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_mount: 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 if (!vn_isdisk(devvp, &error)) { 1104 KKASSERT(error); 1105 kprintf("hammer2_mount: %s not a block device %d\n", 1106 dev, error); 1107 vrele(devvp); 1108 devvp = NULL; 1109 return error; 1110 } 1111 } 1112 1113 /* 1114 * Determine if the device has already been mounted. After this 1115 * check hmp will be non-NULL if we are doing the second or more 1116 * hammer2 mounts from the same device. 1117 */ 1118 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1119 if (devvp) { 1120 /* 1121 * Match the device. Due to the way devfs works, 1122 * we may not be able to directly match the vnode pointer, 1123 * so also check to see if the underlying device matches. 1124 */ 1125 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1126 if (hmp->devvp == devvp) 1127 break; 1128 if (devvp->v_rdev && 1129 hmp->devvp->v_rdev == devvp->v_rdev) { 1130 break; 1131 } 1132 } 1133 1134 /* 1135 * If no match this may be a fresh H2 mount, make sure 1136 * the device is not mounted on anything else. 1137 */ 1138 if (hmp == NULL) 1139 error = vfs_mountedon(devvp); 1140 } else if (error == 0) { 1141 /* 1142 * Match the label to a pmp already probed. 1143 */ 1144 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 1145 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 1146 if (pmp->pfs_names[i] && 1147 strcmp(pmp->pfs_names[i], label) == 0) { 1148 hmp = pmp->pfs_hmps[i]; 1149 break; 1150 } 1151 } 1152 if (hmp) 1153 break; 1154 } 1155 if (hmp == NULL) { 1156 lockmgr(&hammer2_mntlk, LK_RELEASE); 1157 kprintf("hammer2_mount: PFS label \"%s\" not found\n", 1158 label); 1159 return ENOENT; 1160 } 1161 } 1162 1163 /* 1164 * Open the device if this isn't a secondary mount and construct 1165 * the H2 device mount (hmp). 1166 */ 1167 if (hmp == NULL) { 1168 hammer2_chain_t *schain; 1169 hammer2_xid_t xid; 1170 hammer2_xop_head_t xop; 1171 1172 if (error == 0 && vcount(devvp) > 0) { 1173 kprintf("hammer2_mount: Primary device already has references\n"); 1174 error = EBUSY; 1175 } 1176 1177 /* 1178 * Now open the device 1179 */ 1180 if (error == 0) { 1181 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0); 1182 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1183 error = vinvalbuf(devvp, V_SAVE, 0, 0); 1184 if (error == 0) { 1185 error = VOP_OPEN(devvp, 1186 (ronly ? FREAD : FREAD | FWRITE), 1187 FSCRED, NULL); 1188 } 1189 vn_unlock(devvp); 1190 } 1191 if (error && devvp) { 1192 vrele(devvp); 1193 devvp = NULL; 1194 } 1195 if (error) { 1196 lockmgr(&hammer2_mntlk, LK_RELEASE); 1197 return error; 1198 } 1199 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO); 1200 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev); 1201 hmp->ronly = ronly; 1202 hmp->devvp = devvp; 1203 hmp->hflags = info.hflags & HMNT2_DEVFLAGS; 1204 kmalloc_create(&hmp->mchain, "HAMMER2-chains"); 1205 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry); 1206 RB_INIT(&hmp->iotree); 1207 spin_init(&hmp->io_spin, "h2mount_io"); 1208 spin_init(&hmp->list_spin, "h2mount_list"); 1209 1210 lockinit(&hmp->vollk, "h2vol", 0, 0); 1211 lockinit(&hmp->bulklk, "h2bulk", 0, 0); 1212 lockinit(&hmp->bflock, "h2bflk", 0, 0); 1213 1214 /* 1215 * vchain setup. vchain.data is embedded. 1216 * vchain.refs is initialized and will never drop to 0. 1217 * 1218 * NOTE! voldata is not yet loaded. 1219 */ 1220 hmp->vchain.hmp = hmp; 1221 hmp->vchain.refs = 1; 1222 hmp->vchain.data = (void *)&hmp->voldata; 1223 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME; 1224 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1225 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1226 hammer2_chain_core_init(&hmp->vchain); 1227 1228 /* 1229 * fchain setup. fchain.data is embedded. 1230 * fchain.refs is initialized and will never drop to 0. 1231 * 1232 * The data is not used but needs to be initialized to 1233 * pass assertion muster. We use this chain primarily 1234 * as a placeholder for the freemap's top-level radix tree 1235 * so it does not interfere with the volume's topology 1236 * radix tree. 1237 */ 1238 hmp->fchain.hmp = hmp; 1239 hmp->fchain.refs = 1; 1240 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset; 1241 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP; 1242 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1243 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1244 hmp->fchain.bref.methods = 1245 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) | 1246 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE); 1247 hammer2_chain_core_init(&hmp->fchain); 1248 1249 /* 1250 * Install the volume header and initialize fields from 1251 * voldata. 1252 */ 1253 error = hammer2_install_volume_header(hmp); 1254 if (error) { 1255 hammer2_unmount_helper(mp, NULL, hmp); 1256 lockmgr(&hammer2_mntlk, LK_RELEASE); 1257 hammer2_vfs_unmount(mp, MNT_FORCE); 1258 return error; 1259 } 1260 1261 /* 1262 * Really important to get these right or the flush and 1263 * teardown code will get confused. 1264 */ 1265 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL); 1266 spmp = hmp->spmp; 1267 spmp->pfs_hmps[0] = hmp; 1268 1269 /* 1270 * Dummy-up vchain and fchain's modify_tid. mirror_tid 1271 * is inherited from the volume header. 1272 */ 1273 xid = 0; 1274 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1275 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid; 1276 hmp->vchain.pmp = spmp; 1277 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1278 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid; 1279 hmp->fchain.pmp = spmp; 1280 1281 /* 1282 * First locate the super-root inode, which is key 0 1283 * relative to the volume header's blockset. 1284 * 1285 * Then locate the root inode by scanning the directory keyspace 1286 * represented by the label. 1287 */ 1288 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 1289 schain = hammer2_chain_lookup(&parent, &key_dummy, 1290 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY, 1291 &error, 0); 1292 hammer2_chain_lookup_done(parent); 1293 if (schain == NULL) { 1294 kprintf("hammer2_mount: invalid super-root\n"); 1295 hammer2_unmount_helper(mp, NULL, hmp); 1296 lockmgr(&hammer2_mntlk, LK_RELEASE); 1297 hammer2_vfs_unmount(mp, MNT_FORCE); 1298 return EINVAL; 1299 } 1300 if (schain->error) { 1301 kprintf("hammer2_mount: error %s reading super-root\n", 1302 hammer2_error_str(schain->error)); 1303 hammer2_chain_unlock(schain); 1304 hammer2_chain_drop(schain); 1305 schain = NULL; 1306 hammer2_unmount_helper(mp, NULL, hmp); 1307 lockmgr(&hammer2_mntlk, LK_RELEASE); 1308 hammer2_vfs_unmount(mp, MNT_FORCE); 1309 return EINVAL; 1310 } 1311 1312 /* 1313 * The super-root always uses an inode_tid of 1 when 1314 * creating PFSs. 1315 */ 1316 spmp->inode_tid = 1; 1317 spmp->modify_tid = schain->bref.modify_tid + 1; 1318 1319 /* 1320 * Sanity-check schain's pmp and finish initialization. 1321 * Any chain belonging to the super-root topology should 1322 * have a NULL pmp (not even set to spmp). 1323 */ 1324 ripdata = &schain->data->ipdata; 1325 KKASSERT(schain->pmp == NULL); 1326 spmp->pfs_clid = ripdata->meta.pfs_clid; 1327 1328 /* 1329 * Replace the dummy spmp->iroot with a real one. It's 1330 * easier to just do a wholesale replacement than to try 1331 * to update the chain and fixup the iroot fields. 1332 * 1333 * The returned inode is locked with the supplied cluster. 1334 */ 1335 hammer2_dummy_xop_from_chain(&xop, schain); 1336 hammer2_inode_drop(spmp->iroot); 1337 spmp->iroot = NULL; 1338 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1); 1339 spmp->spmp_hmp = hmp; 1340 spmp->pfs_types[0] = ripdata->meta.pfs_type; 1341 spmp->pfs_hmps[0] = hmp; 1342 hammer2_inode_ref(spmp->iroot); 1343 hammer2_inode_unlock(spmp->iroot); 1344 hammer2_cluster_unlock(&xop.cluster); 1345 hammer2_chain_drop(schain); 1346 /* do not call hammer2_cluster_drop() on an embedded cluster */ 1347 schain = NULL; /* now invalid */ 1348 /* leave spmp->iroot with one ref */ 1349 1350 if ((mp->mnt_flag & MNT_RDONLY) == 0) { 1351 error = hammer2_recovery(hmp); 1352 if (error == 0) 1353 error |= hammer2_fixup_pfses(hmp); 1354 /* XXX do something with error */ 1355 } 1356 hammer2_update_pmps(hmp); 1357 hammer2_iocom_init(hmp); 1358 hammer2_bulkfree_init(hmp); 1359 1360 /* 1361 * Ref the cluster management messaging descriptor. The mount 1362 * program deals with the other end of the communications pipe. 1363 * 1364 * Root mounts typically do not supply one. 1365 */ 1366 if (info.cluster_fd >= 0) { 1367 fp = holdfp(curthread, info.cluster_fd, -1); 1368 if (fp) { 1369 hammer2_cluster_reconnect(hmp, fp); 1370 } else { 1371 kprintf("hammer2_mount: bad cluster_fd!