1 /* 2 * Copyright (c) 2011-2015 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 }; 73 74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev); 75 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs); 76 static struct hammer2_mntlist hammer2_mntlist; 77 static struct hammer2_pfslist hammer2_pfslist; 78 static struct lock hammer2_mntlk; 79 80 int hammer2_debug; 81 int hammer2_cluster_enable = 4; 82 int hammer2_hardlink_enable = 1; 83 int hammer2_flush_pipe = 100; 84 int hammer2_synchronous_flush = 1; 85 int hammer2_dio_count; 86 long hammer2_chain_allocs; 87 long hammer2_chain_frees; 88 long hammer2_limit_dirty_chains; 89 long hammer2_count_modified_chains; 90 long hammer2_iod_file_read; 91 long hammer2_iod_meta_read; 92 long hammer2_iod_indr_read; 93 long hammer2_iod_fmap_read; 94 long hammer2_iod_volu_read; 95 long hammer2_iod_file_write; 96 long hammer2_iod_file_wembed; 97 long hammer2_iod_file_wzero; 98 long hammer2_iod_file_wdedup; 99 long hammer2_iod_meta_write; 100 long hammer2_iod_indr_write; 101 long hammer2_iod_fmap_write; 102 long hammer2_iod_volu_write; 103 long hammer2_ioa_file_read; 104 long hammer2_ioa_meta_read; 105 long hammer2_ioa_indr_read; 106 long hammer2_ioa_fmap_read; 107 long hammer2_ioa_volu_read; 108 long hammer2_ioa_fmap_write; 109 long hammer2_ioa_file_write; 110 long hammer2_ioa_meta_write; 111 long hammer2_ioa_indr_write; 112 long hammer2_ioa_volu_write; 113 114 MALLOC_DECLARE(M_HAMMER2_CBUFFER); 115 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer", 116 "Buffer used for compression."); 117 118 MALLOC_DECLARE(M_HAMMER2_DEBUFFER); 119 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer", 120 "Buffer used for decompression."); 121 122 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem"); 123 124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW, 125 &hammer2_debug, 0, ""); 126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW, 127 &hammer2_cluster_enable, 0, ""); 128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW, 129 &hammer2_hardlink_enable, 0, ""); 130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW, 131 &hammer2_flush_pipe, 0, ""); 132 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW, 133 &hammer2_synchronous_flush, 0, ""); 134 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW, 135 &hammer2_chain_allocs, 0, ""); 136 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW, 137 &hammer2_chain_frees, 0, ""); 138 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW, 139 &hammer2_limit_dirty_chains, 0, ""); 140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW, 141 &hammer2_count_modified_chains, 0, ""); 142 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD, 143 &hammer2_dio_count, 0, ""); 144 145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW, 146 &hammer2_iod_file_read, 0, ""); 147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW, 148 &hammer2_iod_meta_read, 0, ""); 149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW, 150 &hammer2_iod_indr_read, 0, ""); 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW, 152 &hammer2_iod_fmap_read, 0, ""); 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW, 154 &hammer2_iod_volu_read, 0, ""); 155 156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW, 157 &hammer2_iod_file_write, 0, ""); 158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW, 159 &hammer2_iod_file_wembed, 0, ""); 160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW, 161 &hammer2_iod_file_wzero, 0, ""); 162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW, 163 &hammer2_iod_file_wdedup, 0, ""); 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW, 165 &hammer2_iod_meta_write, 0, ""); 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW, 167 &hammer2_iod_indr_write, 0, ""); 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW, 169 &hammer2_iod_fmap_write, 0, ""); 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW, 171 &hammer2_iod_volu_write, 0, ""); 172 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW, 174 &hammer2_ioa_file_read, 0, ""); 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW, 176 &hammer2_ioa_meta_read, 0, ""); 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW, 178 &hammer2_ioa_indr_read, 0, ""); 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW, 180 &hammer2_ioa_fmap_read, 0, ""); 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW, 182 &hammer2_ioa_volu_read, 0, ""); 183 184 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW, 185 &hammer2_ioa_file_write, 0, ""); 186 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW, 187 &hammer2_ioa_meta_write, 0, ""); 188 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW, 189 &hammer2_ioa_indr_write, 0, ""); 190 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW, 191 &hammer2_ioa_fmap_write, 0, ""); 192 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW, 193 &hammer2_ioa_volu_write, 0, ""); 194 195 long hammer2_check_icrc32; 196 long hammer2_check_xxhash64; 197 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW, 198 &hammer2_check_icrc32, 0, ""); 199 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW, 200 &hammer2_check_xxhash64, 0, ""); 201 202 static int hammer2_vfs_init(struct vfsconf *conf); 203 static int hammer2_vfs_uninit(struct vfsconf *vfsp); 204 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data, 205 struct ucred *cred); 206 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *, 207 struct vnode *, struct ucred *); 208 static int hammer2_recovery(hammer2_dev_t *hmp); 209 static int hammer2_vfs_unmount(struct mount *mp, int mntflags); 210 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp); 211 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, 212 struct ucred *cred); 213 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, 214 struct ucred *cred); 215 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp, 216 ino_t ino, struct vnode **vpp); 217 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp, 218 struct fid *fhp, struct vnode **vpp); 219 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp); 220 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 221 int *exflagsp, struct ucred **credanonp); 222 223 static int hammer2_install_volume_header(hammer2_dev_t *hmp); 224 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data); 225 226 static void hammer2_update_pmps(hammer2_dev_t *hmp); 227 228 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp); 229 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, 230 hammer2_dev_t *hmp); 231 232 /* 233 * HAMMER2 vfs operations. 234 */ 235 static struct vfsops hammer2_vfsops = { 236 .vfs_init = hammer2_vfs_init, 237 .vfs_uninit = hammer2_vfs_uninit, 238 .vfs_sync = hammer2_vfs_sync, 239 .vfs_mount = hammer2_vfs_mount, 240 .vfs_unmount = hammer2_vfs_unmount, 241 .vfs_root = hammer2_vfs_root, 242 .vfs_statfs = hammer2_vfs_statfs, 243 .vfs_statvfs = hammer2_vfs_statvfs, 244 .vfs_vget = hammer2_vfs_vget, 245 .vfs_vptofh = hammer2_vfs_vptofh, 246 .vfs_fhtovp = hammer2_vfs_fhtovp, 247 .vfs_checkexp = hammer2_vfs_checkexp 248 }; 249 250 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", ""); 251 252 VFS_SET(hammer2_vfsops, hammer2, 0); 253 MODULE_VERSION(hammer2, 1); 254 255 static 256 int 257 hammer2_vfs_init(struct vfsconf *conf) 258 { 259 static struct objcache_malloc_args margs_read; 260 static struct objcache_malloc_args margs_write; 261 static struct objcache_malloc_args margs_vop; 262 263 int error; 264 265 error = 0; 266 267 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref)) 268 error = EINVAL; 269 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data)) 270 error = EINVAL; 271 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data)) 272 error = EINVAL; 273 274 if (error) 275 kprintf("HAMMER2 structure size mismatch; cannot continue.