1 /* 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 2022 Tomohiro Kusumi <tkusumi@netbsd.org> 5 * Copyright (c) 2011-2022 The DragonFly Project. All rights reserved. 6 * 7 * This code is derived from software contributed to The DragonFly Project 8 * by Matthew Dillon <dillon@dragonflybsd.org> 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in 18 * the documentation and/or other materials provided with the 19 * distribution. 20 * 3. Neither the name of The DragonFly Project nor the names of its 21 * contributors may be used to endorse or promote products derived 22 * from this software without specific, prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 26 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 27 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 28 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 29 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 30 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 31 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 32 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 33 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 34 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 35 * SUCH DAMAGE. 36 */ 37 /* 38 #include <sys/param.h> 39 #include <sys/systm.h> 40 #include <sys/kernel.h> 41 #include <sys/nlookup.h> 42 #include <sys/vnode.h> 43 #include <sys/mount.h> 44 #include <sys/fcntl.h> 45 #include <sys/vfsops.h> 46 #include <sys/sysctl.h> 47 #include <sys/socket.h> 48 #include <sys/objcache.h> 49 #include <sys/proc.h> 50 #include <sys/lock.h> 51 #include <sys/file.h> 52 */ 53 54 #include "hammer2.h" 55 56 TAILQ_HEAD(hammer2_mntlist, hammer2_dev); 57 static struct hammer2_mntlist hammer2_mntlist; 58 59 struct hammer2_pfslist hammer2_pfslist; 60 struct hammer2_pfslist hammer2_spmplist; 61 struct lock hammer2_mntlk; 62 63 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT; 64 int hammer2_debug; 65 int hammer2_aux_flags; 66 int hammer2_xop_nthreads; 67 int hammer2_xop_sgroups; 68 int hammer2_xop_xgroups; 69 int hammer2_xop_xbase; 70 int hammer2_xop_mod; 71 long hammer2_debug_inode; 72 int hammer2_cluster_meta_read = 1; /* physical read-ahead */ 73 int hammer2_cluster_data_read = 4; /* physical read-ahead */ 74 int hammer2_cluster_write = 0; /* physical write clustering */ 75 int hammer2_dedup_enable = 1; 76 int hammer2_always_compress = 0; /* always try to compress */ 77 int hammer2_flush_pipe = 100; 78 int hammer2_dio_count; 79 int hammer2_dio_limit = 256; 80 int hammer2_bulkfree_tps = 5000; 81 int hammer2_spread_workers; 82 int hammer2_limit_saved_depth; 83 long hammer2_chain_allocs; 84 long hammer2_limit_saved_chains; 85 long hammer2_limit_dirty_chains; 86 long hammer2_limit_dirty_inodes; 87 long hammer2_count_modified_chains; 88 long hammer2_iod_file_read; 89 long hammer2_iod_meta_read; 90 long hammer2_iod_indr_read; 91 long hammer2_iod_fmap_read; 92 long hammer2_iod_volu_read; 93 long hammer2_iod_file_write; 94 long hammer2_iod_file_wembed; 95 long hammer2_iod_file_wzero; 96 long hammer2_iod_file_wdedup; 97 long hammer2_iod_meta_write; 98 long hammer2_iod_indr_write; 99 long hammer2_iod_fmap_write; 100 long hammer2_iod_volu_write; 101 static long hammer2_iod_inode_creates; 102 static long hammer2_iod_inode_deletes; 103 104 long hammer2_process_icrc32; 105 long hammer2_process_xxhash64; 106 107 int hz; 108 int ticks; 109 int64_t vnode_count; 110 111 MALLOC_DECLARE(M_HAMMER2_CBUFFER); 112 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer", 113 "Buffer used for compression."); 114 115 MALLOC_DECLARE(M_HAMMER2_DEBUFFER); 116 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer", 117 "Buffer used for decompression."); 118 119 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem"); 120 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD, 122 &hammer2_supported_version, 0, ""); 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, aux_flags, CTLFLAG_RW, 124 &hammer2_aux_flags, 0, ""); 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW, 126 &hammer2_debug, 0, ""); 127 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW, 128 &hammer2_debug_inode, 0, ""); 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, spread_workers, CTLFLAG_RW, 130 &hammer2_spread_workers, 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, bulkfree_tps, CTLFLAG_RW, 144 &hammer2_bulkfree_tps, 0, ""); 145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RD, 146 &hammer2_chain_allocs, 0, ""); 147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_saved_chains, CTLFLAG_RW, 148 &hammer2_limit_saved_chains, 0, ""); 149 SYSCTL_INT(_vfs_hammer2, OID_AUTO, limit_saved_depth, CTLFLAG_RW, 150 &hammer2_limit_saved_depth, 0, ""); 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW, 152 &hammer2_limit_dirty_chains, 0, ""); 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW, 154 &hammer2_limit_dirty_inodes, 0, ""); 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RD, 156 &hammer2_count_modified_chains, 0, ""); 157 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD, 158 &hammer2_dio_count, 0, ""); 159 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW, 160 &hammer2_dio_limit, 0, ""); 161 162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RD, 163 &hammer2_iod_file_read, 0, ""); 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RD, 165 &hammer2_iod_meta_read, 0, ""); 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RD, 167 &hammer2_iod_indr_read, 0, ""); 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RD, 169 &hammer2_iod_fmap_read, 0, ""); 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RD, 171 &hammer2_iod_volu_read, 0, ""); 172 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RD, 174 &hammer2_iod_file_write, 0, ""); 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RD, 176 &hammer2_iod_file_wembed, 0, ""); 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RD, 178 &hammer2_iod_file_wzero, 0, ""); 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RD, 180 &hammer2_iod_file_wdedup, 0, ""); 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RD, 182 &hammer2_iod_meta_write, 0, ""); 183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RD, 184 &hammer2_iod_indr_write, 0, ""); 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RD, 186 &hammer2_iod_fmap_write, 0, ""); 187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RD, 188 &hammer2_iod_volu_write, 0, ""); 189 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RD, 190 &hammer2_iod_inode_creates, 0, ""); 191 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RD, 192 &hammer2_iod_inode_deletes, 0, ""); 193 194 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RD, 195 &hammer2_process_icrc32, 0, ""); 196 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RD, 197 &hammer2_process_xxhash64, 0, ""); 198 199 /* 200 static int hammer2_vfs_init(struct vfsconf *conf); 201 static int hammer2_vfs_uninit(struct vfsconf *vfsp); 202 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data, 203 struct ucred *cred); 204 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *, 205 struct ucred *); 206 */ 207 static int hammer2_recovery(hammer2_dev_t *hmp); 208 /* 209 static int hammer2_vfs_unmount(struct mount *mp, int mntflags); 210 static int hammer2_vfs_root(struct mount *mp, struct m_vnode **vpp); 211 */ 212 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, 213 struct ucred *cred); 214 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, 215 struct ucred *cred); 216 /* 217 static int hammer2_vfs_fhtovp(struct mount *mp, struct m_vnode *rootvp, 218 struct fid *fhp, struct m_vnode **vpp); 219 static int hammer2_vfs_vptofh(struct m_vnode *vp, struct fid *fhp); 220 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 221 int *exflagsp, struct ucred **credanonp); 222 static int hammer2_vfs_modifying(struct mount *mp); 223 */ 224 225 static void hammer2_update_pmps(hammer2_dev_t *hmp); 226 227 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp); 228 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, 229 hammer2_dev_t *hmp); 230 static int hammer2_fixup_pfses(hammer2_dev_t *hmp); 231 232 /* 233 * HAMMER2 vfs operations. 234 */ 235 /* 236 static struct vfsops hammer2_vfsops = { 237 .vfs_flags = 0, 238 .vfs_init = hammer2_vfs_init, 239 .vfs_uninit = hammer2_vfs_uninit, 240 .vfs_sync = hammer2_vfs_sync, 241 .vfs_mount = hammer2_vfs_mount, 242 .vfs_unmount = hammer2_vfs_unmount, 243 .vfs_root = hammer2_vfs_root, 244 .vfs_statfs = hammer2_vfs_statfs, 245 .vfs_statvfs = hammer2_vfs_statvfs, 246 .vfs_vget = hammer2_vfs_vget, 247 .vfs_vptofh = hammer2_vfs_vptofh, 248 .vfs_fhtovp = hammer2_vfs_fhtovp, 249 .vfs_checkexp = hammer2_vfs_checkexp, 250 .vfs_modifying = hammer2_vfs_modifying 251 }; 252 */ 253 254 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", ""); 255 256 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE); 257 MODULE_VERSION(hammer2, 1); 258 259 int 260 hammer2_vfs_init(void) 261 { 262 /* 263 static struct objcache_malloc_args margs_read; 264 static struct objcache_malloc_args margs_write; 265 static struct objcache_malloc_args margs_vop; 266 */ 267 268 int error; 269 int mod; 270 271 error = 0; 272 kmalloc_raise_limit(M_HAMMER2, 0); /* unlimited */ 273 274 /* 275 * hammer2_xop_nthreads must be a multiple of ncpus, 276 * minimum 2 * ncpus. 277 */ 278 const int ncpus = 1; 279 mod = ncpus; 280 hammer2_xop_mod = mod; 281 hammer2_xop_nthreads = mod * 2; 282 /* 283 while (hammer2_xop_nthreads / mod < HAMMER2_XOPGROUPS_MIN || 284 hammer2_xop_nthreads < HAMMER2_XOPTHREADS_MIN) 285 { 286 hammer2_xop_nthreads += mod; 287 } 288 hammer2_xop_sgroups = hammer2_xop_nthreads / mod / 2; 289 hammer2_xop_xgroups = hammer2_xop_nthreads / mod - hammer2_xop_sgroups; 290 hammer2_xop_xbase = hammer2_xop_sgroups * mod; 291 */ 292 293 /* 294 * A large DIO cache is needed to retain dedup enablement masks. 295 * The bulkfree code clears related masks as part of the disk block 296 * recycling algorithm, preventing it from being used for a later 297 * dedup. 298 * 299 * NOTE: A large buffer cache can actually interfere with dedup 300 * operation because we dedup based on media physical buffers 301 * and not logical buffers. Try to make the DIO case large 302 * enough to avoid this problem, but also cap it. 303 */ 304 const long nbuf = 100000; /* XXX */ 305 hammer2_dio_limit = nbuf * 2; 306 if (hammer2_dio_limit > 100000) 307 hammer2_dio_limit = 100000; 308 309 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref)) 310 error = EINVAL; 311 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data)) 312 error = EINVAL; 313 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data)) 314 error = EINVAL; 315 316 if (error) { 317 kprintf("HAMMER2 structure size mismatch; cannot continue.\n"); 318 return (error); 319 } 320 321 #if 0 322 margs_read.objsize = 65536; 323 margs_read.mtype = M_HAMMER2_DEBUFFER; 324 325 margs_write.objsize = 32768; 326 margs_write.mtype = M_HAMMER2_CBUFFER; 327 328 margs_vop.objsize = sizeof(hammer2_xop_t); 329 margs_vop.mtype = M_HAMMER2; 330 331 /* 332 * Note thaht for the XOPS cache we want backing store allocations 333 * to use M_ZERO. This is not allowed in objcache_get() (to avoid 334 * confusion), so use the backing store function that does it. This 335 * means that initial XOPS objects are zerod but REUSED objects are 336 * not. So we are responsible for cleaning the object up sufficiently 337 * for our needs before objcache_put()ing it back (typically just the 338 * FIFO indices). 