\n"); 1372 } 1373 } 1374 } else { 1375 spmp = hmp->spmp; 1376 if (info.hflags & HMNT2_DEVFLAGS) { 1377 kprintf("hammer2_mount: Warning: mount flags pertaining " 1378 "to the whole device may only be specified " 1379 "on the first mount of the device: %08x\n", 1380 info.hflags & HMNT2_DEVFLAGS); 1381 } 1382 } 1383 1384 /* 1385 * Force local mount (disassociate all PFSs from their clusters). 1386 * Used primarily for debugging. 1387 */ 1388 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1389 1390 /* 1391 * Lookup the mount point under the media-localized super-root. 1392 * Scanning hammer2_pfslist doesn't help us because it represents 1393 * PFS cluster ids which can aggregate several named PFSs together. 1394 * 1395 * cluster->pmp will incorrectly point to spmp and must be fixed 1396 * up later on. 1397 */ 1398 hammer2_inode_lock(spmp->iroot, 0); 1399 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1400 lhc = hammer2_dirhash(label, strlen(label)); 1401 chain = hammer2_chain_lookup(&parent, &key_next, 1402 lhc, lhc + HAMMER2_DIRHASH_LOMASK, 1403 &error, 0); 1404 while (chain) { 1405 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && 1406 strcmp(label, chain->data->ipdata.filename) == 0) { 1407 break; 1408 } 1409 chain = hammer2_chain_next(&parent, chain, &key_next, 1410 key_next, 1411 lhc + HAMMER2_DIRHASH_LOMASK, 1412 &error, 0); 1413 } 1414 if (parent) { 1415 hammer2_chain_unlock(parent); 1416 hammer2_chain_drop(parent); 1417 } 1418 hammer2_inode_unlock(spmp->iroot); 1419 1420 /* 1421 * PFS could not be found? 1422 */ 1423 if (chain == NULL) { 1424 hammer2_unmount_helper(mp, NULL, hmp); 1425 lockmgr(&hammer2_mntlk, LK_RELEASE); 1426 hammer2_vfs_unmount(mp, MNT_FORCE); 1427 1428 if (error) { 1429 kprintf("hammer2_mount: PFS label I/O error\n"); 1430 return EINVAL; 1431 } else { 1432 kprintf("hammer2_mount: PFS label \"%s\" not found\n", 1433 label); 1434 return ENOENT; 1435 } 1436 } 1437 1438 /* 1439 * Acquire the pmp structure (it should have already been allocated 1440 * via hammer2_update_pmps() so do not pass cluster in to add to 1441 * available chains). 1442 * 1443 * Check if the cluster has already been mounted. A cluster can 1444 * only be mounted once, use null mounts to mount additional copies. 1445 */ 1446 if (chain->error) { 1447 kprintf("hammer2_mount: PFS label I/O error\n"); 1448 } else { 1449 ripdata = &chain->data->ipdata; 1450 bref = chain->bref; 1451 pmp = hammer2_pfsalloc(NULL, ripdata, 1452 bref.modify_tid, force_local); 1453 } 1454 hammer2_chain_unlock(chain); 1455 hammer2_chain_drop(chain); 1456 1457 /* 1458 * Finish the mount 1459 */ 1460 kprintf("hammer2_mount: hmp=%p pmp=%p\n", hmp, pmp); 1461 1462 if (pmp->mp) { 1463 kprintf("hammer2_mount: PFS already mounted!\n"); 1464 hammer2_unmount_helper(mp, NULL, hmp); 1465 lockmgr(&hammer2_mntlk, LK_RELEASE); 1466 hammer2_vfs_unmount(mp, MNT_FORCE); 1467 1468 return EBUSY; 1469 } 1470 1471 pmp->hflags = info.hflags; 1472 mp->mnt_flag |= MNT_LOCAL; 1473 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */ 1474 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */ 1475 1476 /* 1477 * required mount structure initializations 1478 */ 1479 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE; 1480 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE; 1481 1482 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE; 1483 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE; 1484 1485 /* 1486 * Optional fields 1487 */ 1488 mp->mnt_iosize_max = MAXPHYS; 1489 1490 /* 1491 * Connect up mount pointers. 1492 */ 1493 hammer2_mount_helper(mp, pmp); 1494 hmp->devvp->v_rdev->si_mountpoint = mp; 1495 1496 lockmgr(&hammer2_mntlk, LK_RELEASE); 1497 1498 /* 1499 * Finish setup 1500 */ 1501 vfs_getnewfsid(mp); 1502 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops); 1503 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops); 1504 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops); 1505 1506 if (path) { 1507 copyinstr(info.volume, mp->mnt_stat.f_mntfromname, 1508 MNAMELEN - 1, &size); 1509 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size); 1510 } /* else root mount, already in there */ 1511 1512 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname)); 1513 if (path) { 1514 copyinstr(path, mp->mnt_stat.f_mntonname, 1515 sizeof(mp->mnt_stat.f_mntonname) - 1, 1516 &size); 1517 } else { 1518 /* root mount */ 1519 mp->mnt_stat.f_mntonname[0] = '/'; 1520 } 1521 1522 /* 1523 * Initial statfs to prime mnt_stat. 1524 */ 1525 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred); 1526 1527 return 0; 1528 } 1529 1530 /* 1531 * Scan PFSs under the super-root and create hammer2_pfs structures. 1532 */ 1533 static 1534 void 1535 hammer2_update_pmps(hammer2_dev_t *hmp) 1536 { 1537 const hammer2_inode_data_t *ripdata; 1538 hammer2_chain_t *parent; 1539 hammer2_chain_t *chain; 1540 hammer2_blockref_t bref; 1541 hammer2_dev_t *force_local; 1542 hammer2_pfs_t *spmp; 1543 hammer2_pfs_t *pmp; 1544 hammer2_key_t key_next; 1545 int error; 1546 1547 /* 1548 * Force local mount (disassociate all PFSs from their clusters). 1549 * Used primarily for debugging. 1550 */ 1551 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1552 1553 /* 1554 * Lookup mount point under the media-localized super-root. 1555 * 1556 * cluster->pmp will incorrectly point to spmp and must be fixed 1557 * up later on. 1558 */ 1559 spmp = hmp->spmp; 1560 hammer2_inode_lock(spmp->iroot, 0); 1561 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1562 chain = hammer2_chain_lookup(&parent, &key_next, 1563 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 1564 &error, 0); 1565 while (chain) { 1566 if (chain->error) { 1567 kprintf("I/O error scanning PFS labels\n"); 1568 } else if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) { 1569 kprintf("Non inode chain type %d under super-root\n", 1570 chain->bref.type); 1571 } else { 1572 ripdata = &chain->data->ipdata; 1573 bref = chain->bref; 1574 pmp = hammer2_pfsalloc(chain, ripdata, 1575 bref.modify_tid, force_local); 1576 } 1577 chain = hammer2_chain_next(&parent, chain, &key_next, 1578 key_next, HAMMER2_KEY_MAX, 1579 &error, 0); 1580 } 1581 if (parent) { 1582 hammer2_chain_unlock(parent); 1583 hammer2_chain_drop(parent); 1584 } 1585 hammer2_inode_unlock(spmp->iroot); 1586 } 1587 1588 static 1589 int 1590 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused, 1591 struct vnode *devvp, struct ucred *cred) 1592 { 1593 int error; 1594 1595 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) { 1596 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1597 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL); 1598 vn_unlock(devvp); 1599 error = hammer2_recovery(hmp); 1600 if (error == 0) 1601 error |= hammer2_fixup_pfses(hmp); 1602 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1603 if (error == 0) { 1604 VOP_CLOSE(devvp, FREAD, NULL); 1605 hmp->ronly = 0; 1606 } else { 1607 VOP_CLOSE(devvp, FREAD | FWRITE, NULL); 1608 } 1609 vn_unlock(devvp); 1610 } else { 1611 error = 0; 1612 } 1613 return error; 1614 } 1615 1616 static 1617 int 1618 hammer2_vfs_unmount(struct mount *mp, int mntflags) 1619 { 1620 hammer2_pfs_t *pmp; 1621 int flags; 1622 int error = 0; 1623 1624 pmp = MPTOPMP(mp); 1625 1626 if (pmp == NULL) 1627 return(0); 1628 1629 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1630 1631 /* 1632 * If mount initialization proceeded far enough we must flush 1633 * its vnodes and sync the underlying mount points. Three syncs 1634 * are required to fully flush the filesystem (freemap updates lag 1635 * by one flush, and one extra for safety). 