsalloc(chain, ripdata, 1568 bref.modify_tid, force_local); 1569 } 1570 chain = hammer2_chain_next(&parent, chain, &key_next, 1571 key_next, HAMMER2_KEY_MAX, 1572 &error, 0); 1573 } 1574 if (parent) { 1575 hammer2_chain_unlock(parent); 1576 hammer2_chain_drop(parent); 1577 } 1578 hammer2_inode_unlock(spmp->iroot); 1579 } 1580 1581 static 1582 int 1583 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused, 1584 struct vnode *devvp, struct ucred *cred) 1585 { 1586 int error; 1587 1588 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) { 1589 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1590 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL); 1591 vn_unlock(devvp); 1592 error = hammer2_recovery(hmp); 1593 if (error == 0) 1594 error |= hammer2_fixup_pfses(hmp); 1595 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1596 if (error == 0) { 1597 VOP_CLOSE(devvp, FREAD, NULL); 1598 hmp->ronly = 0; 1599 } else { 1600 VOP_CLOSE(devvp, FREAD | FWRITE, NULL); 1601 } 1602 vn_unlock(devvp); 1603 } else { 1604 error = 0; 1605 } 1606 return error; 1607 } 1608 1609 static 1610 int 1611 hammer2_vfs_unmount(struct mount *mp, int mntflags) 1612 { 1613 hammer2_pfs_t *pmp; 1614 int flags; 1615 int error = 0; 1616 1617 pmp = MPTOPMP(mp); 1618 1619 if (pmp == NULL) 1620 return(0); 1621 1622 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1623 1624 /* 1625 * If mount initialization proceeded far enough we must flush 1626 * its vnodes and sync the underlying mount points. Three syncs 1627 * are required to fully flush the filesystem (freemap updates lag 1628 * by one flush, and one extra for safety). 1629 */ 1630 if (mntflags & MNT_FORCE) 1631 flags = FORCECLOSE; 1632 else 1633 flags = 0; 1634 if (pmp->iroot) { 1635 error = vflush(mp, 0, flags); 1636 if (error) 1637 goto failed; 1638 hammer2_vfs_sync(mp, MNT_WAIT); 1639 hammer2_vfs_sync(mp, MNT_WAIT); 1640 hammer2_vfs_sync(mp, MNT_WAIT); 1641 } 1642 1643 /* 1644 * Cleanup the frontend support XOPS threads 1645 */ 1646 hammer2_xop_helper_cleanup(pmp); 1647 1648 if (pmp->mp) 1649 hammer2_unmount_helper(mp, pmp, NULL); 1650 1651 error = 0; 1652 failed: 1653 lockmgr(&hammer2_mntlk, LK_RELEASE); 1654 1655 return (error); 1656 } 1657 1658 /* 1659 * Mount helper, hook the system mount into our PFS. 1660 * The mount lock is held. 1661 * 1662 * We must bump the mount_count on related devices for any 1663 * mounted PFSs. 1664 */ 1665 static 1666 void 1667 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp) 1668 { 1669 hammer2_cluster_t *cluster; 1670 hammer2_chain_t *rchain; 1671 int i; 1672 1673 mp->mnt_data = (qaddr_t)pmp; 1674 pmp->mp = mp; 1675 1676 /* 1677 * After pmp->mp is set we have to adjust hmp->mount_count. 1678 */ 1679 cluster = &pmp->iroot->cluster; 1680 for (i = 0; i < cluster->nchains; ++i) { 1681 rchain = cluster->array[i].chain; 1682 if (rchain == NULL) 1683 continue; 1684 ++rchain->hmp->mount_count; 1685 } 1686 1687 /* 1688 * Create missing Xop threads 1689 */ 1690 hammer2_xop_helper_create(pmp); 1691 } 1692 1693 /* 1694 * Mount helper, unhook the system mount from our PFS. 1695 * The mount lock is held. 1696 * 1697 * If hmp is supplied a mount responsible for being the first to open 1698 * the block device failed and the block device and all PFSs using the 1699 * block device must be cleaned up. 1700 * 1701 * If pmp is supplied multiple devices might be backing the PFS and each 1702 * must be disconnected. This might not be the last PFS using some of the 1703 * underlying devices. Also, we have to adjust our hmp->mount_count 1704 * accounting for the devices backing the pmp which is now undergoing an 1705 * unmount. 1706 */ 1707 static 1708 void 1709 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp) 1710 { 1711 hammer2_cluster_t *cluster; 1712 hammer2_chain_t *rchain; 1713 struct vnode *devvp; 1714 int dumpcnt; 1715 int ronly; 1716 int i; 1717 1718 /* 1719 * If no device supplied this is a high-level unmount and we have to 1720 * to disconnect the mount, adjust mount_count, and locate devices 1721 * that might now have no mounts. 1722 */ 1723 if (pmp) { 1724 KKASSERT(hmp == NULL); 1725 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp); 1726 pmp->mp = NULL; 1727 mp->mnt_data = NULL; 1728 1729 /* 1730 * After pmp->mp is cleared we have to account for 1731 * mount_count. 1732 */ 1733 cluster = &pmp->iroot->cluster; 1734 for (i = 0; i < cluster->nchains; ++i) { 1735 rchain = cluster->array[i].chain; 1736 if (rchain == NULL) 1737 continue; 1738 --rchain->hmp->mount_count; 1739 /* scrapping hmp now may invalidate the pmp */ 1740 } 1741 again: 1742 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1743 if (hmp->mount_count == 0) { 1744 hammer2_unmount_helper(NULL, NULL, hmp); 1745 goto again; 1746 } 1747 } 1748 return; 1749 } 1750 1751 /* 1752 * Try to terminate the block device. We can't terminate it if 1753 * there are still PFSs referencing it. 1754 */ 1755 if (hmp->mount_count) 1756 return; 1757 1758 /* 1759 * Decomission the network before we start messing with the 1760 * device and PFS. 1761 */ 1762 hammer2_iocom_uninit(hmp); 1763 1764 hammer2_bulkfree_uninit(hmp); 1765 hammer2_pfsfree_scan(hmp, 0); 1766 1767 /* 1768 * Cycle the volume data lock as a safety (probably not needed any 1769 * more). To ensure everything is out we need to flush at least 1770 * three times. (1) The running of the sideq can dirty the 1771 * filesystem, (2) A normal flush can dirty the freemap, and 1772 * (3) ensure that the freemap is fully synchronized. 1773 * 1774 * The next mount's recovery scan can clean everything up but we want 1775 * to leave the filesystem in a 100% clean state on a normal unmount. 