1 /* 2 * Copyright (c) 2011-2018 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression) 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in 16 * the documentation and/or other materials provided with the 17 * distribution. 18 * 3. Neither the name of The DragonFly Project nor the names of its 19 * contributors may be used to endorse or promote products derived 20 * from this software without specific, prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 #include <sys/param.h> 36 #include <sys/systm.h> 37 #include <sys/kernel.h> 38 #include <sys/nlookup.h> 39 #include <sys/vnode.h> 40 #include <sys/mount.h> 41 #include <sys/fcntl.h> 42 #include <sys/buf.h> 43 #include <sys/uuid.h> 44 #include <sys/vfsops.h> 45 #include <sys/sysctl.h> 46 #include <sys/socket.h> 47 #include <sys/objcache.h> 48 49 #include <sys/proc.h> 50 #include <sys/namei.h> 51 #include <sys/mountctl.h> 52 #include <sys/dirent.h> 53 #include <sys/uio.h> 54 55 #include <sys/mutex.h> 56 #include <sys/mutex2.h> 57 58 #include "hammer2.h" 59 #include "hammer2_disk.h" 60 #include "hammer2_mount.h" 61 #include "hammer2_lz4.h" 62 63 #include "zlib/hammer2_zlib.h" 64 65 #define REPORT_REFS_ERRORS 1 /* XXX remove me */ 66 67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache"); 68 69 struct hammer2_sync_info { 70 int error; 71 int waitfor; 72 int pass; 73 }; 74 75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev); 76 static struct hammer2_mntlist hammer2_mntlist; 77 78 struct hammer2_pfslist hammer2_pfslist; 79 struct hammer2_pfslist hammer2_spmplist; 80 struct lock hammer2_mntlk; 81 82 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT; 83 int hammer2_debug; 84 int hammer2_xopgroups; 85 long hammer2_debug_inode; 86 int hammer2_cluster_meta_read = 1; /* physical read-ahead */ 87 int hammer2_cluster_data_read = 4; /* physical read-ahead */ 88 int hammer2_cluster_write = 0; /* physical write clustering */ 89 int hammer2_dedup_enable = 1; 90 int hammer2_always_compress = 0; /* always try to compress */ 91 int hammer2_inval_enable = 0; 92 int hammer2_flush_pipe = 100; 93 int hammer2_dio_count; 94 int hammer2_dio_limit = 256; 95 int hammer2_bulkfree_tps = 5000; 96 int hammer2_worker_rmask = 3; 97 long hammer2_chain_allocs; 98 long hammer2_chain_frees; 99 long hammer2_limit_dirty_chains; 100 long hammer2_limit_dirty_inodes; 101 long hammer2_count_modified_chains; 102 long hammer2_iod_invals; 103 long hammer2_iod_file_read; 104 long hammer2_iod_meta_read; 105 long hammer2_iod_indr_read; 106 long hammer2_iod_fmap_read; 107 long hammer2_iod_volu_read; 108 long hammer2_iod_file_write; 109 long hammer2_iod_file_wembed; 110 long hammer2_iod_file_wzero; 111 long hammer2_iod_file_wdedup; 112 long hammer2_iod_meta_write; 113 long hammer2_iod_indr_write; 114 long hammer2_iod_fmap_write; 115 long hammer2_iod_volu_write; 116 long hammer2_iod_inode_creates; 117 long hammer2_iod_inode_deletes; 118 119 MALLOC_DECLARE(M_HAMMER2_CBUFFER); 120 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer", 121 "Buffer used for compression."); 122 123 MALLOC_DECLARE(M_HAMMER2_DEBUFFER); 124 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer", 125 "Buffer used for decompression."); 126 127 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem"); 128 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD, 130 &hammer2_supported_version, 0, ""); 131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW, 132 &hammer2_debug, 0, ""); 133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW, 134 &hammer2_debug_inode, 0, ""); 135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW, 136 &hammer2_cluster_meta_read, 0, ""); 137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW, 138 &hammer2_cluster_data_read, 0, ""); 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW, 140 &hammer2_cluster_write, 0, ""); 141 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW, 142 &hammer2_dedup_enable, 0, ""); 143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW, 144 &hammer2_always_compress, 0, ""); 145 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW, 146 &hammer2_inval_enable, 0, ""); 147 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW, 148 &hammer2_flush_pipe, 0, ""); 149 SYSCTL_INT(_vfs_hammer2, OID_AUTO, worker_rmask, CTLFLAG_RW, 150 &hammer2_worker_rmask, 0, ""); 151 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW, 152 &hammer2_bulkfree_tps, 0, ""); 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW, 154 &hammer2_chain_allocs, 0, ""); 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW, 156 &hammer2_chain_frees, 0, ""); 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW, 158 &hammer2_limit_dirty_chains, 0, ""); 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW, 160 &hammer2_limit_dirty_inodes, 0, ""); 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW, 162 &hammer2_count_modified_chains, 0, ""); 163 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD, 164 &hammer2_dio_count, 0, ""); 165 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW, 166 &hammer2_dio_limit, 0, ""); 167 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW, 169 &hammer2_iod_invals, 0, ""); 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW, 171 &hammer2_iod_file_read, 0, ""); 172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW, 173 &hammer2_iod_meta_read, 0, ""); 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW, 175 &hammer2_iod_indr_read, 0, ""); 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW, 177 &hammer2_iod_fmap_read, 0, ""); 178 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW, 179 &hammer2_iod_volu_read, 0, ""); 180 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW, 182 &hammer2_iod_file_write, 0, ""); 183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW, 184 &hammer2_iod_file_wembed, 0, ""); 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW, 186 &hammer2_iod_file_wzero, 0, ""); 187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW, 188 &hammer2_iod_file_wdedup, 0, ""); 189 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW, 190 &hammer2_iod_meta_write, 0, ""); 191 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW, 192 &hammer2_iod_indr_write, 0, ""); 193 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW, 194 &hammer2_iod_fmap_write, 0, ""); 195 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW, 196 &hammer2_iod_volu_write, 0, ""); 197 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW, 198 &hammer2_iod_inode_creates, 0, ""); 199 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW, 200 &hammer2_iod_inode_deletes, 0, ""); 201 202 long hammer2_process_icrc32; 203 long hammer2_process_xxhash64; 204 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW, 205 &hammer2_process_icrc32, 0, ""); 206 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW, 207 &hammer2_process_xxhash64, 0, ""); 208 209 static int hammer2_vfs_init(struct vfsconf *conf); 210 static int hammer2_vfs_uninit(struct vfsconf *vfsp); 211 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data, 212 struct ucred *cred); 213 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *, 214 struct vnode *, struct ucred *); 215 static int hammer2_recovery(hammer2_dev_t *hmp); 216 static int hammer2_vfs_unmount(struct mount *mp, int mntflags); 217 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp); 218 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, 219 struct ucred *cred); 220 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, 221 struct ucred *cred); 222 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp, 223 struct fid *fhp, struct vnode **vpp); 224 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp); 225 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 226 int *exflagsp, struct ucred **credanonp); 227 static void hammer2_vfs_modifying(struct mount *mp); 228 229 static int hammer2_install_volume_header(hammer2_dev_t *hmp); 230 #if 0 231 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data); 232 #endif 233 234 static void hammer2_update_pmps(hammer2_dev_t *hmp); 235 236 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp); 237 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, 238 hammer2_dev_t *hmp); 239 static int hammer2_fixup_pfses(hammer2_dev_t *hmp); 240 241 /* 242 * HAMMER2 vfs operations. 243 */ 244 static struct vfsops hammer2_vfsops = { 245 .vfs_init = hammer2_vfs_init, 246 .vfs_uninit = hammer2_vfs_uninit, 247 .vfs_sync = hammer2_vfs_sync, 248 .vfs_mount = hammer2_vfs_mount, 249 .vfs_unmount = hammer2_vfs_unmount, 250 .vfs_root = hammer2_vfs_root, 251 .vfs_statfs = hammer2_vfs_statfs, 252 .vfs_statvfs = hammer2_vfs_statvfs, 253 .vfs_vget = hammer2_vfs_vget, 254 .vfs_vptofh = hammer2_vfs_vptofh, 255 .vfs_fhtovp = hammer2_vfs_fhtovp, 256 .vfs_checkexp = hammer2_vfs_checkexp, 257 .vfs_modifying = hammer2_vfs_modifying 258 }; 259 260 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", ""); 261 262 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE); 263 MODULE_VERSION(hammer2, 1); 264 265 static 266 int 267 hammer2_vfs_init(struct vfsconf *conf) 268 { 269 static struct objcache_malloc_args margs_read; 270 static struct objcache_malloc_args margs_write; 271 static struct objcache_malloc_args margs_vop; 272 273 int error; 274 275 error = 0; 276 kmalloc_raise_limit(M_HAMMER2, 0); /* unlimited */ 277 278 /* 279 * hammer2_xopgroups must be even and is most optimal if 280 * 2 x ncpus so strategy functions can be queued to the same 281 * cpu. 282 */ 283 hammer2_xopgroups = HAMMER2_XOPGROUPS_MIN; 284 if (hammer2_xopgroups < ncpus * 2) 285 hammer2_xopgroups = ncpus * 2; 286 287 /* 288 * A large DIO cache is needed to retain dedup enablement masks. 289 * The bulkfree code clears related masks as part of the disk block 290 * recycling algorithm, preventing it from being used for a later 291 * dedup. 292 * 293 * NOTE: A large buffer cache can actually interfere with dedup 294 * operation because we dedup based on media physical buffers 295 * and not logical buffers. Try to make the DIO case large 296 * enough to avoid this problem, but also cap it. 297 */ 298 hammer2_dio_limit = nbuf * 2; 299 if (hammer2_dio_limit > 100000) 300 hammer2_dio_limit = 100000; 301 302 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref)) 303 error = EINVAL; 304 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data)) 305 error = EINVAL; 306 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data)) 307 error = EINVAL; 308 309 if (error) 310 kprintf("HAMMER2 structure size mismatch; cannot continue.\n"); 311 312 margs_read.objsize = 65536; 313 margs_read.mtype = M_HAMMER2_DEBUFFER; 314 315 margs_write.objsize = 32768; 316 margs_write.mtype = M_HAMMER2_CBUFFER; 317 318 margs_vop.objsize = sizeof(hammer2_xop_t); 319 margs_vop.mtype = M_HAMMER2; 320 321 /* 322 * Note thaht for the XOPS cache we want backing store allocations 323 * to use M_ZERO. This is not allowed in objcache_get() (to avoid 324 * confusion), so use the backing store function that does it. This 325 * means that initial XOPS objects are zerod but REUSED objects are 326 * not. So we are responsible for cleaning the object up sufficiently 327 * for our needs before objcache_put()ing it back (typically just the 328 * FIFO indices). 329 */ 330 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc, 331 0, 1, NULL, NULL, NULL, 332 objcache_malloc_alloc, 333 objcache_malloc_free, 334 &margs_read); 335 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc, 336 0, 1, NULL, NULL, NULL, 337 objcache_malloc_alloc, 338 objcache_malloc_free, 339 &margs_write); 340 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc, 341 0, 1, NULL, NULL, NULL, 342 objcache_malloc_alloc_zero, 343 objcache_malloc_free, 344 &margs_vop); 345 346 347 lockinit(&hammer2_mntlk, "mntlk", 0, 0); 348 TAILQ_INIT(&hammer2_mntlist); 349 TAILQ_INIT(&hammer2_pfslist); 350 TAILQ_INIT(&hammer2_spmplist); 351 352 hammer2_limit_dirty_chains = maxvnodes / 10; 353 if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS) 354 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS; 355 if (hammer2_limit_dirty_chains < 1000) 356 hammer2_limit_dirty_chains = 1000; 357 358 hammer2_limit_dirty_inodes = maxvnodes / 25; 359 if (hammer2_limit_dirty_inodes < 100) 360 hammer2_limit_dirty_inodes = 100; 361 if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES) 362 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES; 363 364 return (error); 365 } 366 367 static 368 int 369 hammer2_vfs_uninit(struct vfsconf *vfsp __unused) 370 { 371 objcache_destroy(cache_buffer_read); 372 objcache_destroy(cache_buffer_write); 373 objcache_destroy(cache_xops); 374 return 0; 375 } 376 377 /* 378 * Core PFS allocator. Used to allocate or reference the pmp structure 379 * for PFS cluster mounts and the spmp structure for media (hmp) structures. 