1636 */ 1637 if (mntflags & MNT_FORCE) 1638 flags = FORCECLOSE; 1639 else 1640 flags = 0; 1641 if (pmp->iroot) { 1642 error = vflush(mp, 0, flags); 1643 if (error) 1644 goto failed; 1645 hammer2_vfs_sync(mp, MNT_WAIT); 1646 hammer2_vfs_sync(mp, MNT_WAIT); 1647 hammer2_vfs_sync(mp, MNT_WAIT); 1648 } 1649 1650 /* 1651 * Cleanup the frontend support XOPS threads 1652 */ 1653 hammer2_xop_helper_cleanup(pmp); 1654 1655 if (pmp->mp) 1656 hammer2_unmount_helper(mp, pmp, NULL); 1657 1658 error = 0; 1659 failed: 1660 lockmgr(&hammer2_mntlk, LK_RELEASE); 1661 1662 return (error); 1663 } 1664 1665 /* 1666 * Mount helper, hook the system mount into our PFS. 1667 * The mount lock is held. 1668 * 1669 * We must bump the mount_count on related devices for any 1670 * mounted PFSs. 1671 */ 1672 static 1673 void 1674 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp) 1675 { 1676 hammer2_cluster_t *cluster; 1677 hammer2_chain_t *rchain; 1678 int i; 1679 1680 mp->mnt_data = (qaddr_t)pmp; 1681 pmp->mp = mp; 1682 1683 /* 1684 * After pmp->mp is set we have to adjust hmp->mount_count. 1685 */ 1686 cluster = &pmp->iroot->cluster; 1687 for (i = 0; i < cluster->nchains; ++i) { 1688 rchain = cluster->array[i].chain; 1689 if (rchain == NULL) 1690 continue; 1691 ++rchain->hmp->mount_count; 1692 } 1693 1694 /* 1695 * Create missing Xop threads 1696 */ 1697 hammer2_xop_helper_create(pmp); 1698 } 1699 1700 /* 1701 * Mount helper, unhook the system mount from our PFS. 1702 * The mount lock is held. 1703 * 1704 * If hmp is supplied a mount responsible for being the first to open 1705 * the block device failed and the block device and all PFSs using the 1706 * block device must be cleaned up. 1707 * 1708 * If pmp is supplied multiple devices might be backing the PFS and each 1709 * must be disconnected. This might not be the last PFS using some of the 1710 * underlying devices. Also, we have to adjust our hmp->mount_count 1711 * accounting for the devices backing the pmp which is now undergoing an 1712 * unmount. 1713 */ 1714 static 1715 void 1716 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp) 1717 { 1718 hammer2_cluster_t *cluster; 1719 hammer2_chain_t *rchain; 1720 struct vnode *devvp; 1721 int dumpcnt; 1722 int ronly; 1723 int i; 1724 1725 /* 1726 * If no device supplied this is a high-level unmount and we have to 1727 * to disconnect the mount, adjust mount_count, and locate devices 1728 * that might now have no mounts. 1729 */ 1730 if (pmp) { 1731 KKASSERT(hmp == NULL); 1732 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp); 1733 pmp->mp = NULL; 1734 mp->mnt_data = NULL; 1735 1736 /* 1737 * After pmp->mp is cleared we have to account for 1738 * mount_count. 1739 */ 1740 cluster = &pmp->iroot->cluster; 1741 for (i = 0; i < cluster->nchains; ++i) { 1742 rchain = cluster->array[i].chain; 1743 if (rchain == NULL) 1744 continue; 1745 --rchain->hmp->mount_count; 1746 /* scrapping hmp now may invalidate the pmp */ 1747 } 1748 again: 1749 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1750 if (hmp->mount_count == 0) { 1751 hammer2_unmount_helper(NULL, NULL, hmp); 1752 goto again; 1753 } 1754 } 1755 return; 1756 } 1757 1758 /* 1759 * Try to terminate the block device. We can't terminate it if 1760 * there are still PFSs referencing it. 1761 */ 1762 if (hmp->mount_count) 1763 return; 1764 1765 /* 1766 * Decomission the network before we start messing with the 1767 * device and PFS. 1768 */ 1769 hammer2_iocom_uninit(hmp); 1770 1771 hammer2_bulkfree_uninit(hmp); 1772 hammer2_pfsfree_scan(hmp, 0); 1773 1774 /* 1775 * Cycle the volume data lock as a safety (probably not needed any 1776 * more). To ensure everything is out we need to flush at least 1777 * three times. (1) The running of the sideq can dirty the 1778 * filesystem, (2) A normal flush can dirty the freemap, and 1779 * (3) ensure that the freemap is fully synchronized. 1780 * 1781 * The next mount's recovery scan can clean everything up but we want 1782 * to leave the filesystem in a 100% clean state on a normal unmount. 1783 */ 1784 #if 0 1785 hammer2_voldata_lock(hmp); 1786 hammer2_voldata_unlock(hmp); 1787 #endif 1788 1789 /* 1790 * Flush whatever is left. Unmounted but modified PFS's might still 1791 * have some dirty chains on them. 1792 */ 1793 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS); 1794 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS); 1795 1796 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) { 1797 hammer2_voldata_modify(hmp); 1798 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP | 1799 HAMMER2_FLUSH_ALL); 1800 } 1801 hammer2_chain_unlock(&hmp->fchain); 1802 1803 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) { 1804 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP | 1805 HAMMER2_FLUSH_ALL); 1806 } 1807 hammer2_chain_unlock(&hmp->vchain); 1808 1809 if ((hmp->vchain.flags | hmp->fchain.flags) & 1810 HAMMER2_CHAIN_FLUSH_MASK) { 1811 kprintf("hammer2_unmount: chains left over after final sync\n"); 1812 kprintf(" vchain %08x\n", hmp->vchain.flags); 1813 kprintf(" fchain %08x\n", hmp->fchain.flags); 1814 1815 if (hammer2_debug & 0x0010) 1816 Debugger("entered debugger"); 1817 } 1818 1819 hammer2_pfsfree_scan(hmp, 1); 1820 1821 KKASSERT(hmp->spmp == NULL); 1822 1823 /* 1824 * Finish up with the device vnode 1825 */ 1826 if ((devvp = hmp->devvp) != NULL) { 1827 ronly = hmp->ronly; 1828 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1829 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n", 1830 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree), 1831 ronly); 1832 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0); 1833 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n", 1834 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree)); 1835 devvp->v_rdev->si_mountpoint = NULL; 1836 hmp->devvp = NULL; 1837 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL); 1838 vn_unlock(devvp); 1839 vrele(devvp); 1840 devvp = NULL; 1841 } 1842 1843 /* 1844 * Clear vchain/fchain flags that might prevent final cleanup 1845 * of these chains. 1846 */ 1847 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) { 1848 atomic_add_long(&hammer2_count_modified_chains, -1); 1849 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 1850 hammer2_pfs_memory_wakeup(hmp->vchain.pmp, -1); 1851 } 1852 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) { 1853 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE); 1854 } 1855 1856 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) { 1857 atomic_add_long(&hammer2_count_modified_chains, -1); 1858 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED); 1859 hammer2_pfs_memory_wakeup(hmp->fchain.pmp, -1); 1860 } 1861 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) { 1862 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE); 1863 } 1864 1865 /* 1866 * Final drop of embedded freemap root chain to 1867 * clean up fchain.core (fchain structure is not 1868 * flagged ALLOCATED so it is cleaned out and then 1869 * left to rot). 1870 */ 1871 hammer2_chain_drop(&hmp->fchain); 1872 1873 /* 1874 * Final drop of embedded volume root chain to clean 1875 * up vchain.core (vchain structure is not flagged 1876 * ALLOCATED so it is cleaned out and then left to 1877 * rot). 