1776 */ 1777 #if 0 1778 hammer2_voldata_lock(hmp); 1779 hammer2_voldata_unlock(hmp); 1780 #endif 1781 1782 /* 1783 * Flush whatever is left. Unmounted but modified PFS's might still 1784 * have some dirty chains on them. 1785 */ 1786 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS); 1787 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS); 1788 1789 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) { 1790 hammer2_voldata_modify(hmp); 1791 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP | 1792 HAMMER2_FLUSH_ALL); 1793 } 1794 hammer2_chain_unlock(&hmp->fchain); 1795 1796 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) { 1797 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP | 1798 HAMMER2_FLUSH_ALL); 1799 } 1800 hammer2_chain_unlock(&hmp->vchain); 1801 1802 if ((hmp->vchain.flags | hmp->fchain.flags) & 1803 HAMMER2_CHAIN_FLUSH_MASK) { 1804 kprintf("hammer2_unmount: chains left over after final sync\n"); 1805 kprintf(" vchain %08x\n", hmp->vchain.flags); 1806 kprintf(" fchain %08x\n", hmp->fchain.flags); 1807 1808 if (hammer2_debug & 0x0010) 1809 Debugger("entered debugger"); 1810 } 1811 1812 hammer2_pfsfree_scan(hmp, 1); 1813 1814 KKASSERT(hmp->spmp == NULL); 1815 1816 /* 1817 * Finish up with the device vnode 1818 */ 1819 if ((devvp = hmp->devvp) != NULL) { 1820 ronly = hmp->ronly; 1821 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1822 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n", 1823 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree), 1824 ronly); 1825 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0); 1826 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n", 1827 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree)); 1828 devvp->v_rdev->si_mountpoint = NULL; 1829 hmp->devvp = NULL; 1830 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL); 1831 vn_unlock(devvp); 1832 vrele(devvp); 1833 devvp = NULL; 1834 } 1835 1836 /* 1837 * Clear vchain/fchain flags that might prevent final cleanup 1838 * of these chains. 1839 */ 1840 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) { 1841 atomic_add_long(&hammer2_count_modified_chains, -1); 1842 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 1843 hammer2_pfs_memory_wakeup(hmp->vchain.pmp, -1); 1844 } 1845 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) { 1846 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE); 1847 } 1848 1849 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) { 1850 atomic_add_long(&hammer2_count_modified_chains, -1); 1851 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED); 1852 hammer2_pfs_memory_wakeup(hmp->fchain.pmp, -1); 1853 } 1854 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) { 1855 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE); 1856 } 1857 1858 /* 1859 * Final drop of embedded freemap root chain to 1860 * clean up fchain.core (fchain structure is not 1861 * flagged ALLOCATED so it is cleaned out and then 1862 * left to rot). 1863 */ 1864 hammer2_chain_drop(&hmp->fchain); 1865 1866 /* 1867 * Final drop of embedded volume root chain to clean 1868 * up vchain.core (vchain structure is not flagged 1869 * ALLOCATED so it is cleaned out and then left to 1870 * rot). 1871 */ 1872 dumpcnt = 50; 1873 hammer2_dump_chain(&hmp->vchain, 0, 0, &dumpcnt, 'v', (u_int)-1); 1874 dumpcnt = 50; 1875 hammer2_dump_chain(&hmp->fchain, 0, 0, &dumpcnt, 'f', (u_int)-1); 1876 1877 hammer2_chain_drop(&hmp->vchain); 1878 1879 hammer2_io_cleanup(hmp, &hmp->iotree); 1880 if (hmp->iofree_count) { 1881 kprintf("io_cleanup: %d I/O's left hanging\n", 1882 hmp->iofree_count); 1883 } 1884 1885 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry); 1886 kmalloc_destroy(&hmp->mchain); 1887 kfree(hmp, M_HAMMER2); 1888 } 1889 1890 int 1891 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp, 1892 ino_t ino, struct vnode **vpp) 1893 { 1894 hammer2_xop_lookup_t *xop; 1895 hammer2_pfs_t *pmp; 1896 hammer2_inode_t *ip; 1897 hammer2_tid_t inum; 1898 int error; 1899 1900 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK; 1901 1902 error = 0; 1903 pmp = MPTOPMP(mp); 1904 1905 /* 1906 * Easy if we already have it cached 1907 */ 1908 ip = hammer2_inode_lookup(pmp, inum); 1909 if (ip) { 1910 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED); 1911 *vpp = hammer2_igetv(ip, &error); 1912 hammer2_inode_unlock(ip); 1913 hammer2_inode_drop(ip); /* from lookup */ 1914 1915 return error; 1916 } 1917 1918 /* 1919 * Otherwise we have to find the inode 1920 */ 1921 xop = hammer2_xop_alloc(pmp->iroot, 0); 1922 xop->lhc = inum; 1923 hammer2_xop_start(&xop->head, &hammer2_lookup_desc); 1924 error = hammer2_xop_collect(&xop->head, 0); 1925 1926 if (error == 0) 1927 ip = hammer2_inode_get(pmp, &xop->head, -1, -1); 1928 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1929 1930 if (ip) { 1931 *vpp = hammer2_igetv(ip, &error); 1932 hammer2_inode_unlock(ip); 1933 } else { 1934 *vpp = NULL; 1935 error = ENOENT; 1936 } 1937 return (error); 1938 } 1939 1940 static 1941 int 1942 hammer2_vfs_root(struct mount *mp, struct vnode **vpp) 1943 { 1944 hammer2_pfs_t *pmp; 1945 struct vnode *vp; 1946 int error; 1947 1948 pmp = MPTOPMP(mp); 1949 if (pmp->iroot == NULL) { 1950 kprintf("hammer2 (%s): no root inode\n", 1951 mp->mnt_stat.f_mntfromname); 1952 *vpp = NULL; 1953 return EINVAL; 1954 } 1955 1956 error = 0; 1957 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1958 1959 while (pmp->inode_tid == 0) { 1960 hammer2_xop_ipcluster_t *xop; 1961 const hammer2_inode_meta_t *meta; 1962 1963 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING); 1964 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc); 1965 error = hammer2_xop_collect(&xop->head, 0); 1966 1967 if (error == 0) { 1968 meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta; 1969 pmp->iroot->meta = *meta; 1970 pmp->inode_tid = meta->pfs_inum + 1; 1971 hammer2_xop_pdata(&xop->head); 1972 /* meta invalid */ 1973 1974 if (pmp->inode_tid < HAMMER2_INODE_START) 1975 pmp->inode_tid = HAMMER2_INODE_START; 1976 pmp->modify_tid = 1977 xop->head.cluster.focus->bref.modify_tid + 1; 1978 #if 0 1979 kprintf("PFS: Starting inode %jd\n", 1980 (intmax_t)pmp->inode_tid); 1981 kprintf("PMP focus good set nextino=%ld mod=%016jx\n", 1982 pmp->inode_tid, pmp->modify_tid); 1983 #endif 1984 wakeup(&pmp->iroot); 1985 1986 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1987 1988 /* 1989 * Prime the mount info. 1990 */ 1991 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL); 1992 break; 1993 } 1994 1995 /* 1996 * Loop, try again 1997 */ 1998 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1999 hammer2_inode_unlock(pmp->iroot); 2000 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz); 2001 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 2002 if (error == EINTR) 2003 break; 2004 } 2005 2006 if (error) { 2007 hammer2_inode_unlock(pmp->iroot); 2008 *vpp = NULL; 2009 } else { 2010 vp = hammer2_igetv(pmp->iroot, &error); 2011 hammer2_inode_unlock(pmp->iroot); 