380 * The pmp can be passed in or loaded by this function using the chain and 381 * inode data. 382 * 383 * pmp->modify_tid tracks new modify_tid transaction ids for front-end 384 * transactions. Note that synchronization does not use this field. 385 * (typically frontend operations and synchronization cannot run on the 386 * same PFS node at the same time). 387 * 388 * XXX check locking 389 */ 390 hammer2_pfs_t * 391 hammer2_pfsalloc(hammer2_chain_t *chain, 392 const hammer2_inode_data_t *ripdata, 393 hammer2_tid_t modify_tid, hammer2_dev_t *force_local) 394 { 395 hammer2_pfs_t *pmp; 396 hammer2_inode_t *iroot; 397 int count; 398 int i; 399 int j; 400 401 pmp = NULL; 402 403 /* 404 * Locate or create the PFS based on the cluster id. If ripdata 405 * is NULL this is a spmp which is unique and is always allocated. 406 * 407 * If the device is mounted in local mode all PFSs are considered 408 * independent and not part of any cluster (for debugging only). 409 */ 410 if (ripdata) { 411 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 412 if (force_local != pmp->force_local) 413 continue; 414 if (force_local == NULL && 415 bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid, 416 sizeof(pmp->pfs_clid)) == 0) { 417 break; 418 } else if (force_local && pmp->pfs_names[0] && 419 strcmp(pmp->pfs_names[0], ripdata->filename) == 0) { 420 break; 421 } 422 } 423 } 424 425 if (pmp == NULL) { 426 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO); 427 pmp->force_local = force_local; 428 hammer2_trans_manage_init(pmp); 429 kmalloc_create(&pmp->minode, "HAMMER2-inodes"); 430 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg"); 431 lockinit(&pmp->lock, "pfslk", 0, 0); 432 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0); 433 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum"); 434 spin_init(&pmp->xop_spin, "h2xop"); 435 spin_init(&pmp->lru_spin, "h2lru"); 436 RB_INIT(&pmp->inum_tree); 437 TAILQ_INIT(&pmp->syncq); 438 TAILQ_INIT(&pmp->depq); 439 TAILQ_INIT(&pmp->lru_list); 440 spin_init(&pmp->list_spin, "h2pfsalloc_list"); 441 442 /* 443 * Save the last media transaction id for the flusher. Set 444 * initial 445 */ 446 if (ripdata) { 447 pmp->pfs_clid = ripdata->meta.pfs_clid; 448 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry); 449 } else { 450 pmp->flags |= HAMMER2_PMPF_SPMP; 451 TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry); 452 } 453 454 /* 455 * The synchronization thread may start too early, make 456 * sure it stays frozen until we are ready to let it go. 457 * XXX 458 */ 459 /* 460 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN | 461 HAMMER2_THREAD_REMASTER; 462 */ 463 } 464 465 /* 466 * Create the PFS's root inode and any missing XOP helper threads. 467 */ 468 if ((iroot = pmp->iroot) == NULL) { 469 iroot = hammer2_inode_get(pmp, NULL, 1, -1); 470 if (ripdata) 471 iroot->meta = ripdata->meta; 472 pmp->iroot = iroot; 473 hammer2_inode_ref(iroot); 474 hammer2_inode_unlock(iroot); 475 } 476 477 /* 478 * Stop here if no chain is passed in. 479 */ 480 if (chain == NULL) 481 goto done; 482 483 /* 484 * When a chain is passed in we must add it to the PFS's root 485 * inode, update pmp->pfs_types[], and update the syncronization 486 * threads. 487 * 488 * When forcing local mode, mark the PFS as a MASTER regardless. 489 * 490 * At the moment empty spots can develop due to removals or failures. 491 * Ultimately we want to re-fill these spots but doing so might 492 * confused running code. XXX 493 */ 494 hammer2_inode_ref(iroot); 495 hammer2_mtx_ex(&iroot->lock); 496 j = iroot->cluster.nchains; 497 498 if (j == HAMMER2_MAXCLUSTER) { 499 kprintf("hammer2_mount: cluster full!\n"); 500 /* XXX fatal error? */ 501 } else { 502 KKASSERT(chain->pmp == NULL); 503 chain->pmp = pmp; 504 hammer2_chain_ref(chain); 505 iroot->cluster.array[j].chain = chain; 506 if (force_local) 507 pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER; 508 else 509 pmp->pfs_types[j] = ripdata->meta.pfs_type; 510 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2); 511 pmp->pfs_hmps[j] = chain->hmp; 512 hammer2_spin_ex(&pmp->inum_spin); 513 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset; 514 hammer2_spin_unex(&pmp->inum_spin); 515 516 /* 517 * If the PFS is already mounted we must account 518 * for the mount_count here. 519 */ 520 if (pmp->mp) 521 ++chain->hmp->mount_count; 522 523 /* 524 * May have to fixup dirty chain tracking. Previous 525 * pmp was NULL so nothing to undo. 526 */ 527 if (chain->flags & HAMMER2_CHAIN_MODIFIED) 528 hammer2_pfs_memory_inc(pmp); 529 ++j; 530 } 531 iroot->cluster.nchains = j; 532 533 /* 534 * Update nmasters from any PFS inode which is part of the cluster. 535 * It is possible that this will result in a value which is too 536 * high. MASTER PFSs are authoritative for pfs_nmasters and will 537 * override this value later on. 538 * 539 * (This informs us of masters that might not currently be 540 * discoverable by this mount). 541 */ 542 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) { 543 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters; 544 } 545 546 /* 547 * Count visible masters. Masters are usually added with 548 * ripdata->meta.pfs_nmasters set to 1. This detects when there 549 * are more (XXX and must update the master inodes). 550 */ 551 count = 0; 552 for (i = 0; i < iroot->cluster.nchains; ++i) { 553 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) 554 ++count; 555 } 556 if (pmp->pfs_nmasters < count) 557 pmp->pfs_nmasters = count; 558 559 /* 560 * Create missing synchronization and support threads. 561 * 562 * Single-node masters (including snapshots) have nothing to 563 * synchronize and do not require this thread. 564 * 565 * Multi-node masters or any number of soft masters, slaves, copy, 566 * or other PFS types need the thread. 567 * 568 * Each thread is responsible for its particular cluster index. 569 * We use independent threads so stalls or mismatches related to 570 * any given target do not affect other targets. 571 */ 572 for (i = 0; i < iroot->cluster.nchains; ++i) { 573 /* 574 * Single-node masters (including snapshots) have nothing 575 * to synchronize and will make direct xops support calls, 576 * thus they do not require this thread. 577 * 578 * Note that there can be thousands of snapshots. We do not 579 * want to create thousands of threads. 580 */ 581 if (pmp->pfs_nmasters <= 1 && 582 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) { 583 continue; 584 } 585 586 /* 587 * Sync support thread 588 */ 589 if (pmp->sync_thrs[i].td == NULL) { 590 hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL, 591 "h2nod", i, -1, 592 hammer2_primary_sync_thread); 593 } 594 } 595 596 /* 597 * Create missing Xop threads 598 * 599 * NOTE: We create helper threads for all mounted PFSs or any 600 * PFSs with 2+ nodes (so the sync thread can update them, 601 * even if not mounted). 602 */ 603 if (pmp->mp || iroot->cluster.nchains >= 2) 604 hammer2_xop_helper_create(pmp); 605 606 hammer2_mtx_unlock(&iroot->lock); 607 hammer2_inode_drop(iroot); 608 done: 609 return pmp; 610 } 611 612 /* 613 * Deallocate an element of a probed PFS. If destroying and this is a 614 * MASTER, adjust nmasters. 615 * 616 * This function does not physically destroy the PFS element in its device 617 * under the super-root (see hammer2_ioctl_pfs_delete()). 618 */ 619 void 620 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying) 621 { 622 hammer2_inode_t *iroot; 623 hammer2_chain_t *chain; 624 int j; 625 626 /* 627 * Cleanup our reference on iroot. iroot is (should) not be needed 628 * by the flush code. 629 */ 630 iroot = pmp->iroot; 631 if (iroot) { 632 /* 633 * Stop synchronizing 634 * 635 * XXX flush after acquiring the iroot lock. 636 * XXX clean out the cluster index from all inode structures. 637 */ 638 hammer2_thr_delete(&pmp->sync_thrs[clindex]); 639 640 /* 641 * Remove the cluster index from the group. If destroying 642 * the PFS and this is a master, adjust pfs_nmasters. 643 */ 644 hammer2_mtx_ex(&iroot->lock); 645 chain = iroot->cluster.array[clindex].chain; 646 iroot->cluster.array[clindex].chain = NULL; 647 648 switch(pmp->pfs_types[clindex]) { 649 case HAMMER2_PFSTYPE_MASTER: 650 if (destroying && pmp->pfs_nmasters > 0) 651 --pmp->pfs_nmasters; 652 /* XXX adjust ripdata->meta.pfs_nmasters */ 653 break; 654 default: 655 break; 656 } 657 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE; 658 659 hammer2_mtx_unlock(&iroot->lock); 660 661 /* 662 * Release the chain. 663 */ 664 if (chain) { 665 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 666 hammer2_chain_drop(chain); 667 } 668 669 /* 670 * Terminate all XOP threads for the cluster index. 671 */ 672 if (pmp->xop_groups) { 673 for (j = 0; j < hammer2_xopgroups; ++j) { 674 hammer2_thr_delete( 675 &pmp->xop_groups[j].thrs[clindex]); 676 } 677 } 678 } 679 } 680 681 /* 682 * Destroy a PFS, typically only occurs after the last mount on a device 683 * has gone away. 684 */ 685 static void 686 hammer2_pfsfree(hammer2_pfs_t *pmp) 687 { 688 hammer2_inode_t *iroot; 689 hammer2_chain_t *chain; 690 int chains_still_present = 0; 691 int i; 692 int j; 693 694 /* 695 * Cleanup our reference on iroot. iroot is (should) not be needed 696 * by the flush code. 697 */ 698 if (pmp->flags & HAMMER2_PMPF_SPMP) 699 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry); 700 else 701 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry); 702 703 /* 704 * Cleanup chains remaining on LRU list. 705 */ 706 hammer2_spin_ex(&pmp->lru_spin); 707 while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) { 708 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU); 709 atomic_add_int(&pmp->lru_count, -1); 710 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU); 711 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node); 712 hammer2_chain_ref(chain); 713 hammer2_spin_unex(&pmp->lru_spin); 714 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 715 hammer2_chain_drop(chain); 716 hammer2_spin_ex(&pmp->lru_spin); 717 } 718 hammer2_spin_unex(&pmp->lru_spin); 719 720 /* 721 * Clean up iroot 722 */ 723 iroot = pmp->iroot; 724 if (iroot) { 725 for (i = 0; i < iroot->cluster.nchains; ++i) { 726 hammer2_thr_delete(&pmp->sync_thrs[i]); 727 if (pmp->xop_groups) { 728 for (j = 0; j < hammer2_xopgroups; ++j) 729 hammer2_thr_delete( 730 &pmp->xop_groups[j].thrs[i]); 731 } 732 chain = iroot->cluster.array[i].chain; 733 if (chain && !RB_EMPTY(&chain->core.rbtree)) { 734 kprintf("hammer2: Warning pmp %p still " 735 "has active chains\n", pmp); 736 chains_still_present = 1; 737 } 738 } 739 #if REPORT_REFS_ERRORS 740 if (iroot->refs != 1) 741 kprintf("PMP->IROOT %p REFS WRONG %d\n", 742 iroot, iroot->refs); 743 #else 744 KKASSERT(iroot->refs == 1); 745 #endif 746 /* ref for iroot */ 747 hammer2_inode_drop(iroot); 748 pmp->iroot = NULL; 749 } 750 751 /* 752 * Free remaining pmp resources 753 */ 754 if (chains_still_present) { 755 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp); 756 } else { 757 kmalloc_destroy(&pmp->mmsg); 758 kmalloc_destroy(&pmp->minode); 759 kfree(pmp, M_HAMMER2); 760 } 761 } 762 763 /* 764 * Remove all references to hmp from the pfs list. Any PFS which becomes 765 * empty is terminated and freed. 766 * 767 * XXX inefficient. 768 */ 769 static void 770 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which) 771 { 772 hammer2_pfs_t *pmp; 773 hammer2_inode_t *iroot; 774 hammer2_chain_t *rchain; 775 int i; 776 int j; 777 struct hammer2_pfslist *wlist; 778 779 if (which == 0) 780 wlist = &hammer2_pfslist; 781 else 782 wlist = &hammer2_spmplist; 783 again: 784 TAILQ_FOREACH(pmp, wlist, mntentry) { 785 if ((iroot = pmp->iroot) == NULL) 786 continue; 787 788 /* 789 * Determine if this PFS is affected. If it is we must 790 * freeze all management threads and lock its iroot. 