\n"); 276 277 margs_read.objsize = 65536; 278 margs_read.mtype = M_HAMMER2_DEBUFFER; 279 280 margs_write.objsize = 32768; 281 margs_write.mtype = M_HAMMER2_CBUFFER; 282 283 margs_vop.objsize = sizeof(hammer2_xop_t); 284 margs_vop.mtype = M_HAMMER2; 285 286 /* 287 * Note thaht for the XOPS cache we want backing store allocations 288 * to use M_ZERO. This is not allowed in objcache_get() (to avoid 289 * confusion), so use the backing store function that does it. This 290 * means that initial XOPS objects are zerod but REUSED objects are 291 * not. So we are responsible for cleaning the object up sufficiently 292 * for our needs before objcache_put()ing it back (typically just the 293 * FIFO indices). 294 */ 295 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc, 296 0, 1, NULL, NULL, NULL, 297 objcache_malloc_alloc, 298 objcache_malloc_free, 299 &margs_read); 300 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc, 301 0, 1, NULL, NULL, NULL, 302 objcache_malloc_alloc, 303 objcache_malloc_free, 304 &margs_write); 305 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc, 306 0, 1, NULL, NULL, NULL, 307 objcache_malloc_alloc_zero, 308 objcache_malloc_free, 309 &margs_vop); 310 311 312 lockinit(&hammer2_mntlk, "mntlk", 0, 0); 313 TAILQ_INIT(&hammer2_mntlist); 314 TAILQ_INIT(&hammer2_pfslist); 315 316 hammer2_limit_dirty_chains = desiredvnodes / 10; 317 318 return (error); 319 } 320 321 static 322 int 323 hammer2_vfs_uninit(struct vfsconf *vfsp __unused) 324 { 325 objcache_destroy(cache_buffer_read); 326 objcache_destroy(cache_buffer_write); 327 objcache_destroy(cache_xops); 328 return 0; 329 } 330 331 /* 332 * Core PFS allocator. Used to allocate the pmp structure for PFS cluster 333 * mounts and the spmp structure for media (hmp) structures. 334 * 335 * pmp->modify_tid tracks new modify_tid transaction ids for front-end 336 * transactions. Note that synchronization does not use this field. 337 * (typically frontend operations and synchronization cannot run on the 338 * same PFS node at the same time). 339 * 340 * XXX check locking 341 */ 342 hammer2_pfs_t * 343 hammer2_pfsalloc(hammer2_chain_t *chain, const hammer2_inode_data_t *ripdata, 344 hammer2_tid_t modify_tid, hammer2_dev_t *force_local) 345 { 346 hammer2_inode_t *iroot; 347 hammer2_pfs_t *pmp; 348 int count; 349 int i; 350 int j; 351 352 /* 353 * Locate or create the PFS based on the cluster id. If ripdata 354 * is NULL this is a spmp which is unique and is always allocated. 355 * 356 * If the device is mounted in local mode all PFSs are considered 357 * independent and not part of any cluster (for debugging only). 358 */ 359 if (ripdata) { 360 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 361 if (force_local != pmp->force_local) 362 continue; 363 if (force_local == NULL && 364 bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid, 365 sizeof(pmp->pfs_clid)) == 0) { 366 break; 367 } else if (force_local && pmp->pfs_names[0] && 368 strcmp(pmp->pfs_names[0], ripdata->filename) == 0) { 369 break; 370 } 371 } 372 } else { 373 pmp = NULL; 374 } 375 376 if (pmp == NULL) { 377 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO); 378 pmp->force_local = force_local; 379 hammer2_trans_manage_init(pmp); 380 kmalloc_create(&pmp->minode, "HAMMER2-inodes"); 381 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg"); 382 lockinit(&pmp->lock, "pfslk", 0, 0); 383 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0); 384 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum"); 385 spin_init(&pmp->xop_spin, "h2xop"); 386 RB_INIT(&pmp->inum_tree); 387 TAILQ_INIT(&pmp->sideq); 388 spin_init(&pmp->list_spin, "hm2pfsalloc_list"); 389 390 /* 391 * Distribute backend operations to threads 392 */ 393 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 394 for (j = 0; j < HAMMER2_XOPGROUPS + 395 HAMMER2_SPECTHREADS; ++j) { 396 TAILQ_INIT(&pmp->xopq[i][j]); 397 } 398 } 399 for (i = 0; i < HAMMER2_XOPGROUPS; ++i) 400 hammer2_xop_group_init(pmp, &pmp->xop_groups[i]); 401 402 /* 403 * Save the last media transaction id for the flusher. Set 404 * initial 405 */ 406 if (ripdata) 407 pmp->pfs_clid = ripdata->meta.pfs_clid; 408 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry); 409 410 /* 411 * The synchronization thread may start too early, make 412 * sure it stays frozen until we are ready to let it go. 413 * XXX 414 */ 415 /* 416 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN | 417 HAMMER2_THREAD_REMASTER; 418 */ 419 } 420 421 /* 422 * Create the PFS's root inode. 423 */ 424 if ((iroot = pmp->iroot) == NULL) { 425 iroot = hammer2_inode_get(pmp, NULL, NULL, -1); 426 pmp->iroot = iroot; 427 hammer2_inode_ref(iroot); 428 hammer2_inode_unlock(iroot); 429 } 430 431 /* 432 * Stop here if no chain is passed in. 433 */ 434 if (chain == NULL) 435 goto done; 436 437 /* 438 * When a chain is passed in we must add it to the PFS's root 439 * inode, update pmp->pfs_types[], and update the syncronization 440 * threads. 441 * 442 * When forcing local mode, mark the PFS as a MASTER regardless. 443 * 444 * At the moment empty spots can develop due to removals or failures. 445 * Ultimately we want to re-fill these spots but doing so might 446 * confused running code. XXX 447 */ 448 hammer2_inode_ref(iroot); 449 hammer2_mtx_ex(&iroot->lock); 450 j = iroot->cluster.nchains; 451 452 kprintf("add PFS to pmp %p[%d]\n", pmp, j); 453 454 if (j == HAMMER2_MAXCLUSTER) { 455 kprintf("hammer2_mount: cluster full!\n"); 456 /* XXX fatal error? */ 457 } else { 458 KKASSERT(chain->pmp == NULL); 459 chain->pmp = pmp; 460 hammer2_chain_ref(chain); 461 iroot->cluster.array[j].chain = chain; 462 if (force_local) 463 pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER; 464 else 465 pmp->pfs_types[j] = ripdata->meta.pfs_type; 466 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2); 467 pmp->pfs_hmps[j] = chain->hmp; 468 469 /* 470 * If the PFS is already mounted we must account 471 * for the mount_count here. 472 */ 473 if (pmp->mp) 474 ++chain->hmp->mount_count; 475 476 /* 477 * May have to fixup dirty chain tracking. Previous 478 * pmp was NULL so nothing to undo. 479 */ 480 if (chain->flags & HAMMER2_CHAIN_MODIFIED) 481 hammer2_pfs_memory_inc(pmp); 482 ++j; 483 } 484 iroot->cluster.nchains = j; 485 486 /* 487 * Update nmasters from any PFS inode which is part of the cluster. 488 * It is possible that this will result in a value which is too 489 * high. MASTER PFSs are authoritative for pfs_nmasters and will 490 * override this value later on. 491 * 492 * (This informs us of masters that might not currently be 493 * discoverable by this mount). 494 */ 495 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) { 496 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters; 497 } 498 499 /* 500 * Count visible masters. Masters are usually added with 501 * ripdata->meta.pfs_nmasters set to 1. This detects when there 502 * are more (XXX and must update the master inodes). 503 */ 504 count = 0; 505 for (i = 0; i < iroot->cluster.nchains; ++i) { 506 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) 507 ++count; 508 } 509 if (pmp->pfs_nmasters < count) 510 pmp->pfs_nmasters = count; 511 512 /* 513 * Create missing synchronization and support threads. 514 * 515 * Single-node masters (including snapshots) have nothing to 516 * synchronize and do not require this thread. 