339 */ 340 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc, 341 0, 1, NULL, NULL, NULL, 342 objcache_malloc_alloc, 343 objcache_malloc_free, 344 &margs_read); 345 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc, 346 0, 1, NULL, NULL, NULL, 347 objcache_malloc_alloc, 348 objcache_malloc_free, 349 &margs_write); 350 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc, 351 0, 1, NULL, NULL, NULL, 352 objcache_malloc_alloc_zero, 353 objcache_malloc_free, 354 &margs_vop); 355 #endif 356 357 358 lockinit(&hammer2_mntlk, "mntlk", 0, 0); 359 TAILQ_INIT(&hammer2_mntlist); 360 TAILQ_INIT(&hammer2_pfslist); 361 TAILQ_INIT(&hammer2_spmplist); 362 363 const int maxvnodes = 100000; /* XXX */ 364 hammer2_limit_dirty_chains = maxvnodes / 10; 365 if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS) 366 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS; 367 if (hammer2_limit_dirty_chains < 1000) 368 hammer2_limit_dirty_chains = 1000; 369 370 hammer2_limit_dirty_inodes = maxvnodes / 25; 371 if (hammer2_limit_dirty_inodes < 100) 372 hammer2_limit_dirty_inodes = 100; 373 if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES) 374 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES; 375 376 hammer2_limit_saved_chains = hammer2_limit_dirty_chains * 5; 377 378 return (error); 379 } 380 381 int 382 hammer2_vfs_uninit(void) 383 { 384 /* 385 objcache_destroy(cache_buffer_read); 386 objcache_destroy(cache_buffer_write); 387 objcache_destroy(cache_xops); 388 */ 389 return 0; 390 } 391 392 /* 393 * Core PFS allocator. Used to allocate or reference the pmp structure 394 * for PFS cluster mounts and the spmp structure for media (hmp) structures. 395 * The pmp can be passed in or loaded by this function using the chain and 396 * inode data. 397 * 398 * pmp->modify_tid tracks new modify_tid transaction ids for front-end 399 * transactions. Note that synchronization does not use this field. 400 * (typically frontend operations and synchronization cannot run on the 401 * same PFS node at the same time). 402 * 403 * XXX check locking 404 */ 405 hammer2_pfs_t * 406 hammer2_pfsalloc(hammer2_chain_t *chain, 407 const hammer2_inode_data_t *ripdata, 408 hammer2_dev_t *force_local) 409 { 410 hammer2_pfs_t *pmp; 411 hammer2_inode_t *iroot; 412 int count; 413 int i; 414 int j; 415 416 pmp = NULL; 417 418 /* 419 * Locate or create the PFS based on the cluster id. If ripdata 420 * is NULL this is a spmp which is unique and is always allocated. 421 * 422 * If the device is mounted in local mode all PFSs are considered 423 * independent and not part of any cluster (for debugging only). 424 */ 425 if (ripdata) { 426 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 427 if (force_local != pmp->force_local) 428 continue; 429 if (force_local == NULL && 430 bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid, 431 sizeof(pmp->pfs_clid)) == 0) { 432 break; 433 } else if (force_local && pmp->pfs_names[0] && 434 strcmp(pmp->pfs_names[0], (const char *)ripdata->filename) == 0) { 435 break; 436 } 437 } 438 } 439 440 if (pmp == NULL) { 441 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO); 442 pmp->force_local = force_local; 443 hammer2_trans_manage_init(pmp); 444 kmalloc_create_obj(&pmp->minode, "HAMMER2-inodes", 445 sizeof(struct hammer2_inode)); 446 lockinit(&pmp->lock, "pfslk", 0, 0); 447 hammer2_spin_init(&pmp->inum_spin, "hm2pfsalloc_inum"); 448 hammer2_spin_init(&pmp->xop_spin, "h2xop"); 449 hammer2_spin_init(&pmp->lru_spin, "h2lru"); 450 RB_INIT(&pmp->inum_tree); 451 TAILQ_INIT(&pmp->syncq); 452 TAILQ_INIT(&pmp->depq); 453 TAILQ_INIT(&pmp->lru_list); 454 TAILQ_INIT(&pmp->recq); 455 hammer2_spin_init(&pmp->list_spin, "h2pfsalloc_list"); 456 457 /* 458 * Save the last media transaction id for the flusher. Set 459 * initial 460 */ 461 if (ripdata) { 462 pmp->pfs_clid = ripdata->meta.pfs_clid; 463 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry); 464 } else { 465 pmp->flags |= HAMMER2_PMPF_SPMP; 466 TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry); 467 } 468 469 /* 470 * The synchronization thread may start too early, make 471 * sure it stays frozen until we are ready to let it go. 472 * XXX 473 */ 474 /* 475 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN | 476 HAMMER2_THREAD_REMASTER; 477 */ 478 } 479 480 /* 481 * Create the PFS's root inode and any missing XOP helper threads. 482 */ 483 if ((iroot = pmp->iroot) == NULL) { 484 iroot = hammer2_inode_get(pmp, NULL, 1, -1); 485 if (ripdata) 486 iroot->meta = ripdata->meta; 487 pmp->iroot = iroot; 488 hammer2_inode_ref(iroot); 489 hammer2_inode_unlock(iroot); 490 } 491 492 /* 493 * Stop here if no chain is passed in. 494 */ 495 if (chain == NULL) 496 goto done; 497 498 /* 499 * When a chain is passed in we must add it to the PFS's root 500 * inode, update pmp->pfs_types[], and update the syncronization 501 * threads. 502 * 503 * When forcing local mode, mark the PFS as a MASTER regardless. 504 * 505 * At the moment empty spots can develop due to removals or failures. 506 * Ultimately we want to re-fill these spots but doing so might 507 * confused running code. XXX 508 */ 509 hammer2_inode_ref(iroot); 510 hammer2_mtx_ex(&iroot->lock); 511 j = iroot->cluster.nchains; 512 513 if (j == HAMMER2_MAXCLUSTER) { 514 kprintf("hammer2_pfsalloc: cluster full!\n"); 515 /* XXX fatal error? */ 516 } else { 517 KKASSERT(chain->pmp == NULL); 518 chain->pmp = pmp; 519 hammer2_chain_ref(chain); 520 iroot->cluster.array[j].chain = chain; 521 if (force_local) 522 pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER; 523 else 524 pmp->pfs_types[j] = ripdata->meta.pfs_type; 525 pmp->pfs_names[j] = kstrdup((const char *)ripdata->filename, M_HAMMER2); 526 pmp->pfs_hmps[j] = chain->hmp; 527 hammer2_spin_ex(&pmp->inum_spin); 528 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset; 529 hammer2_spin_unex(&pmp->inum_spin); 530 531 /* 532 * If the PFS is already mounted we must account 533 * for the mount_count here. 534 */ 535 if (pmp->mp) 536 ++chain->hmp->mount_count; 537 538 /* 539 * May have to fixup dirty chain tracking. Previous 540 * pmp was NULL so nothing to undo. 541 */ 542 if (chain->flags & HAMMER2_CHAIN_MODIFIED) 543 hammer2_pfs_memory_inc(pmp); 544 ++j; 545 } 546 iroot->cluster.nchains = j; 547 548 /* 549 * Update nmasters from any PFS inode which is part of the cluster. 550 * It is possible that this will result in a value which is too 551 * high. MASTER PFSs are authoritative for pfs_nmasters and will 552 * override this value later on. 553 * 554 * (This informs us of masters that might not currently be 555 * discoverable by this mount). 556 */ 557 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) { 558 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters; 559 } 560 561 /* 562 * Count visible masters. Masters are usually added with 563 * ripdata->meta.pfs_nmasters set to 1. This detects when there 564 * are more (XXX and must update the master inodes). 565 */ 566 count = 0; 567 for (i = 0; i < iroot->cluster.nchains; ++i) { 568 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) 569 ++count; 570 } 571 if (pmp->pfs_nmasters < count) 572 pmp->pfs_nmasters = count; 573 574 /* 575 * Create missing synchronization and support threads. 576 * 577 * Single-node masters (including snapshots) have nothing to 578 * synchronize and do not require this thread. 579 * 580 * Multi-node masters or any number of soft masters, slaves, copy, 581 * or other PFS types need the thread. 582 * 583 * Each thread is responsible for its particular cluster index. 584 * We use independent threads so stalls or mismatches related to 585 * any given target do not affect other targets. 586 */ 587 for (i = 0; i < iroot->cluster.nchains; ++i) { 588 /* 589 * Single-node masters (including snapshots) have nothing 590 * to synchronize and will make direct xops support calls, 591 * thus they do not require this thread. 592 * 593 * Note that there can be thousands of snapshots. We do not 594 * want to create thousands of threads. 595 */ 596 if (pmp->pfs_nmasters <= 1 && 597 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) { 598 continue; 599 } 600 601 /* 602 * Sync support thread 603 */ 604 /* 605 if (pmp->sync_thrs[i].td == NULL) { 606 hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL, 607 "h2nod", i, -1, 608 hammer2_primary_sync_thread); 609 } 610 */ 611 } 612 613 /* 614 * Create missing Xop threads 615 * 616 * NOTE: We create helper threads for all mounted PFSs or any 617 * PFSs with 2+ nodes (so the sync thread can update them, 618 * even if not mounted). 619 */ 620 if (pmp->mp || iroot->cluster.nchains >= 2) 621 hammer2_xop_helper_create(pmp); 622 623 hammer2_mtx_unlock(&iroot->lock); 624 hammer2_inode_drop(iroot); 625 done: 626 return pmp; 627 } 628 629 /* 630 * Deallocate an element of a probed PFS. If destroying and this is a 631 * MASTER, adjust nmasters. 632 * 633 * This function does not physically destroy the PFS element in its device 634 * under the super-root (see hammer2_ioctl_pfs_delete()). 635 */ 636 void 637 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying) 638 { 639 hammer2_inode_t *iroot; 640 hammer2_chain_t *chain; 641 int j; 642 643 /* 644 * Cleanup our reference on iroot. iroot is (should) not be needed 645 * by the flush code. 646 */ 647 iroot = pmp->iroot; 648 if (iroot) { 649 /* 650 * Stop synchronizing 651 * 652 * XXX flush after acquiring the iroot lock. 653 * XXX clean out the cluster index from all inode structures. 654 */ 655 hammer2_thr_delete(&pmp->sync_thrs[clindex]); 656 657 /* 658 * Remove the cluster index from the group. If destroying 659 * the PFS and this is a master, adjust pfs_nmasters. 660 */ 661 hammer2_mtx_ex(&iroot->lock); 662 chain = iroot->cluster.array[clindex].chain; 663 iroot->cluster.array[clindex].chain = NULL; 664 665 switch(pmp->pfs_types[clindex]) { 666 case HAMMER2_PFSTYPE_MASTER: 667 if (destroying && pmp->pfs_nmasters > 0) 668 --pmp->pfs_nmasters; 669 /* XXX adjust ripdata->meta.pfs_nmasters */ 670 break; 671 default: 672 break; 673 } 674 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE; 675 676 hammer2_mtx_unlock(&iroot->lock); 677 678 /* 679 * Release the chain. 680 */ 681 if (chain) { 682 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 683 hammer2_chain_drop(chain); 684 } 685 686 /* 687 * Terminate all XOP threads for the cluster index. 