791 * 792 * Freezing a management thread forces it idle, operations 793 * in-progress will be aborted and it will have to start 794 * over again when unfrozen, or exit if told to exit. 795 */ 796 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 797 if (pmp->pfs_hmps[i] == hmp) 798 break; 799 } 800 if (i == HAMMER2_MAXCLUSTER) 801 continue; 802 803 hammer2_vfs_sync_pmp(pmp, MNT_WAIT); 804 805 /* 806 * Make sure all synchronization threads are locked 807 * down. 808 */ 809 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 810 if (pmp->pfs_hmps[i] == NULL) 811 continue; 812 hammer2_thr_freeze_async(&pmp->sync_thrs[i]); 813 if (pmp->xop_groups) { 814 for (j = 0; j < hammer2_xopgroups; ++j) { 815 hammer2_thr_freeze_async( 816 &pmp->xop_groups[j].thrs[i]); 817 } 818 } 819 } 820 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 821 if (pmp->pfs_hmps[i] == NULL) 822 continue; 823 hammer2_thr_freeze(&pmp->sync_thrs[i]); 824 if (pmp->xop_groups) { 825 for (j = 0; j < hammer2_xopgroups; ++j) { 826 hammer2_thr_freeze( 827 &pmp->xop_groups[j].thrs[i]); 828 } 829 } 830 } 831 832 /* 833 * Lock the inode and clean out matching chains. 834 * Note that we cannot use hammer2_inode_lock_*() 835 * here because that would attempt to validate the 836 * cluster that we are in the middle of ripping 837 * apart. 838 * 839 * WARNING! We are working directly on the inodes 840 * embedded cluster. 841 */ 842 hammer2_mtx_ex(&iroot->lock); 843 844 /* 845 * Remove the chain from matching elements of the PFS. 846 */ 847 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 848 if (pmp->pfs_hmps[i] != hmp) 849 continue; 850 hammer2_thr_delete(&pmp->sync_thrs[i]); 851 if (pmp->xop_groups) { 852 for (j = 0; j < hammer2_xopgroups; ++j) { 853 hammer2_thr_delete( 854 &pmp->xop_groups[j].thrs[i]); 855 } 856 } 857 rchain = iroot->cluster.array[i].chain; 858 iroot->cluster.array[i].chain = NULL; 859 pmp->pfs_types[i] = 0; 860 if (pmp->pfs_names[i]) { 861 kfree(pmp->pfs_names[i], M_HAMMER2); 862 pmp->pfs_names[i] = NULL; 863 } 864 if (rchain) { 865 hammer2_chain_drop(rchain); 866 /* focus hint */ 867 if (iroot->cluster.focus == rchain) 868 iroot->cluster.focus = NULL; 869 } 870 pmp->pfs_hmps[i] = NULL; 871 } 872 hammer2_mtx_unlock(&iroot->lock); 873 874 /* 875 * Cleanup trailing chains. Gaps may remain. 876 */ 877 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) { 878 if (pmp->pfs_hmps[i]) 879 break; 880 } 881 iroot->cluster.nchains = i + 1; 882 883 /* 884 * If the PMP has no elements remaining we can destroy it. 885 * (this will transition management threads from frozen->exit). 886 */ 887 if (iroot->cluster.nchains == 0) { 888 /* 889 * If this was the hmp's spmp, we need to clean 890 * a little more stuff out. 891 */ 892 if (hmp->spmp == pmp) { 893 hmp->spmp = NULL; 894 hmp->vchain.pmp = NULL; 895 hmp->fchain.pmp = NULL; 896 } 897 898 /* 899 * Free the pmp and restart the loop 900 */ 901 KKASSERT(TAILQ_EMPTY(&pmp->syncq)); 902 KKASSERT(TAILQ_EMPTY(&pmp->depq)); 903 hammer2_pfsfree(pmp); 904 goto again; 905 } 906 907 /* 908 * If elements still remain we need to set the REMASTER 909 * flag and unfreeze it. 910 */ 911 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 912 if (pmp->pfs_hmps[i] == NULL) 913 continue; 914 hammer2_thr_remaster(&pmp->sync_thrs[i]); 915 hammer2_thr_unfreeze(&pmp->sync_thrs[i]); 916 if (pmp->xop_groups) { 917 for (j = 0; j < hammer2_xopgroups; ++j) { 918 hammer2_thr_remaster( 919 &pmp->xop_groups[j].thrs[i]); 920 hammer2_thr_unfreeze( 921 &pmp->xop_groups[j].thrs[i]); 922 } 923 } 924 } 925 } 926 } 927 928 /* 929 * Mount or remount HAMMER2 fileystem from physical media 930 * 931 * mountroot 932 * mp mount point structure 933 * path NULL 934 * data <unused> 935 * cred <unused> 936 * 937 * mount 938 * mp mount point structure 939 * path path to mount point 940 * data pointer to argument structure in user space 941 * volume volume path (device@LABEL form) 942 * hflags user mount flags 943 * cred user credentials 944 * 945 * RETURNS: 0 Success 946 * !0 error number 947 */ 948 static 949 int 950 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data, 951 struct ucred *cred) 952 { 953 struct hammer2_mount_info info; 954 hammer2_pfs_t *pmp; 955 hammer2_pfs_t *spmp; 956 hammer2_dev_t *hmp; 957 hammer2_dev_t *force_local; 958 hammer2_key_t key_next; 959 hammer2_key_t key_dummy; 960 hammer2_key_t lhc; 961 struct vnode *devvp; 962 struct nlookupdata nd; 963 hammer2_chain_t *parent; 964 hammer2_chain_t *chain; 965 const hammer2_inode_data_t *ripdata; 966 hammer2_blockref_t bref; 967 struct file *fp; 968 char devstr[MNAMELEN]; 969 size_t size; 970 size_t done; 971 char *dev; 972 char *label; 973 int ronly = 1; 974 int error; 975 int i; 976 977 hmp = NULL; 978 pmp = NULL; 979 dev = NULL; 980 label = NULL; 981 devvp = NULL; 982 983 if (path == NULL) { 984 /* 985 * Root mount 986 */ 987 bzero(&info, sizeof(info)); 988 info.cluster_fd = -1; 989 ksnprintf(devstr, sizeof(devstr), "%s", 990 mp->mnt_stat.f_mntfromname); 991 kprintf("hammer2_mount: root '%s'\n", devstr); 992 done = strlen(devstr) + 1; 993 } else { 994 /* 995 * Non-root mount or updating a mount 996 */ 997 error = copyin(data, &info, sizeof(info)); 998 if (error) 999 return (error); 1000 1001 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done); 1002 if (error) 1003 return (error); 1004 kprintf("hammer2_mount: '%s'\n", devstr); 1005 } 1006 1007 /* 1008 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA 1009 * if no label specified, based on the partition id. Error out if no 1010 * label or device (with partition id) is specified. This is strictly 1011 * a convenience to match the default label created by newfs_hammer2, 1012 * our preference is that a label always be specified. 1013 * 1014 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command 1015 * that does not specify a device, as long as some H2 label 1016 * has already been mounted from that device. This makes 1017 * mounting snapshots a lot easier. 1018 */ 1019 dev = devstr; 1020 label = strchr(devstr, '@'); 1021 if (label && ((label + 1) - dev) > done) { 1022 kprintf("hammer2: mount: bad label %s/%zd\n", 1023 devstr, done); 1024 return (EINVAL); 1025 } 1026 if (label == NULL || label[1] == 0) { 1027 char slice; 1028 1029 if (label == NULL) 1030 label = devstr + strlen(devstr); 1031 else 1032 *label = '\0'; /* clean up trailing @ */ 1033 1034 slice = label[-1]; 1035 switch(slice) { 1036 case 'a': 1037 label = "BOOT"; 1038 break; 1039 case 'd': 1040 label = "ROOT"; 1041 break; 1042 default: 1043 label = "DATA"; 1044 break; 1045 } 1046 } else { 1047 *label = '\0'; 1048 label++; 1049 } 1050 1051 kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n", 1052 dev, label, (mp->mnt_flag & MNT_RDONLY)); 1053 1054 if (mp->mnt_flag & MNT_UPDATE) { 1055 /* 1056 * Update mount. Note that pmp->iroot->cluster is 1057 * an inode-embedded cluster and thus cannot be 1058 * directly locked. 1059 * 1060 * XXX HAMMER2 needs to implement NFS export via 1061 * mountctl. 1062 */ 1063 hammer2_cluster_t *cluster; 1064 1065 pmp = MPTOPMP(mp); 1066 pmp->hflags = info.hflags; 1067 cluster = &pmp->iroot->cluster; 1068 for (i = 0; i < cluster->nchains; ++i) { 1069 if (cluster->array[i].chain == NULL) 1070 continue; 1071 hmp = cluster->array[i].chain->hmp; 1072 devvp = hmp->devvp; 1073 error = hammer2_remount(hmp, mp, path, 1074 devvp, cred); 1075 if (error) 1076 break; 1077 } 1078 1079 return error; 1080 } 1081 1082 /* 1083 * HMP device mount 1084 * 1085 * If a path is specified and dev is not an empty string, lookup the 1086 * name and verify that it referes to a block device. 1087 * 1088 * If a path is specified and dev is an empty string we fall through 1089 * and locate the label in the hmp search. 1090 */ 1091 if (path && *dev != 0) { 1092 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW); 1093 if (error == 0) 1094 error = nlookup(&nd); 1095 if (error == 0) 1096 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp); 1097 nlookup_done(&nd); 1098 } else if (path == NULL) { 1099 /* root mount */ 1100 cdev_t cdev = kgetdiskbyname(dev); 1101 error = bdevvp(cdev, &devvp); 1102 if (error) 1103 kprintf("hammer2: cannot find '%s'\n", dev); 1104 } else { 1105 /* 1106 * We will locate the hmp using the label in the hmp loop. 1107 */ 1108 error = 0; 1109 } 1110 1111 /* 1112 * Make sure its a block device. Do not check to see if it is 1113 * already mounted until we determine that its a fresh H2 device. 1114 */ 1115 if (error == 0 && devvp) { 1116 vn_isdisk(devvp, &error); 1117 } 1118 1119 /* 1120 * Determine if the device has already been mounted. After this 1121 * check hmp will be non-NULL if we are doing the second or more 1122 * hammer2 mounts from the same device. 1123 */ 1124 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1125 if (devvp) { 1126 /* 1127 * Match the device. Due to the way devfs works, 1128 * we may not be able to directly match the vnode pointer, 1129 * so also check to see if the underlying device matches. 1130 */ 1131 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1132 if (hmp->devvp == devvp) 1133 break; 1134 if (devvp->v_rdev && 1135 hmp->devvp->v_rdev == devvp->v_rdev) { 1136 break; 1137 } 1138 } 1139 1140 /* 1141 * If no match this may be a fresh H2 mount, make sure 1142 * the device is not mounted on anything else. 1143 */ 1144 if (hmp == NULL) 1145 error = vfs_mountedon(devvp); 1146 } else if (error == 0) { 1147 /* 1148 * Match the label to a pmp already probed. 1149 */ 1150 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 1151 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 1152 if (pmp->pfs_names[i] && 1153 strcmp(pmp->pfs_names[i], label) == 0) { 1154 hmp = pmp->pfs_hmps[i]; 1155 break; 1156 } 1157 } 1158 if (hmp) 1159 break; 1160 } 1161 if (hmp == NULL) 1162 error = ENOENT; 1163 } 1164 1165 /* 1166 * Open the device if this isn't a secondary mount and construct 1167 * the H2 device mount (hmp). 1168 */ 1169 if (hmp == NULL) { 1170 hammer2_chain_t *schain; 1171 hammer2_xid_t xid; 1172 hammer2_xop_head_t xop; 1173 1174 if (error == 0 && vcount(devvp) > 0) { 1175 kprintf("Primary device already has references\n"); 1176 error = EBUSY; 1177 } 1178 1179 /* 1180 * Now open the device 1181 */ 1182 if (error == 0) { 1183 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0); 1184 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1185 error = vinvalbuf(devvp, V_SAVE, 0, 0); 1186 if (error == 0) { 1187 error = VOP_OPEN(devvp, 1188 (ronly ? FREAD : FREAD | FWRITE), 1189 FSCRED, NULL); 1190 } 1191 vn_unlock(devvp); 1192 } 1193 if (error && devvp) { 1194 vrele(devvp); 1195 devvp = NULL; 1196 } 1197 if (error) { 1198 lockmgr(&hammer2_mntlk, LK_RELEASE); 1199 return error; 1200 } 1201 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO); 1202 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev); 1203 hmp->ronly = ronly; 1204 hmp->devvp = devvp; 1205 hmp->hflags = info.hflags & HMNT2_DEVFLAGS; 1206 kmalloc_create(&hmp->mchain, "HAMMER2-chains"); 1207 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry); 1208 RB_INIT(&hmp->iotree); 1209 spin_init(&hmp->io_spin, "h2mount_io"); 1210 spin_init(&hmp->list_spin, "h2mount_list"); 1211 1212 lockinit(&hmp->vollk, "h2vol", 0, 0); 1213 lockinit(&hmp->bulklk, "h2bulk", 0, 0); 1214 lockinit(&hmp->bflock, "h2bflk", 0, 0); 1215 1216 /* 1217 * vchain setup. vchain.data is embedded. 1218 * vchain.refs is initialized and will never drop to 0. 1219 * 1220 * NOTE! voldata is not yet loaded. 1221 */ 1222 hmp->vchain.hmp = hmp; 1223 hmp->vchain.refs = 1; 1224 hmp->vchain.data = (void *)&hmp->voldata; 1225 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME; 1226 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1227 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1228 1229 hammer2_chain_core_init(&hmp->vchain); 1230 /* hmp->vchain.u.xxx is left NULL */ 1231 1232 /* 1233 * fchain setup. fchain.data is embedded. 