2012 *vpp = vp; 2013 } 2014 2015 return (error); 2016 } 2017 2018 /* 2019 * Filesystem status 2020 * 2021 * XXX incorporate ipdata->meta.inode_quota and data_quota 2022 */ 2023 static 2024 int 2025 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred) 2026 { 2027 hammer2_pfs_t *pmp; 2028 hammer2_dev_t *hmp; 2029 hammer2_blockref_t bref; 2030 struct statfs tmp; 2031 int i; 2032 2033 /* 2034 * NOTE: iroot might not have validated the cluster yet. 2035 */ 2036 pmp = MPTOPMP(mp); 2037 2038 bzero(&tmp, sizeof(tmp)); 2039 2040 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 2041 hmp = pmp->pfs_hmps[i]; 2042 if (hmp == NULL) 2043 continue; 2044 if (pmp->iroot->cluster.array[i].chain) 2045 bref = pmp->iroot->cluster.array[i].chain->bref; 2046 else 2047 bzero(&bref, sizeof(bref)); 2048 2049 tmp.f_files = bref.embed.stats.inode_count; 2050 tmp.f_ffree = 0; 2051 tmp.f_blocks = hmp->voldata.allocator_size / 2052 mp->mnt_vstat.f_bsize; 2053 tmp.f_bfree = hmp->voldata.allocator_free / 2054 mp->mnt_vstat.f_bsize; 2055 tmp.f_bavail = tmp.f_bfree; 2056 2057 if (cred && cred->cr_uid != 0) { 2058 uint64_t adj; 2059 2060 /* 5% */ 2061 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize; 2062 tmp.f_blocks -= adj; 2063 tmp.f_bfree -= adj; 2064 tmp.f_bavail -= adj; 2065 } 2066 2067 mp->mnt_stat.f_blocks = tmp.f_blocks; 2068 mp->mnt_stat.f_bfree = tmp.f_bfree; 2069 mp->mnt_stat.f_bavail = tmp.f_bavail; 2070 mp->mnt_stat.f_files = tmp.f_files; 2071 mp->mnt_stat.f_ffree = tmp.f_ffree; 2072 2073 *sbp = mp->mnt_stat; 2074 } 2075 return (0); 2076 } 2077 2078 static 2079 int 2080 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred) 2081 { 2082 hammer2_pfs_t *pmp; 2083 hammer2_dev_t *hmp; 2084 hammer2_blockref_t bref; 2085 struct statvfs tmp; 2086 int i; 2087 2088 /* 2089 * NOTE: iroot might not have validated the cluster yet. 2090 */ 2091 pmp = MPTOPMP(mp); 2092 bzero(&tmp, sizeof(tmp)); 2093 2094 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 2095 hmp = pmp->pfs_hmps[i]; 2096 if (hmp == NULL) 2097 continue; 2098 if (pmp->iroot->cluster.array[i].chain) 2099 bref = pmp->iroot->cluster.array[i].chain->bref; 2100 else 2101 bzero(&bref, sizeof(bref)); 2102 2103 tmp.f_files = bref.embed.stats.inode_count; 2104 tmp.f_ffree = 0; 2105 tmp.f_blocks = hmp->voldata.allocator_size / 2106 mp->mnt_vstat.f_bsize; 2107 tmp.f_bfree = hmp->voldata.allocator_free / 2108 mp->mnt_vstat.f_bsize; 2109 tmp.f_bavail = tmp.f_bfree; 2110 2111 if (cred && cred->cr_uid != 0) { 2112 uint64_t adj; 2113 2114 /* 5% */ 2115 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize; 2116 tmp.f_blocks -= adj; 2117 tmp.f_bfree -= adj; 2118 tmp.f_bavail -= adj; 2119 } 2120 2121 mp->mnt_vstat.f_blocks = tmp.f_blocks; 2122 mp->mnt_vstat.f_bfree = tmp.f_bfree; 2123 mp->mnt_vstat.f_bavail = tmp.f_bavail; 2124 mp->mnt_vstat.f_files = tmp.f_files; 2125 mp->mnt_vstat.f_ffree = tmp.f_ffree; 2126 2127 *sbp = mp->mnt_vstat; 2128 } 2129 return (0); 2130 } 2131 2132 /* 2133 * Mount-time recovery (RW mounts) 2134 * 2135 * Updates to the free block table are allowed to lag flushes by one 2136 * transaction. In case of a crash, then on a fresh mount we must do an 2137 * incremental scan of the last committed transaction id and make sure that 2138 * all related blocks have been marked allocated. 2139 */ 2140 struct hammer2_recovery_elm { 2141 TAILQ_ENTRY(hammer2_recovery_elm) entry; 2142 hammer2_chain_t *chain; 2143 hammer2_tid_t sync_tid; 2144 }; 2145 2146 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm); 2147 2148 struct hammer2_recovery_info { 2149 struct hammer2_recovery_list list; 2150 hammer2_tid_t mtid; 2151 int depth; 2152 }; 2153 2154 static int hammer2_recovery_scan(hammer2_dev_t *hmp, 2155 hammer2_chain_t *parent, 2156 struct hammer2_recovery_info *info, 2157 hammer2_tid_t sync_tid); 2158 2159 #define HAMMER2_RECOVERY_MAXDEPTH 10 2160 2161 static 2162 int 2163 hammer2_recovery(hammer2_dev_t *hmp) 2164 { 2165 struct hammer2_recovery_info info; 2166 struct hammer2_recovery_elm *elm; 2167 hammer2_chain_t *parent; 2168 hammer2_tid_t sync_tid; 2169 hammer2_tid_t mirror_tid; 2170 int error; 2171 2172 hammer2_trans_init(hmp->spmp, 0); 2173 2174 sync_tid = hmp->voldata.freemap_tid; 2175 mirror_tid = hmp->voldata.mirror_tid; 2176 2177 kprintf("hammer2_mount \"%s\": ", hmp->devrepname); 2178 if (sync_tid >= mirror_tid) { 2179 kprintf("no recovery needed\n"); 2180 } else { 2181 kprintf("freemap recovery %016jx-%016jx\n", 2182 sync_tid + 1, mirror_tid); 2183 } 2184 2185 TAILQ_INIT(&info.list); 2186 info.depth = 0; 2187 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 2188 error = hammer2_recovery_scan(hmp, parent, &info, sync_tid); 2189 hammer2_chain_lookup_done(parent); 2190 2191 while ((elm = TAILQ_FIRST(&info.list)) != NULL) { 2192 TAILQ_REMOVE(&info.list, elm, entry); 2193 parent = elm->chain; 2194 sync_tid = elm->sync_tid; 2195 kfree(elm, M_HAMMER2); 2196 2197 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2198 error |= hammer2_recovery_scan(hmp, parent, &info, 2199 hmp->voldata.freemap_tid); 2200 hammer2_chain_unlock(parent); 2201 hammer2_chain_drop(parent); /* drop elm->chain ref */ 2202 } 2203 2204 hammer2_trans_done(hmp->spmp, 0); 2205 2206 return error; 2207 } 2208 2209 static 2210 int 2211 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent, 2212 struct hammer2_recovery_info *info, 2213 hammer2_tid_t sync_tid) 2214 { 2215 const hammer2_inode_data_t *ripdata; 2216 hammer2_chain_t *chain; 2217 hammer2_blockref_t bref; 2218 int tmp_error; 2219 int rup_error; 2220 int error; 2221 int first; 2222 2223 /* 2224 * Adjust freemap to ensure that the block(s) are marked allocated. 2225 */ 2226 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) { 2227 hammer2_freemap_adjust(hmp, &parent->bref, 2228 HAMMER2_FREEMAP_DORECOVER); 2229 } 2230 2231 /* 2232 * Check type for recursive scan 2233 */ 2234 switch(parent->bref.type) { 2235 case HAMMER2_BREF_TYPE_VOLUME: 2236 /* data already instantiated */ 2237 break; 2238 case HAMMER2_BREF_TYPE_INODE: 2239 /* 2240 * Must instantiate data for DIRECTDATA test and also 2241 * for recursion. 