688 */ 689 if (pmp->xop_groups) { 690 for (j = 0; j < hammer2_xop_nthreads; ++j) { 691 hammer2_thr_delete( 692 &pmp->xop_groups[j].thrs[clindex]); 693 } 694 } 695 } 696 } 697 698 /* 699 * Destroy a PFS, typically only occurs after the last mount on a device 700 * has gone away. 701 */ 702 static void 703 hammer2_pfsfree(hammer2_pfs_t *pmp) 704 { 705 hammer2_inode_t *iroot, *ip; 706 hammer2_chain_t *chain; 707 int chains_still_present = 0; 708 int i; 709 //int j; 710 711 /* 712 * Cleanup our reference on iroot. iroot is (should) not be needed 713 * by the flush code. 714 */ 715 if (pmp->flags & HAMMER2_PMPF_SPMP) 716 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry); 717 else 718 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry); 719 720 /* 721 * Cleanup chains remaining on LRU list. 722 */ 723 hammer2_spin_ex(&pmp->lru_spin); 724 while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) { 725 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU); 726 atomic_add_int(&pmp->lru_count, -1); 727 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU); 728 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node); 729 hammer2_chain_ref(chain); 730 hammer2_spin_unex(&pmp->lru_spin); 731 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 732 hammer2_chain_drop(chain); 733 hammer2_spin_ex(&pmp->lru_spin); 734 } 735 hammer2_spin_unex(&pmp->lru_spin); 736 737 /* 738 * Clean up iroot 739 */ 740 iroot = pmp->iroot; 741 if (iroot) { 742 for (i = 0; i < iroot->cluster.nchains; ++i) { 743 /* 744 hammer2_thr_delete(&pmp->sync_thrs[i]); 745 if (pmp->xop_groups) { 746 for (j = 0; j < hammer2_xop_nthreads; ++j) 747 hammer2_thr_delete( 748 &pmp->xop_groups[j].thrs[i]); 749 } 750 */ 751 chain = iroot->cluster.array[i].chain; 752 if (chain && !RB_EMPTY(&chain->core.rbtree)) { 753 kprintf("hammer2: Warning pmp %p still " 754 "has active chains\n", pmp); 755 chains_still_present = 1; 756 } 757 } 758 KASSERT(iroot->refs == 1, 759 ("PMP->IROOT %p REFS WRONG %d", iroot, iroot->refs)); 760 761 /* ref for iroot */ 762 hammer2_inode_drop(iroot); 763 pmp->iroot = NULL; 764 } 765 766 /* 767 * Free remaining pmp resources 768 */ 769 if (chains_still_present) { 770 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp); 771 } else { 772 /* 773 * Free inode in reclaim queue. 774 */ 775 while ((ip = TAILQ_FIRST(&pmp->recq)) != NULL) { 776 TAILQ_REMOVE(&pmp->recq, ip, recq_entry); 777 /* 778 * VOP_RECLAIM is currently unused, 779 * so directly free vnode before inode. 780 */ 781 if (ip->vp) { 782 if (ip->vp->v_malloced) 783 freevnode(ip->vp); 784 } else { 785 /* PFS inode ? */ 786 } 787 kfree_obj(ip, pmp->minode); 788 atomic_add_long(&pmp->inmem_inodes, -1); 789 } 790 assert(TAILQ_EMPTY(&pmp->recq)); 791 assert(pmp->inmem_inodes == 0); 792 793 kmalloc_destroy_obj(&pmp->minode); 794 kfree(pmp, M_HAMMER2); 795 } 796 } 797 798 /* 799 * Remove all references to hmp from the pfs list. Any PFS which becomes 800 * empty is terminated and freed. 801 * 802 * XXX inefficient. 803 */ 804 static void 805 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which) 806 { 807 hammer2_pfs_t *pmp; 808 hammer2_inode_t *iroot; 809 hammer2_chain_t *rchain; 810 int i; 811 //int j; 812 struct hammer2_pfslist *wlist; 813 814 if (which == 0) 815 wlist = &hammer2_pfslist; 816 else 817 wlist = &hammer2_spmplist; 818 again: 819 TAILQ_FOREACH(pmp, wlist, mntentry) { 820 if ((iroot = pmp->iroot) == NULL) 821 continue; 822 823 /* 824 * Determine if this PFS is affected. If it is we must 825 * freeze all management threads and lock its iroot. 826 * 827 * Freezing a management thread forces it idle, operations 828 * in-progress will be aborted and it will have to start 829 * over again when unfrozen, or exit if told to exit. 830 */ 831 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 832 if (pmp->pfs_hmps[i] == hmp) 833 break; 834 } 835 if (i == HAMMER2_MAXCLUSTER) 836 continue; 837 838 hammer2_vfs_sync_pmp(pmp, MNT_WAIT); 839 840 /* 841 * Make sure all synchronization threads are locked 842 * down. 843 */ 844 /* 845 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 846 if (pmp->pfs_hmps[i] == NULL) 847 continue; 848 hammer2_thr_freeze_async(&pmp->sync_thrs[i]); 849 if (pmp->xop_groups) { 850 for (j = 0; j < hammer2_xop_nthreads; ++j) { 851 hammer2_thr_freeze_async( 852 &pmp->xop_groups[j].thrs[i]); 853 } 854 } 855 } 856 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 857 if (pmp->pfs_hmps[i] == NULL) 858 continue; 859 hammer2_thr_freeze(&pmp->sync_thrs[i]); 860 if (pmp->xop_groups) { 861 for (j = 0; j < hammer2_xop_nthreads; ++j) { 862 hammer2_thr_freeze( 863 &pmp->xop_groups[j].thrs[i]); 864 } 865 } 866 } 867 */ 868 869 /* 870 * Lock the inode and clean out matching chains. 871 * Note that we cannot use hammer2_inode_lock_*() 872 * here because that would attempt to validate the 873 * cluster that we are in the middle of ripping 874 * apart. 875 * 876 * WARNING! We are working directly on the inodes 877 * embedded cluster. 878 */ 879 hammer2_mtx_ex(&iroot->lock); 880 881 /* 882 * Remove the chain from matching elements of the PFS. 883 */ 884 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 885 if (pmp->pfs_hmps[i] != hmp) 886 continue; 887 /* 888 hammer2_thr_delete(&pmp->sync_thrs[i]); 889 if (pmp->xop_groups) { 890 for (j = 0; j < hammer2_xop_nthreads; ++j) { 891 hammer2_thr_delete( 892 &pmp->xop_groups[j].thrs[i]); 893 } 894 } 895 */ 896 rchain = iroot->cluster.array[i].chain; 897 iroot->cluster.array[i].chain = NULL; 898 pmp->pfs_types[i] = HAMMER2_PFSTYPE_NONE; 899 if (pmp->pfs_names[i]) { 900 kfree(pmp->pfs_names[i], M_HAMMER2); 901 pmp->pfs_names[i] = NULL; 902 } 903 if (rchain) { 904 hammer2_chain_drop(rchain); 905 /* focus hint */ 906 if (iroot->cluster.focus == rchain) 907 iroot->cluster.focus = NULL; 908 } 909 pmp->pfs_hmps[i] = NULL; 910 } 911 hammer2_mtx_unlock(&iroot->lock); 912 913 /* 914 * Cleanup trailing chains. Gaps may remain. 915 */ 916 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) { 917 if (pmp->pfs_hmps[i]) 918 break; 919 } 920 iroot->cluster.nchains = i + 1; 921 922 /* 923 * If the PMP has no elements remaining we can destroy it. 924 * (this will transition management threads from frozen->exit). 925 */ 926 if (iroot->cluster.nchains == 0) { 927 /* 928 * If this was the hmp's spmp, we need to clean 929 * a little more stuff out. 930 */ 931 if (hmp->spmp == pmp) { 932 hmp->spmp = NULL; 933 hmp->vchain.pmp = NULL; 934 hmp->fchain.pmp = NULL; 935 } 936 937 /* 938 * Free the pmp and restart the loop 939 */ 940 KKASSERT(TAILQ_EMPTY(&pmp->syncq)); 941 KKASSERT(TAILQ_EMPTY(&pmp->depq)); 942 hammer2_pfsfree(pmp); 943 goto again; 944 } 945 946 /* 947 * If elements still remain we need to set the REMASTER 948 * flag and unfreeze it. 949 */ 950 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 951 if (pmp->pfs_hmps[i] == NULL) 952 continue; 953 /* 954 hammer2_thr_remaster(&pmp->sync_thrs[i]); 955 hammer2_thr_unfreeze(&pmp->sync_thrs[i]); 956 if (pmp->xop_groups) { 957 for (j = 0; j < hammer2_xop_nthreads; ++j) { 958 hammer2_thr_remaster( 959 &pmp->xop_groups[j].thrs[i]); 960 hammer2_thr_unfreeze( 961 &pmp->xop_groups[j].thrs[i]); 962 } 963 } 964 */ 965 } 966 } 967 } 968 969 /* 970 * Mount or remount HAMMER2 fileystem from physical media 971 * 972 * mountroot 973 * mp mount point structure 974 * path NULL 975 * data <unused> 976 * cred <unused> 977 * 978 * mount 979 * mp mount point structure 980 * path path to mount point 981 * data pointer to argument structure in user space 982 * volume volume path (device@LABEL form) 983 * hflags user mount flags 984 * cred user credentials 985 * 986 * RETURNS: 0 Success 987 * !0 error number 988 */ 989 int 990 hammer2_vfs_mount(struct m_vnode *makefs_devvp, struct mount *mp, 991 const char *label, const struct hammer2_mount_info *mi) 992 { 993 struct hammer2_mount_info info = *mi; 994 hammer2_pfs_t *pmp; 995 hammer2_pfs_t *spmp; 996 hammer2_dev_t *hmp, *hmp_tmp; 997 hammer2_dev_t *force_local; 998 hammer2_key_t key_next; 999 hammer2_key_t key_dummy; 1000 hammer2_key_t lhc; 1001 hammer2_chain_t *parent; 1002 hammer2_chain_t *chain; 1003 const hammer2_inode_data_t *ripdata; 1004 hammer2_devvp_list_t devvpl; 1005 hammer2_devvp_t *e, *e_tmp; 1006 char *devstr; 1007 int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0); 1008 int error; 1009 int i; 1010 1011 hmp = NULL; 1012 pmp = NULL; 1013 devstr = NULL; 1014 1015 kprintf("hammer2_mount: device=\"%s\" label=\"%s\" rdonly=%d\n", 1016 devstr, label, ronly); 1017 1018 /* 1019 * Initialize all device vnodes. 1020 */ 1021 TAILQ_INIT(&devvpl); 1022 error = hammer2_init_devvp(makefs_devvp, &devvpl); 1023 if (error) { 1024 kprintf("hammer2: failed to initialize devvp in %s\n", devstr); 1025 hammer2_cleanup_devvp(&devvpl); 1026 return error; 1027 } 1028 1029 /* 1030 * Determine if the device has already been mounted. After this 1031 * check hmp will be non-NULL if we are doing the second or more 1032 * hammer2 mounts from the same device. 1033 */ 1034 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1035 if (!TAILQ_EMPTY(&devvpl)) { 1036 /* 1037 * Match the device. Due to the way devfs works, 1038 * we may not be able to directly match the vnode pointer, 1039 * so also check to see if the underlying device matches. 1040 */ 1041 TAILQ_FOREACH(hmp_tmp, &hammer2_mntlist, mntentry) { 1042 TAILQ_FOREACH(e_tmp, &hmp_tmp->devvpl, entry) { 1043 int devvp_found = 0; 1044 TAILQ_FOREACH(e, &devvpl, entry) { 1045 KKASSERT(e->devvp); 1046 if (e_tmp->devvp == e->devvp) 1047 devvp_found = 1; 1048 /* 1049 if (e_tmp->devvp->v_rdev && 1050 e_tmp->devvp->v_rdev == e->devvp->v_rdev) 1051 devvp_found = 1; 1052 */ 1053 } 1054 if (!devvp_found) 1055 goto next_hmp; 1056 } 1057 hmp = hmp_tmp; 1058 kprintf("hammer2_mount: hmp=%p matched\n", hmp); 1059 break; 1060 next_hmp: 1061 continue; 1062 } 1063 1064 /* 1065 * If no match this may be a fresh H2 mount, make sure 1066 * the device is not mounted on anything else. 1067 */ 1068 if (hmp == NULL) { 1069 TAILQ_FOREACH(e, &devvpl, entry) { 1070 struct m_vnode *devvp = e->devvp; 1071 KKASSERT(devvp); 1072 error = vfs_mountedon(devvp); 1073 if (error) { 1074 kprintf("hammer2_mount: %s mounted %d\n", 1075 e->path, error); 1076 hammer2_cleanup_devvp(&devvpl); 1077 lockmgr(&hammer2_mntlk, LK_RELEASE); 1078 return error; 1079 } 1080 } 1081 } 1082 } else { 1083 /* 1084 * Match the label to a pmp already probed. 1085 */ 1086 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 1087 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 1088 if (pmp->pfs_names[i] && 1089 strcmp(pmp->pfs_names[i], label) == 0) { 1090 hmp = pmp->pfs_hmps[i]; 1091 break; 1092 } 1093 } 1094 if (hmp) 1095 break; 1096 } 1097 if (hmp == NULL) { 1098 kprintf("hammer2_mount: PFS label \"%s\" not found\n", 1099 label); 1100 hammer2_cleanup_devvp(&devvpl); 1101 lockmgr(&hammer2_mntlk, LK_RELEASE); 1102 return ENOENT; 1103 } 1104 } 1105 1106 /* 1107 * Open the device if this isn't a secondary mount and construct 1108 * the H2 device mount (hmp). 1109 */ 1110 if (hmp == NULL) { 1111 hammer2_chain_t *schain; 1112 hammer2_xop_head_t xop; 1113 1114 /* 1115 * Now open the device 1116 */ 1117 KKASSERT(!TAILQ_EMPTY(&devvpl)); 1118 error = hammer2_open_devvp(&devvpl, ronly); 1119 if (error) { 1120 hammer2_close_devvp(&devvpl, ronly); 1121 hammer2_cleanup_devvp(&devvpl); 1122 lockmgr(&hammer2_mntlk, LK_RELEASE); 1123 return error; 1124 } 1125 1126 /* 1127 * Construct volumes and link with device vnodes. 1128 */ 1129 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO); 1130 hmp->devvp = NULL; 1131 error = hammer2_init_vfsvolumes(mp, &devvpl, hmp->volumes, 1132 &hmp->voldata, &hmp->volhdrno, 1133 &hmp->devvp); 1134 if (error) { 1135 hammer2_close_devvp(&devvpl, ronly); 1136 hammer2_cleanup_devvp(&devvpl); 1137 lockmgr(&hammer2_mntlk, LK_RELEASE); 1138 kfree(hmp, M_HAMMER2); 1139 return error; 1140 } 1141 if (!hmp->devvp) { 1142 kprintf("hammer2: failed to initialize root volume\n"); 1143 hammer2_unmount_helper(mp, NULL, hmp); 1144 lockmgr(&hammer2_mntlk, LK_RELEASE); 1145 hammer2_vfs_unmount(mp, MNT_FORCE); 1146 return EINVAL; 1147 } 1148 1149 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", devstr); 1150 hmp->ronly = ronly; 1151 hmp->hflags = info.hflags & HMNT2_DEVFLAGS; 1152 kmalloc_create_obj(&hmp->mchain, "HAMMER2-chains", 1153 sizeof(struct hammer2_chain)); 1154 kmalloc_create_obj(&hmp->mio, "HAMMER2-dio", 1155 sizeof(struct hammer2_io)); 1156 kmalloc_create(&hmp->mmsg, "HAMMER2-msg"); 1157 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry); 1158 RB_INIT(&hmp->iotree); 1159 hammer2_spin_init(&hmp->io_spin, "h2mount_io"); 1160 hammer2_spin_init(&hmp->list_spin, "h2mount_list"); 1161 1162 lockinit(&hmp->vollk, "h2vol", 0, 0); 1163 lockinit(&hmp->bulklk, "h2bulk", 0, 0); 1164 lockinit(&hmp->bflock, "h2bflk", 0, 0); 1165 1166 /* 1167 * vchain setup. vchain.data is embedded. 