1234 * fchain.refs is initialized and will never drop to 0. 1235 * 1236 * The data is not used but needs to be initialized to 1237 * pass assertion muster. We use this chain primarily 1238 * as a placeholder for the freemap's top-level RBTREE 1239 * so it does not interfere with the volume's topology 1240 * RBTREE. 1241 */ 1242 hmp->fchain.hmp = hmp; 1243 hmp->fchain.refs = 1; 1244 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset; 1245 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP; 1246 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1247 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1248 hmp->fchain.bref.methods = 1249 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) | 1250 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE); 1251 1252 hammer2_chain_core_init(&hmp->fchain); 1253 /* hmp->fchain.u.xxx is left NULL */ 1254 1255 /* 1256 * Install the volume header and initialize fields from 1257 * voldata. 1258 */ 1259 error = hammer2_install_volume_header(hmp); 1260 if (error) { 1261 hammer2_unmount_helper(mp, NULL, hmp); 1262 lockmgr(&hammer2_mntlk, LK_RELEASE); 1263 hammer2_vfs_unmount(mp, MNT_FORCE); 1264 return error; 1265 } 1266 1267 /* 1268 * Really important to get these right or the flush and 1269 * teardown code will get confused. 1270 */ 1271 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL); 1272 spmp = hmp->spmp; 1273 spmp->pfs_hmps[0] = hmp; 1274 1275 /* 1276 * Dummy-up vchain and fchain's modify_tid. mirror_tid 1277 * is inherited from the volume header. 1278 */ 1279 xid = 0; 1280 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1281 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid; 1282 hmp->vchain.pmp = spmp; 1283 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1284 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid; 1285 hmp->fchain.pmp = spmp; 1286 1287 /* 1288 * First locate the super-root inode, which is key 0 1289 * relative to the volume header's blockset. 1290 * 1291 * Then locate the root inode by scanning the directory keyspace 1292 * represented by the label. 1293 */ 1294 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 1295 schain = hammer2_chain_lookup(&parent, &key_dummy, 1296 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY, 1297 &error, 0); 1298 hammer2_chain_lookup_done(parent); 1299 if (schain == NULL) { 1300 kprintf("hammer2_mount: invalid super-root\n"); 1301 hammer2_unmount_helper(mp, NULL, hmp); 1302 lockmgr(&hammer2_mntlk, LK_RELEASE); 1303 hammer2_vfs_unmount(mp, MNT_FORCE); 1304 return EINVAL; 1305 } 1306 if (schain->error) { 1307 kprintf("hammer2_mount: error %s reading super-root\n", 1308 hammer2_error_str(schain->error)); 1309 hammer2_chain_unlock(schain); 1310 hammer2_chain_drop(schain); 1311 schain = NULL; 1312 hammer2_unmount_helper(mp, NULL, hmp); 1313 lockmgr(&hammer2_mntlk, LK_RELEASE); 1314 hammer2_vfs_unmount(mp, MNT_FORCE); 1315 return EINVAL; 1316 } 1317 1318 /* 1319 * The super-root always uses an inode_tid of 1 when 1320 * creating PFSs. 1321 */ 1322 spmp->inode_tid = 1; 1323 spmp->modify_tid = schain->bref.modify_tid + 1; 1324 1325 /* 1326 * Sanity-check schain's pmp and finish initialization. 1327 * Any chain belonging to the super-root topology should 1328 * have a NULL pmp (not even set to spmp). 1329 */ 1330 ripdata = &hammer2_chain_rdata(schain)->ipdata; 1331 KKASSERT(schain->pmp == NULL); 1332 spmp->pfs_clid = ripdata->meta.pfs_clid; 1333 1334 /* 1335 * Replace the dummy spmp->iroot with a real one. It's 1336 * easier to just do a wholesale replacement than to try 1337 * to update the chain and fixup the iroot fields. 1338 * 1339 * The returned inode is locked with the supplied cluster. 1340 */ 1341 hammer2_dummy_xop_from_chain(&xop, schain); 1342 hammer2_inode_drop(spmp->iroot); 1343 spmp->iroot = NULL; 1344 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1); 1345 spmp->spmp_hmp = hmp; 1346 spmp->pfs_types[0] = ripdata->meta.pfs_type; 1347 spmp->pfs_hmps[0] = hmp; 1348 hammer2_inode_ref(spmp->iroot); 1349 hammer2_inode_unlock(spmp->iroot); 1350 hammer2_cluster_unlock(&xop.cluster); 1351 hammer2_chain_drop(schain); 1352 /* do not call hammer2_cluster_drop() on an embedded cluster */ 1353 schain = NULL; /* now invalid */ 1354 /* leave spmp->iroot with one ref */ 1355 1356 if ((mp->mnt_flag & MNT_RDONLY) == 0) { 1357 error = hammer2_recovery(hmp); 1358 if (error == 0) 1359 error |= hammer2_fixup_pfses(hmp); 1360 /* XXX do something with error */ 1361 } 1362 hammer2_update_pmps(hmp); 1363 hammer2_iocom_init(hmp); 1364 hammer2_bulkfree_init(hmp); 1365 1366 /* 1367 * Ref the cluster management messaging descriptor. The mount 1368 * program deals with the other end of the communications pipe. 1369 * 1370 * Root mounts typically do not supply one. 1371 */ 1372 if (info.cluster_fd >= 0) { 1373 fp = holdfp(curthread, info.cluster_fd, -1); 1374 if (fp) { 1375 hammer2_cluster_reconnect(hmp, fp); 1376 } else { 1377 kprintf("hammer2_mount: bad cluster_fd!\n"); 1378 } 1379 } 1380 } else { 1381 spmp = hmp->spmp; 1382 if (info.hflags & HMNT2_DEVFLAGS) { 1383 kprintf("hammer2: Warning: mount flags pertaining " 1384 "to the whole device may only be specified " 1385 "on the first mount of the device: %08x\n", 1386 info.hflags & HMNT2_DEVFLAGS); 1387 } 1388 } 1389 1390 /* 1391 * Force local mount (disassociate all PFSs from their clusters). 1392 * Used primarily for debugging. 1393 */ 1394 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1395 1396 /* 1397 * Lookup the mount point under the media-localized super-root. 1398 * Scanning hammer2_pfslist doesn't help us because it represents 1399 * PFS cluster ids which can aggregate several named PFSs together. 1400 * 1401 * cluster->pmp will incorrectly point to spmp and must be fixed 1402 * up later on. 1403 */ 1404 hammer2_inode_lock(spmp->iroot, 0); 1405 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1406 lhc = hammer2_dirhash(label, strlen(label)); 1407 chain = hammer2_chain_lookup(&parent, &key_next, 1408 lhc, lhc + HAMMER2_DIRHASH_LOMASK, 1409 &error, 0); 1410 while (chain) { 1411 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && 1412 strcmp(label, chain->data->ipdata.filename) == 0) { 1413 break; 1414 } 1415 chain = hammer2_chain_next(&parent, chain, &key_next, 1416 key_next, 1417 lhc + HAMMER2_DIRHASH_LOMASK, 1418 &error, 0); 1419 } 1420 if (parent) { 1421 hammer2_chain_unlock(parent); 1422 hammer2_chain_drop(parent); 1423 } 1424 hammer2_inode_unlock(spmp->iroot); 1425 1426 /* 1427 * PFS could not be found? 1428 */ 1429 if (chain == NULL) { 1430 if (error) 1431 kprintf("hammer2_mount: PFS label I/O error\n"); 1432 else 1433 kprintf("hammer2_mount: PFS label not found\n"); 1434 hammer2_unmount_helper(mp, NULL, hmp); 1435 lockmgr(&hammer2_mntlk, LK_RELEASE); 1436 hammer2_vfs_unmount(mp, MNT_FORCE); 1437 1438 return EINVAL; 1439 } 1440 1441 /* 1442 * Acquire the pmp structure (it should have already been allocated 1443 * via hammer2_update_pmps() so do not pass cluster in to add to 1444 * available chains). 1445 * 1446 * Check if the cluster has already been mounted. A cluster can 1447 * only be mounted once, use null mounts to mount additional copies. 1448 */ 1449 if (chain->error) { 1450 kprintf("hammer2_mount: PFS label I/O error\n"); 1451 } else { 1452 ripdata = &chain->data->ipdata; 1453 bref = chain->bref; 1454 pmp = hammer2_pfsalloc(NULL, ripdata, 1455 bref.modify_tid, force_local); 1456 } 1457 hammer2_chain_unlock(chain); 1458 hammer2_chain_drop(chain); 1459 1460 /* 1461 * Finish the mount 1462 */ 1463 kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp); 1464 1465 if (pmp->mp) { 1466 kprintf("hammer2_mount: PFS already mounted!\n"); 1467 hammer2_unmount_helper(mp, NULL, hmp); 1468 lockmgr(&hammer2_mntlk, LK_RELEASE); 1469 hammer2_vfs_unmount(mp, MNT_FORCE); 1470 1471 return EBUSY; 1472 } 1473 1474 pmp->hflags = info.hflags; 1475 mp->mnt_flag |= MNT_LOCAL; 1476 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */ 1477 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */ 1478 1479 /* 1480 * required mount structure initializations 1481 */ 1482 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE; 1483 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE; 1484 1485 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE; 1486 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE; 1487 1488 /* 1489 * Optional fields 1490 */ 1491 mp->mnt_iosize_max = MAXPHYS; 1492 1493 /* 1494 * Connect up mount pointers. 1495 */ 1496 hammer2_mount_helper(mp, pmp); 1497 1498 lockmgr(&hammer2_mntlk, LK_RELEASE); 1499 1500 /* 1501 * Finish setup 1502 */ 1503 vfs_getnewfsid(mp); 1504 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops); 1505 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops); 1506 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops); 1507 1508 if (path) { 1509 copyinstr(info.volume, mp->mnt_stat.f_mntfromname, 1510 MNAMELEN - 1, &size); 1511 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size); 1512 } /* else root mount, already in there */ 1513 1514 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname)); 1515 if (path) { 1516 copyinstr(path, mp->mnt_stat.f_mntonname, 1517 sizeof(mp->mnt_stat.f_mntonname) - 1, 1518 &size); 1519 } else { 1520 /* root mount */ 1521 mp->mnt_stat.f_mntonname[0] = '/'; 1522 } 1523 1524 /* 1525 * Initial statfs to prime mnt_stat. 1526 */ 1527 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred); 1528 1529 return 0; 1530 } 1531 1532 /* 1533 * Scan PFSs under the super-root and create hammer2_pfs structures. 1534 */ 1535 static 1536 void 1537 hammer2_update_pmps(hammer2_dev_t *hmp) 1538 { 1539 const hammer2_inode_data_t *ripdata; 1540 hammer2_chain_t *parent; 1541 hammer2_chain_t *chain; 1542 hammer2_blockref_t bref; 1543 hammer2_dev_t *force_local; 1544 hammer2_pfs_t *spmp; 1545 hammer2_pfs_t *pmp; 1546 hammer2_key_t key_next; 1547 int error; 1548 1549 /* 1550 * Force local mount (disassociate all PFSs from their clusters). 1551 * Used primarily for debugging. 1552 */ 1553 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1554 1555 /* 1556 * Lookup mount point under the media-localized super-root. 1557 * 1558 * cluster->pmp will incorrectly point to spmp and must be fixed 1559 * up later on. 1560 */ 1561 spmp = hmp->spmp; 1562 hammer2_inode_lock(spmp->iroot, 0); 1563 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1564 chain = hammer2_chain_lookup(&parent, &key_next, 1565 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 1566 &error, 0); 1567 while (chain) { 1568 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) 1569 continue; 1570 if (chain->error) { 1571 kprintf("I/O error scanning PFS labels\n"); 1572 } else { 1573 ripdata = &chain->data->ipdata; 1574 bref = chain->bref; 1575 1576 pmp = hammer2_pfsalloc(chain, ripdata, 1577 bref.modify_tid, force_local); 1578 } 1579 chain = hammer2_chain_next(&parent, chain, &key_next, 1580 key_next, HAMMER2_KEY_MAX, 1581 &error, 0); 1582 } 1583 if (parent) { 1584 hammer2_chain_unlock(parent); 1585 hammer2_chain_drop(parent); 1586 } 1587 hammer2_inode_unlock(spmp->iroot); 1588 } 1589 1590 static 1591 int 1592 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused, 1593 struct vnode *devvp, struct ucred *cred) 1594 { 1595 int error; 1596 1597 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) { 1598 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1599 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL); 1600 vn_unlock(devvp); 1601 error = hammer2_recovery(hmp); 1602 if (error == 0) 1603 error |= hammer2_fixup_pfses(hmp); 1604 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1605 if (error == 0) { 1606 VOP_CLOSE(devvp, FREAD, NULL); 1607 hmp->ronly = 0; 1608 } else { 1609 VOP_CLOSE(devvp, FREAD | FWRITE, NULL); 1610 } 1611 vn_unlock(devvp); 1612 } else { 1613 error = 0; 1614 } 1615 return error; 1616 } 1617 1618 static 1619 int 1620 hammer2_vfs_unmount(struct mount *mp, int mntflags) 1621 { 1622 hammer2_pfs_t *pmp; 1623 int flags; 1624 int error = 0; 1625 1626 pmp = MPTOPMP(mp); 1627 1628 if (pmp == NULL) 1629 return(0); 1630 1631 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1632 1633 /* 1634 * If mount initialization proceeded far enough we must flush 1635 * its vnodes and sync the underlying mount points. Three syncs 1636 * are required to fully flush the filesystem (freemap updates lag 1637 * by one flush, and one extra for safety). 