517 * 518 * Multi-node masters or any number of soft masters, slaves, copy, 519 * or other PFS types need the thread. 520 * 521 * Each thread is responsible for its particular cluster index. 522 * We use independent threads so stalls or mismatches related to 523 * any given target do not affect other targets. 524 */ 525 for (i = 0; i < iroot->cluster.nchains; ++i) { 526 /* 527 * Single-node masters (including snapshots) have nothing 528 * to synchronize and will make direct xops support calls, 529 * thus they do not require this thread. 530 * 531 * Note that there can be thousands of snapshots. We do not 532 * want to create thousands of threads. 533 */ 534 if (pmp->pfs_nmasters <= 1 && 535 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) { 536 continue; 537 } 538 539 /* 540 * Sync support thread 541 */ 542 if (pmp->sync_thrs[i].td == NULL) { 543 hammer2_thr_create(&pmp->sync_thrs[i], pmp, 544 "h2nod", i, -1, 545 hammer2_primary_sync_thread); 546 } 547 } 548 549 /* 550 * Create missing Xop threads 551 */ 552 if (pmp->mp) 553 hammer2_xop_helper_create(pmp); 554 555 hammer2_mtx_unlock(&iroot->lock); 556 hammer2_inode_drop(iroot); 557 done: 558 return pmp; 559 } 560 561 /* 562 * Destroy a PFS, typically only occurs after the last mount on a device 563 * has gone away. 564 */ 565 static void 566 hammer2_pfsfree(hammer2_pfs_t *pmp) 567 { 568 hammer2_inode_t *iroot; 569 int i; 570 int j; 571 572 /* 573 * Cleanup our reference on iroot. iroot is (should) not be needed 574 * by the flush code. 575 */ 576 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry); 577 578 iroot = pmp->iroot; 579 if (iroot) { 580 for (i = 0; i < iroot->cluster.nchains; ++i) { 581 hammer2_thr_delete(&pmp->sync_thrs[i]); 582 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) 583 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]); 584 } 585 #if REPORT_REFS_ERRORS 586 if (pmp->iroot->refs != 1) 587 kprintf("PMP->IROOT %p REFS WRONG %d\n", 588 pmp->iroot, pmp->iroot->refs); 589 #else 590 KKASSERT(pmp->iroot->refs == 1); 591 #endif 592 /* ref for pmp->iroot */ 593 hammer2_inode_drop(pmp->iroot); 594 pmp->iroot = NULL; 595 } 596 597 kmalloc_destroy(&pmp->mmsg); 598 kmalloc_destroy(&pmp->minode); 599 600 kfree(pmp, M_HAMMER2); 601 } 602 603 /* 604 * Remove all references to hmp from the pfs list. Any PFS which becomes 605 * empty is terminated and freed. 606 * 607 * XXX inefficient. 608 */ 609 static void 610 hammer2_pfsfree_scan(hammer2_dev_t *hmp) 611 { 612 hammer2_pfs_t *pmp; 613 hammer2_inode_t *iroot; 614 hammer2_chain_t *rchain; 615 int didfreeze; 616 int i; 617 int j; 618 619 again: 620 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 621 if ((iroot = pmp->iroot) == NULL) 622 continue; 623 if (hmp->spmp == pmp) { 624 kprintf("unmount hmp %p remove spmp %p\n", 625 hmp, pmp); 626 hmp->spmp = NULL; 627 } 628 629 /* 630 * Determine if this PFS is affected. If it is we must 631 * freeze all management threads and lock its iroot. 632 * 633 * Freezing a management thread forces it idle, operations 634 * in-progress will be aborted and it will have to start 635 * over again when unfrozen, or exit if told to exit. 636 */ 637 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 638 if (pmp->pfs_hmps[i] == hmp) 639 break; 640 } 641 if (i != HAMMER2_MAXCLUSTER) { 642 /* 643 * Make sure all synchronization threads are locked 644 * down. 645 */ 646 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 647 if (pmp->pfs_hmps[i] == NULL) 648 continue; 649 hammer2_thr_freeze_async(&pmp->sync_thrs[i]); 650 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) { 651 hammer2_thr_freeze_async( 652 &pmp->xop_groups[j].thrs[i]); 653 } 654 } 655 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 656 if (pmp->pfs_hmps[i] == NULL) 657 continue; 658 hammer2_thr_freeze(&pmp->sync_thrs[i]); 659 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) { 660 hammer2_thr_freeze( 661 &pmp->xop_groups[j].thrs[i]); 662 } 663 } 664 665 /* 666 * Lock the inode and clean out matching chains. 667 * Note that we cannot use hammer2_inode_lock_*() 668 * here because that would attempt to validate the 669 * cluster that we are in the middle of ripping 670 * apart. 671 * 672 * WARNING! We are working directly on the inodes 673 * embedded cluster. 674 */ 675 hammer2_mtx_ex(&iroot->lock); 676 677 /* 678 * Remove the chain from matching elements of the PFS. 679 */ 680 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 681 if (pmp->pfs_hmps[i] != hmp) 682 continue; 683 hammer2_thr_delete(&pmp->sync_thrs[i]); 684 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) { 685 hammer2_thr_delete( 686 &pmp->xop_groups[j].thrs[i]); 687 } 688 rchain = iroot->cluster.array[i].chain; 689 iroot->cluster.array[i].chain = NULL; 690 pmp->pfs_types[i] = 0; 691 if (pmp->pfs_names[i]) { 692 kfree(pmp->pfs_names[i], M_HAMMER2); 693 pmp->pfs_names[i] = NULL; 694 } 695 if (rchain) { 696 hammer2_chain_drop(rchain); 697 /* focus hint */ 698 if (iroot->cluster.focus == rchain) 699 iroot->cluster.focus = NULL; 700 } 701 pmp->pfs_hmps[i] = NULL; 702 } 703 hammer2_mtx_unlock(&iroot->lock); 704 didfreeze = 1; /* remaster, unfreeze down below */ 705 } else { 706 didfreeze = 0; 707 } 708 709 /* 710 * Cleanup trailing chains. Gaps may remain. 711 */ 712 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) { 713 if (pmp->pfs_hmps[i]) 714 break; 715 } 716 iroot->cluster.nchains = i + 1; 717 718 /* 719 * If the PMP has no elements remaining we can destroy it. 720 * (this will transition management threads from frozen->exit). 721 */ 722 if (iroot->cluster.nchains == 0) { 723 kprintf("unmount hmp %p last ref to PMP=%p\n", 724 hmp, pmp); 725 hammer2_pfsfree(pmp); 726 goto again; 727 } 728 729 /* 730 * If elements still remain we need to set the REMASTER 731 * flag and unfreeze it. 732 */ 733 if (didfreeze) { 734 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 735 if (pmp->pfs_hmps[i] == NULL) 736 continue; 737 hammer2_thr_remaster(&pmp->sync_thrs[i]); 738 hammer2_thr_unfreeze(&pmp->sync_thrs[i]); 739 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) { 740 hammer2_thr_remaster( 741 &pmp->xop_groups[j].thrs[i]); 742 hammer2_thr_unfreeze( 743 &pmp->xop_groups[j].thrs[i]); 744 } 745 } 746 } 747 } 748 } 749 750 /* 751 * Mount or remount HAMMER2 fileystem from physical media 752 * 753 * mountroot 754 * mp mount point structure 755 * path NULL 756 * data <unused> 757 * cred <unused> 758 * 759 * mount 760 * mp mount point structure 761 * path path to mount point 762 * data pointer to argument structure in user space 763 * volume volume path (device@LABEL form) 764 * hflags user mount flags 765 * cred user credentials 766 * 767 * RETURNS: 0 Success 768 * !0 error number 769 */ 770 static 771 int 772 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data, 773 struct ucred *cred) 774 { 775 struct hammer2_mount_info info; 776 hammer2_pfs_t *pmp; 777 hammer2_pfs_t *spmp; 778 hammer2_dev_t *hmp; 779 hammer2_dev_t *force_local; 780 hammer2_key_t key_next; 781 hammer2_key_t key_dummy; 782 hammer2_key_t lhc; 783 struct vnode *devvp; 784 struct nlookupdata nd; 785 hammer2_chain_t *parent; 786 hammer2_chain_t *chain; 787 hammer2_cluster_t *cluster; 788 const hammer2_inode_data_t *ripdata; 789 hammer2_blockref_t bref; 790 struct file *fp; 791 char devstr[MNAMELEN]; 792 size_t size; 793 size_t done; 794 char *dev; 795 char *label; 796 int ronly = 1; 797 int error; 798 int cache_index; 799 int i; 800 801 hmp = NULL; 802 pmp = NULL; 803 dev = NULL; 804 label = NULL; 805 devvp = NULL; 806 cache_index = -1; 807 808 kprintf("hammer2_mount\n"); 809 810 if (path == NULL) { 811 /* 812 * Root mount 813 */ 814 bzero(&info, sizeof(info)); 815 info.cluster_fd = -1; 816 ksnprintf(devstr, sizeof(devstr), "%s", 817 mp->mnt_stat.f_mntfromname); 818 kprintf("hammer2_mount: root '%s'\n", devstr); 819 } else { 820 /* 821 * Non-root mount or updating a mount 822 */ 823 error = copyin(data, &info, sizeof(info)); 824 if (error) 825 return (error); 826 827 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done); 828 if (error) 829 return (error); 830 } 831 832 /* Extract device and label */ 833 dev = devstr; 834 label = strchr(devstr, '@'); 835 if (label == NULL || 836 ((label + 1) - dev) > done) { 837 return (EINVAL); 838 } 839 *label = '\0'; 840 label++; 841 if (*label == '\0') 842 return (EINVAL); 843 844 if (mp->mnt_flag & MNT_UPDATE) { 845 /* 846 * Update mount. Note that pmp->iroot->cluster is 847 * an inode-embedded cluster and thus cannot be 848 * directly locked. 