1168 * vchain.refs is initialized and will never drop to 0. 1169 */ 1170 hmp->vchain.hmp = hmp; 1171 hmp->vchain.refs = 1; 1172 hmp->vchain.data = (void *)&hmp->voldata; 1173 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME; 1174 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1175 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1176 hammer2_chain_init(&hmp->vchain); 1177 1178 /* 1179 * fchain setup. fchain.data is embedded. 1180 * fchain.refs is initialized and will never drop to 0. 1181 * 1182 * The data is not used but needs to be initialized to 1183 * pass assertion muster. We use this chain primarily 1184 * as a placeholder for the freemap's top-level radix tree 1185 * so it does not interfere with the volume's topology 1186 * radix tree. 1187 */ 1188 hmp->fchain.hmp = hmp; 1189 hmp->fchain.refs = 1; 1190 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset; 1191 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP; 1192 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1193 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1194 hmp->fchain.bref.methods = 1195 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) | 1196 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE); 1197 hammer2_chain_init(&hmp->fchain); 1198 1199 /* 1200 * Initialize volume header related fields. 1201 */ 1202 KKASSERT(hmp->voldata.magic == HAMMER2_VOLUME_ID_HBO || 1203 hmp->voldata.magic == HAMMER2_VOLUME_ID_ABO); 1204 hmp->volsync = hmp->voldata; 1205 hmp->free_reserved = hmp->voldata.allocator_size / 20; 1206 /* 1207 * Must use hmp instead of volume header for these two 1208 * in order to handle volume versions transparently. 1209 */ 1210 if (hmp->voldata.version >= HAMMER2_VOL_VERSION_MULTI_VOLUMES) { 1211 hmp->nvolumes = hmp->voldata.nvolumes; 1212 hmp->total_size = hmp->voldata.total_size; 1213 } else { 1214 hmp->nvolumes = 1; 1215 hmp->total_size = hmp->voldata.volu_size; 1216 } 1217 KKASSERT(hmp->nvolumes > 0); 1218 1219 /* 1220 * Move devvpl entries to hmp. 1221 */ 1222 TAILQ_INIT(&hmp->devvpl); 1223 while ((e = TAILQ_FIRST(&devvpl)) != NULL) { 1224 TAILQ_REMOVE(&devvpl, e, entry); 1225 TAILQ_INSERT_TAIL(&hmp->devvpl, e, entry); 1226 } 1227 KKASSERT(TAILQ_EMPTY(&devvpl)); 1228 KKASSERT(!TAILQ_EMPTY(&hmp->devvpl)); 1229 1230 /* 1231 * Really important to get these right or the flush and 1232 * teardown code will get confused. 1233 */ 1234 hmp->spmp = hammer2_pfsalloc(NULL, NULL, NULL); 1235 spmp = hmp->spmp; 1236 spmp->pfs_hmps[0] = hmp; 1237 1238 /* 1239 * Dummy-up vchain and fchain's modify_tid. mirror_tid 1240 * is inherited from the volume header. 1241 */ 1242 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1243 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid; 1244 hmp->vchain.pmp = spmp; 1245 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1246 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid; 1247 hmp->fchain.pmp = spmp; 1248 1249 /* 1250 * First locate the super-root inode, which is key 0 1251 * relative to the volume header's blockset. 1252 * 1253 * Then locate the root inode by scanning the directory keyspace 1254 * represented by the label. 1255 */ 1256 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 1257 schain = hammer2_chain_lookup(&parent, &key_dummy, 1258 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY, 1259 &error, 0); 1260 hammer2_chain_lookup_done(parent); 1261 if (schain == NULL) { 1262 kprintf("hammer2_mount: invalid super-root\n"); 1263 hammer2_unmount_helper(mp, NULL, hmp); 1264 lockmgr(&hammer2_mntlk, LK_RELEASE); 1265 hammer2_vfs_unmount(mp, MNT_FORCE); 1266 return EINVAL; 1267 } 1268 if (schain->error) { 1269 kprintf("hammer2_mount: error %s reading super-root\n", 1270 hammer2_error_str(schain->error)); 1271 hammer2_chain_unlock(schain); 1272 hammer2_chain_drop(schain); 1273 schain = NULL; 1274 hammer2_unmount_helper(mp, NULL, hmp); 1275 lockmgr(&hammer2_mntlk, LK_RELEASE); 1276 hammer2_vfs_unmount(mp, MNT_FORCE); 1277 return EINVAL; 1278 } 1279 1280 /* 1281 * The super-root always uses an inode_tid of 1 when 1282 * creating PFSs. 1283 */ 1284 spmp->inode_tid = 1; 1285 spmp->modify_tid = schain->bref.modify_tid + 1; 1286 1287 /* 1288 * Sanity-check schain's pmp and finish initialization. 1289 * Any chain belonging to the super-root topology should 1290 * have a NULL pmp (not even set to spmp). 1291 */ 1292 ripdata = &schain->data->ipdata; 1293 KKASSERT(schain->pmp == NULL); 1294 spmp->pfs_clid = ripdata->meta.pfs_clid; 1295 1296 /* 1297 * Replace the dummy spmp->iroot with a real one. It's 1298 * easier to just do a wholesale replacement than to try 1299 * to update the chain and fixup the iroot fields. 1300 * 1301 * The returned inode is locked with the supplied cluster. 1302 */ 1303 hammer2_dummy_xop_from_chain(&xop, schain); 1304 hammer2_inode_drop(spmp->iroot); 1305 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1); 1306 spmp->spmp_hmp = hmp; 1307 spmp->pfs_types[0] = ripdata->meta.pfs_type; 1308 spmp->pfs_hmps[0] = hmp; 1309 hammer2_inode_ref(spmp->iroot); 1310 hammer2_inode_unlock(spmp->iroot); 1311 hammer2_cluster_unlock(&xop.cluster); 1312 hammer2_chain_drop(schain); 1313 /* do not call hammer2_cluster_drop() on an embedded cluster */ 1314 schain = NULL; /* now invalid */ 1315 /* leave spmp->iroot with one ref */ 1316 1317 if (!hmp->ronly) { 1318 error = hammer2_recovery(hmp); 1319 if (error == 0) 1320 error |= hammer2_fixup_pfses(hmp); 1321 /* XXX do something with error */ 1322 } 1323 hammer2_update_pmps(hmp); 1324 hammer2_iocom_init(hmp); 1325 hammer2_bulkfree_init(hmp); 1326 1327 /* 1328 * Ref the cluster management messaging descriptor. The mount 1329 * program deals with the other end of the communications pipe. 1330 * 1331 * Root mounts typically do not supply one. 1332 */ 1333 /* 1334 if (info.cluster_fd >= 0) { 1335 fp = holdfp(curthread, info.cluster_fd, -1); 1336 if (fp) { 1337 hammer2_cluster_reconnect(hmp, fp); 1338 } else { 1339 kprintf("hammer2_mount: bad cluster_fd!\n"); 1340 } 1341 } 1342 */ 1343 } else { 1344 /* hmp->devvp_list is already constructed. */ 1345 hammer2_cleanup_devvp(&devvpl); 1346 spmp = hmp->spmp; 1347 if (info.hflags & HMNT2_DEVFLAGS) { 1348 kprintf("hammer2_mount: Warning: mount flags pertaining " 1349 "to the whole device may only be specified " 1350 "on the first mount of the device: %08x\n", 1351 info.hflags & HMNT2_DEVFLAGS); 1352 } 1353 } 1354 1355 /* 1356 * Force local mount (disassociate all PFSs from their clusters). 1357 * Used primarily for debugging. 1358 */ 1359 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1360 1361 /* 1362 * Lookup the mount point under the media-localized super-root. 1363 * Scanning hammer2_pfslist doesn't help us because it represents 1364 * PFS cluster ids which can aggregate several named PFSs together. 1365 * 1366 * cluster->pmp will incorrectly point to spmp and must be fixed 1367 * up later on. 1368 */ 1369 hammer2_inode_lock(spmp->iroot, 0); 1370 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1371 lhc = hammer2_dirhash(label, strlen(label)); 1372 chain = hammer2_chain_lookup(&parent, &key_next, 1373 lhc, lhc + HAMMER2_DIRHASH_LOMASK, 1374 &error, 0); 1375 while (chain) { 1376 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && 1377 strcmp(label, (char *)chain->data->ipdata.filename) == 0) { 1378 break; 1379 } 1380 chain = hammer2_chain_next(&parent, chain, &key_next, 1381 key_next, 1382 lhc + HAMMER2_DIRHASH_LOMASK, 1383 &error, 0); 1384 } 1385 if (parent) { 1386 hammer2_chain_unlock(parent); 1387 hammer2_chain_drop(parent); 1388 } 1389 hammer2_inode_unlock(spmp->iroot); 1390 1391 /* 1392 * PFS could not be found? 1393 */ 1394 if (chain == NULL) { 1395 hammer2_unmount_helper(mp, NULL, hmp); 1396 lockmgr(&hammer2_mntlk, LK_RELEASE); 1397 hammer2_vfs_unmount(mp, MNT_FORCE); 1398 1399 if (error) { 1400 kprintf("hammer2_mount: PFS label I/O error\n"); 1401 return EINVAL; 1402 } else { 1403 kprintf("hammer2_mount: PFS label \"%s\" not found\n", 1404 label); 1405 return ENOENT; 1406 } 1407 } 1408 1409 /* 1410 * Acquire the pmp structure (it should have already been allocated 1411 * via hammer2_update_pmps()). 1412 */ 1413 if (chain->error) { 1414 kprintf("hammer2_mount: PFS label I/O error\n"); 1415 } else { 1416 ripdata = &chain->data->ipdata; 1417 pmp = hammer2_pfsalloc(NULL, ripdata, force_local); 1418 } 1419 hammer2_chain_unlock(chain); 1420 hammer2_chain_drop(chain); 1421 1422 /* 1423 * PFS to mount must exist at this point. 1424 */ 1425 if (pmp == NULL) { 1426 kprintf("hammer2_mount: Failed to acquire PFS structure\n"); 1427 hammer2_unmount_helper(mp, NULL, hmp); 1428 lockmgr(&hammer2_mntlk, LK_RELEASE); 1429 hammer2_vfs_unmount(mp, MNT_FORCE); 1430 return EINVAL; 1431 } 1432 1433 /* 1434 * Finish the mount 1435 */ 1436 kprintf("hammer2_mount: hmp=%p pmp=%p\n", hmp, pmp); 1437 1438 /* Check if the pmp has already been mounted. */ 1439 if (pmp->mp) { 1440 kprintf("hammer2_mount: PFS already mounted!\n"); 1441 hammer2_unmount_helper(mp, NULL, hmp); 1442 lockmgr(&hammer2_mntlk, LK_RELEASE); 1443 hammer2_vfs_unmount(mp, MNT_FORCE); 1444 return EBUSY; 1445 } 1446 1447 pmp->hflags = info.hflags; 1448 mp->mnt_flag |= MNT_LOCAL; 1449 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */ 1450 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */ 1451 1452 /* 1453 * required mount structure initializations 1454 */ 1455 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE; 1456 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE; 1457 1458 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE; 1459 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE; 1460 1461 /* 1462 * Optional fields 1463 */ 1464 mp->mnt_iosize_max = MAXPHYS; 1465 1466 /* 1467 * Connect up mount pointers. 1468 */ 1469 hammer2_mount_helper(mp, pmp); 1470 lockmgr(&hammer2_mntlk, LK_RELEASE); 1471 1472 #if 0 1473 /* 1474 * Finish setup 1475 */ 1476 vfs_getnewfsid(mp); 1477 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops); 1478 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops); 1479 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops); 1480 1481 if (path) { 1482 copyinstr(info.volume, mp->mnt_stat.f_mntfromname, 1483 MNAMELEN - 1, &size); 1484 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size); 1485 } /* else root mount, already in there */ 1486 1487 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname)); 1488 if (path) { 1489 copyinstr(path, mp->mnt_stat.f_mntonname, 1490 sizeof(mp->mnt_stat.f_mntonname) - 1, 1491 &size); 1492 } else { 1493 /* root mount */ 1494 mp->mnt_stat.f_mntonname[0] = '/'; 1495 } 1496 #endif 1497 1498 /* 1499 * Initial statfs to prime mnt_stat. 1500 */ 1501 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL); 1502 hammer2_vfs_statvfs(mp, &mp->mnt_vstat, NULL); 1503 1504 return 0; 1505 } 1506 1507 /* 1508 * Scan PFSs under the super-root and create hammer2_pfs structures. 