1878 */ 1879 dumpcnt = 50; 1880 hammer2_dump_chain(&hmp->vchain, 0, 0, &dumpcnt, 'v', (u_int)-1); 1881 dumpcnt = 50; 1882 hammer2_dump_chain(&hmp->fchain, 0, 0, &dumpcnt, 'f', (u_int)-1); 1883 1884 hammer2_chain_drop(&hmp->vchain); 1885 1886 hammer2_io_cleanup(hmp, &hmp->iotree); 1887 if (hmp->iofree_count) { 1888 kprintf("io_cleanup: %d I/O's left hanging\n", 1889 hmp->iofree_count); 1890 } 1891 1892 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry); 1893 kmalloc_destroy(&hmp->mchain); 1894 kfree(hmp, M_HAMMER2); 1895 } 1896 1897 int 1898 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp, 1899 ino_t ino, struct vnode **vpp) 1900 { 1901 hammer2_xop_lookup_t *xop; 1902 hammer2_pfs_t *pmp; 1903 hammer2_inode_t *ip; 1904 hammer2_tid_t inum; 1905 int error; 1906 1907 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK; 1908 1909 error = 0; 1910 pmp = MPTOPMP(mp); 1911 1912 /* 1913 * Easy if we already have it cached 1914 */ 1915 ip = hammer2_inode_lookup(pmp, inum); 1916 if (ip) { 1917 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED); 1918 *vpp = hammer2_igetv(ip, &error); 1919 hammer2_inode_unlock(ip); 1920 hammer2_inode_drop(ip); /* from lookup */ 1921 1922 return error; 1923 } 1924 1925 /* 1926 * Otherwise we have to find the inode 1927 */ 1928 xop = hammer2_xop_alloc(pmp->iroot, 0); 1929 xop->lhc = inum; 1930 hammer2_xop_start(&xop->head, &hammer2_lookup_desc); 1931 error = hammer2_xop_collect(&xop->head, 0); 1932 1933 if (error == 0) 1934 ip = hammer2_inode_get(pmp, &xop->head, -1, -1); 1935 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1936 1937 if (ip) { 1938 *vpp = hammer2_igetv(ip, &error); 1939 hammer2_inode_unlock(ip); 1940 } else { 1941 *vpp = NULL; 1942 error = ENOENT; 1943 } 1944 return (error); 1945 } 1946 1947 static 1948 int 1949 hammer2_vfs_root(struct mount *mp, struct vnode **vpp) 1950 { 1951 hammer2_pfs_t *pmp; 1952 struct vnode *vp; 1953 int error; 1954 1955 pmp = MPTOPMP(mp); 1956 if (pmp->iroot == NULL) { 1957 kprintf("hammer2 (%s): no root inode\n", 1958 mp->mnt_stat.f_mntfromname); 1959 *vpp = NULL; 1960 return EINVAL; 1961 } 1962 1963 error = 0; 1964 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1965 1966 while (pmp->inode_tid == 0) { 1967 hammer2_xop_ipcluster_t *xop; 1968 const hammer2_inode_meta_t *meta; 1969 1970 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING); 1971 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc); 1972 error = hammer2_xop_collect(&xop->head, 0); 1973 1974 if (error == 0) { 1975 meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta; 1976 pmp->iroot->meta = *meta; 1977 pmp->inode_tid = meta->pfs_inum + 1; 1978 hammer2_xop_pdata(&xop->head); 1979 /* meta invalid */ 1980 1981 if (pmp->inode_tid < HAMMER2_INODE_START) 1982 pmp->inode_tid = HAMMER2_INODE_START; 1983 pmp->modify_tid = 1984 xop->head.cluster.focus->bref.modify_tid + 1; 1985 #if 0 1986 kprintf("PFS: Starting inode %jd\n", 1987 (intmax_t)pmp->inode_tid); 1988 kprintf("PMP focus good set nextino=%ld mod=%016jx\n", 1989 pmp->inode_tid, pmp->modify_tid); 1990 #endif 1991 wakeup(&pmp->iroot); 1992 1993 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1994 1995 /* 1996 * Prime the mount info. 1997 */ 1998 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL); 1999 break; 2000 } 2001 2002 /* 2003 * Loop, try again 2004 */ 2005 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 2006 hammer2_inode_unlock(pmp->iroot); 2007 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz); 2008 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 2009 if (error == EINTR) 2010 break; 2011 } 2012 2013 if (error) { 2014 hammer2_inode_unlock(pmp->iroot); 2015 *vpp = NULL; 2016 } else { 2017 vp = hammer2_igetv(pmp->iroot, &error); 2018 hammer2_inode_unlock(pmp->iroot); 2019 *vpp = vp; 2020 } 2021 2022 return (error); 2023 } 2024 2025 /* 2026 * Filesystem status 2027 * 2028 * XXX incorporate ipdata->meta.inode_quota and data_quota 2029 */ 2030 static 2031 int 2032 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred) 2033 { 2034 hammer2_pfs_t *pmp; 2035 hammer2_dev_t *hmp; 2036 hammer2_blockref_t bref; 2037 struct statfs tmp; 2038 int i; 2039 2040 /* 2041 * NOTE: iroot might not have validated the cluster yet. 2042 */ 2043 pmp = MPTOPMP(mp); 2044 2045 bzero(&tmp, sizeof(tmp)); 2046 2047 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 2048 hmp = pmp->pfs_hmps[i]; 2049 if (hmp == NULL) 2050 continue; 2051 if (pmp->iroot->cluster.array[i].chain) 2052 bref = pmp->iroot->cluster.array[i].chain->bref; 2053 else 2054 bzero(&bref, sizeof(bref)); 2055 2056 tmp.f_files = bref.embed.stats.inode_count; 2057 tmp.f_ffree = 0; 2058 tmp.f_blocks = hmp->voldata.allocator_size / 2059 mp->mnt_vstat.f_bsize; 2060 tmp.f_bfree = hmp->voldata.allocator_free / 2061 mp->mnt_vstat.f_bsize; 2062 tmp.f_bavail = tmp.f_bfree; 2063 2064 if (cred && cred->cr_uid != 0) { 2065 uint64_t adj; 2066 2067 /* 5% */ 2068 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize; 2069 tmp.f_blocks -= adj; 2070 tmp.f_bfree -= adj; 2071 tmp.f_bavail -= adj; 2072 } 2073 2074 mp->mnt_stat.f_blocks = tmp.f_blocks; 2075 mp->mnt_stat.f_bfree = tmp.f_bfree; 2076 mp->mnt_stat.f_bavail = tmp.f_bavail; 2077 mp->mnt_stat.f_files = tmp.f_files; 2078 mp->mnt_stat.f_ffree = tmp.f_ffree; 2079 2080 *sbp = mp->mnt_stat; 2081 } 2082 return (0); 2083 } 2084 2085 static 2086 int 2087 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred) 2088 { 2089 hammer2_pfs_t *pmp; 2090 hammer2_dev_t *hmp; 2091 hammer2_blockref_t bref; 2092 struct statvfs tmp; 2093 int i; 2094 2095 /* 2096 * NOTE: iroot might not have validated the cluster yet. 2097 */ 2098 pmp = MPTOPMP(mp); 2099 bzero(&tmp, sizeof(tmp)); 2100 2101 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 2102 hmp = pmp->pfs_hmps[i]; 2103 if (hmp == NULL) 2104 continue; 2105 if (pmp->iroot->cluster.array[i].chain) 2106 bref = pmp->iroot->cluster.array[i].chain->bref; 2107 else 2108 bzero(&bref, sizeof(bref)); 2109 2110 tmp.f_files = bref.embed.stats.inode_count; 2111 tmp.f_ffree = 0; 2112 tmp.f_blocks = hmp->voldata.allocator_size / 2113 mp->mnt_vstat.f_bsize; 2114 tmp.f_bfree = hmp->voldata.allocator_free / 2115 mp->mnt_vstat.f_bsize; 2116 tmp.f_bavail = tmp.f_bfree; 2117 2118 if (cred && cred->cr_uid != 0) { 2119 uint64_t adj; 2120 2121 /* 5% */ 2122 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize; 2123 tmp.f_blocks -= adj; 2124 tmp.f_bfree -= adj; 2125 tmp.f_bavail -= adj; 2126 } 2127 2128 mp->mnt_vstat.f_blocks = tmp.f_blocks; 2129 mp->mnt_vstat.f_bfree = tmp.f_bfree; 2130 mp->mnt_vstat.f_bavail = tmp.f_bavail; 2131 mp->mnt_vstat.f_files = tmp.f_files; 2132 mp->mnt_vstat.f_ffree = tmp.f_ffree; 2133 2134 *sbp = mp->mnt_vstat; 2135 } 2136 return (0); 2137 } 2138 2139 /* 2140 * Mount-time recovery (RW mounts) 2141 * 2142 * Updates to the free block table are allowed to lag flushes by one 2143 * transaction. In case of a crash, then on a fresh mount we must do an 2144 * incremental scan of the last committed transaction id and make sure that 2145 * all related blocks have been marked allocated. 2146 */ 2147 struct hammer2_recovery_elm { 2148 TAILQ_ENTRY(hammer2_recovery_elm) entry; 2149 hammer2_chain_t *chain; 2150 hammer2_tid_t sync_tid; 2151 }; 2152 2153 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm); 2154 2155 struct hammer2_recovery_info { 2156 struct hammer2_recovery_list list; 2157 hammer2_tid_t mtid; 2158 int depth; 2159 }; 2160 2161 static int hammer2_recovery_scan(hammer2_dev_t *hmp, 2162 hammer2_chain_t *parent, 2163 struct hammer2_recovery_info *info, 2164 hammer2_tid_t sync_tid); 2165 2166 #define HAMMER2_RECOVERY_MAXDEPTH 10 2167 2168 static 2169 int 2170 hammer2_recovery(hammer2_dev_t *hmp) 2171 { 2172 struct hammer2_recovery_info info; 2173 struct hammer2_recovery_elm *elm; 2174 hammer2_chain_t *parent; 2175 hammer2_tid_t sync_tid; 2176 hammer2_tid_t mirror_tid; 2177 int error; 2178 2179 hammer2_trans_init(hmp->spmp, 0); 2180 2181 sync_tid = hmp->voldata.freemap_tid; 2182 mirror_tid = hmp->voldata.mirror_tid; 2183 2184 kprintf("hammer2_mount: \"%s\": ", hmp->devrepname); 2185 if (sync_tid >= mirror_tid) { 2186 kprintf("no recovery needed\n"); 2187 } else { 2188 kprintf("freemap recovery %016jx-%016jx\n", 2189 sync_tid + 1, mirror_tid); 2190 } 2191 2192 TAILQ_INIT(&info.list); 2193 info.depth = 0; 2194 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 2195 error = hammer2_recovery_scan(hmp, parent, &info, sync_tid); 2196 hammer2_chain_lookup_done(parent); 2197 2198 while ((elm = TAILQ_FIRST(&info.list)) != NULL) { 2199 TAILQ_REMOVE(&info.list, elm, entry); 2200 parent = elm->chain; 2201 sync_tid = elm->sync_tid; 2202 kfree(elm, M_HAMMER2); 2203 2204 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2205 error |= hammer2_recovery_scan(hmp, parent, &info, 2206 hmp->voldata.freemap_tid); 2207 hammer2_chain_unlock(parent); 2208 hammer2_chain_drop(parent); /* drop elm->chain ref */ 2209 } 2210 2211 hammer2_trans_done(hmp->spmp, 0); 2212 2213 return error; 2214 } 2215 2216 static 2217 int 2218 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent, 2219 struct hammer2_recovery_info *info, 2220 hammer2_tid_t sync_tid) 2221 { 2222 const hammer2_inode_data_t *ripdata; 2223 hammer2_chain_t *chain; 2224 hammer2_blockref_t bref; 2225 int tmp_error; 2226 int rup_error; 2227 int error; 2228 int first; 2229 2230 /* 2231 * Adjust freemap to ensure that the block(s) are marked allocated. 