2242 */ 2243 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2244 ripdata = &parent->data->ipdata; 2245 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) { 2246 /* not applicable to recovery scan */ 2247 hammer2_chain_unlock(parent); 2248 return 0; 2249 } 2250 hammer2_chain_unlock(parent); 2251 break; 2252 case HAMMER2_BREF_TYPE_INDIRECT: 2253 /* 2254 * Must instantiate data for recursion 2255 */ 2256 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2257 hammer2_chain_unlock(parent); 2258 break; 2259 case HAMMER2_BREF_TYPE_DIRENT: 2260 case HAMMER2_BREF_TYPE_DATA: 2261 case HAMMER2_BREF_TYPE_FREEMAP: 2262 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2263 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 2264 /* not applicable to recovery scan */ 2265 return 0; 2266 break; 2267 default: 2268 return HAMMER2_ERROR_BADBREF; 2269 } 2270 2271 /* 2272 * Defer operation if depth limit reached. 2273 */ 2274 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) { 2275 struct hammer2_recovery_elm *elm; 2276 2277 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK); 2278 elm->chain = parent; 2279 elm->sync_tid = sync_tid; 2280 hammer2_chain_ref(parent); 2281 TAILQ_INSERT_TAIL(&info->list, elm, entry); 2282 /* unlocked by caller */ 2283 2284 return(0); 2285 } 2286 2287 2288 /* 2289 * Recursive scan of the last flushed transaction only. We are 2290 * doing this without pmp assignments so don't leave the chains 2291 * hanging around after we are done with them. 2292 * 2293 * error Cumulative error this level only 2294 * rup_error Cumulative error for recursion 2295 * tmp_error Specific non-cumulative recursion error 2296 */ 2297 chain = NULL; 2298 first = 1; 2299 rup_error = 0; 2300 error = 0; 2301 2302 for (;;) { 2303 error |= hammer2_chain_scan(parent, &chain, &bref, 2304 &first, 2305 HAMMER2_LOOKUP_NODATA); 2306 2307 /* 2308 * Problem during scan or EOF 2309 */ 2310 if (error) 2311 break; 2312 2313 /* 2314 * If this is a leaf 2315 */ 2316 if (chain == NULL) { 2317 if (bref.mirror_tid > sync_tid) { 2318 hammer2_freemap_adjust(hmp, &bref, 2319 HAMMER2_FREEMAP_DORECOVER); 2320 } 2321 continue; 2322 } 2323 2324 /* 2325 * This may or may not be a recursive node. 2326 */ 2327 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 2328 if (bref.mirror_tid > sync_tid) { 2329 ++info->depth; 2330 tmp_error = hammer2_recovery_scan(hmp, chain, 2331 info, sync_tid); 2332 --info->depth; 2333 } else { 2334 tmp_error = 0; 2335 } 2336 2337 /* 2338 * Flush the recovery at the PFS boundary to stage it for 2339 * the final flush of the super-root topology. 2340 */ 2341 if (tmp_error == 0 && 2342 (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) && 2343 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) { 2344 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 2345 HAMMER2_FLUSH_ALL); 2346 } 2347 rup_error |= tmp_error; 2348 } 2349 return ((error | rup_error) & ~HAMMER2_ERROR_EOF); 2350 } 2351 2352 /* 2353 * This fixes up an error introduced in earlier H2 implementations where 2354 * moving a PFS inode into an indirect block wound up causing the 2355 * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared. 2356 */ 2357 static 2358 int 2359 hammer2_fixup_pfses(hammer2_dev_t *hmp) 2360 { 2361 const hammer2_inode_data_t *ripdata; 2362 hammer2_chain_t *parent; 2363 hammer2_chain_t *chain; 2364 hammer2_key_t key_next; 2365 hammer2_pfs_t *spmp; 2366 int error; 2367 2368 error = 0; 2369 2370 /* 2371 * Lookup mount point under the media-localized super-root. 2372 * 2373 * cluster->pmp will incorrectly point to spmp and must be fixed 2374 * up later on. 2375 */ 2376 spmp = hmp->spmp; 2377 hammer2_inode_lock(spmp->iroot, 0); 2378 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 2379 chain = hammer2_chain_lookup(&parent, &key_next, 2380 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 2381 &error, 0); 2382 while (chain) { 2383 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) 2384 continue; 2385 if (chain->error) { 2386 kprintf("I/O error scanning PFS labels\n"); 2387 error |= chain->error; 2388 } else if ((chain->bref.flags & 2389 HAMMER2_BREF_FLAG_PFSROOT) == 0) { 2390 int error2; 2391 2392 ripdata = &chain->data->ipdata; 2393 hammer2_trans_init(hmp->spmp, 0); 2394 error2 = hammer2_chain_modify(chain, 2395 chain->bref.modify_tid, 2396 0, 0); 2397 if (error2 == 0) { 2398 kprintf("hammer2: Correct mis-flagged PFS %s\n", 2399 ripdata->filename); 2400 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT; 2401 } else { 2402 error |= error2; 2403 } 2404 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 2405 HAMMER2_FLUSH_ALL); 2406 hammer2_trans_done(hmp->spmp, 0); 2407 } 2408 chain = hammer2_chain_next(&parent, chain, &key_next, 2409 key_next, HAMMER2_KEY_MAX, 2410 &error, 0); 2411 } 2412 if (parent) { 2413 hammer2_chain_unlock(parent); 2414 hammer2_chain_drop(parent); 2415 } 2416 hammer2_inode_unlock(spmp->iroot); 2417 2418 return error; 2419 } 2420 2421 /* 2422 * Sync a mount point; this is called periodically on a per-mount basis from 2423 * the filesystem syncer, and whenever a user issues a sync. 2424 */ 2425 int 2426 hammer2_vfs_sync(struct mount *mp, int waitfor) 2427 { 2428 int error; 2429 2430 error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor); 2431 2432 return error; 2433 } 2434 2435 /* 2436 * Because frontend operations lock vnodes before we get a chance to 2437 * lock the related inode, we can't just acquire a vnode lock without 2438 * risking a deadlock. The frontend may be holding a vnode lock while 2439 * also blocked on our SYNCQ flag while trying to get the inode lock. 2440 * 2441 * To deal with this situation we can check the vnode lock situation 2442 * after locking the inode and perform a work-around. 2443 */ 2444 int 2445 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor) 2446 { 2447 struct mount *mp; 2448 /*hammer2_xop_flush_t *xop;*/ 2449 /*struct hammer2_sync_info info;*/ 2450 hammer2_inode_t *ip; 2451 hammer2_depend_t *depend; 2452 hammer2_depend_t *depend_next; 2453 struct vnode *vp; 2454 uint32_t pass2; 2455 int error; 2456 int wakecount; 2457 int dorestart; 2458 2459 mp = pmp->mp; 2460 2461 /* 2462 * Move all inodes on sideq to syncq. This will clear sideq. 2463 * This should represent all flushable inodes. These inodes 2464 * will already have refs due to being on syncq or sideq. We 2465 * must do this all at once with the spinlock held to ensure that 2466 * all inode dependencies are part of the same flush. 