849 * 850 * XXX HAMMER2 needs to implement NFS export via 851 * mountctl. 852 */ 853 pmp = MPTOPMP(mp); 854 pmp->hflags = info.hflags; 855 cluster = &pmp->iroot->cluster; 856 for (i = 0; i < cluster->nchains; ++i) { 857 if (cluster->array[i].chain == NULL) 858 continue; 859 hmp = cluster->array[i].chain->hmp; 860 devvp = hmp->devvp; 861 error = hammer2_remount(hmp, mp, path, 862 devvp, cred); 863 if (error) 864 break; 865 } 866 867 return error; 868 } 869 870 /* 871 * HMP device mount 872 * 873 * Lookup name and verify it refers to a block device. 874 */ 875 if (path) { 876 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW); 877 if (error == 0) 878 error = nlookup(&nd); 879 if (error == 0) 880 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp); 881 nlookup_done(&nd); 882 } else { 883 /* root mount */ 884 cdev_t cdev = kgetdiskbyname(dev); 885 error = bdevvp(cdev, &devvp); 886 if (error) 887 kprintf("hammer2: cannot find '%s'\n", dev); 888 } 889 890 if (error == 0) { 891 if (vn_isdisk(devvp, &error)) 892 error = vfs_mountedon(devvp); 893 } 894 895 /* 896 * Determine if the device has already been mounted. After this 897 * check hmp will be non-NULL if we are doing the second or more 898 * hammer2 mounts from the same device. 899 */ 900 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 901 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 902 if (hmp->devvp == devvp) 903 break; 904 } 905 906 /* 907 * Open the device if this isn't a secondary mount and construct 908 * the H2 device mount (hmp). 909 */ 910 if (hmp == NULL) { 911 hammer2_chain_t *schain; 912 hammer2_xid_t xid; 913 914 if (error == 0 && vcount(devvp) > 0) 915 error = EBUSY; 916 917 /* 918 * Now open the device 919 */ 920 if (error == 0) { 921 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0); 922 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 923 error = vinvalbuf(devvp, V_SAVE, 0, 0); 924 if (error == 0) { 925 error = VOP_OPEN(devvp, 926 ronly ? FREAD : FREAD | FWRITE, 927 FSCRED, NULL); 928 } 929 vn_unlock(devvp); 930 } 931 if (error && devvp) { 932 vrele(devvp); 933 devvp = NULL; 934 } 935 if (error) { 936 lockmgr(&hammer2_mntlk, LK_RELEASE); 937 return error; 938 } 939 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO); 940 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev); 941 hmp->ronly = ronly; 942 hmp->devvp = devvp; 943 hmp->hflags = info.hflags & HMNT2_DEVFLAGS; 944 kmalloc_create(&hmp->mchain, "HAMMER2-chains"); 945 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry); 946 RB_INIT(&hmp->iotree); 947 spin_init(&hmp->io_spin, "hm2mount_io"); 948 spin_init(&hmp->list_spin, "hm2mount_list"); 949 TAILQ_INIT(&hmp->flushq); 950 951 lockinit(&hmp->vollk, "h2vol", 0, 0); 952 lockinit(&hmp->bulklk, "h2bulk", 0, 0); 953 954 /* 955 * vchain setup. vchain.data is embedded. 956 * vchain.refs is initialized and will never drop to 0. 957 * 958 * NOTE! voldata is not yet loaded. 959 */ 960 hmp->vchain.hmp = hmp; 961 hmp->vchain.refs = 1; 962 hmp->vchain.data = (void *)&hmp->voldata; 963 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME; 964 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 965 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 966 967 hammer2_chain_core_init(&hmp->vchain); 968 /* hmp->vchain.u.xxx is left NULL */ 969 970 /* 971 * fchain setup. fchain.data is embedded. 972 * fchain.refs is initialized and will never drop to 0. 973 * 974 * The data is not used but needs to be initialized to 975 * pass assertion muster. We use this chain primarily 976 * as a placeholder for the freemap's top-level RBTREE 977 * so it does not interfere with the volume's topology 978 * RBTREE. 979 */ 980 hmp->fchain.hmp = hmp; 981 hmp->fchain.refs = 1; 982 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset; 983 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP; 984 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 985 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 986 hmp->fchain.bref.methods = 987 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) | 988 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE); 989 990 hammer2_chain_core_init(&hmp->fchain); 991 /* hmp->fchain.u.xxx is left NULL */ 992 993 /* 994 * Install the volume header and initialize fields from 995 * voldata. 996 */ 997 error = hammer2_install_volume_header(hmp); 998 if (error) { 999 hammer2_unmount_helper(mp, NULL, hmp); 1000 lockmgr(&hammer2_mntlk, LK_RELEASE); 1001 hammer2_vfs_unmount(mp, MNT_FORCE); 1002 return error; 1003 } 1004 1005 /* 1006 * Really important to get these right or flush will get 1007 * confused. 1008 */ 1009 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL); 1010 kprintf("alloc spmp %p tid %016jx\n", 1011 hmp->spmp, hmp->voldata.mirror_tid); 1012 spmp = hmp->spmp; 1013 1014 /* 1015 * Dummy-up vchain and fchain's modify_tid. mirror_tid 1016 * is inherited from the volume header. 1017 */ 1018 xid = 0; 1019 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1020 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid; 1021 hmp->vchain.pmp = spmp; 1022 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1023 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid; 1024 hmp->fchain.pmp = spmp; 1025 1026 /* 1027 * First locate the super-root inode, which is key 0 1028 * relative to the volume header's blockset. 1029 * 1030 * Then locate the root inode by scanning the directory keyspace 1031 * represented by the label. 1032 */ 1033 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 1034 schain = hammer2_chain_lookup(&parent, &key_dummy, 1035 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY, 1036 &cache_index, 0); 1037 hammer2_chain_lookup_done(parent); 1038 if (schain == NULL) { 1039 kprintf("hammer2_mount: invalid super-root\n"); 1040 hammer2_unmount_helper(mp, NULL, hmp); 1041 lockmgr(&hammer2_mntlk, LK_RELEASE); 1042 hammer2_vfs_unmount(mp, MNT_FORCE); 1043 return EINVAL; 1044 } 1045 if (schain->error) { 1046 kprintf("hammer2_mount: error %s reading super-root\n", 1047 hammer2_error_str(schain->error)); 1048 hammer2_chain_unlock(schain); 1049 hammer2_chain_drop(schain); 1050 schain = NULL; 1051 hammer2_unmount_helper(mp, NULL, hmp); 1052 lockmgr(&hammer2_mntlk, LK_RELEASE); 1053 hammer2_vfs_unmount(mp, MNT_FORCE); 1054 return EINVAL; 1055 } 1056 1057 /* 1058 * The super-root always uses an inode_tid of 1 when 1059 * creating PFSs. 1060 */ 1061 spmp->inode_tid = 1; 1062 spmp->modify_tid = schain->bref.modify_tid + 1; 1063 1064 /* 1065 * Sanity-check schain's pmp and finish initialization. 1066 * Any chain belonging to the super-root topology should 1067 * have a NULL pmp (not even set to spmp). 