2232 */ 2233 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) { 2234 hammer2_freemap_adjust(hmp, &parent->bref, 2235 HAMMER2_FREEMAP_DORECOVER); 2236 } 2237 2238 /* 2239 * Check type for recursive scan 2240 */ 2241 switch(parent->bref.type) { 2242 case HAMMER2_BREF_TYPE_VOLUME: 2243 /* data already instantiated */ 2244 break; 2245 case HAMMER2_BREF_TYPE_INODE: 2246 /* 2247 * Must instantiate data for DIRECTDATA test and also 2248 * for recursion. 2249 */ 2250 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2251 ripdata = &parent->data->ipdata; 2252 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) { 2253 /* not applicable to recovery scan */ 2254 hammer2_chain_unlock(parent); 2255 return 0; 2256 } 2257 hammer2_chain_unlock(parent); 2258 break; 2259 case HAMMER2_BREF_TYPE_INDIRECT: 2260 /* 2261 * Must instantiate data for recursion 2262 */ 2263 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2264 hammer2_chain_unlock(parent); 2265 break; 2266 case HAMMER2_BREF_TYPE_DIRENT: 2267 case HAMMER2_BREF_TYPE_DATA: 2268 case HAMMER2_BREF_TYPE_FREEMAP: 2269 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2270 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 2271 /* not applicable to recovery scan */ 2272 return 0; 2273 break; 2274 default: 2275 return HAMMER2_ERROR_BADBREF; 2276 } 2277 2278 /* 2279 * Defer operation if depth limit reached. 2280 */ 2281 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) { 2282 struct hammer2_recovery_elm *elm; 2283 2284 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK); 2285 elm->chain = parent; 2286 elm->sync_tid = sync_tid; 2287 hammer2_chain_ref(parent); 2288 TAILQ_INSERT_TAIL(&info->list, elm, entry); 2289 /* unlocked by caller */ 2290 2291 return(0); 2292 } 2293 2294 2295 /* 2296 * Recursive scan of the last flushed transaction only. We are 2297 * doing this without pmp assignments so don't leave the chains 2298 * hanging around after we are done with them. 2299 * 2300 * error Cumulative error this level only 2301 * rup_error Cumulative error for recursion 2302 * tmp_error Specific non-cumulative recursion error 2303 */ 2304 chain = NULL; 2305 first = 1; 2306 rup_error = 0; 2307 error = 0; 2308 2309 for (;;) { 2310 error |= hammer2_chain_scan(parent, &chain, &bref, 2311 &first, 2312 HAMMER2_LOOKUP_NODATA); 2313 2314 /* 2315 * Problem during scan or EOF 2316 */ 2317 if (error) 2318 break; 2319 2320 /* 2321 * If this is a leaf 2322 */ 2323 if (chain == NULL) { 2324 if (bref.mirror_tid > sync_tid) { 2325 hammer2_freemap_adjust(hmp, &bref, 2326 HAMMER2_FREEMAP_DORECOVER); 2327 } 2328 continue; 2329 } 2330 2331 /* 2332 * This may or may not be a recursive node. 2333 */ 2334 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 2335 if (bref.mirror_tid > sync_tid) { 2336 ++info->depth; 2337 tmp_error = hammer2_recovery_scan(hmp, chain, 2338 info, sync_tid); 2339 --info->depth; 2340 } else { 2341 tmp_error = 0; 2342 } 2343 2344 /* 2345 * Flush the recovery at the PFS boundary to stage it for 2346 * the final flush of the super-root topology. 2347 */ 2348 if (tmp_error == 0 && 2349 (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) && 2350 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) { 2351 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 2352 HAMMER2_FLUSH_ALL); 2353 } 2354 rup_error |= tmp_error; 2355 } 2356 return ((error | rup_error) & ~HAMMER2_ERROR_EOF); 2357 } 2358 2359 /* 2360 * This fixes up an error introduced in earlier H2 implementations where 2361 * moving a PFS inode into an indirect block wound up causing the 2362 * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared. 2363 */ 2364 static 2365 int 2366 hammer2_fixup_pfses(hammer2_dev_t *hmp) 2367 { 2368 const hammer2_inode_data_t *ripdata; 2369 hammer2_chain_t *parent; 2370 hammer2_chain_t *chain; 2371 hammer2_key_t key_next; 2372 hammer2_pfs_t *spmp; 2373 int error; 2374 2375 error = 0; 2376 2377 /* 2378 * Lookup mount point under the media-localized super-root. 2379 * 2380 * cluster->pmp will incorrectly point to spmp and must be fixed 2381 * up later on. 2382 */ 2383 spmp = hmp->spmp; 2384 hammer2_inode_lock(spmp->iroot, 0); 2385 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 2386 chain = hammer2_chain_lookup(&parent, &key_next, 2387 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 2388 &error, 0); 2389 while (chain) { 2390 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) 2391 continue; 2392 if (chain->error) { 2393 kprintf("I/O error scanning PFS labels\n"); 2394 error |= chain->error; 2395 } else if ((chain->bref.flags & 2396 HAMMER2_BREF_FLAG_PFSROOT) == 0) { 2397 int error2; 2398 2399 ripdata = &chain->data->ipdata; 2400 hammer2_trans_init(hmp->spmp, 0); 2401 error2 = hammer2_chain_modify(chain, 2402 chain->bref.modify_tid, 2403 0, 0); 2404 if (error2 == 0) { 2405 kprintf("hammer2: Correct mis-flagged PFS %s\n", 2406 ripdata->filename); 2407 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT; 2408 } else { 2409 error |= error2; 2410 } 2411 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 2412 HAMMER2_FLUSH_ALL); 2413 hammer2_trans_done(hmp->spmp, 0); 2414 } 2415 chain = hammer2_chain_next(&parent, chain, &key_next, 2416 key_next, HAMMER2_KEY_MAX, 2417 &error, 0); 2418 } 2419 if (parent) { 2420 hammer2_chain_unlock(parent); 2421 hammer2_chain_drop(parent); 2422 } 2423 hammer2_inode_unlock(spmp->iroot); 2424 2425 return error; 2426 } 2427 2428 /* 2429 * Sync a mount point; this is called periodically on a per-mount basis from 2430 * the filesystem syncer, and whenever a user issues a sync. 2431 */ 2432 int 2433 hammer2_vfs_sync(struct mount *mp, int waitfor) 2434 { 2435 int error; 2436 2437 error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor); 2438 2439 return error; 2440 } 2441 2442 /* 2443 * Because frontend operations lock vnodes before we get a chance to 2444 * lock the related inode, we can't just acquire a vnode lock without 2445 * risking a deadlock. The frontend may be holding a vnode lock while 2446 * also blocked on our SYNCQ flag while trying to get the inode lock. 2447 * 2448 * To deal with this situation we can check the vnode lock situation 2449 * after locking the inode and perform a work-around. 2450 */ 2451 int 2452 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor) 2453 { 2454 struct mount *mp; 2455 /*hammer2_xop_flush_t *xop;*/ 2456 /*struct hammer2_sync_info info;*/ 2457 hammer2_inode_t *ip; 2458 hammer2_depend_t *depend; 2459 hammer2_depend_t *depend_next; 2460 struct vnode *vp; 2461 uint32_t pass2; 2462 int error; 2463 int wakecount; 2464 int dorestart; 2465 2466 mp = pmp->mp; 2467 2468 /* 2469 * Move all inodes on sideq to syncq. This will clear sideq. 2470 * This should represent all flushable inodes. These inodes 2471 * will already have refs due to being on syncq or sideq. We 2472 * must do this all at once with the spinlock held to ensure that 2473 * all inode dependencies are part of the same flush. 2474 * 2475 * We should be able to do this asynchronously from frontend 2476 * operations because we will be locking the inodes later on 2477 * to actually flush them, and that will partition any frontend 2478 * op using the same inode. Either it has already locked the 2479 * inode and we will block, or it has not yet locked the inode 2480 * and it will block until we are finished flushing that inode. 2481 * 2482 * When restarting, only move the inodes flagged as PASS2 from 2483 * SIDEQ to SYNCQ. PASS2 propagation by inode_lock4() and 2484 * inode_depend() are atomic with the spin-lock. 2485 */ 2486 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH); 2487 #ifdef HAMMER2_DEBUG_SYNC 2488 kprintf("FILESYSTEM SYNC BOUNDARY\n"); 2489 #endif 2490 dorestart = 0; 2491 2492 /* 2493 * Move inodes from depq to syncq, releasing the related 2494 * depend structures. 