2467 * 2468 * We should be able to do this asynchronously from frontend 2469 * operations because we will be locking the inodes later on 2470 * to actually flush them, and that will partition any frontend 2471 * op using the same inode. Either it has already locked the 2472 * inode and we will block, or it has not yet locked the inode 2473 * and it will block until we are finished flushing that inode. 2474 * 2475 * When restarting, only move the inodes flagged as PASS2 from 2476 * SIDEQ to SYNCQ. PASS2 propagation by inode_lock4() and 2477 * inode_depend() are atomic with the spin-lock. 2478 */ 2479 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH); 2480 #ifdef HAMMER2_DEBUG_SYNC 2481 kprintf("FILESYSTEM SYNC BOUNDARY\n"); 2482 #endif 2483 dorestart = 0; 2484 2485 /* 2486 * Move inodes from depq to syncq, releasing the related 2487 * depend structures. 2488 */ 2489 restart: 2490 #ifdef HAMMER2_DEBUG_SYNC 2491 kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart); 2492 #endif 2493 hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/); 2494 hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN); 2495 2496 /* 2497 * Move inodes from depq to syncq. When restarting, only depq's 2498 * marked pass2 are moved. 2499 */ 2500 hammer2_spin_ex(&pmp->list_spin); 2501 depend_next = TAILQ_FIRST(&pmp->depq); 2502 wakecount = 0; 2503 2504 while ((depend = depend_next) != NULL) { 2505 depend_next = TAILQ_NEXT(depend, entry); 2506 if (dorestart && depend->pass2 == 0) 2507 continue; 2508 TAILQ_FOREACH(ip, &depend->sideq, entry) { 2509 KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ); 2510 atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ); 2511 atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ); 2512 ip->depend = NULL; 2513 } 2514 2515 /* 2516 * NOTE: pmp->sideq_count includes both sideq and syncq 2517 */ 2518 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry); 2519 2520 depend->count = 0; 2521 depend->pass2 = 0; 2522 TAILQ_REMOVE(&pmp->depq, depend, entry); 2523 } 2524 2525 hammer2_spin_unex(&pmp->list_spin); 2526 hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/ 2527 HAMMER2_TRANS_WAITING); 2528 dorestart = 0; 2529 2530 /* 2531 * sideq_count may have dropped enough to allow us to unstall 2532 * the frontend. 2533 */ 2534 hammer2_pfs_memory_wakeup(pmp, 0); 2535 2536 /* 2537 * Now run through all inodes on syncq. 2538 * 2539 * Flush transactions only interlock with other flush transactions. 2540 * Any conflicting frontend operations will block on the inode, but 2541 * may hold a vnode lock while doing so. 2542 */ 2543 hammer2_spin_ex(&pmp->list_spin); 2544 while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) { 2545 /* 2546 * Remove the inode from the SYNCQ, transfer the syncq ref 2547 * to us. We must clear SYNCQ to allow any potential 2548 * front-end deadlock to proceed. We must set PASS2 so 2549 * the dependency code knows what to do. 2550 */ 2551 pass2 = ip->flags; 2552 cpu_ccfence(); 2553 if (atomic_cmpset_int(&ip->flags, 2554 pass2, 2555 (pass2 & ~(HAMMER2_INODE_SYNCQ | 2556 HAMMER2_INODE_SYNCQ_WAKEUP)) | 2557 HAMMER2_INODE_SYNCQ_PASS2) == 0) { 2558 continue; 2559 } 2560 TAILQ_REMOVE(&pmp->syncq, ip, entry); 2561 --pmp->sideq_count; 2562 hammer2_spin_unex(&pmp->list_spin); 2563 2564 /* 2565 * Tickle anyone waiting on ip->flags or the hysteresis 2566 * on the dirty inode count. 2567 */ 2568 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP) 2569 wakeup(&ip->flags); 2570 if (++wakecount >= hammer2_limit_dirty_inodes / 20 + 1) { 2571 wakecount = 0; 2572 hammer2_pfs_memory_wakeup(pmp, 0); 2573 } 2574 2575 /* 2576 * Relock the inode, and we inherit a ref from the above. 2577 * We will check for a race after we acquire the vnode. 2578 */ 2579 hammer2_mtx_ex(&ip->lock); 2580 2581 /* 2582 * We need the vp in order to vfsync() dirty buffers, so if 2583 * one isn't attached we can skip it. 2584 * 2585 * Ordering the inode lock and then the vnode lock has the 2586 * potential to deadlock. If we had left SYNCQ set that could 2587 * also deadlock us against the frontend even if we don't hold 2588 * any locks, but the latter is not a problem now since we 2589 * cleared it. igetv will temporarily release the inode lock 2590 * in a safe manner to work-around the deadlock. 2591 * 2592 * Unfortunately it is still possible to deadlock when the 2593 * frontend obtains multiple inode locks, because all the 2594 * related vnodes are already locked (nor can the vnode locks 2595 * be released and reacquired without messing up RECLAIM and 2596 * INACTIVE sequencing). 2597 * 2598 * The solution for now is to move the vp back onto SIDEQ 2599 * and set dorestart, which will restart the flush after we 2600 * exhaust the current SYNCQ. Note that additional 2601 * dependencies may build up, so we definitely need to move 2602 * the whole SIDEQ back to SYNCQ when we restart. 2603 */ 2604 vp = ip->vp; 2605 if (vp) { 2606 if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) { 2607 /* 2608 * Failed to get the vnode, requeue the inode 2609 * (PASS2 is already set so it will be found 2610 * again on the restart). 2611 * 2612 * Then unlock, possibly sleep, and retry 2613 * later. We sleep if PASS2 was *previously* 2614 * set, before we set it again above. 2615 */ 2616 vp = NULL; 2617 dorestart = 1; 2618 #ifdef HAMMER2_DEBUG_SYNC 2619 kprintf("inum %ld (sync delayed by vnode)\n", 2620 (long)ip->meta.inum); 2621 #endif 2622 hammer2_inode_delayed_sideq(ip); 2623 2624 hammer2_mtx_unlock(&ip->lock); 2625 hammer2_inode_drop(ip); 2626 2627 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) { 2628 tsleep(&dorestart, 0, "h2syndel", 2); 2629 } 2630 hammer2_spin_ex(&pmp->list_spin); 2631 continue; 2632 } 2633 } else { 2634 vp = NULL; 2635 } 2636 2637 /* 2638 * If the inode wound up on a SIDEQ again it will already be 2639 * prepped for another PASS2. In this situation if we flush 2640 * it now we will just wind up flushing it again in the same 2641 * syncer run, so we might as well not flush it now. 2642 */ 2643 if (ip->flags & HAMMER2_INODE_SIDEQ) { 2644 hammer2_mtx_unlock(&ip->lock); 2645 hammer2_inode_drop(ip); 2646 if (vp) 2647 vput(vp); 2648 dorestart = 1; 2649 hammer2_spin_ex(&pmp->list_spin); 2650 continue; 2651 } 2652 2653 /* 2654 * Ok we have the inode exclusively locked and if vp is 2655 * not NULL that will also be exclusively locked. Do the 2656 * meat of the flush. 