1068 */ 1069 ripdata = &hammer2_chain_rdata(schain)->ipdata; 1070 KKASSERT(schain->pmp == NULL); 1071 spmp->pfs_clid = ripdata->meta.pfs_clid; 1072 1073 /* 1074 * Replace the dummy spmp->iroot with a real one. It's 1075 * easier to just do a wholesale replacement than to try 1076 * to update the chain and fixup the iroot fields. 1077 * 1078 * The returned inode is locked with the supplied cluster. 1079 */ 1080 cluster = hammer2_cluster_from_chain(schain); 1081 hammer2_inode_drop(spmp->iroot); 1082 spmp->iroot = NULL; 1083 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1); 1084 spmp->spmp_hmp = hmp; 1085 spmp->pfs_types[0] = ripdata->meta.pfs_type; 1086 spmp->pfs_hmps[0] = hmp; 1087 hammer2_inode_ref(spmp->iroot); 1088 hammer2_inode_unlock(spmp->iroot); 1089 hammer2_cluster_unlock(cluster); 1090 hammer2_cluster_drop(cluster); 1091 schain = NULL; 1092 /* leave spmp->iroot with one ref */ 1093 1094 if ((mp->mnt_flag & MNT_RDONLY) == 0) { 1095 error = hammer2_recovery(hmp); 1096 /* XXX do something with error */ 1097 } 1098 hammer2_update_pmps(hmp); 1099 hammer2_iocom_init(hmp); 1100 1101 /* 1102 * Ref the cluster management messaging descriptor. The mount 1103 * program deals with the other end of the communications pipe. 1104 * 1105 * Root mounts typically do not supply one. 1106 */ 1107 if (info.cluster_fd >= 0) { 1108 fp = holdfp(curproc->p_fd, info.cluster_fd, -1); 1109 if (fp) { 1110 hammer2_cluster_reconnect(hmp, fp); 1111 } else { 1112 kprintf("hammer2_mount: bad cluster_fd!\n"); 1113 } 1114 } 1115 } else { 1116 spmp = hmp->spmp; 1117 if (info.hflags & HMNT2_DEVFLAGS) { 1118 kprintf("hammer2: Warning: mount flags pertaining " 1119 "to the whole device may only be specified " 1120 "on the first mount of the device: %08x\n", 1121 info.hflags & HMNT2_DEVFLAGS); 1122 } 1123 } 1124 1125 /* 1126 * Force local mount (disassociate all PFSs from their clusters). 1127 * Used primarily for debugging. 1128 */ 1129 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1130 1131 /* 1132 * Lookup the mount point under the media-localized super-root. 1133 * Scanning hammer2_pfslist doesn't help us because it represents 1134 * PFS cluster ids which can aggregate several named PFSs together. 1135 * 1136 * cluster->pmp will incorrectly point to spmp and must be fixed 1137 * up later on. 1138 */ 1139 hammer2_inode_lock(spmp->iroot, 0); 1140 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1141 lhc = hammer2_dirhash(label, strlen(label)); 1142 chain = hammer2_chain_lookup(&parent, &key_next, 1143 lhc, lhc + HAMMER2_DIRHASH_LOMASK, 1144 &cache_index, 0); 1145 while (chain) { 1146 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && 1147 strcmp(label, chain->data->ipdata.filename) == 0) { 1148 break; 1149 } 1150 chain = hammer2_chain_next(&parent, chain, &key_next, 1151 key_next, 1152 lhc + HAMMER2_DIRHASH_LOMASK, 1153 &cache_index, 0); 1154 } 1155 if (parent) { 1156 hammer2_chain_unlock(parent); 1157 hammer2_chain_drop(parent); 1158 } 1159 hammer2_inode_unlock(spmp->iroot); 1160 1161 /* 1162 * PFS could not be found? 1163 */ 1164 if (chain == NULL) { 1165 kprintf("hammer2_mount: PFS label not found\n"); 1166 hammer2_unmount_helper(mp, NULL, hmp); 1167 lockmgr(&hammer2_mntlk, LK_RELEASE); 1168 hammer2_vfs_unmount(mp, MNT_FORCE); 1169 1170 return EINVAL; 1171 } 1172 1173 /* 1174 * Acquire the pmp structure (it should have already been allocated 1175 * via hammer2_update_pmps() so do not pass cluster in to add to 1176 * available chains). 1177 * 1178 * Check if the cluster has already been mounted. A cluster can 1179 * only be mounted once, use null mounts to mount additional copies. 1180 */ 1181 ripdata = &chain->data->ipdata; 1182 bref = chain->bref; 1183 pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid, force_local); 1184 hammer2_chain_unlock(chain); 1185 hammer2_chain_drop(chain); 1186 1187 if (pmp->mp) { 1188 kprintf("hammer2_mount: PFS already mounted!\n"); 1189 hammer2_unmount_helper(mp, NULL, hmp); 1190 lockmgr(&hammer2_mntlk, LK_RELEASE); 1191 hammer2_vfs_unmount(mp, MNT_FORCE); 1192 1193 return EBUSY; 1194 } 1195 1196 /* 1197 * Finish the mount 1198 */ 1199 kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp); 1200 1201 pmp->hflags = info.hflags; 1202 mp->mnt_flag = MNT_LOCAL; 1203 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */ 1204 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */ 1205 1206 /* 1207 * required mount structure initializations 1208 */ 1209 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE; 1210 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE; 1211 1212 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE; 1213 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE; 1214 1215 /* 1216 * Optional fields 1217 */ 1218 mp->mnt_iosize_max = MAXPHYS; 1219 1220 /* 1221 * Connect up mount pointers. 1222 */ 1223 hammer2_mount_helper(mp, pmp); 1224 1225 lockmgr(&hammer2_mntlk, LK_RELEASE); 1226 1227 /* 1228 * Finish setup 1229 */ 1230 vfs_getnewfsid(mp); 1231 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops); 1232 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops); 1233 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops); 1234 1235 if (path) { 1236 copyinstr(info.volume, mp->mnt_stat.f_mntfromname, 1237 MNAMELEN - 1, &size); 1238 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size); 1239 } /* else root mount, already in there */ 1240 1241 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname)); 1242 if (path) { 1243 copyinstr(path, mp->mnt_stat.f_mntonname, 1244 sizeof(mp->mnt_stat.f_mntonname) - 1, 1245 &size); 1246 } else { 1247 /* root mount */ 1248 mp->mnt_stat.f_mntonname[0] = '/'; 1249 } 1250 1251 /* 1252 * Initial statfs to prime mnt_stat. 1253 */ 1254 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred); 1255 1256 return 0; 1257 } 1258 1259 /* 1260 * Scan PFSs under the super-root and create hammer2_pfs structures. 1261 */ 1262 static 1263 void 1264 hammer2_update_pmps(hammer2_dev_t *hmp) 1265 { 1266 const hammer2_inode_data_t *ripdata; 1267 hammer2_chain_t *parent; 1268 hammer2_chain_t *chain; 1269 hammer2_blockref_t bref; 1270 hammer2_dev_t *force_local; 1271 hammer2_pfs_t *spmp; 1272 hammer2_pfs_t *pmp; 1273 hammer2_key_t key_next; 1274 int cache_index = -1; 1275 1276 /* 1277 * Force local mount (disassociate all PFSs from their clusters). 1278 * Used primarily for debugging. 1279 */ 1280 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1281 1282 /* 1283 * Lookup mount point under the media-localized super-root. 1284 * 1285 * cluster->pmp will incorrectly point to spmp and must be fixed 1286 * up later on. 1287 */ 1288 spmp = hmp->spmp; 1289 hammer2_inode_lock(spmp->iroot, 0); 1290 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1291 chain = hammer2_chain_lookup(&parent, &key_next, 1292 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 1293 &cache_index, 0); 1294 while (chain) { 1295 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) 1296 continue; 1297 ripdata = &chain->data->ipdata; 1298 bref = chain->bref; 1299 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename); 1300 1301 pmp = hammer2_pfsalloc(chain, ripdata, 1302 bref.modify_tid, force_local); 1303 chain = hammer2_chain_next(&parent, chain, &key_next, 1304 key_next, HAMMER2_KEY_MAX, 1305 &cache_index, 0); 1306 } 1307 if (parent) { 1308 hammer2_chain_unlock(parent); 1309 hammer2_chain_drop(parent); 1310 } 1311 hammer2_inode_unlock(spmp->iroot); 1312 } 1313 1314 static 1315 int 1316 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused, 1317 struct vnode *devvp, struct ucred *cred) 1318 { 1319 int error; 1320 1321 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) { 1322 error = hammer2_recovery(hmp); 1323 } else { 1324 error = 0; 1325 } 1326 return error; 1327 } 1328 1329 static 1330 int 1331 hammer2_vfs_unmount(struct mount *mp, int mntflags) 1332 { 1333 hammer2_pfs_t *pmp; 1334 int flags; 1335 int error = 0; 1336 1337 pmp = MPTOPMP(mp); 1338 1339 if (pmp == NULL) 1340 return(0); 1341 1342 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1343 1344 /* 1345 * If mount initialization proceeded far enough we must flush 1346 * its vnodes and sync the underlying mount points. Three syncs 1347 * are required to fully flush the filesystem (freemap updates lag 1348 * by one flush, and one extra for safety). 