2495 */ 2496 restart: 2497 #ifdef HAMMER2_DEBUG_SYNC 2498 kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart); 2499 #endif 2500 hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/); 2501 hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN); 2502 2503 /* 2504 * Move inodes from depq to syncq. When restarting, only depq's 2505 * marked pass2 are moved. 2506 */ 2507 hammer2_spin_ex(&pmp->list_spin); 2508 depend_next = TAILQ_FIRST(&pmp->depq); 2509 wakecount = 0; 2510 2511 while ((depend = depend_next) != NULL) { 2512 depend_next = TAILQ_NEXT(depend, entry); 2513 if (dorestart && depend->pass2 == 0) 2514 continue; 2515 TAILQ_FOREACH(ip, &depend->sideq, entry) { 2516 KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ); 2517 atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ); 2518 atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ); 2519 ip->depend = NULL; 2520 } 2521 2522 /* 2523 * NOTE: pmp->sideq_count includes both sideq and syncq 2524 */ 2525 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry); 2526 2527 depend->count = 0; 2528 depend->pass2 = 0; 2529 TAILQ_REMOVE(&pmp->depq, depend, entry); 2530 } 2531 2532 hammer2_spin_unex(&pmp->list_spin); 2533 hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/ 2534 HAMMER2_TRANS_WAITING); 2535 dorestart = 0; 2536 2537 /* 2538 * sideq_count may have dropped enough to allow us to unstall 2539 * the frontend. 2540 */ 2541 hammer2_pfs_memory_wakeup(pmp, 0); 2542 2543 /* 2544 * Now run through all inodes on syncq. 2545 * 2546 * Flush transactions only interlock with other flush transactions. 2547 * Any conflicting frontend operations will block on the inode, but 2548 * may hold a vnode lock while doing so. 2549 */ 2550 hammer2_spin_ex(&pmp->list_spin); 2551 while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) { 2552 /* 2553 * Remove the inode from the SYNCQ, transfer the syncq ref 2554 * to us. We must clear SYNCQ to allow any potential 2555 * front-end deadlock to proceed. We must set PASS2 so 2556 * the dependency code knows what to do. 2557 */ 2558 pass2 = ip->flags; 2559 cpu_ccfence(); 2560 if (atomic_cmpset_int(&ip->flags, 2561 pass2, 2562 (pass2 & ~(HAMMER2_INODE_SYNCQ | 2563 HAMMER2_INODE_SYNCQ_WAKEUP)) | 2564 HAMMER2_INODE_SYNCQ_PASS2) == 0) { 2565 continue; 2566 } 2567 TAILQ_REMOVE(&pmp->syncq, ip, entry); 2568 --pmp->sideq_count; 2569 hammer2_spin_unex(&pmp->list_spin); 2570 2571 /* 2572 * Tickle anyone waiting on ip->flags or the hysteresis 2573 * on the dirty inode count. 2574 */ 2575 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP) 2576 wakeup(&ip->flags); 2577 if (++wakecount >= hammer2_limit_dirty_inodes / 20 + 1) { 2578 wakecount = 0; 2579 hammer2_pfs_memory_wakeup(pmp, 0); 2580 } 2581 2582 /* 2583 * Relock the inode, and we inherit a ref from the above. 2584 * We will check for a race after we acquire the vnode. 2585 */ 2586 hammer2_mtx_ex(&ip->lock); 2587 2588 /* 2589 * We need the vp in order to vfsync() dirty buffers, so if 2590 * one isn't attached we can skip it. 2591 * 2592 * Ordering the inode lock and then the vnode lock has the 2593 * potential to deadlock. If we had left SYNCQ set that could 2594 * also deadlock us against the frontend even if we don't hold 2595 * any locks, but the latter is not a problem now since we 2596 * cleared it. igetv will temporarily release the inode lock 2597 * in a safe manner to work-around the deadlock. 2598 * 2599 * Unfortunately it is still possible to deadlock when the 2600 * frontend obtains multiple inode locks, because all the 2601 * related vnodes are already locked (nor can the vnode locks 2602 * be released and reacquired without messing up RECLAIM and 2603 * INACTIVE sequencing). 2604 * 2605 * The solution for now is to move the vp back onto SIDEQ 2606 * and set dorestart, which will restart the flush after we 2607 * exhaust the current SYNCQ. Note that additional 2608 * dependencies may build up, so we definitely need to move 2609 * the whole SIDEQ back to SYNCQ when we restart. 2610 */ 2611 vp = ip->vp; 2612 if (vp) { 2613 if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) { 2614 /* 2615 * Failed to get the vnode, requeue the inode 2616 * (PASS2 is already set so it will be found 2617 * again on the restart). 2618 * 2619 * Then unlock, possibly sleep, and retry 2620 * later. We sleep if PASS2 was *previously* 2621 * set, before we set it again above. 2622 */ 2623 vp = NULL; 2624 dorestart = 1; 2625 #ifdef HAMMER2_DEBUG_SYNC 2626 kprintf("inum %ld (sync delayed by vnode)\n", 2627 (long)ip->meta.inum); 2628 #endif 2629 hammer2_inode_delayed_sideq(ip); 2630 2631 hammer2_mtx_unlock(&ip->lock); 2632 hammer2_inode_drop(ip); 2633 2634 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) { 2635 tsleep(&dorestart, 0, "h2syndel", 2); 2636 } 2637 hammer2_spin_ex(&pmp->list_spin); 2638 continue; 2639 } 2640 } else { 2641 vp = NULL; 2642 } 2643 2644 /* 2645 * If the inode wound up on a SIDEQ again it will already be 2646 * prepped for another PASS2. In this situation if we flush 2647 * it now we will just wind up flushing it again in the same 2648 * syncer run, so we might as well not flush it now. 2649 */ 2650 if (ip->flags & HAMMER2_INODE_SIDEQ) { 2651 hammer2_mtx_unlock(&ip->lock); 2652 hammer2_inode_drop(ip); 2653 if (vp) 2654 vput(vp); 2655 dorestart = 1; 2656 hammer2_spin_ex(&pmp->list_spin); 2657 continue; 2658 } 2659 2660 /* 2661 * Ok we have the inode exclusively locked and if vp is 2662 * not NULL that will also be exclusively locked. Do the 2663 * meat of the flush. 2664 * 2665 * vp token needed for v_rbdirty_tree check / vclrisdirty 2666 * sequencing. Though we hold the vnode exclusively so 2667 * we shouldn't need to hold the token also in this case. 2668 */ 2669 if (vp) { 2670 vfsync(vp, MNT_WAIT, 1, NULL, NULL); 2671 bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */ 2672 } 2673 2674 /* 2675 * If the inode has not yet been inserted into the tree 2676 * we must do so. Then sync and flush it. The flush should 2677 * update the parent. 2678 */ 2679 if (ip->flags & HAMMER2_INODE_DELETING) { 2680 #ifdef HAMMER2_DEBUG_SYNC 2681 kprintf("inum %ld destroy\n", (long)ip->meta.inum); 2682 #endif 2683 hammer2_inode_chain_des(ip); 2684 atomic_add_long(&hammer2_iod_inode_deletes, 1); 2685 } else if (ip->flags & HAMMER2_INODE_CREATING) { 2686 #ifdef HAMMER2_DEBUG_SYNC 2687 kprintf("inum %ld insert\n", (long)ip->meta.inum); 2688 #endif 2689 hammer2_inode_chain_ins(ip); 2690 atomic_add_long(&hammer2_iod_inode_creates, 1); 2691 } 2692 #ifdef HAMMER2_DEBUG_SYNC 2693 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum); 2694 #endif 2695 2696 /* 2697 * Because I kinda messed up the design and index the inodes 2698 * under the root inode, along side the directory entries, 2699 * we can't flush the inode index under the iroot until the 2700 * end. If we do it now we might miss effects created by 2701 * other inodes on the SYNCQ. 2702 * 2703 * Do a normal (non-FSSYNC) flush instead, which allows the 2704 * vnode code to work the same. We don't want to force iroot 2705 * back onto the SIDEQ, and we also don't want the flush code 2706 * to update pfs_iroot_blocksets until the final flush later. 2707 * 2708 * XXX at the moment this will likely result in a double-flush 2709 * of the iroot chain. 2710 */ 2711 hammer2_inode_chain_sync(ip); 2712 if (ip == pmp->iroot) { 2713 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP); 2714 } else { 2715 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP | 2716 HAMMER2_XOP_FSSYNC); 2717 } 2718 if (vp) { 2719 lwkt_gettoken(&vp->v_token); 2720 if ((ip->flags & (HAMMER2_INODE_MODIFIED | 2721 HAMMER2_INODE_RESIZED | 2722 HAMMER2_INODE_DIRTYDATA)) == 0 && 2723 RB_EMPTY(&vp->v_rbdirty_tree) && 2724 !bio_track_active(&vp->v_track_write)) { 2725 vclrisdirty(vp); 2726 } else { 2727 hammer2_inode_delayed_sideq(ip); 2728 } 2729 lwkt_reltoken(&vp->v_token); 2730 vput(vp); 2731 vp = NULL; /* safety */ 2732 } 2733 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2); 2734 hammer2_inode_unlock(ip); /* unlock+drop */ 2735 /* ip pointer invalid */ 2736 2737 /* 2738 * If the inode got dirted after we dropped our locks, 2739 * it will have already been moved back to the SIDEQ. 