2657 * 2658 * vp token needed for v_rbdirty_tree check / vclrisdirty 2659 * sequencing. Though we hold the vnode exclusively so 2660 * we shouldn't need to hold the token also in this case. 2661 */ 2662 if (vp) { 2663 vfsync(vp, MNT_WAIT, 1, NULL, NULL); 2664 bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */ 2665 } 2666 2667 /* 2668 * If the inode has not yet been inserted into the tree 2669 * we must do so. Then sync and flush it. The flush should 2670 * update the parent. 2671 */ 2672 if (ip->flags & HAMMER2_INODE_DELETING) { 2673 #ifdef HAMMER2_DEBUG_SYNC 2674 kprintf("inum %ld destroy\n", (long)ip->meta.inum); 2675 #endif 2676 hammer2_inode_chain_des(ip); 2677 atomic_add_long(&hammer2_iod_inode_deletes, 1); 2678 } else if (ip->flags & HAMMER2_INODE_CREATING) { 2679 #ifdef HAMMER2_DEBUG_SYNC 2680 kprintf("inum %ld insert\n", (long)ip->meta.inum); 2681 #endif 2682 hammer2_inode_chain_ins(ip); 2683 atomic_add_long(&hammer2_iod_inode_creates, 1); 2684 } 2685 #ifdef HAMMER2_DEBUG_SYNC 2686 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum); 2687 #endif 2688 2689 /* 2690 * Because I kinda messed up the design and index the inodes 2691 * under the root inode, along side the directory entries, 2692 * we can't flush the inode index under the iroot until the 2693 * end. If we do it now we might miss effects created by 2694 * other inodes on the SYNCQ. 2695 * 2696 * Do a normal (non-FSSYNC) flush instead, which allows the 2697 * vnode code to work the same. We don't want to force iroot 2698 * back onto the SIDEQ, and we also don't want the flush code 2699 * to update pfs_iroot_blocksets until the final flush later. 2700 * 2701 * XXX at the moment this will likely result in a double-flush 2702 * of the iroot chain. 2703 */ 2704 hammer2_inode_chain_sync(ip); 2705 if (ip == pmp->iroot) { 2706 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP); 2707 } else { 2708 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP | 2709 HAMMER2_XOP_FSSYNC); 2710 } 2711 if (vp) { 2712 lwkt_gettoken(&vp->v_token); 2713 if ((ip->flags & (HAMMER2_INODE_MODIFIED | 2714 HAMMER2_INODE_RESIZED | 2715 HAMMER2_INODE_DIRTYDATA)) == 0 && 2716 RB_EMPTY(&vp->v_rbdirty_tree) && 2717 !bio_track_active(&vp->v_track_write)) { 2718 vclrisdirty(vp); 2719 } else { 2720 hammer2_inode_delayed_sideq(ip); 2721 } 2722 lwkt_reltoken(&vp->v_token); 2723 vput(vp); 2724 vp = NULL; /* safety */ 2725 } 2726 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2); 2727 hammer2_inode_unlock(ip); /* unlock+drop */ 2728 /* ip pointer invalid */ 2729 2730 /* 2731 * If the inode got dirted after we dropped our locks, 2732 * it will have already been moved back to the SIDEQ. 2733 */ 2734 hammer2_spin_ex(&pmp->list_spin); 2735 } 2736 hammer2_spin_unex(&pmp->list_spin); 2737 hammer2_pfs_memory_wakeup(pmp, 0); 2738 2739 if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) { 2740 #ifdef HAMMER2_DEBUG_SYNC 2741 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n"); 2742 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/ 2743 #endif 2744 dorestart = 1; 2745 goto restart; 2746 } 2747 #ifdef HAMMER2_DEBUG_SYNC 2748 kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n"); 2749 /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/ 2750 #endif 2751 2752 /* 2753 * We have to flush the PFS root last, even if it does not appear to 2754 * be dirty, because all the inodes in the PFS are indexed under it. 2755 * The normal flushing of iroot above would only occur if directory 2756 * entries under the root were changed. 2757 * 2758 * Specifying VOLHDR will cause an additionl flush of hmp->spmp 2759 * for the media making up the cluster. 2760 */ 2761 if ((ip = pmp->iroot) != NULL) { 2762 hammer2_inode_ref(ip); 2763 hammer2_mtx_ex(&ip->lock); 2764 hammer2_inode_chain_sync(ip); 2765 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP | 2766 HAMMER2_XOP_FSSYNC | 2767 HAMMER2_XOP_VOLHDR); 2768 hammer2_inode_unlock(ip); /* unlock+drop */ 2769 } 2770 #ifdef HAMMER2_DEBUG_SYNC 2771 kprintf("FILESYSTEM SYNC STAGE 2 DONE\n"); 2772 #endif 2773 2774 /* 2775 * device bioq sync 2776 */ 2777 hammer2_bioq_sync(pmp); 2778 2779 #if 0 2780 /* 2781 * Generally speaking we now want to flush the media topology from 2782 * the iroot through to the inodes. The flush stops at any inode 2783 * boundary, which allows the frontend to continue running concurrent 2784 * modifying operations on inodes (including kernel flushes of 2785 * buffers) without interfering with the main sync. 2786 * 2787 * Use the XOP interface to concurrently flush all nodes to 2788 * synchronize the PFSROOT subtopology to the media. A standard 2789 * end-of-scan ENOENT error indicates cluster sufficiency. 2790 * 2791 * Note that this flush will not be visible on crash recovery until 2792 * we flush the super-root topology in the next loop. 2793 * 2794 * XXX For now wait for all flushes to complete. 2795 */ 2796 if (mp && (ip = pmp->iroot) != NULL) { 2797 /* 2798 * If unmounting try to flush everything including any 2799 * sub-trees under inodes, just in case there is dangling 2800 * modified data, as a safety. Otherwise just flush up to 2801 * the inodes in this stage. 2802 */ 2803 kprintf("MP & IROOT\n"); 2804 #ifdef HAMMER2_DEBUG_SYNC 2805 kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n"); 2806 #endif 2807 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 2808 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 2809 HAMMER2_XOP_VOLHDR | 2810 HAMMER2_XOP_FSSYNC | 2811 HAMMER2_XOP_INODE_STOP); 2812 } else { 2813 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 2814 HAMMER2_XOP_INODE_STOP | 2815 HAMMER2_XOP_VOLHDR | 2816 HAMMER2_XOP_FSSYNC | 2817 HAMMER2_XOP_INODE_STOP); 2818 } 2819 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc); 2820 error = hammer2_xop_collect(&xop->head, 2821 HAMMER2_XOP_COLLECT_WAITALL); 2822 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 2823 #ifdef HAMMER2_DEBUG_SYNC 2824 kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n"); 2825 #endif 2826 if (error == HAMMER2_ERROR_ENOENT) 2827 error = 0; 2828 else 2829 error = hammer2_error_to_errno(error); 2830 } else { 2831 error = 0; 2832 } 2833 #endif 2834 error = 0; /* XXX */ 2835 hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH); 2836 2837 return (error); 2838 } 2839 2840 static 