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