2740 */ 2741 hammer2_spin_ex(&pmp->list_spin); 2742 } 2743 hammer2_spin_unex(&pmp->list_spin); 2744 hammer2_pfs_memory_wakeup(pmp, 0); 2745 2746 if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) { 2747 #ifdef HAMMER2_DEBUG_SYNC 2748 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n"); 2749 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/ 2750 #endif 2751 dorestart = 1; 2752 goto restart; 2753 } 2754 #ifdef HAMMER2_DEBUG_SYNC 2755 kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n"); 2756 /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/ 2757 #endif 2758 2759 /* 2760 * We have to flush the PFS root last, even if it does not appear to 2761 * be dirty, because all the inodes in the PFS are indexed under it. 2762 * The normal flushing of iroot above would only occur if directory 2763 * entries under the root were changed. 2764 * 2765 * Specifying VOLHDR will cause an additionl flush of hmp->spmp 2766 * for the media making up the cluster. 2767 */ 2768 if ((ip = pmp->iroot) != NULL) { 2769 hammer2_inode_ref(ip); 2770 hammer2_mtx_ex(&ip->lock); 2771 hammer2_inode_chain_sync(ip); 2772 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP | 2773 HAMMER2_XOP_FSSYNC | 2774 HAMMER2_XOP_VOLHDR); 2775 hammer2_inode_unlock(ip); /* unlock+drop */ 2776 } 2777 #ifdef HAMMER2_DEBUG_SYNC 2778 kprintf("FILESYSTEM SYNC STAGE 2 DONE\n"); 2779 #endif 2780 2781 /* 2782 * device bioq sync 2783 */ 2784 hammer2_bioq_sync(pmp); 2785 2786 #if 0 2787 /* 2788 * Generally speaking we now want to flush the media topology from 2789 * the iroot through to the inodes. The flush stops at any inode 2790 * boundary, which allows the frontend to continue running concurrent 2791 * modifying operations on inodes (including kernel flushes of 2792 * buffers) without interfering with the main sync. 2793 * 2794 * Use the XOP interface to concurrently flush all nodes to 2795 * synchronize the PFSROOT subtopology to the media. A standard 2796 * end-of-scan ENOENT error indicates cluster sufficiency. 2797 * 2798 * Note that this flush will not be visible on crash recovery until 2799 * we flush the super-root topology in the next loop. 2800 * 2801 * XXX For now wait for all flushes to complete. 2802 */ 2803 if (mp && (ip = pmp->iroot) != NULL) { 2804 /* 2805 * If unmounting try to flush everything including any 2806 * sub-trees under inodes, just in case there is dangling 2807 * modified data, as a safety. Otherwise just flush up to 2808 * the inodes in this stage. 2809 */ 2810 kprintf("MP & IROOT\n"); 2811 #ifdef HAMMER2_DEBUG_SYNC 2812 kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n"); 2813 #endif 2814 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 2815 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 2816 HAMMER2_XOP_VOLHDR | 2817 HAMMER2_XOP_FSSYNC | 2818 HAMMER2_XOP_INODE_STOP); 2819 } else { 2820 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 2821 HAMMER2_XOP_INODE_STOP | 2822 HAMMER2_XOP_VOLHDR | 2823 HAMMER2_XOP_FSSYNC | 2824 HAMMER2_XOP_INODE_STOP); 2825 } 2826 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc); 2827 error = hammer2_xop_collect(&xop->head, 2828 HAMMER2_XOP_COLLECT_WAITALL); 2829 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 2830 #ifdef HAMMER2_DEBUG_SYNC 2831 kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n"); 2832 #endif 2833 if (error == HAMMER2_ERROR_ENOENT) 2834 error = 0; 2835 else 2836 error = hammer2_error_to_errno(error); 2837 } else { 2838 error = 0; 2839 } 2840 #endif 2841 error = 0; /* XXX */ 2842 hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH); 2843 2844 return (error); 2845 } 2846 2847 static 2848 int 2849 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp) 2850 { 2851 hammer2_inode_t *ip; 2852 2853 KKASSERT(MAXFIDSZ >= 16); 2854 ip = VTOI(vp); 2855 fhp->fid_len = offsetof(struct fid, fid_data[16]); 2856 fhp->fid_ext = 0; 2857 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum; 2858 ((hammer2_tid_t *)fhp->fid_data)[1] = 0; 2859 2860 return 0; 2861 } 2862 2863 static 2864 int 2865 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp, 2866 struct fid *fhp, struct vnode **vpp) 2867 { 2868 hammer2_pfs_t *pmp; 2869 hammer2_tid_t inum; 2870 int error; 2871 2872 pmp = MPTOPMP(mp); 2873 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK; 2874 if (vpp) { 2875 if (inum == 1) 2876 error = hammer2_vfs_root(mp, vpp); 2877 else 2878 error = hammer2_vfs_vget(mp, NULL, inum, vpp); 2879 } else { 2880 error = 0; 2881 } 2882 if (error) 2883 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error); 2884 return error; 2885 } 2886 2887 static 2888 int 2889 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 2890 int *exflagsp, struct ucred **credanonp) 2891 { 2892 hammer2_pfs_t *pmp; 2893 struct netcred *np; 2894 int error; 2895 2896 pmp = MPTOPMP(mp); 2897 np = vfs_export_lookup(mp, &pmp->export, nam); 2898 if (np) { 2899 *exflagsp = np->netc_exflags; 2900 *credanonp = &np->netc_anon; 2901 error = 0; 2902 } else { 2903 error = EACCES; 2904 } 2905 return error; 2906 } 2907 2908 /* 2909 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume 2910 * header into the HMP 2911 */ 2912 static 2913 int 2914 hammer2_install_volume_header(hammer2_dev_t *hmp) 2915 { 2916 hammer2_volume_data_t *vd; 2917 struct buf *bp; 2918 hammer2_crc32_t crc0, crc, bcrc0, bcrc; 2919 int error_reported; 2920 int error; 2921 int valid; 2922 int i; 2923 2924 error_reported = 0; 2925 error = 0; 2926 valid = 0; 2927 bp = NULL; 2928 2929 /* 2930 * There are up to 4 copies of the volume header (syncs iterate 2931 * between them so there is no single master). We don't trust the 2932 * volu_size field so we don't know precisely how large the filesystem 2933 * is, so depend on the OS to return an error if we go beyond the 2934 * block device's EOF. 2935 */ 2936 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) { 2937 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64, 2938 HAMMER2_VOLUME_BYTES, &bp); 2939 if (error) { 2940 brelse(bp); 2941 bp = NULL; 2942 continue; 2943 } 2944 2945 vd = (struct hammer2_volume_data *) bp->b_data; 2946 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) && 2947 (vd->magic != HAMMER2_VOLUME_ID_ABO)) { 2948 kprintf("hammer2: volume header #%d: bad magic\n", i); 2949 brelse(bp); 2950 bp = NULL; 2951 continue; 2952 } 2953 2954 if (vd->magic == HAMMER2_VOLUME_ID_ABO) { 2955 /* XXX: Reversed-endianness filesystem */ 2956 kprintf("hammer2: volume header #%d: reverse-endian " 2957 "filesystem detected\n", i); 2958 brelse(bp); 2959 bp = NULL; 2960 continue; 2961 } 2962 2963 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0]; 2964 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF, 2965 HAMMER2_VOLUME_ICRC0_SIZE); 2966 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1]; 2967 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF, 2968 HAMMER2_VOLUME_ICRC1_SIZE); 2969 if ((crc0 != crc) || (bcrc0 != bcrc)) { 2970 kprintf("hammer2: volume header #%d: volume header crc " 2971 "mismatch %08x/%08x\n", 2972 i, crc0, crc); 2973 error_reported = 1; 2974 brelse(bp); 2975 bp = NULL; 2976 continue; 2977 } 2978 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) { 2979 valid = 1; 2980 hmp->voldata = *vd; 2981 hmp->volhdrno = i; 2982 } 2983 brelse(bp); 2984 bp = NULL; 2985 } 2986 if (valid) { 2987 hmp->volsync = hmp->voldata; 2988 hmp->free_reserved = hmp->voldata.allocator_size / 20; 2989 error = 0; 2990 if (error_reported || bootverbose || 1) { /* 1/DEBUG */ 2991 kprintf("hammer2: using volume header #%d\n", 2992 hmp->volhdrno); 2993 } 2994 } else { 2995 error = EINVAL; 2996 kprintf("hammer2: no valid volume headers found!\n"); 2997 } 2998 return (error); 2999 } 3000 3001 /* 3002 * This handles hysteresis on regular file flushes. Because the BIOs are 3003 * routed to a thread it is possible for an excessive number to build up 3004 * and cause long front-end stalls long before the runningbuffspace limit 3005 * is hit, so we implement hammer2_flush_pipe to control the 3006 * hysteresis. 