1349 */ 1350 if (mntflags & MNT_FORCE) 1351 flags = FORCECLOSE; 1352 else 1353 flags = 0; 1354 if (pmp->iroot) { 1355 error = vflush(mp, 0, flags); 1356 if (error) 1357 goto failed; 1358 hammer2_vfs_sync(mp, MNT_WAIT); 1359 hammer2_vfs_sync(mp, MNT_WAIT); 1360 hammer2_vfs_sync(mp, MNT_WAIT); 1361 } 1362 1363 /* 1364 * Cleanup the frontend support XOPS threads 1365 */ 1366 hammer2_xop_helper_cleanup(pmp); 1367 1368 /* 1369 * Cleanup our reference on ihidden. 1370 */ 1371 if (pmp->ihidden) { 1372 hammer2_inode_drop(pmp->ihidden); 1373 pmp->ihidden = NULL; 1374 } 1375 if (pmp->mp) 1376 hammer2_unmount_helper(mp, pmp, NULL); 1377 1378 error = 0; 1379 failed: 1380 lockmgr(&hammer2_mntlk, LK_RELEASE); 1381 1382 return (error); 1383 } 1384 1385 /* 1386 * Mount helper, hook the system mount into our PFS. 1387 * The mount lock is held. 1388 * 1389 * We must bump the mount_count on related devices for any 1390 * mounted PFSs. 1391 */ 1392 static 1393 void 1394 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp) 1395 { 1396 hammer2_cluster_t *cluster; 1397 hammer2_chain_t *rchain; 1398 int i; 1399 1400 mp->mnt_data = (qaddr_t)pmp; 1401 pmp->mp = mp; 1402 1403 /* 1404 * After pmp->mp is set we have to adjust hmp->mount_count. 1405 */ 1406 cluster = &pmp->iroot->cluster; 1407 for (i = 0; i < cluster->nchains; ++i) { 1408 rchain = cluster->array[i].chain; 1409 if (rchain == NULL) 1410 continue; 1411 ++rchain->hmp->mount_count; 1412 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n", 1413 rchain->hmp, rchain->hmp->mount_count); 1414 } 1415 1416 /* 1417 * Create missing Xop threads 1418 */ 1419 hammer2_xop_helper_create(pmp); 1420 } 1421 1422 /* 1423 * Mount helper, unhook the system mount from our PFS. 1424 * The mount lock is held. 1425 * 1426 * If hmp is supplied a mount responsible for being the first to open 1427 * the block device failed and the block device and all PFSs using the 1428 * block device must be cleaned up. 1429 * 1430 * If pmp is supplied multiple devices might be backing the PFS and each 1431 * must be disconnect. This might not be the last PFS using some of the 1432 * underlying devices. Also, we have to adjust our hmp->mount_count 1433 * accounting for the devices backing the pmp which is now undergoing an 1434 * unmount. 1435 */ 1436 static 1437 void 1438 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp) 1439 { 1440 hammer2_cluster_t *cluster; 1441 hammer2_chain_t *rchain; 1442 struct vnode *devvp; 1443 int dumpcnt; 1444 int ronly = 0; 1445 int i; 1446 1447 /* 1448 * If no device supplied this is a high-level unmount and we have to 1449 * to disconnect the mount, adjust mount_count, and locate devices 1450 * that might now have no mounts. 1451 */ 1452 if (pmp) { 1453 KKASSERT(hmp == NULL); 1454 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp); 1455 pmp->mp = NULL; 1456 mp->mnt_data = NULL; 1457 1458 /* 1459 * After pmp->mp is cleared we have to account for 1460 * mount_count. 1461 */ 1462 cluster = &pmp->iroot->cluster; 1463 for (i = 0; i < cluster->nchains; ++i) { 1464 rchain = cluster->array[i].chain; 1465 if (rchain == NULL) 1466 continue; 1467 --rchain->hmp->mount_count; 1468 kprintf("hammer2_unmount hmp=%p --mount_count=%d\n", 1469 rchain->hmp, rchain->hmp->mount_count); 1470 /* scrapping hmp now may invalidate the pmp */ 1471 } 1472 again: 1473 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1474 if (hmp->mount_count == 0) { 1475 hammer2_unmount_helper(NULL, NULL, hmp); 1476 goto again; 1477 } 1478 } 1479 return; 1480 } 1481 1482 /* 1483 * Try to terminate the block device. We can't terminate it if 1484 * there are still PFSs referencing it. 1485 */ 1486 kprintf("hammer2_unmount hmp=%p mount_count=%d\n", 1487 hmp, hmp->mount_count); 1488 if (hmp->mount_count) 1489 return; 1490 1491 hammer2_pfsfree_scan(hmp); 1492 hammer2_dev_exlock(hmp); /* XXX order */ 1493 1494 /* 1495 * Cycle the volume data lock as a safety (probably not needed any 1496 * more). To ensure everything is out we need to flush at least 1497 * three times. (1) The running of the sideq can dirty the 1498 * filesystem, (2) A normal flush can dirty the freemap, and 1499 * (3) ensure that the freemap is fully synchronized. 1500 * 1501 * The next mount's recovery scan can clean everything up but we want 1502 * to leave the filesystem in a 100% clean state on a normal unmount. 1503 */ 1504 #if 0 1505 hammer2_voldata_lock(hmp); 1506 hammer2_voldata_unlock(hmp); 1507 #endif 1508 hammer2_iocom_uninit(hmp); 1509 1510 if ((hmp->vchain.flags | hmp->fchain.flags) & 1511 HAMMER2_CHAIN_FLUSH_MASK) { 1512 kprintf("hammer2_unmount: chains left over " 1513 "after final sync\n"); 1514 kprintf(" vchain %08x\n", hmp->vchain.flags); 1515 kprintf(" fchain %08x\n", hmp->fchain.flags); 1516 1517 if (hammer2_debug & 0x0010) 1518 Debugger("entered debugger"); 1519 } 1520 1521 KKASSERT(hmp->spmp == NULL); 1522 1523 /* 1524 * Finish up with the device vnode 1525 */ 1526 if ((devvp = hmp->devvp) != NULL) { 1527 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1528 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0); 1529 hmp->devvp = NULL; 1530 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL); 1531 vn_unlock(devvp); 1532 vrele(devvp); 1533 devvp = NULL; 1534 } 1535 1536 /* 1537 * Clear vchain/fchain flags that might prevent final cleanup 1538 * of these chains. 1539 */ 1540 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) { 1541 atomic_add_long(&hammer2_count_modified_chains, -1); 1542 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 1543 hammer2_pfs_memory_wakeup(hmp->vchain.pmp); 1544 } 1545 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) { 1546 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE); 1547 } 1548 1549 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) { 1550 atomic_add_long(&hammer2_count_modified_chains, -1); 1551 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED); 1552 hammer2_pfs_memory_wakeup(hmp->fchain.pmp); 1553 } 1554 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) { 1555 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE); 1556 } 1557 1558 /* 1559 * Final drop of embedded freemap root chain to 1560 * clean up fchain.core (fchain structure is not 1561 * flagged ALLOCATED so it is cleaned out and then 1562 * left to rot). 1563 */ 1564 hammer2_chain_drop(&hmp->fchain); 1565 1566 /* 1567 * Final drop of embedded volume root chain to clean 1568 * up vchain.core (vchain structure is not flagged 1569 * ALLOCATED so it is cleaned out and then left to 1570 * rot). 