1509 */ 1510 static 1511 void 1512 hammer2_update_pmps(hammer2_dev_t *hmp) 1513 { 1514 const hammer2_inode_data_t *ripdata; 1515 hammer2_chain_t *parent; 1516 hammer2_chain_t *chain; 1517 hammer2_dev_t *force_local; 1518 hammer2_pfs_t *spmp; 1519 hammer2_key_t key_next; 1520 int error; 1521 1522 /* 1523 * Force local mount (disassociate all PFSs from their clusters). 1524 * Used primarily for debugging. 1525 */ 1526 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1527 1528 /* 1529 * Lookup mount point under the media-localized super-root. 1530 * 1531 * cluster->pmp will incorrectly point to spmp and must be fixed 1532 * up later on. 1533 */ 1534 spmp = hmp->spmp; 1535 hammer2_inode_lock(spmp->iroot, 0); 1536 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1537 chain = hammer2_chain_lookup(&parent, &key_next, 1538 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 1539 &error, 0); 1540 while (chain) { 1541 if (chain->error) { 1542 kprintf("I/O error scanning PFS labels\n"); 1543 } else if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) { 1544 kprintf("Non inode chain type %d under super-root\n", 1545 chain->bref.type); 1546 } else { 1547 ripdata = &chain->data->ipdata; 1548 hammer2_pfsalloc(chain, ripdata, force_local); 1549 } 1550 chain = hammer2_chain_next(&parent, chain, &key_next, 1551 key_next, HAMMER2_KEY_MAX, 1552 &error, 0); 1553 } 1554 if (parent) { 1555 hammer2_chain_unlock(parent); 1556 hammer2_chain_drop(parent); 1557 } 1558 hammer2_inode_unlock(spmp->iroot); 1559 } 1560 1561 #if 0 1562 static 1563 int 1564 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused, 1565 struct ucred *cred) 1566 { 1567 hammer2_volume_t *vol; 1568 struct m_vnode *devvp; 1569 int i, error, result = 0; 1570 1571 if (!(hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR))) 1572 return 0; 1573 1574 for (i = 0; i < hmp->nvolumes; ++i) { 1575 vol = &hmp->volumes[i]; 1576 devvp = vol->dev->devvp; 1577 KKASSERT(devvp); 1578 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1579 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL); 1580 vn_unlock(devvp); 1581 error = 0; 1582 if (vol->id == HAMMER2_ROOT_VOLUME) { 1583 error = hammer2_recovery(hmp); 1584 if (error == 0) 1585 error |= hammer2_fixup_pfses(hmp); 1586 } 1587 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1588 if (error == 0) { 1589 VOP_CLOSE(devvp, FREAD, NULL); 1590 } else { 1591 VOP_CLOSE(devvp, FREAD | FWRITE, NULL); 1592 } 1593 vn_unlock(devvp); 1594 result |= error; 1595 } 1596 if (result == 0) { 1597 kprintf("hammer2: enable read/write\n"); 1598 hmp->ronly = 0; 1599 } 1600 1601 return result; 1602 } 1603 #endif 1604 1605 int 1606 hammer2_vfs_unmount(struct mount *mp, int mntflags) 1607 { 1608 hammer2_pfs_t *pmp; 1609 int flags; 1610 int error = 0; 1611 1612 pmp = MPTOPMP(mp); 1613 1614 if (pmp == NULL) 1615 return(0); 1616 1617 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1618 1619 /* 1620 * If mount initialization proceeded far enough we must flush 1621 * its vnodes and sync the underlying mount points. Three syncs 1622 * are required to fully flush the filesystem (freemap updates lag 1623 * by one flush, and one extra for safety). 1624 */ 1625 if (mntflags & MNT_FORCE) 1626 flags = FORCECLOSE; 1627 else 1628 flags = 0; 1629 if (pmp->iroot) { 1630 error = vflush(mp, 0, flags); 1631 if (error) 1632 goto failed; 1633 hammer2_vfs_sync(mp, MNT_WAIT); 1634 hammer2_vfs_sync(mp, MNT_WAIT); 1635 hammer2_vfs_sync(mp, MNT_WAIT); 1636 } 1637 1638 /* 1639 * Cleanup the frontend support XOPS threads 1640 */ 1641 hammer2_xop_helper_cleanup(pmp); 1642 1643 if (pmp->mp) 1644 hammer2_unmount_helper(mp, pmp, NULL); 1645 1646 error = 0; 1647 failed: 1648 lockmgr(&hammer2_mntlk, LK_RELEASE); 1649 1650 return (error); 1651 } 1652 1653 /* 1654 * Mount helper, hook the system mount into our PFS. 1655 * The mount lock is held. 1656 * 1657 * We must bump the mount_count on related devices for any 1658 * mounted PFSs. 1659 */ 1660 static 1661 void 1662 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp) 1663 { 1664 hammer2_cluster_t *cluster; 1665 hammer2_chain_t *rchain; 1666 int i; 1667 1668 mp->mnt_data = (qaddr_t)pmp; 1669 pmp->mp = mp; 1670 1671 /* 1672 * After pmp->mp is set we have to adjust hmp->mount_count. 1673 */ 1674 cluster = &pmp->iroot->cluster; 1675 for (i = 0; i < cluster->nchains; ++i) { 1676 rchain = cluster->array[i].chain; 1677 if (rchain == NULL) 1678 continue; 1679 ++rchain->hmp->mount_count; 1680 } 1681 1682 /* 1683 * Create missing Xop threads 1684 */ 1685 hammer2_xop_helper_create(pmp); 1686 } 1687 1688 /* 1689 * Unmount helper, unhook the system mount from our PFS. 1690 * The mount lock is held. 1691 * 1692 * If hmp is supplied a mount responsible for being the first to open 1693 * the block device failed and the block device and all PFSs using the 1694 * block device must be cleaned up. 1695 * 1696 * If pmp is supplied multiple devices might be backing the PFS and each 1697 * must be disconnected. This might not be the last PFS using some of the 1698 * underlying devices. Also, we have to adjust our hmp->mount_count 1699 * accounting for the devices backing the pmp which is now undergoing an 1700 * unmount. 1701 */ 1702 static 1703 void 1704 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp) 1705 { 1706 hammer2_cluster_t *cluster; 1707 hammer2_chain_t *rchain; 1708 int dumpcnt; 1709 int i; 1710 1711 /* 1712 * If no device supplied this is a high-level unmount and we have to 1713 * to disconnect the mount, adjust mount_count, and locate devices 1714 * that might now have no mounts. 1715 */ 1716 if (pmp) { 1717 KKASSERT(hmp == NULL); 1718 KKASSERT(MPTOPMP(mp) == pmp); 1719 pmp->mp = NULL; 1720 mp->mnt_data = NULL; 1721 1722 /* 1723 * After pmp->mp is cleared we have to account for 1724 * mount_count. 1725 */ 1726 cluster = &pmp->iroot->cluster; 1727 for (i = 0; i < cluster->nchains; ++i) { 1728 rchain = cluster->array[i].chain; 1729 if (rchain == NULL) 1730 continue; 1731 --rchain->hmp->mount_count; 1732 /* scrapping hmp now may invalidate the pmp */ 1733 } 1734 again: 1735 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1736 if (hmp->mount_count == 0) { 1737 hammer2_unmount_helper(NULL, NULL, hmp); 1738 goto again; 1739 } 1740 } 1741 return; 1742 } 1743 1744 /* 1745 * Try to terminate the block device. We can't terminate it if 1746 * there are still PFSs referencing it. 1747 */ 1748 if (hmp->mount_count) 1749 return; 1750 1751 /* 1752 * Decomission the network before we start messing with the 1753 * device and PFS. 1754 */ 1755 hammer2_iocom_uninit(hmp); 1756 1757 hammer2_bulkfree_uninit(hmp); 1758 hammer2_pfsfree_scan(hmp, 0); 1759 1760 /* 1761 * Cycle the volume data lock as a safety (probably not needed any 1762 * more). To ensure everything is out we need to flush at least 1763 * three times. (1) The running of the sideq can dirty the 1764 * filesystem, (2) A normal flush can dirty the freemap, and 1765 * (3) ensure that the freemap is fully synchronized. 1766 * 1767 * The next mount's recovery scan can clean everything up but we want 1768 * to leave the filesystem in a 100% clean state on a normal unmount. 1769 */ 1770 #if 0 1771 hammer2_voldata_lock(hmp); 1772 hammer2_voldata_unlock(hmp); 1773 #endif 1774 1775 /* 1776 * Flush whatever is left. Unmounted but modified PFS's might still 1777 * have some dirty chains on them. 1778 */ 1779 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS); 1780 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS); 1781 1782 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) { 1783 hammer2_voldata_modify(hmp); 1784 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP | 1785 HAMMER2_FLUSH_ALL); 1786 } 1787 hammer2_chain_unlock(&hmp->fchain); 1788 1789 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) { 1790 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP | 1791 HAMMER2_FLUSH_ALL); 1792 } 1793 hammer2_chain_unlock(&hmp->vchain); 1794 1795 if ((hmp->vchain.flags | hmp->fchain.flags) & 1796 HAMMER2_CHAIN_FLUSH_MASK) { 1797 kprintf("hammer2_unmount: chains left over after final sync\n"); 1798 kprintf(" vchain %08x\n", hmp->vchain.flags); 1799 kprintf(" fchain %08x\n", hmp->fchain.flags); 1800 1801 if (hammer2_debug & 0x0010) 1802 Debugger("entered debugger"); 1803 } 1804 1805 hammer2_pfsfree_scan(hmp, 1); 1806 1807 KKASSERT(hmp->spmp == NULL); 1808 1809 /* 1810 * Finish up with the device vnode 1811 */ 1812 if (!TAILQ_EMPTY(&hmp->devvpl)) { 1813 hammer2_close_devvp(&hmp->devvpl, hmp->ronly); 1814 hammer2_cleanup_devvp(&hmp->devvpl); 1815 } 1816 KKASSERT(TAILQ_EMPTY(&hmp->devvpl)); 1817 1818 /* 1819 * Clear vchain/fchain flags that might prevent final cleanup 1820 * of these chains. 1821 */ 1822 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) { 1823 atomic_add_long(&hammer2_count_modified_chains, -1); 1824 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 1825 hammer2_pfs_memory_wakeup(hmp->vchain.pmp, -1); 1826 } 1827 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) { 1828 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE); 1829 } 1830 1831 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) { 1832 atomic_add_long(&hammer2_count_modified_chains, -1); 1833 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED); 1834 hammer2_pfs_memory_wakeup(hmp->fchain.pmp, -1); 1835 } 1836 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) { 1837 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE); 1838 } 1839 1840 dumpcnt = 50; 1841 hammer2_dump_chain(&hmp->vchain, 0, 0, &dumpcnt, 'v', (u_int)-1); 1842 dumpcnt = 50; 1843 hammer2_dump_chain(&hmp->fchain, 0, 0, &dumpcnt, 'f', (u_int)-1); 1844 1845 /* 1846 * Final drop of embedded freemap root chain to 1847 * clean up fchain.core (fchain structure is not 1848 * flagged ALLOCATED so it is cleaned out and then 1849 * left to rot). 1850 */ 1851 hammer2_chain_drop(&hmp->fchain); 1852 1853 /* 1854 * Final drop of embedded volume root chain to clean 1855 * up vchain.core (vchain structure is not flagged 1856 * ALLOCATED so it is cleaned out and then left to 1857 * rot). 