1638 */ 1639 if (mntflags & MNT_FORCE) 1640 flags = FORCECLOSE; 1641 else 1642 flags = 0; 1643 if (pmp->iroot) { 1644 error = vflush(mp, 0, flags); 1645 if (error) 1646 goto failed; 1647 hammer2_vfs_sync(mp, MNT_WAIT); 1648 hammer2_vfs_sync(mp, MNT_WAIT); 1649 hammer2_vfs_sync(mp, MNT_WAIT); 1650 } 1651 1652 /* 1653 * Cleanup the frontend support XOPS threads 1654 */ 1655 hammer2_xop_helper_cleanup(pmp); 1656 1657 if (pmp->mp) 1658 hammer2_unmount_helper(mp, pmp, NULL); 1659 1660 error = 0; 1661 failed: 1662 lockmgr(&hammer2_mntlk, LK_RELEASE); 1663 1664 return (error); 1665 } 1666 1667 /* 1668 * Mount helper, hook the system mount into our PFS. 1669 * The mount lock is held. 1670 * 1671 * We must bump the mount_count on related devices for any 1672 * mounted PFSs. 1673 */ 1674 static 1675 void 1676 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp) 1677 { 1678 hammer2_cluster_t *cluster; 1679 hammer2_chain_t *rchain; 1680 int i; 1681 1682 mp->mnt_data = (qaddr_t)pmp; 1683 pmp->mp = mp; 1684 1685 /* 1686 * After pmp->mp is set we have to adjust hmp->mount_count. 1687 */ 1688 cluster = &pmp->iroot->cluster; 1689 for (i = 0; i < cluster->nchains; ++i) { 1690 rchain = cluster->array[i].chain; 1691 if (rchain == NULL) 1692 continue; 1693 ++rchain->hmp->mount_count; 1694 } 1695 1696 /* 1697 * Create missing Xop threads 1698 */ 1699 hammer2_xop_helper_create(pmp); 1700 } 1701 1702 /* 1703 * Mount helper, unhook the system mount from our PFS. 1704 * The mount lock is held. 1705 * 1706 * If hmp is supplied a mount responsible for being the first to open 1707 * the block device failed and the block device and all PFSs using the 1708 * block device must be cleaned up. 1709 * 1710 * If pmp is supplied multiple devices might be backing the PFS and each 1711 * must be disconnected. This might not be the last PFS using some of the 1712 * underlying devices. Also, we have to adjust our hmp->mount_count 1713 * accounting for the devices backing the pmp which is now undergoing an 1714 * unmount. 1715 */ 1716 static 1717 void 1718 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp) 1719 { 1720 hammer2_cluster_t *cluster; 1721 hammer2_chain_t *rchain; 1722 struct vnode *devvp; 1723 int dumpcnt; 1724 int ronly; 1725 int i; 1726 1727 /* 1728 * If no device supplied this is a high-level unmount and we have to 1729 * to disconnect the mount, adjust mount_count, and locate devices 1730 * that might now have no mounts. 1731 */ 1732 if (pmp) { 1733 KKASSERT(hmp == NULL); 1734 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp); 1735 pmp->mp = NULL; 1736 mp->mnt_data = NULL; 1737 1738 /* 1739 * After pmp->mp is cleared we have to account for 1740 * mount_count. 1741 */ 1742 cluster = &pmp->iroot->cluster; 1743 for (i = 0; i < cluster->nchains; ++i) { 1744 rchain = cluster->array[i].chain; 1745 if (rchain == NULL) 1746 continue; 1747 --rchain->hmp->mount_count; 1748 /* scrapping hmp now may invalidate the pmp */ 1749 } 1750 again: 1751 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1752 if (hmp->mount_count == 0) { 1753 hammer2_unmount_helper(NULL, NULL, hmp); 1754 goto again; 1755 } 1756 } 1757 return; 1758 } 1759 1760 /* 1761 * Try to terminate the block device. We can't terminate it if 1762 * there are still PFSs referencing it. 1763 */ 1764 if (hmp->mount_count) 1765 return; 1766 1767 /* 1768 * Decomission the network before we start messing with the 1769 * device and PFS. 1770 */ 1771 hammer2_iocom_uninit(hmp); 1772 1773 hammer2_bulkfree_uninit(hmp); 1774 hammer2_pfsfree_scan(hmp, 0); 1775 #if 0 1776 hammer2_dev_exlock(hmp); /* XXX order */ 1777 #endif 1778 1779 /* 1780 * Cycle the volume data lock as a safety (probably not needed any 1781 * more). To ensure everything is out we need to flush at least 1782 * three times. (1) The running of the sideq can dirty the 1783 * filesystem, (2) A normal flush can dirty the freemap, and 1784 * (3) ensure that the freemap is fully synchronized. 1785 * 1786 * The next mount's recovery scan can clean everything up but we want 1787 * to leave the filesystem in a 100% clean state on a normal unmount. 1788 */ 1789 #if 0 1790 hammer2_voldata_lock(hmp); 1791 hammer2_voldata_unlock(hmp); 1792 #endif 1793 1794 /* 1795 * Flush whatever is left. Unmounted but modified PFS's might still 1796 * have some dirty chains on them. 1797 */ 1798 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS); 1799 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS); 1800 1801 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) { 1802 hammer2_voldata_modify(hmp); 1803 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP | 1804 HAMMER2_FLUSH_ALL); 1805 } 1806 hammer2_chain_unlock(&hmp->fchain); 1807 1808 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) { 1809 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP | 1810 HAMMER2_FLUSH_ALL); 1811 } 1812 hammer2_chain_unlock(&hmp->vchain); 1813 1814 if ((hmp->vchain.flags | hmp->fchain.flags) & 1815 HAMMER2_CHAIN_FLUSH_MASK) { 1816 kprintf("hammer2_unmount: chains left over " 1817 "after final sync\n"); 1818 kprintf(" vchain %08x\n", hmp->vchain.flags); 1819 kprintf(" fchain %08x\n", hmp->fchain.flags); 1820 1821 if (hammer2_debug & 0x0010) 1822 Debugger("entered debugger"); 1823 } 1824 1825 hammer2_pfsfree_scan(hmp, 1); 1826 1827 KKASSERT(hmp->spmp == NULL); 1828 1829 /* 1830 * Finish up with the device vnode 1831 */ 1832 if ((devvp = hmp->devvp) != NULL) { 1833 ronly = hmp->ronly; 1834 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1835 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n", 1836 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree), 1837 ronly); 1838 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0); 1839 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n", 1840 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree)); 1841 hmp->devvp = NULL; 1842 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL); 1843 vn_unlock(devvp); 1844 vrele(devvp); 1845 devvp = NULL; 1846 } 1847 1848 /* 1849 * Clear vchain/fchain flags that might prevent final cleanup 1850 * of these chains. 1851 */ 1852 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) { 1853 atomic_add_long(&hammer2_count_modified_chains, -1); 1854 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 1855 hammer2_pfs_memory_wakeup(hmp->vchain.pmp); 1856 } 1857 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) { 1858 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE); 1859 } 1860 1861 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) { 1862 atomic_add_long(&hammer2_count_modified_chains, -1); 1863 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED); 1864 hammer2_pfs_memory_wakeup(hmp->fchain.pmp); 1865 } 1866 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) { 1867 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE); 1868 } 1869 1870 /* 1871 * Final drop of embedded freemap root chain to 1872 * clean up fchain.core (fchain structure is not 1873 * flagged ALLOCATED so it is cleaned out and then 1874 * left to rot). 1875 */ 1876 hammer2_chain_drop(&hmp->fchain); 1877 1878 /* 1879 * Final drop of embedded volume root chain to clean 1880 * up vchain.core (vchain structure is not flagged 1881 * ALLOCATED so it is cleaned out and then left to 1882 * rot). 1883 */ 1884 dumpcnt = 50; 1885 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1); 1886 dumpcnt = 50; 1887 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1); 1888 #if 0 1889 hammer2_dev_unlock(hmp); 1890 #endif 1891 hammer2_chain_drop(&hmp->vchain); 1892 1893 hammer2_io_cleanup(hmp, &hmp->iotree); 1894 if (hmp->iofree_count) { 1895 kprintf("io_cleanup: %d I/O's left hanging\n", 1896 hmp->iofree_count); 1897 } 1898 1899 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry); 1900 kmalloc_destroy(&hmp->mchain); 1901 kfree(hmp, M_HAMMER2); 1902 } 1903 1904 int 1905 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp, 1906 ino_t ino, struct vnode **vpp) 1907 { 1908 hammer2_xop_lookup_t *xop; 1909 hammer2_pfs_t *pmp; 1910 hammer2_inode_t *ip; 1911 hammer2_tid_t inum; 1912 int error; 1913 1914 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK; 1915 1916 error = 0; 1917 pmp = MPTOPMP(mp); 1918 1919 /* 1920 * Easy if we already have it cached 1921 */ 1922 ip = hammer2_inode_lookup(pmp, inum); 1923 if (ip) { 1924 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED); 1925 *vpp = hammer2_igetv(ip, &error); 1926 hammer2_inode_unlock(ip); 1927 hammer2_inode_drop(ip); /* from lookup */ 1928 1929 return error; 1930 } 1931 1932 /* 1933 * Otherwise we have to find the inode 1934 */ 1935 xop = hammer2_xop_alloc(pmp->iroot, 0); 1936 xop->lhc = inum; 1937 hammer2_xop_start(&xop->head, &hammer2_lookup_desc); 1938 error = hammer2_xop_collect(&xop->head, 0); 1939 1940 if (error == 0) 1941 ip = hammer2_inode_get(pmp, &xop->head, -1, -1); 1942 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1943 1944 if (ip) { 1945 *vpp = hammer2_igetv(ip, &error); 1946 hammer2_inode_unlock(ip); 1947 } else { 1948 *vpp = NULL; 1949 error = ENOENT; 1950 } 1951 return (error); 1952 } 1953 1954 static 1955 int 1956 hammer2_vfs_root(struct mount *mp, struct vnode **vpp) 1957 { 1958 hammer2_pfs_t *pmp; 1959 struct vnode *vp; 1960 int error; 1961 1962 pmp = MPTOPMP(mp); 1963 if (pmp->iroot == NULL) { 1964 kprintf("hammer2 (%s): no root inode\n", 1965 mp->mnt_stat.f_mntfromname); 1966 *vpp = NULL; 1967 return EINVAL; 1968 } 1969 1970 error = 0; 1971 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1972 1973 while (pmp->inode_tid == 0) { 1974 hammer2_xop_ipcluster_t *xop; 1975 const hammer2_inode_meta_t *meta; 1976 1977 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING); 1978 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc); 1979 error = hammer2_xop_collect(&xop->head, 0); 1980 1981 if (error == 0) { 1982 meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta; 1983 pmp->iroot->meta = *meta; 1984 pmp->inode_tid = meta->pfs_inum + 1; 1985 hammer2_xop_pdata(&xop->head); 1986 /* meta invalid */ 1987 1988 if (pmp->inode_tid < HAMMER2_INODE_START) 1989 pmp->inode_tid = HAMMER2_INODE_START; 1990 pmp->modify_tid = 1991 xop->head.cluster.focus->bref.modify_tid + 1; 1992 #if 0 1993 kprintf("PFS: Starting inode %jd\n", 1994 (intmax_t)pmp->inode_tid); 1995 kprintf("PMP focus good set nextino=%ld mod=%016jx\n", 1996 pmp->inode_tid, pmp->modify_tid); 1997 #endif 1998 wakeup(&pmp->iroot); 1999 2000 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 2001 2002 /* 2003 * Prime the mount info. 2004 */ 2005 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL); 2006 break; 2007 } 2008 2009 /* 2010 * Loop, try again 2011 */ 2012 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 2013 hammer2_inode_unlock(pmp->iroot); 2014 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz); 2015 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 2016 if (error == EINTR) 2017 break; 2018 } 2019 2020 if (error) { 2021 hammer2_inode_unlock(pmp->iroot); 2022 *vpp = NULL; 2023 } else { 2024 vp = hammer2_igetv(pmp->iroot, &error); 2025 hammer2_inode_unlock(pmp->iroot); 2026 *vpp = vp; 2027 } 2028 2029 return (error); 2030 } 2031 2032 /* 2033 * Filesystem status 2034 * 2035 * XXX incorporate ipdata->meta.inode_quota and data_quota 2036 */ 2037 static 2038 int 2039 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred) 2040 { 2041 hammer2_pfs_t *pmp; 2042 hammer2_dev_t *hmp; 2043 hammer2_blockref_t bref; 2044 struct statfs tmp; 2045 int i; 2046 2047 /* 2048 * NOTE: iroot might not have validated the cluster yet. 2049 */ 2050 pmp = MPTOPMP(mp); 2051 2052 bzero(&tmp, sizeof(tmp)); 2053 2054 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 2055 hmp = pmp->pfs_hmps[i]; 2056 if (hmp == NULL) 2057 continue; 2058 if (pmp->iroot->cluster.array[i].chain) 2059 bref = pmp->iroot->cluster.array[i].chain->bref; 2060 else 2061 bzero(&bref, sizeof(bref)); 2062 2063 tmp.f_files = bref.embed.stats.inode_count; 2064 tmp.f_ffree = 0; 2065 