2841 int 2842 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp) 2843 { 2844 hammer2_inode_t *ip; 2845 2846 KKASSERT(MAXFIDSZ >= 16); 2847 ip = VTOI(vp); 2848 fhp->fid_len = offsetof(struct fid, fid_data[16]); 2849 fhp->fid_ext = 0; 2850 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum; 2851 ((hammer2_tid_t *)fhp->fid_data)[1] = 0; 2852 2853 return 0; 2854 } 2855 2856 static 2857 int 2858 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp, 2859 struct fid *fhp, struct vnode **vpp) 2860 { 2861 hammer2_pfs_t *pmp; 2862 hammer2_tid_t inum; 2863 int error; 2864 2865 pmp = MPTOPMP(mp); 2866 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK; 2867 if (vpp) { 2868 if (inum == 1) 2869 error = hammer2_vfs_root(mp, vpp); 2870 else 2871 error = hammer2_vfs_vget(mp, NULL, inum, vpp); 2872 } else { 2873 error = 0; 2874 } 2875 if (error) 2876 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error); 2877 return error; 2878 } 2879 2880 static 2881 int 2882 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 2883 int *exflagsp, struct ucred **credanonp) 2884 { 2885 hammer2_pfs_t *pmp; 2886 struct netcred *np; 2887 int error; 2888 2889 pmp = MPTOPMP(mp); 2890 np = vfs_export_lookup(mp, &pmp->export, nam); 2891 if (np) { 2892 *exflagsp = np->netc_exflags; 2893 *credanonp = &np->netc_anon; 2894 error = 0; 2895 } else { 2896 error = EACCES; 2897 } 2898 return error; 2899 } 2900 2901 /* 2902 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume 2903 * header into the HMP 2904 */ 2905 static 2906 int 2907 hammer2_install_volume_header(hammer2_dev_t *hmp) 2908 { 2909 hammer2_volume_data_t *vd; 2910 struct buf *bp; 2911 hammer2_crc32_t crc0, crc, bcrc0, bcrc; 2912 int error_reported; 2913 int error; 2914 int valid; 2915 int i; 2916 2917 error_reported = 0; 2918 error = 0; 2919 valid = 0; 2920 bp = NULL; 2921 2922 /* 2923 * There are up to 4 copies of the volume header (syncs iterate 2924 * between them so there is no single master). We don't trust the 2925 * volu_size field so we don't know precisely how large the filesystem 2926 * is, so depend on the OS to return an error if we go beyond the 2927 * block device's EOF. 2928 */ 2929 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) { 2930 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64, 2931 HAMMER2_VOLUME_BYTES, &bp); 2932 if (error) { 2933 brelse(bp); 2934 bp = NULL; 2935 continue; 2936 } 2937 2938 vd = (struct hammer2_volume_data *) bp->b_data; 2939 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) && 2940 (vd->magic != HAMMER2_VOLUME_ID_ABO)) { 2941 kprintf("hammer2: volume header #%d: bad magic\n", i); 2942 brelse(bp); 2943 bp = NULL; 2944 continue; 2945 } 2946 2947 if (vd->magic == HAMMER2_VOLUME_ID_ABO) { 2948 /* XXX: Reversed-endianness filesystem */ 2949 kprintf("hammer2: volume header #%d: reverse-endian " 2950 "filesystem detected\n", i); 2951 brelse(bp); 2952 bp = NULL; 2953 continue; 2954 } 2955 2956 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0]; 2957 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF, 2958 HAMMER2_VOLUME_ICRC0_SIZE); 2959 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1]; 2960 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF, 2961 HAMMER2_VOLUME_ICRC1_SIZE); 2962 if ((crc0 != crc) || (bcrc0 != bcrc)) { 2963 kprintf("hammer2: volume header #%d: volume header crc " 2964 "mismatch %08x/%08x\n", 2965 i, crc0, crc); 2966 error_reported = 1; 2967 brelse(bp); 2968 bp = NULL; 2969 continue; 2970 } 2971 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) { 2972 valid = 1; 2973 hmp->voldata = *vd; 2974 hmp->volhdrno = i; 2975 } 2976 brelse(bp); 2977 bp = NULL; 2978 } 2979 if (valid) { 2980 hmp->volsync = hmp->voldata; 2981 hmp->free_reserved = hmp->voldata.allocator_size / 20; 2982 error = 0; 2983 if (error_reported || bootverbose || 1) { /* 1/DEBUG */ 2984 kprintf("hammer2: using volume header #%d\n", 2985 hmp->volhdrno); 2986 } 2987 } else { 2988 error = EINVAL; 2989 kprintf("hammer2: no valid volume headers found!\n"); 2990 } 2991 return (error); 2992 } 2993 2994 /* 2995 * This handles hysteresis on regular file flushes. Because the BIOs are 2996 * routed to a thread it is possible for an excessive number to build up 2997 * and cause long front-end stalls long before the runningbuffspace limit 2998 * is hit, so we implement hammer2_flush_pipe to control the 2999 * hysteresis. 3000 * 3001 * This is a particular problem when compression is used. 3002 */ 3003 void 3004 hammer2_lwinprog_ref(hammer2_pfs_t *pmp) 3005 { 3006 atomic_add_int(&pmp->count_lwinprog, 1); 3007 } 3008 3009 void 3010 hammer2_lwinprog_drop(hammer2_pfs_t *pmp) 3011 { 3012 int lwinprog; 3013 3014 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1); 3015 if ((lwinprog & HAMMER2_LWINPROG_WAITING) && 3016 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) { 3017 atomic_clear_int(&pmp->count_lwinprog, 3018 HAMMER2_LWINPROG_WAITING); 3019 wakeup(&pmp->count_lwinprog); 3020 } 3021 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) && 3022 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) { 3023 atomic_clear_int(&pmp->count_lwinprog, 3024 HAMMER2_LWINPROG_WAITING0); 3025 wakeup(&pmp->count_lwinprog); 3026 } 3027 } 3028 3029 void 3030 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe) 3031 { 3032 int lwinprog; 3033 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING : 3034 HAMMER2_LWINPROG_WAITING0; 3035 3036 for (;;) { 3037 lwinprog = pmp->count_lwinprog; 3038 cpu_ccfence(); 3039 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 3040 break; 3041 tsleep_interlock(&pmp->count_lwinprog, 0); 3042 atomic_set_int(&pmp->count_lwinprog, lwflag); 3043 lwinprog = pmp->count_lwinprog; 3044 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 3045 break; 3046 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz); 3047 } 3048 } 3049 3050 /* 3051 * It is possible for an excessive number of dirty chains or dirty inodes 3052 * to build up. When this occurs we start an asynchronous filesystem sync. 3053 * If the level continues to build up, we stall, waiting for it to drop, 3054 * with some hysteresis. 