1858 */ 1859 hammer2_chain_drop(&hmp->vchain); 1860 1861 hammer2_io_cleanup(hmp, &hmp->iotree); 1862 if (hmp->iofree_count) { 1863 kprintf("io_cleanup: %d I/O's left hanging\n", 1864 hmp->iofree_count); 1865 } 1866 1867 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry); 1868 kmalloc_destroy_obj(&hmp->mchain); 1869 kmalloc_destroy_obj(&hmp->mio); 1870 kmalloc_destroy(&hmp->mmsg); 1871 kfree(hmp, M_HAMMER2); 1872 } 1873 1874 int 1875 hammer2_vfs_vget(struct mount *mp, struct m_vnode *dvp, 1876 ino_t ino, struct m_vnode **vpp) 1877 { 1878 hammer2_xop_lookup_t *xop; 1879 hammer2_pfs_t *pmp; 1880 hammer2_inode_t *ip; 1881 hammer2_tid_t inum; 1882 int error; 1883 1884 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK; 1885 1886 error = 0; 1887 pmp = MPTOPMP(mp); 1888 1889 /* 1890 * Easy if we already have it cached 1891 */ 1892 ip = hammer2_inode_lookup(pmp, inum); 1893 if (ip) { 1894 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED); 1895 *vpp = hammer2_igetv(ip, &error); 1896 hammer2_inode_unlock(ip); 1897 hammer2_inode_drop(ip); /* from lookup */ 1898 1899 return error; 1900 } 1901 1902 /* 1903 * Otherwise we have to find the inode 1904 */ 1905 xop = hammer2_xop_alloc(pmp->iroot, 0); 1906 xop->lhc = inum; 1907 hammer2_xop_start(&xop->head, &hammer2_lookup_desc); 1908 error = hammer2_xop_collect(&xop->head, 0); 1909 1910 if (error == 0) 1911 ip = hammer2_inode_get(pmp, &xop->head, -1, -1); 1912 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1913 1914 if (ip) { 1915 *vpp = hammer2_igetv(ip, &error); 1916 hammer2_inode_unlock(ip); 1917 } else { 1918 *vpp = NULL; 1919 error = ENOENT; 1920 } 1921 return (error); 1922 } 1923 1924 int 1925 hammer2_vfs_root(struct mount *mp, struct m_vnode **vpp) 1926 { 1927 hammer2_pfs_t *pmp; 1928 struct m_vnode *vp; 1929 int error; 1930 1931 pmp = MPTOPMP(mp); 1932 if (pmp->iroot == NULL) { 1933 kprintf("hammer2 (%s): no root inode\n", 1934 mp->mnt_stat.f_mntfromname); 1935 *vpp = NULL; 1936 return EINVAL; 1937 } 1938 1939 error = 0; 1940 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1941 1942 while (pmp->inode_tid == 0) { 1943 hammer2_xop_ipcluster_t *xop; 1944 const hammer2_inode_meta_t *meta; 1945 1946 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING); 1947 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc); 1948 error = hammer2_xop_collect(&xop->head, 0); 1949 1950 if (error == 0) { 1951 meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta; 1952 pmp->iroot->meta = *meta; 1953 pmp->inode_tid = meta->pfs_inum + 1; 1954 hammer2_xop_pdata(&xop->head); 1955 /* meta invalid */ 1956 1957 if (pmp->inode_tid < HAMMER2_INODE_START) 1958 pmp->inode_tid = HAMMER2_INODE_START; 1959 pmp->modify_tid = 1960 xop->head.cluster.focus->bref.modify_tid + 1; 1961 #if 0 1962 kprintf("PFS: Starting inode %jd\n", 1963 (intmax_t)pmp->inode_tid); 1964 kprintf("PMP focus good set nextino=%ld mod=%016jx\n", 1965 pmp->inode_tid, pmp->modify_tid); 1966 #endif 1967 //wakeup(&pmp->iroot); XXX 1968 1969 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1970 1971 /* 1972 * Prime the mount info. 1973 */ 1974 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL); 1975 break; 1976 } 1977 1978 /* 1979 * Loop, try again 1980 */ 1981 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1982 hammer2_inode_unlock(pmp->iroot); 1983 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz); 1984 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1985 if (error == EINTR) 1986 break; 1987 } 1988 1989 if (error) { 1990 hammer2_inode_unlock(pmp->iroot); 1991 *vpp = NULL; 1992 } else { 1993 vp = hammer2_igetv(pmp->iroot, &error); 1994 hammer2_inode_unlock(pmp->iroot); 1995 *vpp = vp; 1996 } 1997 1998 return (error); 1999 } 2000 2001 /* 2002 * Filesystem status 2003 * 2004 * XXX incorporate ipdata->meta.inode_quota and data_quota 2005 */ 2006 static 2007 int 2008 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred) 2009 { 2010 hammer2_pfs_t *pmp; 2011 hammer2_dev_t *hmp; 2012 hammer2_blockref_t bref; 2013 struct statfs tmp; 2014 int i; 2015 2016 /* 2017 * NOTE: iroot might not have validated the cluster yet. 2018 */ 2019 pmp = MPTOPMP(mp); 2020 2021 bzero(&tmp, sizeof(tmp)); 2022 2023 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 2024 hmp = pmp->pfs_hmps[i]; 2025 if (hmp == NULL) 2026 continue; 2027 if (pmp->iroot->cluster.array[i].chain) 2028 bref = pmp->iroot->cluster.array[i].chain->bref; 2029 else 2030 bzero(&bref, sizeof(bref)); 2031 2032 tmp.f_files = bref.embed.stats.inode_count; 2033 tmp.f_ffree = 0; 2034 tmp.f_blocks = hmp->voldata.allocator_size / 2035 mp->mnt_vstat.f_bsize; 2036 tmp.f_bfree = hmp->voldata.allocator_free / 2037 mp->mnt_vstat.f_bsize; 2038 tmp.f_bavail = tmp.f_bfree; 2039 2040 if (cred && cred->cr_uid != 0) { 2041 uint64_t adj; 2042 2043 /* 5% */ 2044 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize; 2045 tmp.f_blocks -= adj; 2046 tmp.f_bfree -= adj; 2047 tmp.f_bavail -= adj; 2048 } 2049 2050 mp->mnt_stat.f_blocks = tmp.f_blocks; 2051 mp->mnt_stat.f_bfree = tmp.f_bfree; 2052 mp->mnt_stat.f_bavail = tmp.f_bavail; 2053 mp->mnt_stat.f_files = tmp.f_files; 2054 mp->mnt_stat.f_ffree = tmp.f_ffree; 2055 2056 *sbp = mp->mnt_stat; 2057 } 2058 return (0); 2059 } 2060 2061 static 2062 int 2063 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred) 2064 { 2065 hammer2_pfs_t *pmp; 2066 hammer2_dev_t *hmp; 2067 hammer2_blockref_t bref; 2068 struct statvfs tmp; 2069 int i; 2070 2071 /* 2072 * NOTE: iroot might not have validated the cluster yet. 2073 */ 2074 pmp = MPTOPMP(mp); 2075 bzero(&tmp, sizeof(tmp)); 2076 2077 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 2078 hmp = pmp->pfs_hmps[i]; 2079 if (hmp == NULL) 2080 continue; 2081 if (pmp->iroot->cluster.array[i].chain) 2082 bref = pmp->iroot->cluster.array[i].chain->bref; 2083 else 2084 bzero(&bref, sizeof(bref)); 2085 2086 tmp.f_files = bref.embed.stats.inode_count; 2087 tmp.f_ffree = 0; 2088 tmp.f_blocks = hmp->voldata.allocator_size / 2089 mp->mnt_vstat.f_bsize; 2090 tmp.f_bfree = hmp->voldata.allocator_free / 2091 mp->mnt_vstat.f_bsize; 2092 tmp.f_bavail = tmp.f_bfree; 2093 2094 if (cred && cred->cr_uid != 0) { 2095 uint64_t adj; 2096 2097 /* 5% */ 2098 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize; 2099 tmp.f_blocks -= adj; 2100 tmp.f_bfree -= adj; 2101 tmp.f_bavail -= adj; 2102 } 2103 2104 mp->mnt_vstat.f_blocks = tmp.f_blocks; 2105 mp->mnt_vstat.f_bfree = tmp.f_bfree; 2106 mp->mnt_vstat.f_bavail = tmp.f_bavail; 2107 mp->mnt_vstat.f_files = tmp.f_files; 2108 mp->mnt_vstat.f_ffree = tmp.f_ffree; 2109 2110 *sbp = mp->mnt_vstat; 2111 } 2112 return (0); 2113 } 2114 2115 /* 2116 * Mount-time recovery (RW mounts) 2117 * 2118 * Updates to the free block table are allowed to lag flushes by one 2119 * transaction. In case of a crash, then on a fresh mount we must do an 2120 * incremental scan of the last committed transaction id and make sure that 2121 * all related blocks have been marked allocated. 2122 */ 2123 struct hammer2_recovery_elm { 2124 TAILQ_ENTRY(hammer2_recovery_elm) entry; 2125 hammer2_chain_t *chain; 2126 hammer2_tid_t sync_tid; 2127 }; 2128 2129 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm); 2130 2131 struct hammer2_recovery_info { 2132 struct hammer2_recovery_list list; 2133 hammer2_tid_t mtid; 2134 int depth; 2135 }; 2136 2137 static int hammer2_recovery_scan(hammer2_dev_t *hmp, 2138 hammer2_chain_t *parent, 2139 struct hammer2_recovery_info *info, 2140 hammer2_tid_t sync_tid); 2141 2142 #define HAMMER2_RECOVERY_MAXDEPTH 10 2143 2144 static 2145 int 2146 hammer2_recovery(hammer2_dev_t *hmp) 2147 { 2148 struct hammer2_recovery_info info; 2149 struct hammer2_recovery_elm *elm; 2150 hammer2_chain_t *parent; 2151 hammer2_tid_t sync_tid; 2152 hammer2_tid_t mirror_tid; 2153 int error; 2154 2155 hammer2_trans_init(hmp->spmp, 0); 2156 2157 sync_tid = hmp->voldata.freemap_tid; 2158 mirror_tid = hmp->voldata.mirror_tid; 2159 2160 kprintf("hammer2_mount: \"%s\": ", hmp->devrepname); 2161 if (sync_tid >= mirror_tid) { 2162 kprintf("no recovery needed\n"); 2163 } else { 2164 kprintf("freemap recovery %016jx-%016jx\n", 2165 sync_tid + 1, mirror_tid); 2166 } 2167 2168 TAILQ_INIT(&info.list); 2169 info.depth = 0; 2170 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 2171 error = hammer2_recovery_scan(hmp, parent, &info, sync_tid); 2172 hammer2_chain_lookup_done(parent); 2173 2174 while ((elm = TAILQ_FIRST(&info.list)) != NULL) { 2175 TAILQ_REMOVE(&info.list, elm, entry); 2176 parent = elm->chain; 2177 sync_tid = elm->sync_tid; 2178 kfree(elm, M_HAMMER2); 2179 2180 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2181 error |= hammer2_recovery_scan(hmp, parent, &info, 2182 hmp->voldata.freemap_tid); 2183 hammer2_chain_unlock(parent); 2184 hammer2_chain_drop(parent); /* drop elm->chain ref */ 2185 } 2186 2187 hammer2_trans_done(hmp->spmp, 0); 2188 2189 return error; 2190 } 2191 2192 static 2193 int 2194 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent, 2195 struct hammer2_recovery_info *info, 2196 hammer2_tid_t sync_tid) 2197 { 2198 const hammer2_inode_data_t *ripdata; 2199 hammer2_chain_t *chain; 2200 hammer2_blockref_t bref; 2201 int tmp_error; 2202 int rup_error; 2203 int error; 2204 int first; 2205 2206 /* 2207 * Adjust freemap to ensure that the block(s) are marked allocated. 2208 */ 2209 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) { 2210 hammer2_freemap_adjust(hmp, &parent->bref, 2211 HAMMER2_FREEMAP_DORECOVER); 2212 } 2213 2214 /* 2215 * Check type for recursive scan 2216 */ 2217 switch(parent->bref.type) { 2218 case HAMMER2_BREF_TYPE_VOLUME: 2219 /* data already instantiated */ 2220 break; 2221 case HAMMER2_BREF_TYPE_INODE: 2222 /* 2223 * Must instantiate data for DIRECTDATA test and also 2224 * for recursion. 2225 */ 2226 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2227 ripdata = &parent->data->ipdata; 2228 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) { 2229 /* not applicable to recovery scan */ 2230 hammer2_chain_unlock(parent); 2231 return 0; 2232 } 2233 hammer2_chain_unlock(parent); 2234 break; 2235 case HAMMER2_BREF_TYPE_INDIRECT: 2236 /* 2237 * Must instantiate data for recursion 2238 */ 2239 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2240 hammer2_chain_unlock(parent); 2241 break; 2242 case HAMMER2_BREF_TYPE_DIRENT: 2243 case HAMMER2_BREF_TYPE_DATA: 2244 case HAMMER2_BREF_TYPE_FREEMAP: 2245 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2246 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 2247 /* not applicable to recovery scan */ 2248 return 0; 2249 break; 2250 default: 2251 return HAMMER2_ERROR_BADBREF; 2252 } 2253 2254 /* 2255 * Defer operation if depth limit reached. 2256 */ 2257 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) { 2258 struct hammer2_recovery_elm *elm; 2259 2260 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK); 2261 elm->chain = parent; 2262 elm->sync_tid = sync_tid; 2263 hammer2_chain_ref(parent); 2264 TAILQ_INSERT_TAIL(&info->list, elm, entry); 2265 /* unlocked by caller */ 2266 2267 return(0); 2268 } 2269 2270 2271 /* 2272 * Recursive scan of the last flushed transaction only. We are 2273 * doing this without pmp assignments so don't leave the chains 2274 * hanging around after we are done with them. 2275 * 2276 * error Cumulative error this level only 2277 * rup_error Cumulative error for recursion 2278 * tmp_error Specific non-cumulative recursion error 2279 */ 2280 chain = NULL; 2281 first = 1; 2282 rup_error = 0; 2283 error = 0; 2284 2285 for (;;) { 2286 error |= hammer2_chain_scan(parent, &chain, &bref, 2287 &first, 2288 HAMMER2_LOOKUP_NODATA); 2289 2290 /* 2291 * Problem during scan or EOF 2292 */ 2293 if (error) 2294 break; 2295 2296 /* 2297 * If this is a leaf 2298 */ 2299 if (chain == NULL) { 2300 if (bref.mirror_tid > sync_tid) { 2301 hammer2_freemap_adjust(hmp, &bref, 2302 HAMMER2_FREEMAP_DORECOVER); 2303 } 2304 continue; 2305 } 2306 2307 /* 2308 * This may or may not be a recursive node. 2309 */ 2310 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 2311 if (bref.mirror_tid > sync_tid) { 2312 ++info->depth; 2313 tmp_error = hammer2_recovery_scan(hmp, chain, 2314 info, sync_tid); 2315 --info->depth; 2316 } else { 2317 tmp_error = 0; 2318 } 2319 2320 /* 2321 * Flush the recovery at the PFS boundary to stage it for 2322 * the final flush of the super-root topology. 