tmp.f_blocks = hmp->voldata.allocator_size / 2066 mp->mnt_vstat.f_bsize; 2067 tmp.f_bfree = hmp->voldata.allocator_free / 2068 mp->mnt_vstat.f_bsize; 2069 tmp.f_bavail = tmp.f_bfree; 2070 2071 if (cred && cred->cr_uid != 0) { 2072 uint64_t adj; 2073 2074 /* 5% */ 2075 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize; 2076 tmp.f_blocks -= adj; 2077 tmp.f_bfree -= adj; 2078 tmp.f_bavail -= adj; 2079 } 2080 2081 mp->mnt_stat.f_blocks = tmp.f_blocks; 2082 mp->mnt_stat.f_bfree = tmp.f_bfree; 2083 mp->mnt_stat.f_bavail = tmp.f_bavail; 2084 mp->mnt_stat.f_files = tmp.f_files; 2085 mp->mnt_stat.f_ffree = tmp.f_ffree; 2086 2087 *sbp = mp->mnt_stat; 2088 } 2089 return (0); 2090 } 2091 2092 static 2093 int 2094 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred) 2095 { 2096 hammer2_pfs_t *pmp; 2097 hammer2_dev_t *hmp; 2098 hammer2_blockref_t bref; 2099 struct statvfs tmp; 2100 int i; 2101 2102 /* 2103 * NOTE: iroot might not have validated the cluster yet. 2104 */ 2105 pmp = MPTOPMP(mp); 2106 bzero(&tmp, sizeof(tmp)); 2107 2108 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 2109 hmp = pmp->pfs_hmps[i]; 2110 if (hmp == NULL) 2111 continue; 2112 if (pmp->iroot->cluster.array[i].chain) 2113 bref = pmp->iroot->cluster.array[i].chain->bref; 2114 else 2115 bzero(&bref, sizeof(bref)); 2116 2117 tmp.f_files = bref.embed.stats.inode_count; 2118 tmp.f_ffree = 0; 2119 tmp.f_blocks = hmp->voldata.allocator_size / 2120 mp->mnt_vstat.f_bsize; 2121 tmp.f_bfree = hmp->voldata.allocator_free / 2122 mp->mnt_vstat.f_bsize; 2123 tmp.f_bavail = tmp.f_bfree; 2124 2125 if (cred && cred->cr_uid != 0) { 2126 uint64_t adj; 2127 2128 /* 5% */ 2129 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize; 2130 tmp.f_blocks -= adj; 2131 tmp.f_bfree -= adj; 2132 tmp.f_bavail -= adj; 2133 } 2134 2135 mp->mnt_vstat.f_blocks = tmp.f_blocks; 2136 mp->mnt_vstat.f_bfree = tmp.f_bfree; 2137 mp->mnt_vstat.f_bavail = tmp.f_bavail; 2138 mp->mnt_vstat.f_files = tmp.f_files; 2139 mp->mnt_vstat.f_ffree = tmp.f_ffree; 2140 2141 *sbp = mp->mnt_vstat; 2142 } 2143 return (0); 2144 } 2145 2146 /* 2147 * Mount-time recovery (RW mounts) 2148 * 2149 * Updates to the free block table are allowed to lag flushes by one 2150 * transaction. In case of a crash, then on a fresh mount we must do an 2151 * incremental scan of the last committed transaction id and make sure that 2152 * all related blocks have been marked allocated. 2153 * 2154 * The super-root topology and each PFS has its own transaction id domain, 2155 * so we must track PFS boundary transitions. 2156 */ 2157 struct hammer2_recovery_elm { 2158 TAILQ_ENTRY(hammer2_recovery_elm) entry; 2159 hammer2_chain_t *chain; 2160 hammer2_tid_t sync_tid; 2161 }; 2162 2163 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm); 2164 2165 struct hammer2_recovery_info { 2166 struct hammer2_recovery_list list; 2167 hammer2_tid_t mtid; 2168 int depth; 2169 }; 2170 2171 static int hammer2_recovery_scan(hammer2_dev_t *hmp, 2172 hammer2_chain_t *parent, 2173 struct hammer2_recovery_info *info, 2174 hammer2_tid_t sync_tid); 2175 2176 #define HAMMER2_RECOVERY_MAXDEPTH 10 2177 2178 static 2179 int 2180 hammer2_recovery(hammer2_dev_t *hmp) 2181 { 2182 struct hammer2_recovery_info info; 2183 struct hammer2_recovery_elm *elm; 2184 hammer2_chain_t *parent; 2185 hammer2_tid_t sync_tid; 2186 hammer2_tid_t mirror_tid; 2187 int error; 2188 2189 hammer2_trans_init(hmp->spmp, 0); 2190 2191 sync_tid = hmp->voldata.freemap_tid; 2192 mirror_tid = hmp->voldata.mirror_tid; 2193 2194 kprintf("hammer2 mount \"%s\": ", hmp->devrepname); 2195 if (sync_tid >= mirror_tid) { 2196 kprintf(" no recovery needed\n"); 2197 } else { 2198 kprintf(" freemap recovery %016jx-%016jx\n", 2199 sync_tid + 1, mirror_tid); 2200 } 2201 2202 TAILQ_INIT(&info.list); 2203 info.depth = 0; 2204 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 2205 error = hammer2_recovery_scan(hmp, parent, &info, sync_tid); 2206 hammer2_chain_lookup_done(parent); 2207 2208 while ((elm = TAILQ_FIRST(&info.list)) != NULL) { 2209 TAILQ_REMOVE(&info.list, elm, entry); 2210 parent = elm->chain; 2211 sync_tid = elm->sync_tid; 2212 kfree(elm, M_HAMMER2); 2213 2214 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2215 error |= hammer2_recovery_scan(hmp, parent, &info, 2216 hmp->voldata.freemap_tid); 2217 hammer2_chain_unlock(parent); 2218 hammer2_chain_drop(parent); /* drop elm->chain ref */ 2219 } 2220 2221 hammer2_trans_done(hmp->spmp, 0); 2222 2223 return error; 2224 } 2225 2226 static 2227 int 2228 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent, 2229 struct hammer2_recovery_info *info, 2230 hammer2_tid_t sync_tid) 2231 { 2232 const hammer2_inode_data_t *ripdata; 2233 hammer2_chain_t *chain; 2234 hammer2_blockref_t bref; 2235 int tmp_error; 2236 int rup_error; 2237 int error; 2238 int first; 2239 2240 /* 2241 * Adjust freemap to ensure that the block(s) are marked allocated. 2242 */ 2243 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) { 2244 hammer2_freemap_adjust(hmp, &parent->bref, 2245 HAMMER2_FREEMAP_DORECOVER); 2246 } 2247 2248 /* 2249 * Check type for recursive scan 2250 */ 2251 switch(parent->bref.type) { 2252 case HAMMER2_BREF_TYPE_VOLUME: 2253 /* data already instantiated */ 2254 break; 2255 case HAMMER2_BREF_TYPE_INODE: 2256 /* 2257 * Must instantiate data for DIRECTDATA test and also 2258 * for recursion. 2259 */ 2260 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2261 ripdata = &hammer2_chain_rdata(parent)->ipdata; 2262 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) { 2263 /* not applicable to recovery scan */ 2264 hammer2_chain_unlock(parent); 2265 return 0; 2266 } 2267 hammer2_chain_unlock(parent); 2268 break; 2269 case HAMMER2_BREF_TYPE_INDIRECT: 2270 /* 2271 * Must instantiate data for recursion 2272 */ 2273 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2274 hammer2_chain_unlock(parent); 2275 break; 2276 case HAMMER2_BREF_TYPE_DIRENT: 2277 case HAMMER2_BREF_TYPE_DATA: 2278 case HAMMER2_BREF_TYPE_FREEMAP: 2279 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2280 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 2281 /* not applicable to recovery scan */ 2282 return 0; 2283 break; 2284 default: 2285 return HAMMER2_ERROR_BADBREF; 2286 } 2287 2288 /* 2289 * Defer operation if depth limit reached or if we are crossing a 2290 * PFS boundary. 2291 */ 2292 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) { 2293 struct hammer2_recovery_elm *elm; 2294 2295 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK); 2296 elm->chain = parent; 2297 elm->sync_tid = sync_tid; 2298 hammer2_chain_ref(parent); 2299 TAILQ_INSERT_TAIL(&info->list, elm, entry); 2300 /* unlocked by caller */ 2301 2302 return(0); 2303 } 2304 2305 2306 /* 2307 * Recursive scan of the last flushed transaction only. We are 2308 * doing this without pmp assignments so don't leave the chains 2309 * hanging around after we are done with them. 2310 * 2311 * error Cumulative error this level only 2312 * rup_error Cumulative error for recursion 2313 * tmp_error Specific non-cumulative recursion error 2314 */ 2315 chain = NULL; 2316 first = 1; 2317 rup_error = 0; 2318 error = 0; 2319 2320 for (;;) { 2321 error |= hammer2_chain_scan(parent, &chain, &bref, 2322 &first, 2323 HAMMER2_LOOKUP_NODATA); 2324 2325 /* 2326 * Problem during scan or EOF 2327 */ 2328 if (error) 2329 break; 2330 2331 /* 2332 * If this is a leaf 2333 */ 2334 if (chain == NULL) { 2335 if (bref.mirror_tid > sync_tid) { 2336 hammer2_freemap_adjust(hmp, &bref, 2337 HAMMER2_FREEMAP_DORECOVER); 2338 } 2339 continue; 2340 } 2341 2342 /* 2343 * This may or may not be a recursive node. 2344 */ 2345 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 2346 if (bref.mirror_tid > sync_tid) { 2347 ++info->depth; 2348 tmp_error = hammer2_recovery_scan(hmp, chain, 2349 info, sync_tid); 2350 --info->depth; 2351 } else { 2352 tmp_error = 0; 2353 } 2354 2355 /* 2356 * Flush the recovery at the PFS boundary to stage it for 2357 * the final flush of the super-root topology. 2358 */ 2359 if (tmp_error == 0 && 2360 (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) && 2361 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) { 2362 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 2363 HAMMER2_FLUSH_ALL); 2364 } 2365 rup_error |= tmp_error; 2366 } 2367 return ((error | rup_error) & ~HAMMER2_ERROR_EOF); 2368 } 2369 2370 /* 2371 * This fixes up an error introduced in earlier H2 implementations where 2372 * moving a PFS inode into an indirect block wound up causing the 2373 * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared. 2374 */ 2375 static 2376 int 2377 hammer2_fixup_pfses(hammer2_dev_t *hmp) 2378 { 2379 const hammer2_inode_data_t *ripdata; 2380 hammer2_chain_t *parent; 2381 hammer2_chain_t *chain; 2382 hammer2_key_t key_next; 2383 hammer2_pfs_t *spmp; 2384 int error; 2385 2386 error = 0; 2387 2388 /* 2389 * Lookup mount point under the media-localized super-root. 2390 * 2391 * cluster->pmp will incorrectly point to spmp and must be fixed 2392 * up later on. 2393 */ 2394 spmp = hmp->spmp; 2395 hammer2_inode_lock(spmp->iroot, 0); 2396 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 2397 chain = hammer2_chain_lookup(&parent, &key_next, 2398 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 2399 &error, 0); 2400 while (chain) { 2401 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) 2402 continue; 2403 if (chain->error) { 2404 kprintf("I/O error scanning PFS labels\n"); 2405 error |= chain->error; 2406 } else if ((chain->bref.flags & 2407 HAMMER2_BREF_FLAG_PFSROOT) == 0) { 2408 int error2; 2409 2410 ripdata = &chain->data->ipdata; 2411 hammer2_trans_init(hmp->spmp, 0); 2412 error2 = hammer2_chain_modify(chain, 2413 chain->bref.modify_tid, 2414 0, 0); 2415 if (error2 == 0) { 2416 kprintf("hammer2: Correct mis-flagged PFS %s\n", 2417 ripdata->filename); 2418 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT; 2419 } else { 2420 error |= error2; 2421 } 2422 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 2423 HAMMER2_FLUSH_ALL); 2424 hammer2_trans_done(hmp->spmp, 0); 2425 } 2426 chain = hammer2_chain_next(&parent, chain, &key_next, 2427 key_next, HAMMER2_KEY_MAX, 2428 &error, 0); 2429 } 2430 if (parent) { 2431 hammer2_chain_unlock(parent); 2432 hammer2_chain_drop(parent); 2433 } 2434 hammer2_inode_unlock(spmp->iroot); 2435 2436 return error; 2437 } 2438 2439 /* 2440 * Sync a mount point; this is called periodically on a per-mount basis from 2441 * the filesystem syncer, and whenever a user issues a sync. 2442 */ 2443 int 2444 hammer2_vfs_sync(struct mount *mp, int waitfor) 2445 { 2446 int error; 2447 2448 error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor); 2449 2450 return error; 2451 } 2452 2453 /* 2454 * Because frontend operations lock vnodes before we get a chance to 2455 * lock the related inode, we can't just acquire a vnode lock without 2456 * risking a deadlock. The frontend may be holding a vnode lock while 2457 * also blocked on our SYNCQ flag while trying to get the inode lock. 2458 * 2459 * To deal with this situation we can check the vnode lock situation 2460 * after locking the inode and perform a work-around. 2461 */ 2462 int 2463 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor) 2464 { 2465 struct mount *mp; 2466 /*hammer2_xop_flush_t *xop;*/ 2467 /*struct hammer2_sync_info info;*/ 2468 hammer2_inode_t *ip; 2469 hammer2_depend_t *depend; 2470 hammer2_depend_t *depend_next; 2471 struct vnode *vp; 2472 uint32_t pass2; 2473 int error; 2474 int dorestart; 2475 2476 mp = pmp->mp; 2477 2478 /* 2479 * Move all inodes on sideq to syncq. This will clear sideq. 2480 * This should represent all flushable inodes. These inodes 2481 * will already have refs due to being on syncq or sideq. We 2482 * must do this all at once with the spinlock held to ensure that 2483 * all inode dependencies are part of the same flush. 2484 * 2485 * We should be able to do this asynchronously from frontend 2486 * operations because we will be locking the inodes later on 2487 * to actually flush them, and that will partition any frontend 2488 * op using the same inode. Either it has already locked the 2489 * inode and we will block, or it has not yet locked the inode 2490 * and it will block until we are finished flushing that inode. 