3055 * 3056 * This relies on the kernel calling hammer2_vfs_modifying() prior to 3057 * obtaining any vnode locks before making a modifying VOP call. 3058 */ 3059 static int 3060 hammer2_vfs_modifying(struct mount *mp) 3061 { 3062 if (mp->mnt_flag & MNT_RDONLY) 3063 return EROFS; 3064 hammer2_pfs_memory_wait(MPTOPMP(mp)); 3065 3066 return 0; 3067 } 3068 3069 /* 3070 * Initiate an asynchronous filesystem sync and, with hysteresis, 3071 * stall if the internal data structure count becomes too bloated. 3072 */ 3073 void 3074 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp) 3075 { 3076 uint32_t waiting; 3077 int pcatch; 3078 int error; 3079 3080 if (pmp == NULL || pmp->mp == NULL) 3081 return; 3082 3083 for (;;) { 3084 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK; 3085 cpu_ccfence(); 3086 3087 /* 3088 * Start the syncer running at 1/2 the limit 3089 */ 3090 if (waiting > hammer2_limit_dirty_chains / 2 || 3091 pmp->sideq_count > hammer2_limit_dirty_inodes / 2) { 3092 trigger_syncer(pmp->mp); 3093 } 3094 3095 /* 3096 * Stall at the limit waiting for the counts to drop. 3097 * This code will typically be woken up once the count 3098 * drops below 3/4 the limit, or in one second. 3099 */ 3100 if (waiting < hammer2_limit_dirty_chains && 3101 pmp->sideq_count < hammer2_limit_dirty_inodes) { 3102 break; 3103 } 3104 3105 pcatch = curthread->td_proc ? PCATCH : 0; 3106 3107 tsleep_interlock(&pmp->inmem_dirty_chains, pcatch); 3108 atomic_set_int(&pmp->inmem_dirty_chains, 3109 HAMMER2_DIRTYCHAIN_WAITING); 3110 if (waiting < hammer2_limit_dirty_chains && 3111 pmp->sideq_count < hammer2_limit_dirty_inodes) { 3112 break; 3113 } 3114 trigger_syncer(pmp->mp); 3115 error = tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED | pcatch, 3116 "h2memw", hz); 3117 if (error == ERESTART) 3118 break; 3119 } 3120 } 3121 3122 /* 3123 * Wake up any stalled frontend ops waiting, with hysteresis, using 3124 * 2/3 of the limit. 3125 */ 3126 void 3127 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp, int count) 3128 { 3129 uint32_t waiting; 3130 3131 if (pmp) { 3132 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, count); 3133 /* don't need --waiting to test flag */ 3134 3135 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) && 3136 (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <= 3137 hammer2_limit_dirty_chains * 2 / 3 && 3138 pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) { 3139 atomic_clear_int(&pmp->inmem_dirty_chains, 3140 HAMMER2_DIRTYCHAIN_WAITING); 3141 wakeup(&pmp->inmem_dirty_chains); 3142 } 3143 } 3144 } 3145 3146 void 3147 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp) 3148 { 3149 if (pmp) { 3150 atomic_add_int(&pmp->inmem_dirty_chains, 1); 3151 } 3152 } 3153 3154 /* 3155 * Volume header data locks 3156 */ 3157 void 3158 hammer2_voldata_lock(hammer2_dev_t *hmp) 3159 { 3160 lockmgr(&hmp->vollk, LK_EXCLUSIVE); 3161 } 3162 3163 void 3164 hammer2_voldata_unlock(hammer2_dev_t *hmp) 3165 { 3166 lockmgr(&hmp->vollk, LK_RELEASE); 3167 } 3168 3169 void 3170 hammer2_voldata_modify(hammer2_dev_t *hmp) 3171 { 3172 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) { 3173 atomic_add_long(&hammer2_count_modified_chains, 1); 3174 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 3175 hammer2_pfs_memory_inc(hmp->vchain.pmp); 3176 } 3177 } 3178 3179 /* 3180 * Returns 0 if the filesystem has tons of free space 3181 * Returns 1 if the filesystem has less than 10% remaining 3182 * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining. 3183 */ 3184 int 3185 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred) 3186 { 3187 hammer2_pfs_t *pmp; 3188 hammer2_dev_t *hmp; 3189 hammer2_off_t free_reserved; 3190 hammer2_off_t free_nominal; 3191 int i; 3192 3193 pmp = ip->pmp; 3194 3195 if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) { 3196 free_reserved = HAMMER2_SEGSIZE; 3197 free_nominal = 0x7FFFFFFFFFFFFFFFLLU; 3198 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 3199 hmp = pmp->pfs_hmps[i]; 3200 if (hmp == NULL) 3201 continue; 3202 if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER && 3203 pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER) 3204 continue; 3205 3206 if (free_nominal > hmp->voldata.allocator_free) 3207 free_nominal = hmp->voldata.allocator_free; 3208 if (free_reserved < hmp->free_reserved) 3209 free_reserved = hmp->free_reserved; 3210 } 3211 3212 /* 3213 * SMP races ok 3214 */ 3215 pmp->free_reserved = free_reserved; 3216 pmp->free_nominal = free_nominal; 3217 pmp->free_ticks = ticks; 3218 } else { 3219 free_reserved = pmp->free_reserved; 3220 free_nominal = pmp->free_nominal; 3221 } 3222 if (cred && cred->cr_uid != 0) { 3223 if ((int64_t)(free_nominal - bytes) < 3224 (int64_t)free_reserved) { 3225 return 2; 3226 } 3227 } else { 3228 if ((int64_t)(free_nominal - bytes) < 3229 (int64_t)free_reserved / 2) { 3230 return 2; 3231 } 3232 } 3233 if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2) 3234 return 1; 3235 return 0; 3236 } 3237 3238 /* 3239 * Debugging 3240 */ 3241 void 3242 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int bi, int *countp, 3243 char pfx, u_int flags) 3244 { 3245 hammer2_chain_t *scan; 3246 hammer2_chain_t *parent; 3247 3248 --*countp; 3249 if (*countp == 0) { 3250 kprintf("%*.*s...\n", tab, tab, ""); 3251 return; 3252 } 3253 if (*countp < 0) 3254 return; 3255 kprintf("%*.*s%c-chain %p %s.%-3d %016jx %016jx/%-2d mir=%016jx\n", 3256 tab, tab, "", pfx, chain, 3257 hammer2_bref_type_str(chain->bref.type), bi, 3258 chain->bref.data_off, chain->bref.key, chain->bref.keybits, 3259 chain->bref.mirror_tid); 3260 3261 kprintf("%*.*s [%08x] (%s) refs=%d", 3262 tab, tab, "", 3263 chain->flags, 3264 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE && 3265 chain->data) ? (char *)chain->data->ipdata.filename : "?"), 3266 chain->refs); 3267 3268 parent = chain->parent; 3269 if (parent) 3270 kprintf("\n%*.*s p=%p [pflags %08x prefs %d]", 3271 tab, tab, "", 3272 parent, parent->flags, parent->refs); 3273 if (RB_EMPTY(&chain->core.rbtree)) { 3274 kprintf("\n"); 3275 } else { 3276 int bi = 0; 3277 kprintf(" {\n"); 3278 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) { 3279 if ((scan->flags & flags) || flags == (u_int)-1) { 3280 hammer2_dump_chain(scan, tab + 4, bi, countp, 3281 'a', flags); 3282 } 3283 bi++; 3284 } 3285 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data) 3286 kprintf("%*.*s}(%s)\n", tab, tab, "", 3287 chain->data->ipdata.filename); 3288 else 3289 kprintf("%*.*s}\n", tab, tab, ""); 3290 } 3291 } 3292