2323 */ 2324 if (tmp_error == 0 && 2325 (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) && 2326 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) { 2327 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 2328 HAMMER2_FLUSH_ALL); 2329 } 2330 rup_error |= tmp_error; 2331 } 2332 return ((error | rup_error) & ~HAMMER2_ERROR_EOF); 2333 } 2334 2335 /* 2336 * This fixes up an error introduced in earlier H2 implementations where 2337 * moving a PFS inode into an indirect block wound up causing the 2338 * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared. 2339 */ 2340 static 2341 int 2342 hammer2_fixup_pfses(hammer2_dev_t *hmp) 2343 { 2344 const hammer2_inode_data_t *ripdata; 2345 hammer2_chain_t *parent; 2346 hammer2_chain_t *chain; 2347 hammer2_key_t key_next; 2348 hammer2_pfs_t *spmp; 2349 int error; 2350 2351 error = 0; 2352 2353 /* 2354 * Lookup mount point under the media-localized super-root. 2355 * 2356 * cluster->pmp will incorrectly point to spmp and must be fixed 2357 * up later on. 2358 */ 2359 spmp = hmp->spmp; 2360 hammer2_inode_lock(spmp->iroot, 0); 2361 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 2362 chain = hammer2_chain_lookup(&parent, &key_next, 2363 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 2364 &error, 0); 2365 while (chain) { 2366 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) 2367 continue; 2368 if (chain->error) { 2369 kprintf("I/O error scanning PFS labels\n"); 2370 error |= chain->error; 2371 } else if ((chain->bref.flags & 2372 HAMMER2_BREF_FLAG_PFSROOT) == 0) { 2373 int error2; 2374 2375 ripdata = &chain->data->ipdata; 2376 hammer2_trans_init(hmp->spmp, 0); 2377 error2 = hammer2_chain_modify(chain, 2378 chain->bref.modify_tid, 2379 0, 0); 2380 if (error2 == 0) { 2381 kprintf("hammer2: Correct mis-flagged PFS %s\n", 2382 ripdata->filename); 2383 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT; 2384 } else { 2385 error |= error2; 2386 } 2387 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 2388 HAMMER2_FLUSH_ALL); 2389 hammer2_trans_done(hmp->spmp, 0); 2390 } 2391 chain = hammer2_chain_next(&parent, chain, &key_next, 2392 key_next, HAMMER2_KEY_MAX, 2393 &error, 0); 2394 } 2395 if (parent) { 2396 hammer2_chain_unlock(parent); 2397 hammer2_chain_drop(parent); 2398 } 2399 hammer2_inode_unlock(spmp->iroot); 2400 2401 return error; 2402 } 2403 2404 /* 2405 * Sync a mount point; this is called periodically on a per-mount basis from 2406 * the filesystem syncer, and whenever a user issues a sync. 2407 */ 2408 int 2409 hammer2_vfs_sync(struct mount *mp, int waitfor) 2410 { 2411 int error; 2412 2413 error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor); 2414 2415 return error; 2416 } 2417 2418 /* 2419 * Because frontend operations lock vnodes before we get a chance to 2420 * lock the related inode, we can't just acquire a vnode lock without 2421 * risking a deadlock. The frontend may be holding a vnode lock while 2422 * also blocked on our SYNCQ flag while trying to get the inode lock. 2423 * 2424 * To deal with this situation we can check the vnode lock situation 2425 * after locking the inode and perform a work-around. 2426 */ 2427 int 2428 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor) 2429 { 2430 hammer2_inode_t *ip; 2431 hammer2_depend_t *depend; 2432 hammer2_depend_t *depend_next; 2433 struct m_vnode *vp; 2434 uint32_t pass2; 2435 int error; 2436 int wakecount; 2437 int dorestart; 2438 2439 /* 2440 * Move all inodes on sideq to syncq. This will clear sideq. 2441 * This should represent all flushable inodes. These inodes 2442 * will already have refs due to being on syncq or sideq. We 2443 * must do this all at once with the spinlock held to ensure that 2444 * all inode dependencies are part of the same flush. 2445 * 2446 * We should be able to do this asynchronously from frontend 2447 * operations because we will be locking the inodes later on 2448 * to actually flush them, and that will partition any frontend 2449 * op using the same inode. Either it has already locked the 2450 * inode and we will block, or it has not yet locked the inode 2451 * and it will block until we are finished flushing that inode. 2452 * 2453 * When restarting, only move the inodes flagged as PASS2 from 2454 * SIDEQ to SYNCQ. PASS2 propagation by inode_lock4() and 2455 * inode_depend() are atomic with the spin-lock. 2456 */ 2457 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH); 2458 #ifdef HAMMER2_DEBUG_SYNC 2459 kprintf("FILESYSTEM SYNC BOUNDARY\n"); 2460 #endif 2461 dorestart = 0; 2462 2463 /* 2464 * Move inodes from depq to syncq, releasing the related 2465 * depend structures. 2466 */ 2467 restart: 2468 #ifdef HAMMER2_DEBUG_SYNC 2469 kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart); 2470 #endif 2471 hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/); 2472 hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN); 2473 2474 /* 2475 * Move inodes from depq to syncq. When restarting, only depq's 2476 * marked pass2 are moved. 2477 */ 2478 hammer2_spin_ex(&pmp->list_spin); 2479 depend_next = TAILQ_FIRST(&pmp->depq); 2480 wakecount = 0; 2481 2482 while ((depend = depend_next) != NULL) { 2483 depend_next = TAILQ_NEXT(depend, entry); 2484 if (dorestart && depend->pass2 == 0) 2485 continue; 2486 TAILQ_FOREACH(ip, &depend->sideq, entry) { 2487 KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ); 2488 atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ); 2489 atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ); 2490 ip->depend = NULL; 2491 } 2492 2493 /* 2494 * NOTE: pmp->sideq_count includes both sideq and syncq 2495 */ 2496 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry); 2497 2498 depend->count = 0; 2499 depend->pass2 = 0; 2500 TAILQ_REMOVE(&pmp->depq, depend, entry); 2501 } 2502 2503 hammer2_spin_unex(&pmp->list_spin); 2504 hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/ 2505 HAMMER2_TRANS_WAITING); 2506 dorestart = 0; 2507 2508 /* 2509 * sideq_count may have dropped enough to allow us to unstall 2510 * the frontend. 2511 */ 2512 hammer2_pfs_memory_wakeup(pmp, 0); 2513 2514 /* 2515 * Now run through all inodes on syncq. 2516 * 2517 * Flush transactions only interlock with other flush transactions. 2518 * Any conflicting frontend operations will block on the inode, but 2519 * may hold a vnode lock while doing so. 2520 */ 2521 hammer2_spin_ex(&pmp->list_spin); 2522 while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) { 2523 /* 2524 * Remove the inode from the SYNCQ, transfer the syncq ref 2525 * to us. We must clear SYNCQ to allow any potential 2526 * front-end deadlock to proceed. We must set PASS2 so 2527 * the dependency code knows what to do. 2528 */ 2529 pass2 = ip->flags; 2530 cpu_ccfence(); 2531 if (atomic_cmpset_int(&ip->flags, 2532 pass2, 2533 (pass2 & ~(HAMMER2_INODE_SYNCQ | 2534 HAMMER2_INODE_SYNCQ_WAKEUP)) | 2535 HAMMER2_INODE_SYNCQ_PASS2) == 0) 2536 { 2537 continue; 2538 } 2539 TAILQ_REMOVE(&pmp->syncq, ip, entry); 2540 --pmp->sideq_count; 2541 hammer2_spin_unex(&pmp->list_spin); 2542 2543 /* 2544 * Tickle anyone waiting on ip->flags or the hysteresis 2545 * on the dirty inode count. 2546 */ 2547 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP) 2548 wakeup(&ip->flags); 2549 if (++wakecount >= hammer2_limit_dirty_inodes / 20 + 1) { 2550 wakecount = 0; 2551 hammer2_pfs_memory_wakeup(pmp, 0); 2552 } 2553 2554 /* 2555 * Relock the inode, and we inherit a ref from the above. 2556 * We will check for a race after we acquire the vnode. 2557 */ 2558 hammer2_mtx_ex(&ip->lock); 2559 2560 /* 2561 * We need the vp in order to vfsync() dirty buffers, so if 2562 * one isn't attached we can skip it. 2563 * 2564 * Ordering the inode lock and then the vnode lock has the 2565 * potential to deadlock. If we had left SYNCQ set that could 2566 * also deadlock us against the frontend even if we don't hold 2567 * any locks, but the latter is not a problem now since we 2568 * cleared it. igetv will temporarily release the inode lock 2569 * in a safe manner to work-around the deadlock. 2570 * 2571 * Unfortunately it is still possible to deadlock when the 2572 * frontend obtains multiple inode locks, because all the 2573 * related vnodes are already locked (nor can the vnode locks 2574 * be released and reacquired without messing up RECLAIM and 2575 * INACTIVE sequencing). 2576 * 2577 * The solution for now is to move the vp back onto SIDEQ 2578 * and set dorestart, which will restart the flush after we 2579 * exhaust the current SYNCQ. Note that additional 2580 * dependencies may build up, so we definitely need to move 2581 * the whole SIDEQ back to SYNCQ when we restart. 2582 */ 2583 vp = ip->vp; 2584 if (vp) { 2585 if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) { 2586 /* 2587 * Failed to get the vnode, requeue the inode 2588 * (PASS2 is already set so it will be found 2589 * again on the restart). 2590 * 2591 * Then unlock, possibly sleep, and retry 2592 * later. We sleep if PASS2 was *previously* 2593 * set, before we set it again above. 2594 */ 2595 vp = NULL; 2596 dorestart = 1; 2597 #ifdef HAMMER2_DEBUG_SYNC 2598 kprintf("inum %ld (sync delayed by vnode)\n", 2599 (long)ip->meta.inum); 2600 #endif 2601 hammer2_inode_delayed_sideq(ip); 2602 2603 hammer2_mtx_unlock(&ip->lock); 2604 hammer2_inode_drop(ip); 2605 2606 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) { 2607 tsleep(&dorestart, 0, "h2syndel", 2); 2608 } 2609 hammer2_spin_ex(&pmp->list_spin); 2610 continue; 2611 } 2612 } else { 2613 vp = NULL; 2614 } 2615 2616 /* 2617 * If the inode wound up on a SIDEQ again it will already be 2618 * prepped for another PASS2. In this situation if we flush 2619 * it now we will just wind up flushing it again in the same 2620 * syncer run, so we might as well not flush it now. 2621 */ 2622 if (ip->flags & HAMMER2_INODE_SIDEQ) { 2623 hammer2_mtx_unlock(&ip->lock); 2624 hammer2_inode_drop(ip); 2625 if (vp) 2626 vput(vp); 2627 dorestart = 1; 2628 hammer2_spin_ex(&pmp->list_spin); 2629 continue; 2630 } 2631 2632 /* 2633 * Ok we have the inode exclusively locked and if vp is 2634 * not NULL that will also be exclusively locked. Do the 2635 * meat of the flush. 2636 * 2637 * vp token needed for v_rbdirty_tree check / vclrisdirty 2638 * sequencing. Though we hold the vnode exclusively so 2639 * we shouldn't need to hold the token also in this case. 2640 */ 2641 if (vp) { 2642 vfsync(vp, MNT_WAIT, 1, NULL, NULL); 2643 bio_track_wait(NULL, 0, 0); /* XXX */ 2644 } 2645 2646 /* 2647 * If the inode has not yet been inserted into the tree 2648 * we must do so. Then sync and flush it. The flush should 2649 * update the parent. 2650 */ 2651 if (ip->flags & HAMMER2_INODE_DELETING) { 2652 #ifdef HAMMER2_DEBUG_SYNC 2653 kprintf("inum %ld destroy\n", (long)ip->meta.inum); 2654 #endif 2655 hammer2_inode_chain_des(ip); 2656 atomic_add_long(&hammer2_iod_inode_deletes, 1); 2657 } else if (ip->flags & HAMMER2_INODE_CREATING) { 2658 #ifdef HAMMER2_DEBUG_SYNC 2659 kprintf("inum %ld insert\n", (long)ip->meta.inum); 2660 #endif 2661 hammer2_inode_chain_ins(ip); 2662 atomic_add_long(&hammer2_iod_inode_creates, 1); 2663 } 2664 #ifdef HAMMER2_DEBUG_SYNC 2665 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum); 2666 #endif 2667 2668 /* 2669 * Because I kinda messed up the design and index the inodes 2670 * under the root inode, along side the directory entries, 2671 * we can't flush the inode index under the iroot until the 2672 * end. If we do it now we might miss effects created by 2673 * other inodes on the SYNCQ. 2674 * 2675 * Do a normal (non-FSSYNC) flush instead, which allows the 2676 * vnode code to work the same. We don't want to force iroot 2677 * back onto the SIDEQ, and we also don't want the flush code 2678 * to update pfs_iroot_blocksets until the final flush later. 2679 * 2680 * XXX at the moment this will likely result in a double-flush 2681 * of the iroot chain. 