2491 * 2492 * When restarting, only move the inodes flagged as PASS2 from 2493 * SIDEQ to SYNCQ. PASS2 propagation by inode_lock4() and 2494 * inode_depend() are atomic with the spin-lock. 2495 */ 2496 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH); 2497 #ifdef HAMMER2_DEBUG_SYNC 2498 kprintf("FILESYSTEM SYNC BOUNDARY\n"); 2499 #endif 2500 dorestart = 0; 2501 2502 /* 2503 * Move inodes from depq to syncq, releasing the related 2504 * depend structures. 2505 */ 2506 restart: 2507 #ifdef HAMMER2_DEBUG_SYNC 2508 kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart); 2509 #endif 2510 hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/); 2511 hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN); 2512 2513 /* 2514 * Move inodes from depq to syncq. When restarting, only depq's 2515 * marked pass2 are moved. 2516 */ 2517 hammer2_spin_ex(&pmp->list_spin); 2518 depend_next = TAILQ_FIRST(&pmp->depq); 2519 2520 while ((depend = depend_next) != NULL) { 2521 depend_next = TAILQ_NEXT(depend, entry); 2522 if (dorestart && depend->pass2 == 0) 2523 continue; 2524 TAILQ_FOREACH(ip, &depend->sideq, entry) { 2525 KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ); 2526 atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ); 2527 atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ); 2528 ip->depend = NULL; 2529 } 2530 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry); 2531 pmp->sideq_count -= depend->count; 2532 depend->count = 0; 2533 depend->pass2 = 0; 2534 TAILQ_REMOVE(&pmp->depq, depend, entry); 2535 } 2536 2537 hammer2_spin_unex(&pmp->list_spin); 2538 hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/ 2539 HAMMER2_TRANS_WAITING); 2540 dorestart = 0; 2541 2542 /* 2543 * sideq_count may have dropped enough to allow us to unstall 2544 * the frontend. 2545 */ 2546 hammer2_pfs_memory_inc(pmp); 2547 hammer2_pfs_memory_wakeup(pmp); 2548 2549 /* 2550 * Now run through all inodes on syncq. 2551 * 2552 * Flush transactions only interlock with other flush transactions. 2553 * Any conflicting frontend operations will block on the inode, but 2554 * may hold a vnode lock while doing so. 2555 */ 2556 hammer2_spin_ex(&pmp->list_spin); 2557 while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) { 2558 /* 2559 * Remove the inode from the SYNCQ, transfer the syncq ref 2560 * to us. We must clear SYNCQ to allow any potential 2561 * front-end deadlock to proceed. We must set PASS2 so 2562 * the dependency code knows what to do. 2563 */ 2564 pass2 = ip->flags; 2565 cpu_ccfence(); 2566 if (atomic_cmpset_int(&ip->flags, 2567 pass2, 2568 (pass2 & ~(HAMMER2_INODE_SYNCQ | 2569 HAMMER2_INODE_SYNCQ_WAKEUP)) | 2570 HAMMER2_INODE_SYNCQ_PASS2) == 0) { 2571 continue; 2572 } 2573 TAILQ_REMOVE(&pmp->syncq, ip, entry); 2574 hammer2_spin_unex(&pmp->list_spin); 2575 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP) 2576 wakeup(&ip->flags); 2577 2578 /* 2579 * Relock the inode, and we inherit a ref from the above. 2580 * We will check for a race after we acquire the vnode. 2581 */ 2582 hammer2_mtx_ex(&ip->lock); 2583 2584 /* 2585 * We need the vp in order to vfsync() dirty buffers, so if 2586 * one isn't attached we can skip it. 2587 * 2588 * Ordering the inode lock and then the vnode lock has the 2589 * potential to deadlock. If we had left SYNCQ set that could 2590 * also deadlock us against the frontend even if we don't hold 2591 * any locks, but the latter is not a problem now since we 2592 * cleared it. igetv will temporarily release the inode lock 2593 * in a safe manner to work-around the deadlock. 2594 * 2595 * Unfortunately it is still possible to deadlock when the 2596 * frontend obtains multiple inode locks, because all the 2597 * related vnodes are already locked (nor can the vnode locks 2598 * be released and reacquired without messing up RECLAIM and 2599 * INACTIVE sequencing). 2600 * 2601 * The solution for now is to move the vp back onto SIDEQ 2602 * and set dorestart, which will restart the flush after we 2603 * exhaust the current SYNCQ. Note that additional 2604 * dependencies may build up, so we definitely need to move 2605 * the whole SIDEQ back to SYNCQ when we restart. 2606 */ 2607 vp = ip->vp; 2608 if (vp) { 2609 if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) { 2610 /* 2611 * Failed to get the vnode, requeue the inode 2612 * (PASS2 is already set so it will be found 2613 * again on the restart). 2614 * 2615 * Then unlock, possibly sleep, and retry 2616 * later. We sleep if PASS2 was *previously* 2617 * set, before we set it again above. 2618 */ 2619 vp = NULL; 2620 dorestart = 1; 2621 #ifdef HAMMER2_DEBUG_SYNC 2622 kprintf("inum %ld (sync delayed by vnode)\n", 2623 (long)ip->meta.inum); 2624 #endif 2625 hammer2_inode_delayed_sideq(ip); 2626 2627 hammer2_mtx_unlock(&ip->lock); 2628 hammer2_inode_drop(ip); 2629 2630 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) { 2631 tsleep(&dorestart, 0, "h2syndel", 2); 2632 } 2633 hammer2_spin_ex(&pmp->list_spin); 2634 continue; 2635 } 2636 } else { 2637 vp = NULL; 2638 } 2639 2640 /* 2641 * If the inode wound up on a SIDEQ again it will already be 2642 * prepped for another PASS2. In this situation if we flush 2643 * it now we will just wind up flushing it again in the same 2644 * syncer run, so we might as well not flush it now. 2645 */ 2646 if (ip->flags & HAMMER2_INODE_SIDEQ) { 2647 hammer2_mtx_unlock(&ip->lock); 2648 hammer2_inode_drop(ip); 2649 if (vp) 2650 vput(vp); 2651 dorestart = 1; 2652 hammer2_spin_ex(&pmp->list_spin); 2653 continue; 2654 } 2655 2656 /* 2657 * Ok we have the inode exclusively locked and if vp is 2658 * not NULL that will also be exclusively locked. Do the 2659 * meat of the flush. 2660 * 2661 * vp token needed for v_rbdirty_tree check / vclrisdirty 2662 * sequencing. Though we hold the vnode exclusively so 2663 * we shouldn't need to hold the token also in this case. 2664 */ 2665 if (vp) { 2666 vfsync(vp, MNT_WAIT, 1, NULL, NULL); 2667 bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */ 2668 } 2669 2670 /* 2671 * If the inode has not yet been inserted into the tree 2672 * we must do so. Then sync and flush it. The flush should 2673 * update the parent. 2674 */ 2675 if (ip->flags & HAMMER2_INODE_DELETING) { 2676 #ifdef HAMMER2_DEBUG_SYNC 2677 kprintf("inum %ld destroy\n", (long)ip->meta.inum); 2678 #endif 2679 hammer2_inode_chain_des(ip); 2680 atomic_add_long(&hammer2_iod_inode_deletes, 1); 2681 } else if (ip->flags & HAMMER2_INODE_CREATING) { 2682 #ifdef HAMMER2_DEBUG_SYNC 2683 kprintf("inum %ld insert\n", (long)ip->meta.inum); 2684 #endif 2685 hammer2_inode_chain_ins(ip); 2686 atomic_add_long(&hammer2_iod_inode_creates, 1); 2687 } 2688 #ifdef HAMMER2_DEBUG_SYNC 2689 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum); 2690 #endif 2691 2692 /* 2693 * Because I kinda messed up the design and index the inodes 2694 * under the root inode, along side the directory entries, 2695 * we can't flush the inode index under the iroot until the 2696 * end. If we do it now we might miss effects created by 2697 * other inodes on the SYNCQ. 2698 * 2699 * Do a normal (non-FSSYNC) flush instead, which allows the 2700 * vnode code to work the same. We don't want to force iroot 2701 * back onto the SIDEQ, and we also don't want the flush code 2702 * to update pfs_iroot_blocksets until the final flush later. 2703 * 2704 * XXX at the moment this will likely result in a double-flush 2705 * of the iroot chain. 2706 */ 2707 hammer2_inode_chain_sync(ip); 2708 if (ip == pmp->iroot) { 2709 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP); 2710 } else { 2711 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP | 2712 HAMMER2_XOP_FSSYNC); 2713 } 2714 if (vp) { 2715 lwkt_gettoken(&vp->v_token); 2716 if ((ip->flags & (HAMMER2_INODE_MODIFIED | 2717 HAMMER2_INODE_RESIZED | 2718 HAMMER2_INODE_DIRTYDATA)) == 0 && 2719 RB_EMPTY(&vp->v_rbdirty_tree) && 2720 !bio_track_active(&vp->v_track_write)) { 2721 vclrisdirty(vp); 2722 } else { 2723 hammer2_inode_delayed_sideq(ip); 2724 } 2725 lwkt_reltoken(&vp->v_token); 2726 vput(vp); 2727 vp = NULL; /* safety */ 2728 } 2729 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2); 2730 hammer2_inode_unlock(ip); /* unlock+drop */ 2731 /* ip pointer invalid */ 2732 2733 /* 2734 * If the inode got dirted after we dropped our locks, 2735 * it will have already been moved back to the SIDEQ. 2736 */ 2737 hammer2_spin_ex(&pmp->list_spin); 2738 } 2739 hammer2_spin_unex(&pmp->list_spin); 2740 if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) { 2741 #ifdef HAMMER2_DEBUG_SYNC 2742 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n"); 2743 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/ 2744 #endif 2745 dorestart = 1; 2746 goto restart; 2747 } 2748 #ifdef HAMMER2_DEBUG_SYNC 2749 kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n"); 2750 /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/ 2751 #endif 2752 2753 /* 2754 * We have to flush the PFS root last, even if it does not appear to 2755 * be dirty, because all the inodes in the PFS are indexed under it. 2756 * The normal flushing of iroot above would only occur if directory 2757 * entries under the root were changed. 2758 * 2759 * Specifying VOLHDR will cause an additionl flush of hmp->spmp 2760 * for the media making up the cluster. 2761 */ 2762 if ((ip = pmp->iroot) != NULL) { 2763 hammer2_inode_ref(ip); 2764 hammer2_mtx_ex(&ip->lock); 2765 hammer2_inode_chain_sync(ip); 2766 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP | 2767 HAMMER2_XOP_FSSYNC | 2768 HAMMER2_XOP_VOLHDR); 2769 hammer2_inode_unlock(ip); /* unlock+drop */ 2770 } 2771 #ifdef HAMMER2_DEBUG_SYNC 2772 kprintf("FILESYSTEM SYNC STAGE 2 DONE\n"); 2773 #endif 2774 2775 /* 2776 * device bioq sync 2777 */ 2778 hammer2_bioq_sync(pmp); 2779 2780 #if 0 2781 info.pass = 1; 2782 info.waitfor = MNT_WAIT; 2783 vsyncscan(mp, flags, hammer2_sync_scan2, &info); 2784 2785 info.pass = 2; 2786 info.waitfor = MNT_WAIT; 2787 vsyncscan(mp, flags, hammer2_sync_scan2, &info); 2788 #endif 2789 #if 0 2790 /* 2791 * Generally speaking we now want to flush the media topology from 2792 * the iroot through to the inodes. The flush stops at any inode 2793 * boundary, which allows the frontend to continue running concurrent 2794 * modifying operations on inodes (including kernel flushes of 2795 * buffers) without interfering with the main sync. 2796 * 2797 * Use the XOP interface to concurrently flush all nodes to 2798 * synchronize the PFSROOT subtopology to the media. A standard 2799 * end-of-scan ENOENT error indicates cluster sufficiency. 2800 * 2801 * Note that this flush will not be visible on crash recovery until 2802 * we flush the super-root topology in the next loop. 2803 * 2804 * XXX For now wait for all flushes to complete. 2805 */ 2806 if (mp && (ip = pmp->iroot) != NULL) { 2807 /* 2808 * If unmounting try to flush everything including any 2809 * sub-trees under inodes, just in case there is dangling 2810 * modified data, as a safety. Otherwise just flush up to 2811 * the inodes in this stage. 