3007 * 3008 * This is a particular problem when compression is used. 3009 */ 3010 void 3011 hammer2_lwinprog_ref(hammer2_pfs_t *pmp) 3012 { 3013 atomic_add_int(&pmp->count_lwinprog, 1); 3014 } 3015 3016 void 3017 hammer2_lwinprog_drop(hammer2_pfs_t *pmp) 3018 { 3019 int lwinprog; 3020 3021 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1); 3022 if ((lwinprog & HAMMER2_LWINPROG_WAITING) && 3023 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) { 3024 atomic_clear_int(&pmp->count_lwinprog, 3025 HAMMER2_LWINPROG_WAITING); 3026 wakeup(&pmp->count_lwinprog); 3027 } 3028 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) && 3029 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) { 3030 atomic_clear_int(&pmp->count_lwinprog, 3031 HAMMER2_LWINPROG_WAITING0); 3032 wakeup(&pmp->count_lwinprog); 3033 } 3034 } 3035 3036 void 3037 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe) 3038 { 3039 int lwinprog; 3040 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING : 3041 HAMMER2_LWINPROG_WAITING0; 3042 3043 for (;;) { 3044 lwinprog = pmp->count_lwinprog; 3045 cpu_ccfence(); 3046 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 3047 break; 3048 tsleep_interlock(&pmp->count_lwinprog, 0); 3049 atomic_set_int(&pmp->count_lwinprog, lwflag); 3050 lwinprog = pmp->count_lwinprog; 3051 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 3052 break; 3053 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz); 3054 } 3055 } 3056 3057 /* 3058 * It is possible for an excessive number of dirty chains or dirty inodes 3059 * to build up. When this occurs we start an asynchronous filesystem sync. 3060 * If the level continues to build up, we stall, waiting for it to drop, 3061 * with some hysteresis. 3062 * 3063 * This relies on the kernel calling hammer2_vfs_modifying() prior to 3064 * obtaining any vnode locks before making a modifying VOP call. 3065 */ 3066 static int 3067 hammer2_vfs_modifying(struct mount *mp) 3068 { 3069 if (mp->mnt_flag & MNT_RDONLY) 3070 return EROFS; 3071 hammer2_pfs_memory_wait(MPTOPMP(mp)); 3072 3073 return 0; 3074 } 3075 3076 /* 3077 * Initiate an asynchronous filesystem sync and, with hysteresis, 3078 * stall if the internal data structure count becomes too bloated. 3079 */ 3080 void 3081 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp) 3082 { 3083 uint32_t waiting; 3084 int pcatch; 3085 int error; 3086 3087 if (pmp == NULL || pmp->mp == NULL) 3088 return; 3089 3090 for (;;) { 3091 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK; 3092 cpu_ccfence(); 3093 3094 /* 3095 * Start the syncer running at 1/2 the limit 3096 */ 3097 if (waiting > hammer2_limit_dirty_chains / 2 || 3098 pmp->sideq_count > hammer2_limit_dirty_inodes / 2) { 3099 trigger_syncer(pmp->mp); 3100 } 3101 3102 /* 3103 * Stall at the limit waiting for the counts to drop. 3104 * This code will typically be woken up once the count 3105 * drops below 3/4 the limit, or in one second. 3106 */ 3107 if (waiting < hammer2_limit_dirty_chains && 3108 pmp->sideq_count < hammer2_limit_dirty_inodes) { 3109 break; 3110 } 3111 3112 pcatch = curthread->td_proc ? PCATCH : 0; 3113 3114 tsleep_interlock(&pmp->inmem_dirty_chains, pcatch); 3115 atomic_set_int(&pmp->inmem_dirty_chains, 3116 HAMMER2_DIRTYCHAIN_WAITING); 3117 if (waiting < hammer2_limit_dirty_chains && 3118 pmp->sideq_count < hammer2_limit_dirty_inodes) { 3119 break; 3120 } 3121 trigger_syncer(pmp->mp); 3122 error = tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED | pcatch, 3123 "h2memw", hz); 3124 if (error == ERESTART) 3125 break; 3126 } 3127 } 3128 3129 /* 3130 * Wake up any stalled frontend ops waiting, with hysteresis, using 3131 * 2/3 of the limit. 3132 */ 3133 void 3134 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp, int count) 3135 { 3136 uint32_t waiting; 3137 3138 if (pmp) { 3139 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, count); 3140 /* don't need --waiting to test flag */ 3141 3142 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) && 3143 (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <= 3144 hammer2_limit_dirty_chains * 2 / 3 && 3145 pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) { 3146 atomic_clear_int(&pmp->inmem_dirty_chains, 3147 HAMMER2_DIRTYCHAIN_WAITING); 3148 wakeup(&pmp->inmem_dirty_chains); 3149 } 3150 } 3151 } 3152 3153 void 3154 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp) 3155 { 3156 if (pmp) { 3157 atomic_add_int(&pmp->inmem_dirty_chains, 1); 3158 } 3159 } 3160 3161 /* 3162 * Volume header data locks 3163 */ 3164 void 3165 hammer2_voldata_lock(hammer2_dev_t *hmp) 3166 { 3167 lockmgr(&hmp->vollk, LK_EXCLUSIVE); 3168 } 3169 3170 void 3171 hammer2_voldata_unlock(hammer2_dev_t *hmp) 3172 { 3173 lockmgr(&hmp->vollk, LK_RELEASE); 3174 } 3175 3176 void 3177 hammer2_voldata_modify(hammer2_dev_t *hmp) 3178 { 3179 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) { 3180 atomic_add_long(&hammer2_count_modified_chains, 1); 3181 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 3182 hammer2_pfs_memory_inc(hmp->vchain.pmp); 3183 } 3184 } 3185 3186 /* 3187 * Returns 0 if the filesystem has tons of free space 3188 * Returns 1 if the filesystem has less than 10% remaining 3189 * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining. 3190 */ 3191 int 3192 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred) 3193 { 3194 hammer2_pfs_t *pmp; 3195 hammer2_dev_t *hmp; 3196 hammer2_off_t free_reserved; 3197 hammer2_off_t free_nominal; 3198 int i; 3199 3200 pmp = ip->pmp; 3201 3202 if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) { 3203 free_reserved = HAMMER2_SEGSIZE; 3204 free_nominal = 0x7FFFFFFFFFFFFFFFLLU; 3205 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 3206 hmp = pmp->pfs_hmps[i]; 3207 if (hmp == NULL) 3208 continue; 3209 if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER && 3210 pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER) 3211 continue; 3212 3213 if (free_nominal > hmp->voldata.allocator_free) 3214 free_nominal = hmp->voldata.allocator_free; 3215 if (free_reserved < hmp->free_reserved) 3216 free_reserved = hmp->free_reserved; 3217 } 3218 3219 /* 3220 * SMP races ok 3221 */ 3222 pmp->free_reserved = free_reserved; 3223 pmp->free_nominal = free_nominal; 3224 pmp->free_ticks = ticks; 3225 } else { 3226 free_reserved = pmp->free_reserved; 3227 free_nominal = pmp->free_nominal; 3228 } 3229 if (cred && cred->cr_uid != 0) { 3230 if ((int64_t)(free_nominal - bytes) < 3231 (int64_t)free_reserved) { 3232 return 2; 3233 } 3234 } else { 3235 if ((int64_t)(free_nominal - bytes) < 3236 (int64_t)free_reserved / 2) { 3237 return 2; 3238 } 3239 } 3240 if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2) 3241 return 1; 3242 return 0; 3243 } 3244 3245 /* 3246 * Debugging 3247 */ 3248 void 3249 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int bi, int *countp, 3250 char pfx, u_int flags) 3251 { 3252 hammer2_chain_t *scan; 3253 hammer2_chain_t *parent; 3254 3255 --*countp; 3256 if (*countp == 0) { 3257 kprintf("%*.*s...\n", tab, tab, ""); 3258 return; 3259 } 3260 if (*countp < 0) 3261 return; 3262 kprintf("%*.*s%c-chain %p %s.%-3d %016jx %016jx/%-2d mir=%016jx\n", 3263 tab, tab, "", pfx, chain, 3264 hammer2_bref_type_str(chain->bref.type), bi, 3265 chain->bref.data_off, chain->bref.key, chain->bref.keybits, 3266 chain->bref.mirror_tid); 3267 3268 kprintf("%*.*s [%08x] (%s) refs=%d", 3269 tab, tab, "", 3270 chain->flags, 3271 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE && 3272 chain->data) ? (char *)chain->data->ipdata.filename : "?"), 3273 chain->refs); 3274 3275 parent = chain->parent; 3276 if (parent) 3277 kprintf("\n%*.*s p=%p [pflags %08x prefs %d]", 3278 tab, tab, "", 3279 parent, parent->flags, parent->refs); 3280 if (RB_EMPTY(&chain->core.rbtree)) { 3281 kprintf("\n"); 3282 } else { 3283 int bi = 0; 3284 kprintf(" {\n"); 3285 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) { 3286 if ((scan->flags & flags) || flags == (u_int)-1) { 3287 hammer2_dump_chain(scan, tab + 4, bi, countp, 3288 'a', flags); 3289 } 3290 bi++; 3291 } 3292 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data) 3293 kprintf("%*.*s}(%s)\n", tab, tab, "", 3294 chain->data->ipdata.filename); 3295 else 3296 kprintf("%*.*s}\n", tab, tab, ""); 3297 } 3298 } 3299