2682 */ 2683 hammer2_inode_chain_sync(ip); 2684 if (ip == pmp->iroot) { 2685 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP); 2686 } else { 2687 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP | 2688 HAMMER2_XOP_FSSYNC); 2689 } 2690 if (vp) { 2691 lwkt_gettoken(NULL); 2692 if ((ip->flags & (HAMMER2_INODE_MODIFIED | 2693 HAMMER2_INODE_RESIZED | 2694 HAMMER2_INODE_DIRTYDATA)) == 0) { 2695 //RB_EMPTY(&vp->v_rbdirty_tree) && 2696 //!bio_track_active(&vp->v_track_write)) { 2697 vclrisdirty(vp); 2698 } else { 2699 hammer2_inode_delayed_sideq(ip); 2700 } 2701 lwkt_reltoken(NULL); 2702 vput(vp); 2703 vp = NULL; /* safety */ 2704 } 2705 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2); 2706 hammer2_inode_unlock(ip); /* unlock+drop */ 2707 /* ip pointer invalid */ 2708 2709 /* 2710 * If the inode got dirted after we dropped our locks, 2711 * it will have already been moved back to the SIDEQ. 2712 */ 2713 hammer2_spin_ex(&pmp->list_spin); 2714 } 2715 hammer2_spin_unex(&pmp->list_spin); 2716 hammer2_pfs_memory_wakeup(pmp, 0); 2717 2718 if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) { 2719 #ifdef HAMMER2_DEBUG_SYNC 2720 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n"); 2721 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/ 2722 #endif 2723 dorestart = 1; 2724 goto restart; 2725 } 2726 #ifdef HAMMER2_DEBUG_SYNC 2727 kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n"); 2728 /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/ 2729 #endif 2730 2731 /* 2732 * We have to flush the PFS root last, even if it does not appear to 2733 * be dirty, because all the inodes in the PFS are indexed under it. 2734 * The normal flushing of iroot above would only occur if directory 2735 * entries under the root were changed. 2736 * 2737 * Specifying VOLHDR will cause an additionl flush of hmp->spmp 2738 * for the media making up the cluster. 2739 */ 2740 if ((ip = pmp->iroot) != NULL) { 2741 hammer2_inode_ref(ip); 2742 hammer2_mtx_ex(&ip->lock); 2743 hammer2_inode_chain_sync(ip); 2744 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP | 2745 HAMMER2_XOP_FSSYNC | 2746 HAMMER2_XOP_VOLHDR); 2747 hammer2_inode_unlock(ip); /* unlock+drop */ 2748 } 2749 #ifdef HAMMER2_DEBUG_SYNC 2750 kprintf("FILESYSTEM SYNC STAGE 2 DONE\n"); 2751 #endif 2752 2753 /* 2754 * device bioq sync 2755 */ 2756 hammer2_bioq_sync(pmp); 2757 2758 error = 0; /* XXX */ 2759 hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH); 2760 2761 return (error); 2762 } 2763 2764 #if 0 2765 static 2766 int 2767 hammer2_vfs_vptofh(struct m_vnode *vp, struct fid *fhp) 2768 { 2769 hammer2_inode_t *ip; 2770 2771 KKASSERT(MAXFIDSZ >= 16); 2772 ip = VTOI(vp); 2773 fhp->fid_len = offsetof(struct fid, fid_data[16]); 2774 fhp->fid_ext = 0; 2775 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum; 2776 ((hammer2_tid_t *)fhp->fid_data)[1] = 0; 2777 2778 return 0; 2779 } 2780 2781 static 2782 int 2783 hammer2_vfs_fhtovp(struct mount *mp, struct m_vnode *rootvp, 2784 struct fid *fhp, struct m_vnode **vpp) 2785 { 2786 hammer2_tid_t inum; 2787 int error; 2788 2789 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK; 2790 if (vpp) { 2791 if (inum == 1) 2792 error = hammer2_vfs_root(mp, vpp); 2793 else 2794 error = hammer2_vfs_vget(mp, NULL, inum, vpp); 2795 } else { 2796 error = 0; 2797 } 2798 return error; 2799 } 2800 2801 static 2802 int 2803 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 2804 int *exflagsp, struct ucred **credanonp) 2805 { 2806 hammer2_pfs_t *pmp; 2807 struct netcred *np; 2808 int error; 2809 2810 pmp = MPTOPMP(mp); 2811 np = vfs_export_lookup(mp, &pmp->export, nam); 2812 if (np) { 2813 *exflagsp = np->netc_exflags; 2814 *credanonp = &np->netc_anon; 2815 error = 0; 2816 } else { 2817 error = EACCES; 2818 } 2819 return error; 2820 } 2821 #endif 2822 2823 /* 2824 * This handles hysteresis on regular file flushes. Because the BIOs are 2825 * routed to a thread it is possible for an excessive number to build up 2826 * and cause long front-end stalls long before the runningbuffspace limit 2827 * is hit, so we implement hammer2_flush_pipe to control the 2828 * hysteresis. 2829 * 2830 * This is a particular problem when compression is used. 2831 */ 2832 void 2833 hammer2_lwinprog_ref(hammer2_pfs_t *pmp) 2834 { 2835 atomic_add_int(&pmp->count_lwinprog, 1); 2836 } 2837 2838 void 2839 hammer2_lwinprog_drop(hammer2_pfs_t *pmp) 2840 { 2841 #if 0 2842 int lwinprog; 2843 2844 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1); 2845 if ((lwinprog & HAMMER2_LWINPROG_WAITING) && 2846 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) { 2847 atomic_clear_int(&pmp->count_lwinprog, 2848 HAMMER2_LWINPROG_WAITING); 2849 wakeup(&pmp->count_lwinprog); 2850 } 2851 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) && 2852 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) { 2853 atomic_clear_int(&pmp->count_lwinprog, 2854 HAMMER2_LWINPROG_WAITING0); 2855 wakeup(&pmp->count_lwinprog); 2856 } 2857 #endif 2858 } 2859 2860 void 2861 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe) 2862 { 2863 #if 0 2864 int lwinprog; 2865 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING : 2866 HAMMER2_LWINPROG_WAITING0; 2867 2868 for (;;) { 2869 lwinprog = pmp->count_lwinprog; 2870 cpu_ccfence(); 2871 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 2872 break; 2873 tsleep_interlock(&pmp->count_lwinprog, 0); 2874 atomic_set_int(&pmp->count_lwinprog, lwflag); 2875 lwinprog = pmp->count_lwinprog; 2876 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 2877 break; 2878 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz); 2879 } 2880 #endif 2881 } 2882 2883 #if 0 2884 /* 2885 * It is possible for an excessive number of dirty chains or dirty inodes 2886 * to build up. When this occurs we start an asynchronous filesystem sync. 2887 * If the level continues to build up, we stall, waiting for it to drop, 2888 * with some hysteresis. 2889 * 2890 * This relies on the kernel calling hammer2_vfs_modifying() prior to 2891 * obtaining any vnode locks before making a modifying VOP call. 2892 */ 2893 static int 2894 hammer2_vfs_modifying(struct mount *mp) 2895 { 2896 if (mp->mnt_flag & MNT_RDONLY) 2897 return EROFS; 2898 hammer2_pfs_memory_wait(MPTOPMP(mp)); 2899 2900 return 0; 2901 } 2902 #endif 2903 2904 /* 2905 * Initiate an asynchronous filesystem sync and, with hysteresis, 2906 * stall if the internal data structure count becomes too bloated. 2907 */ 2908 void 2909 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp) 2910 { 2911 uint32_t waiting; 2912 int pcatch; 2913 int error; 2914 2915 if (pmp == NULL || pmp->mp == NULL) 2916 return; 2917 2918 for (;;) { 2919 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK; 2920 cpu_ccfence(); 2921 2922 /* 2923 * Start the syncer running at 1/2 the limit 2924 */ 2925 if (waiting > hammer2_limit_dirty_chains / 2 || 2926 pmp->sideq_count > hammer2_limit_dirty_inodes / 2) { 2927 trigger_syncer(pmp->mp); 2928 } 2929 2930 /* 2931 * Stall at the limit waiting for the counts to drop. 2932 * This code will typically be woken up once the count 2933 * drops below 3/4 the limit, or in one second. 2934 */ 2935 if (waiting < hammer2_limit_dirty_chains && 2936 pmp->sideq_count < hammer2_limit_dirty_inodes) { 2937 break; 2938 } 2939 2940 pcatch = curthread->td_proc ? PCATCH : 0; 2941 2942 tsleep_interlock(&pmp->inmem_dirty_chains, pcatch); 2943 atomic_set_int(&pmp->inmem_dirty_chains, 2944 HAMMER2_DIRTYCHAIN_WAITING); 2945 if (waiting < hammer2_limit_dirty_chains && 2946 pmp->sideq_count < hammer2_limit_dirty_inodes) { 2947 break; 2948 } 2949 trigger_syncer(pmp->mp); 2950 error = tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED | pcatch, 2951 "h2memw", hz); 2952 if (error == ERESTART) 2953 break; 2954 } 2955 } 2956 2957 /* 2958 * Wake up any stalled frontend ops waiting, with hysteresis, using 2959 * 2/3 of the limit. 2960 */ 2961 void 2962 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp, int count) 2963 { 2964 uint32_t waiting; 2965 2966 if (pmp) { 2967 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, count); 2968 /* don't need --waiting to test flag */ 2969 2970 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) && 2971 (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <= 2972 hammer2_limit_dirty_chains * 2 / 3 && 2973 pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) { 2974 atomic_clear_int(&pmp->inmem_dirty_chains, 2975 HAMMER2_DIRTYCHAIN_WAITING); 2976 wakeup(&pmp->inmem_dirty_chains); 2977 } 2978 } 2979 } 2980 2981 void 2982 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp) 2983 { 2984 if (pmp) { 2985 atomic_add_int(&pmp->inmem_dirty_chains, 1); 2986 } 2987 } 2988 2989 /* 2990 * Volume header data locks 2991 */ 2992 void 2993 hammer2_voldata_lock(hammer2_dev_t *hmp) 2994 { 2995 lockmgr(&hmp->vollk, LK_EXCLUSIVE); 2996 } 2997 2998 void 2999 hammer2_voldata_unlock(hammer2_dev_t *hmp) 3000 { 3001 lockmgr(&hmp->vollk, LK_RELEASE); 3002 } 3003 3004 /* 3005 * Caller indicates that the volume header is being modified. Flag 3006 * the related chain and adjust its transaction id. 3007 * 3008 * The transaction id is set to voldata.mirror_tid + 1, similar to 3009 * what hammer2_chain_modify() does. Be very careful here, volume 3010 * data can be updated independently of the rest of the filesystem. 3011 */ 3012 void 3013 hammer2_voldata_modify(hammer2_dev_t *hmp) 3014 { 3015 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) { 3016 atomic_add_long(&hammer2_count_modified_chains, 1); 3017 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 3018 hammer2_pfs_memory_inc(hmp->vchain.pmp); 3019 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid + 1; 3020 } 3021 } 3022 3023 /* 3024 * Returns 0 if the filesystem has tons of free space 3025 * Returns 1 if the filesystem has less than 10% remaining 3026 * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining. 3027 */ 3028 int 3029 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred) 3030 { 3031 hammer2_pfs_t *pmp; 3032 hammer2_dev_t *hmp; 3033 hammer2_off_t free_reserved; 3034 hammer2_off_t free_nominal; 3035 int i; 3036 3037 pmp = ip->pmp; 3038 3039 if (/*XXX*/ 1 || pmp->free_ticks == 0 || pmp->free_ticks != ticks) { 3040 free_reserved = HAMMER2_SEGSIZE; 3041 free_nominal = 0x7FFFFFFFFFFFFFFFLLU; 3042 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 3043 hmp = pmp->pfs_hmps[i]; 3044 if (hmp == NULL) 3045 continue; 3046 if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER && 3047 pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER) 3048 continue; 3049 3050 if (free_nominal > hmp->voldata.allocator_free) 3051 free_nominal = hmp->voldata.allocator_free; 3052 if (free_reserved < hmp->free_reserved) 3053 free_reserved = hmp->free_reserved; 3054 } 3055 3056 /* 3057 * SMP races ok 3058 */ 3059 pmp->free_reserved = free_reserved; 3060 pmp->free_nominal = free_nominal; 3061 pmp->free_ticks = ticks; 3062 } else { 3063 free_reserved = pmp->free_reserved; 3064 free_nominal = pmp->free_nominal; 3065 } 3066 if (cred && cred->cr_uid != 0) { 3067 if ((int64_t)(free_nominal - bytes) < 3068 (int64_t)free_reserved) { 3069 return 2; 3070 } 3071 } else { 3072 if ((int64_t)(free_nominal - bytes) < 3073 (int64_t)free_reserved / 2) { 3074 return 2; 3075 } 3076 } 3077 if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2) 3078 return 1; 3079 return 0; 3080 } 3081