2812 */ 2813 kprintf("MP & IROOT\n"); 2814 #ifdef HAMMER2_DEBUG_SYNC 2815 kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n"); 2816 #endif 2817 if (mp->mnt_kern_flag & MNTK_UNMOUNT) { 2818 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 2819 HAMMER2_XOP_VOLHDR | 2820 HAMMER2_XOP_FSSYNC | 2821 HAMMER2_XOP_INODE_STOP); 2822 } else { 2823 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 2824 HAMMER2_XOP_INODE_STOP | 2825 HAMMER2_XOP_VOLHDR | 2826 HAMMER2_XOP_FSSYNC | 2827 HAMMER2_XOP_INODE_STOP); 2828 } 2829 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc); 2830 error = hammer2_xop_collect(&xop->head, 2831 HAMMER2_XOP_COLLECT_WAITALL); 2832 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 2833 #ifdef HAMMER2_DEBUG_SYNC 2834 kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n"); 2835 #endif 2836 if (error == HAMMER2_ERROR_ENOENT) 2837 error = 0; 2838 else 2839 error = hammer2_error_to_errno(error); 2840 } else { 2841 error = 0; 2842 } 2843 #endif 2844 error = 0; /* XXX */ 2845 hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH); 2846 2847 return (error); 2848 } 2849 2850 static 2851 int 2852 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp) 2853 { 2854 hammer2_inode_t *ip; 2855 2856 KKASSERT(MAXFIDSZ >= 16); 2857 ip = VTOI(vp); 2858 fhp->fid_len = offsetof(struct fid, fid_data[16]); 2859 fhp->fid_ext = 0; 2860 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum; 2861 ((hammer2_tid_t *)fhp->fid_data)[1] = 0; 2862 2863 return 0; 2864 } 2865 2866 static 2867 int 2868 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp, 2869 struct fid *fhp, struct vnode **vpp) 2870 { 2871 hammer2_pfs_t *pmp; 2872 hammer2_tid_t inum; 2873 int error; 2874 2875 pmp = MPTOPMP(mp); 2876 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK; 2877 if (vpp) { 2878 if (inum == 1) 2879 error = hammer2_vfs_root(mp, vpp); 2880 else 2881 error = hammer2_vfs_vget(mp, NULL, inum, vpp); 2882 } else { 2883 error = 0; 2884 } 2885 if (error) 2886 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error); 2887 return error; 2888 } 2889 2890 static 2891 int 2892 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 2893 int *exflagsp, struct ucred **credanonp) 2894 { 2895 hammer2_pfs_t *pmp; 2896 struct netcred *np; 2897 int error; 2898 2899 pmp = MPTOPMP(mp); 2900 np = vfs_export_lookup(mp, &pmp->export, nam); 2901 if (np) { 2902 *exflagsp = np->netc_exflags; 2903 *credanonp = &np->netc_anon; 2904 error = 0; 2905 } else { 2906 error = EACCES; 2907 } 2908 return error; 2909 } 2910 2911 /* 2912 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume 2913 * header into the HMP 2914 * 2915 * XXX read four volhdrs and use the one with the highest TID whos CRC 2916 * matches. 2917 * 2918 * XXX check iCRCs. 2919 * 2920 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to 2921 * nonexistant locations. 2922 * 2923 * XXX Record selected volhdr and ring updates to each of 4 volhdrs 2924 */ 2925 static 2926 int 2927 hammer2_install_volume_header(hammer2_dev_t *hmp) 2928 { 2929 hammer2_volume_data_t *vd; 2930 struct buf *bp; 2931 hammer2_crc32_t crc0, crc, bcrc0, bcrc; 2932 int error_reported; 2933 int error; 2934 int valid; 2935 int i; 2936 2937 error_reported = 0; 2938 error = 0; 2939 valid = 0; 2940 bp = NULL; 2941 2942 /* 2943 * There are up to 4 copies of the volume header (syncs iterate 2944 * between them so there is no single master). We don't trust the 2945 * volu_size field so we don't know precisely how large the filesystem 2946 * is, so depend on the OS to return an error if we go beyond the 2947 * block device's EOF. 2948 */ 2949 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) { 2950 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64, 2951 HAMMER2_VOLUME_BYTES, &bp); 2952 if (error) { 2953 brelse(bp); 2954 bp = NULL; 2955 continue; 2956 } 2957 2958 vd = (struct hammer2_volume_data *) bp->b_data; 2959 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) && 2960 (vd->magic != HAMMER2_VOLUME_ID_ABO)) { 2961 brelse(bp); 2962 bp = NULL; 2963 continue; 2964 } 2965 2966 if (vd->magic == HAMMER2_VOLUME_ID_ABO) { 2967 /* XXX: Reversed-endianness filesystem */ 2968 kprintf("hammer2: reverse-endian filesystem detected"); 2969 brelse(bp); 2970 bp = NULL; 2971 continue; 2972 } 2973 2974 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0]; 2975 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF, 2976 HAMMER2_VOLUME_ICRC0_SIZE); 2977 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1]; 2978 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF, 2979 HAMMER2_VOLUME_ICRC1_SIZE); 2980 if ((crc0 != crc) || (bcrc0 != bcrc)) { 2981 kprintf("hammer2 volume header crc " 2982 "mismatch copy #%d %08x/%08x\n", 2983 i, crc0, crc); 2984 error_reported = 1; 2985 brelse(bp); 2986 bp = NULL; 2987 continue; 2988 } 2989 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) { 2990 valid = 1; 2991 hmp->voldata = *vd; 2992 hmp->volhdrno = i; 2993 } 2994 brelse(bp); 2995 bp = NULL; 2996 } 2997 if (valid) { 2998 hmp->volsync = hmp->voldata; 2999 hmp->free_reserved = hmp->voldata.allocator_size / 20; 3000 error = 0; 3001 if (error_reported || bootverbose || 1) { /* 1/DEBUG */ 3002 kprintf("hammer2: using volume header #%d\n", 3003 hmp->volhdrno); 3004 } 3005 } else { 3006 error = EINVAL; 3007 kprintf("hammer2: no valid volume headers found!\n"); 3008 } 3009 return (error); 3010 } 3011 3012 /* 3013 * This handles hysteresis on regular file flushes. Because the BIOs are 3014 * routed to a thread it is possible for an excessive number to build up 3015 * and cause long front-end stalls long before the runningbuffspace limit 3016 * is hit, so we implement hammer2_flush_pipe to control the 3017 * hysteresis. 3018 * 3019 * This is a particular problem when compression is used. 3020 */ 3021 void 3022 hammer2_lwinprog_ref(hammer2_pfs_t *pmp) 3023 { 3024 atomic_add_int(&pmp->count_lwinprog, 1); 3025 } 3026 3027 void 3028 hammer2_lwinprog_drop(hammer2_pfs_t *pmp) 3029 { 3030 int lwinprog; 3031 3032 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1); 3033 if ((lwinprog & HAMMER2_LWINPROG_WAITING) && 3034 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) { 3035 atomic_clear_int(&pmp->count_lwinprog, 3036 HAMMER2_LWINPROG_WAITING); 3037 wakeup(&pmp->count_lwinprog); 3038 } 3039 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) && 3040 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) { 3041 atomic_clear_int(&pmp->count_lwinprog, 3042 HAMMER2_LWINPROG_WAITING0); 3043 wakeup(&pmp->count_lwinprog); 3044 } 3045 } 3046 3047 void 3048 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe) 3049 { 3050 int lwinprog; 3051 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING : 3052 HAMMER2_LWINPROG_WAITING0; 3053 3054 for (;;) { 3055 lwinprog = pmp->count_lwinprog; 3056 cpu_ccfence(); 3057 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 3058 break; 3059 tsleep_interlock(&pmp->count_lwinprog, 0); 3060 atomic_set_int(&pmp->count_lwinprog, lwflag); 3061 lwinprog = pmp->count_lwinprog; 3062 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 3063 break; 3064 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz); 3065 } 3066 } 3067 3068 /* 3069 * It is possible for an excessive number of dirty chains or dirty inodes 3070 * to build up. When this occurs we start an asynchronous filesystem sync. 3071 * If the level continues to build up, we stall, waiting for it to drop, 3072 * with some hysteresis. 3073 * 3074 * We limit the stall to two seconds per call. 3075 * 3076 * This relies on the kernel calling hammer2_vfs_modifying() prior to 3077 * obtaining any vnode locks before making a modifying VOP call. 3078 */ 3079 static void 3080 hammer2_vfs_modifying(struct mount *mp) 3081 { 3082 hammer2_pfs_memory_wait(MPTOPMP(mp)); 3083 } 3084 3085 /* 3086 * Initiate an asynchronous filesystem sync and, with hysteresis, 3087 * stall if the internal data structure count becomes too bloated. 3088 */ 3089 void 3090 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp) 3091 { 3092 uint32_t waiting; 3093 int loops; 3094 3095 if (pmp == NULL || pmp->mp == NULL) 3096 return; 3097 3098 for (loops = 0; loops < 2; ++loops) { 3099 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK; 3100 cpu_ccfence(); 3101 3102 /* 3103 * Start the syncer running at 1/2 the limit 3104 */ 3105 if (waiting > hammer2_limit_dirty_chains / 2 || 3106 pmp->sideq_count > hammer2_limit_dirty_inodes / 2) { 3107 trigger_syncer(pmp->mp); 3108 } 3109 3110 /* 3111 * Stall at the limit waiting for the counts to drop. 3112 * This code will typically be woken up once the count 3113 * drops below 3/4 the limit, or in one second. 3114 */ 3115 if (waiting < hammer2_limit_dirty_chains && 3116 pmp->sideq_count < hammer2_limit_dirty_inodes) { 3117 break; 3118 } 3119 tsleep_interlock(&pmp->inmem_dirty_chains, 0); 3120 atomic_set_int(&pmp->inmem_dirty_chains, 3121 HAMMER2_DIRTYCHAIN_WAITING); 3122 if (waiting < hammer2_limit_dirty_chains && 3123 pmp->sideq_count < hammer2_limit_dirty_inodes) { 3124 break; 3125 } 3126 trigger_syncer(pmp->mp); 3127 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED, "h2memw", hz); 3128 #if 0 3129 limit = pmp->mp->mnt_nvnodelistsize / 10; 3130 if (limit < hammer2_limit_dirty_chains) 3131 limit = hammer2_limit_dirty_chains; 3132 if (limit < 1000) 3133 limit = 1000; 3134 #endif 3135 } 3136 } 3137 3138 void 3139 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp) 3140 { 3141 if (pmp) { 3142 atomic_add_int(&pmp->inmem_dirty_chains, 1); 3143 } 3144 } 3145 3146 void 3147 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp) 3148 { 3149 uint32_t waiting; 3150 3151 if (pmp) { 3152 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1); 3153 /* don't need --waiting to test flag */ 3154 3155 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) && 3156 (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <= 3157 hammer2_limit_dirty_chains * 2 / 3 && 3158 pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) { 3159 atomic_clear_int(&pmp->inmem_dirty_chains, 3160 HAMMER2_DIRTYCHAIN_WAITING); 3161 wakeup(&pmp->inmem_dirty_chains); 3162 } 3163 } 3164 } 3165 3166 /* 3167 * Returns 0 if the filesystem has tons of free space 3168 * Returns 1 if the filesystem has less than 10% remaining 3169 * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining. 3170 */ 3171 int 3172 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred) 3173 { 3174 hammer2_pfs_t *pmp; 3175 hammer2_dev_t *hmp; 3176 hammer2_off_t free_reserved; 3177 hammer2_off_t free_nominal; 3178 int i; 3179 3180 pmp = ip->pmp; 3181 3182 if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) { 3183 free_reserved = HAMMER2_SEGSIZE; 3184 free_nominal = 0x7FFFFFFFFFFFFFFFLLU; 3185 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 3186 hmp = pmp->pfs_hmps[i]; 3187 if (hmp == NULL) 3188 continue; 3189 if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER && 3190 pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER) 3191 continue; 3192 3193 if (free_nominal > hmp->voldata.allocator_free) 3194 free_nominal = hmp->voldata.allocator_free; 3195 if (free_reserved < hmp->free_reserved) 3196 free_reserved = hmp->free_reserved; 3197 } 3198 3199 /* 3200 * SMP races ok 3201 */ 3202 pmp->free_reserved = free_reserved; 3203 pmp->free_nominal = free_nominal; 3204 pmp->free_ticks = ticks; 3205 } else { 3206 free_reserved = pmp->free_reserved; 3207 free_nominal = pmp->free_nominal; 3208 } 3209 if (cred && cred->cr_uid != 0) { 3210 if ((int64_t)(free_nominal - bytes) < 3211 (int64_t)free_reserved) { 3212 return 2; 3213 } 3214 } else { 3215 if ((int64_t)(free_nominal - bytes) < 3216 (int64_t)free_reserved / 2) { 3217 return 2; 3218 } 3219 } 3220 if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2) 3221 return 1; 3222 return 0; 3223 } 3224 3225 /* 3226 * Debugging 3227 */ 3228 void 3229 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx, 3230 u_int flags) 3231 { 3232 hammer2_chain_t *scan; 3233 hammer2_chain_t *parent; 3234 3235 --*countp; 3236 if (*countp == 0) { 3237 kprintf("%*.*s...\n", tab, tab, ""); 3238 return; 3239 } 3240 if (*countp < 0) 3241 return; 3242 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n", 3243 tab, tab, "", pfx, 3244 chain, chain->bref.type, 3245 chain->bref.key, chain->bref.keybits, 3246 chain->bref.mirror_tid); 3247 3248 kprintf("%*.*s [%08x] (%s) refs=%d", 3249 tab, tab, "", 3250 chain->flags, 3251 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE && 3252 chain->data) ? (char *)chain->data->ipdata.filename : "?"), 3253 chain->refs); 3254 3255 parent = chain->parent; 3256 if (parent) 3257 kprintf("\n%*.*s p=%p [pflags %08x prefs %d", 3258 tab, tab, "", 3259 parent, parent->flags, parent->refs); 3260 if (RB_EMPTY(&chain->core.rbtree)) { 3261 kprintf("\n"); 3262 } else { 3263 kprintf(" {\n"); 3264 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) { 3265 if ((scan->flags & flags) || flags == (u_int)-1) { 3266 hammer2_dump_chain(scan, tab + 4, countp, 'a', 3267 flags); 3268 } 3269 } 3270 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data) 3271 kprintf("%*.*s}(%s)\n", tab, tab, "", 3272 chain->data->ipdata.filename); 3273 else 3274 kprintf("%*.*s}\n", tab, tab, ""); 3275 } 3276 } 3277