1571 */ 1572 dumpcnt = 50; 1573 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v'); 1574 dumpcnt = 50; 1575 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f'); 1576 hammer2_dev_unlock(hmp); 1577 hammer2_chain_drop(&hmp->vchain); 1578 1579 hammer2_io_cleanup(hmp, &hmp->iotree); 1580 if (hmp->iofree_count) { 1581 kprintf("io_cleanup: %d I/O's left hanging\n", 1582 hmp->iofree_count); 1583 } 1584 1585 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry); 1586 kmalloc_destroy(&hmp->mchain); 1587 kfree(hmp, M_HAMMER2); 1588 } 1589 1590 static 1591 int 1592 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp, 1593 ino_t ino, struct vnode **vpp) 1594 { 1595 kprintf("hammer2_vget\n"); 1596 return (EOPNOTSUPP); 1597 } 1598 1599 static 1600 int 1601 hammer2_vfs_root(struct mount *mp, struct vnode **vpp) 1602 { 1603 hammer2_pfs_t *pmp; 1604 int error; 1605 struct vnode *vp; 1606 1607 pmp = MPTOPMP(mp); 1608 if (pmp->iroot == NULL) { 1609 *vpp = NULL; 1610 return EINVAL; 1611 } 1612 1613 error = 0; 1614 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1615 1616 while (pmp->inode_tid == 0) { 1617 hammer2_xop_ipcluster_t *xop; 1618 hammer2_inode_meta_t *meta; 1619 1620 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING); 1621 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster); 1622 error = hammer2_xop_collect(&xop->head, 0); 1623 1624 if (error == 0) { 1625 meta = &xop->head.cluster.focus->data->ipdata.meta; 1626 pmp->iroot->meta = *meta; 1627 pmp->inode_tid = meta->pfs_inum + 1; 1628 if (pmp->inode_tid < HAMMER2_INODE_START) 1629 pmp->inode_tid = HAMMER2_INODE_START; 1630 pmp->modify_tid = 1631 xop->head.cluster.focus->bref.modify_tid + 1; 1632 kprintf("PFS: Starting inode %jd\n", 1633 (intmax_t)pmp->inode_tid); 1634 kprintf("PMP focus good set nextino=%ld mod=%016jx\n", 1635 pmp->inode_tid, pmp->modify_tid); 1636 wakeup(&pmp->iroot); 1637 1638 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1639 1640 /* 1641 * Prime the mount info. 1642 */ 1643 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL); 1644 1645 /* 1646 * With the cluster operational, check for and 1647 * install ihidden if needed. The install_hidden 1648 * code needs to get a transaction so we must unlock 1649 * iroot around it. 1650 * 1651 * This is only applicable PFS mounts, there is no 1652 * hidden directory in the spmp. 1653 */ 1654 hammer2_inode_unlock(pmp->iroot); 1655 hammer2_inode_install_hidden(pmp); 1656 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1657 1658 break; 1659 } 1660 1661 /* 1662 * Loop, try again 1663 */ 1664 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1665 hammer2_inode_unlock(pmp->iroot); 1666 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz); 1667 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1668 if (error == EINTR) 1669 break; 1670 } 1671 1672 if (error) { 1673 hammer2_inode_unlock(pmp->iroot); 1674 *vpp = NULL; 1675 } else { 1676 vp = hammer2_igetv(pmp->iroot, &error); 1677 hammer2_inode_unlock(pmp->iroot); 1678 *vpp = vp; 1679 } 1680 1681 return (error); 1682 } 1683 1684 /* 1685 * Filesystem status 1686 * 1687 * XXX incorporate ipdata->meta.inode_quota and data_quota 1688 */ 1689 static 1690 int 1691 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred) 1692 { 1693 hammer2_pfs_t *pmp; 1694 hammer2_dev_t *hmp; 1695 hammer2_blockref_t bref; 1696 int i; 1697 1698 /* 1699 * NOTE: iroot might not have validated the cluster yet. 1700 */ 1701 pmp = MPTOPMP(mp); 1702 1703 mp->mnt_stat.f_files = 0; 1704 mp->mnt_stat.f_ffree = 0; 1705 mp->mnt_stat.f_blocks = 0; 1706 mp->mnt_stat.f_bfree = 0; 1707 mp->mnt_stat.f_bavail = 0; 1708 1709 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 1710 hmp = pmp->pfs_hmps[i]; 1711 if (hmp == NULL) 1712 continue; 1713 if (pmp->iroot->cluster.array[i].chain) 1714 bref = pmp->iroot->cluster.array[i].chain->bref; 1715 else 1716 bzero(&bref, sizeof(bref)); 1717 1718 mp->mnt_stat.f_files = bref.inode_count; 1719 mp->mnt_stat.f_ffree = 0; 1720 mp->mnt_stat.f_blocks = (bref.data_count + 1721 hmp->voldata.allocator_free) / 1722 mp->mnt_vstat.f_bsize; 1723 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free / 1724 mp->mnt_vstat.f_bsize; 1725 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree; 1726 1727 *sbp = mp->mnt_stat; 1728 } 1729 return (0); 1730 } 1731 1732 static 1733 int 1734 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred) 1735 { 1736 hammer2_pfs_t *pmp; 1737 hammer2_dev_t *hmp; 1738 hammer2_blockref_t bref; 1739 int i; 1740 1741 /* 1742 * NOTE: iroot might not have validated the cluster yet. 1743 */ 1744 pmp = MPTOPMP(mp); 1745 1746 mp->mnt_vstat.f_bsize = 0; 1747 mp->mnt_vstat.f_files = 0; 1748 mp->mnt_vstat.f_ffree = 0; 1749 mp->mnt_vstat.f_blocks = 0; 1750 mp->mnt_vstat.f_bfree = 0; 1751 mp->mnt_vstat.f_bavail = 0; 1752 1753 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 1754 hmp = pmp->pfs_hmps[i]; 1755 if (hmp == NULL) 1756 continue; 1757 if (pmp->iroot->cluster.array[i].chain) 1758 bref = pmp->iroot->cluster.array[i].chain->bref; 1759 else 1760 bzero(&bref, sizeof(bref)); 1761 1762 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE; 1763 mp->mnt_vstat.f_files = bref.inode_count; 1764 mp->mnt_vstat.f_ffree = 0; 1765 mp->mnt_vstat.f_blocks = (bref.data_count + 1766 hmp->voldata.allocator_free) / 1767 mp->mnt_vstat.f_bsize; 1768 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free / 1769 mp->mnt_vstat.f_bsize; 1770 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree; 1771 1772 *sbp = mp->mnt_vstat; 1773 } 1774 return (0); 1775 } 1776 1777 /* 1778 * Mount-time recovery (RW mounts) 1779 * 1780 * Updates to the free block table are allowed to lag flushes by one 1781 * transaction. In case of a crash, then on a fresh mount we must do an 1782 * incremental scan of the last committed transaction id and make sure that 1783 * all related blocks have been marked allocated. 1784 * 1785 * The super-root topology and each PFS has its own transaction id domain, 1786 * so we must track PFS boundary transitions. 1787 */ 1788 struct hammer2_recovery_elm { 1789 TAILQ_ENTRY(hammer2_recovery_elm) entry; 1790 hammer2_chain_t *chain; 1791 hammer2_tid_t sync_tid; 1792 }; 1793 1794 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm); 1795 1796 struct hammer2_recovery_info { 1797 struct hammer2_recovery_list list; 1798 hammer2_tid_t mtid; 1799 int depth; 1800 }; 1801 1802 static int hammer2_recovery_scan(hammer2_dev_t *hmp, 1803 hammer2_chain_t *parent, 1804 struct hammer2_recovery_info *info, 1805 hammer2_tid_t sync_tid); 1806 1807 #define HAMMER2_RECOVERY_MAXDEPTH 10 1808 1809 static 1810 int 1811 hammer2_recovery(hammer2_dev_t *hmp) 1812 { 1813 struct hammer2_recovery_info info; 1814 struct hammer2_recovery_elm *elm; 1815 hammer2_chain_t *parent; 1816 hammer2_tid_t sync_tid; 1817 hammer2_tid_t mirror_tid; 1818 int error; 1819 int cumulative_error = 0; 1820 1821 hammer2_trans_init(hmp->spmp, 0); 1822 1823 sync_tid = hmp->voldata.freemap_tid; 1824 mirror_tid = hmp->voldata.mirror_tid; 1825 1826 kprintf("hammer2 mount \"%s\": ", hmp->devrepname); 1827 if (sync_tid >= mirror_tid) { 1828 kprintf(" no recovery needed\n"); 1829 } else { 1830 kprintf(" freemap recovery %016jx-%016jx\n", 1831 sync_tid + 1, mirror_tid); 1832 } 1833 1834 TAILQ_INIT(&info.list); 1835 info.depth = 0; 1836 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 1837 cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid); 1838 hammer2_chain_lookup_done(parent); 1839 1840 while ((elm = TAILQ_FIRST(&info.list)) != NULL) { 1841 TAILQ_REMOVE(&info.list, elm, entry); 1842 parent = elm->chain; 1843 sync_tid = elm->sync_tid; 1844 kfree(elm, M_HAMMER2); 1845 1846 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 1847 error = hammer2_recovery_scan(hmp, parent, &info, 1848 hmp->voldata.freemap_tid); 1849 hammer2_chain_unlock(parent); 1850 hammer2_chain_drop(parent); /* drop elm->chain ref */ 1851 if (error) 1852 cumulative_error = error; 1853 } 1854 hammer2_trans_done(hmp->spmp); 1855 1856 return cumulative_error; 1857 } 1858 1859 static 1860 int 1861 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent, 1862 struct hammer2_recovery_info *info, 1863 hammer2_tid_t sync_tid) 1864 { 1865 const hammer2_inode_data_t *ripdata; 1866 hammer2_chain_t *chain; 1867 hammer2_blockref_t bref; 1868 int cache_index; 1869 int cumulative_error = 0; 1870 int error; 1871 int first; 1872 1873 /* 1874 * Adjust freemap to ensure that the block(s) are marked allocated. 1875 */ 1876 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) { 1877 hammer2_freemap_adjust(hmp, &parent->bref, 1878 HAMMER2_FREEMAP_DORECOVER); 1879 } 1880 1881 /* 1882 * Check type for recursive scan 1883 */ 1884 switch(parent->bref.type) { 1885 case HAMMER2_BREF_TYPE_VOLUME: 1886 /* data already instantiated */ 1887 break; 1888 case HAMMER2_BREF_TYPE_INODE: 1889 /* 1890 * Must instantiate data for DIRECTDATA test and also 1891 * for recursion. 1892 */ 1893 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 1894 ripdata = &hammer2_chain_rdata(parent)->ipdata; 1895 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) { 1896 /* not applicable to recovery scan */ 1897 hammer2_chain_unlock(parent); 1898 return 0; 1899 } 1900 hammer2_chain_unlock(parent); 1901 break; 1902 case HAMMER2_BREF_TYPE_INDIRECT: 1903 /* 1904 * Must instantiate data for recursion 1905 */ 1906 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 1907 hammer2_chain_unlock(parent); 1908 break; 1909 case HAMMER2_BREF_TYPE_DATA: 1910 case HAMMER2_BREF_TYPE_FREEMAP: 1911 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1912 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 1913 /* not applicable to recovery scan */ 1914 return 0; 1915 break; 1916 default: 1917 return EDOM; 1918 } 1919 1920 /* 1921 * Defer operation if depth limit reached or if we are crossing a 1922 * PFS boundary. 1923 */ 1924 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) { 1925 struct hammer2_recovery_elm *elm; 1926 1927 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK); 1928 elm->chain = parent; 1929 elm->sync_tid = sync_tid; 1930 hammer2_chain_ref(parent); 1931 TAILQ_INSERT_TAIL(&info->list, elm, entry); 1932 /* unlocked by caller */ 1933 1934 return(0); 1935 } 1936 1937 1938 /* 1939 * Recursive scan of the last flushed transaction only. We are 1940 * doing this without pmp assignments so don't leave the chains 1941 * hanging around after we are done with them. 1942 */ 1943 cache_index = 0; 1944 chain = NULL; 1945 first = 1; 1946 1947 while (hammer2_chain_scan(parent, &chain, &bref, 1948 &first, &cache_index, 1949 HAMMER2_LOOKUP_NODATA) != NULL) { 1950 /* 1951 * If this is a leaf 1952 */ 1953 if (chain == NULL) { 1954 if (bref.mirror_tid > sync_tid) { 1955 hammer2_freemap_adjust(hmp, &bref, 1956 HAMMER2_FREEMAP_DORECOVER); 1957 } 1958 continue; 1959 } 1960 1961 /* 1962 * This may or may not be a recursive node. 1963 */ 1964 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 1965 if (bref.mirror_tid > sync_tid) { 1966 ++info->depth; 1967 error = hammer2_recovery_scan(hmp, chain, 1968 info, sync_tid); 1969 --info->depth; 1970 if (error) 1971 cumulative_error = error; 1972 } 1973 1974 /* 1975 * Flush the recovery at the PFS boundary to stage it for 1976 * the final flush of the super-root topology. 1977 */ 1978 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) && 1979 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) { 1980 hammer2_flush(chain, HAMMER2_FLUSH_TOP); 1981 } 1982 } 1983 1984 return cumulative_error; 1985 } 1986 1987 /* 1988 * Sync a mount point; this is called on a per-mount basis from the 1989 * filesystem syncer process periodically and whenever a user issues 1990 * a sync. 1991 */ 1992 int 1993 hammer2_vfs_sync(struct mount *mp, int waitfor) 1994 { 1995 hammer2_xop_flush_t *xop; 1996 struct hammer2_sync_info info; 1997 hammer2_inode_t *iroot; 1998 hammer2_pfs_t *pmp; 1999 int flags; 2000 int error; 2001 2002 pmp = MPTOPMP(mp); 2003 iroot = pmp->iroot; 2004 KKASSERT(iroot); 2005 KKASSERT(iroot->pmp == pmp); 2006 2007 /* 2008 * We can't acquire locks on existing vnodes while in a transaction 2009 * without risking a deadlock. This assumes that vfsync() can be 2010 * called without the vnode locked (which it can in DragonFly). 2011 * Otherwise we'd have to implement a multi-pass or flag the lock 2012 * failures and retry. 2013 * 2014 * The reclamation code interlocks with the sync list's token 2015 * (by removing the vnode from the scan list) before unlocking 2016 * the inode, giving us time to ref the inode. 2017 */ 2018 /*flags = VMSC_GETVP;*/ 2019 flags = 0; 2020 if (waitfor & MNT_LAZY) 2021 flags |= VMSC_ONEPASS; 2022 2023 #if 0 2024 /* 2025 * Preflush the vnodes using a normal transaction before interlocking 2026 * with a flush transaction. 2027 */ 2028 hammer2_trans_init(pmp, 0); 2029 info.error = 0; 2030 info.waitfor = MNT_NOWAIT; 2031 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info); 2032 hammer2_trans_done(pmp); 2033 #endif 2034 2035 /* 2036 * Start our flush transaction. This does not return until all 2037 * concurrent transactions have completed and will prevent any 2038 * new transactions from running concurrently, except for the 2039 * buffer cache transactions. 2040 * 2041 * For efficiency do an async pass before making sure with a 2042 * synchronous pass on all related buffer cache buffers. It 2043 * should theoretically not be possible for any new file buffers 2044 * to be instantiated during this sequence. 2045 */ 2046 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH | 2047 HAMMER2_TRANS_PREFLUSH); 2048 hammer2_inode_run_sideq(pmp); 2049 2050 info.error = 0; 2051 info.waitfor = MNT_NOWAIT; 2052 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info); 2053 info.waitfor = MNT_WAIT; 2054 vsyncscan(mp, flags, hammer2_sync_scan2, &info); 2055 2056 /* 2057 * Clear PREFLUSH. This prevents (or asserts on) any new logical 2058 * buffer cache flushes which occur during the flush. Device buffers 2059 * are not affected. 2060 */ 2061 hammer2_bioq_sync(pmp); 2062 hammer2_trans_clear_preflush(pmp); 2063 2064 /* 2065 * Use the XOP interface to concurrently flush all nodes to 2066 * synchronize the PFSROOT subtopology to the media. A standard 2067 * end-of-scan ENOENT error indicates cluster sufficiency. 2068 * 2069 * Note that this flush will not be visible on crash recovery until 2070 * we flush the super-root topology in the next loop. 2071 * 2072 * XXX For now wait for all flushes to complete. 2073 */ 2074 if (iroot) { 2075 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING); 2076 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush); 2077 error = hammer2_xop_collect(&xop->head, 2078 HAMMER2_XOP_COLLECT_WAITALL); 2079 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 2080 if (error == ENOENT) 2081 error = 0; 2082 } else { 2083 error = 0; 2084 } 2085 hammer2_trans_done(pmp); 2086 2087 return (error); 2088 } 2089 2090 /* 2091 * Sync passes. 2092 * 2093 * Note that we ignore the tranasction mtid we got above. Instead, 2094 * each vfsync below will ultimately get its own via TRANS_BUFCACHE 2095 * transactions. 2096 */ 2097 static int 2098 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data) 2099 { 2100 struct hammer2_sync_info *info = data; 2101 hammer2_inode_t *ip; 2102 int error; 2103 2104 /* 2105 * Degenerate cases. Note that ip == NULL typically means the 2106 * syncer vnode itself and we don't want to vclrisdirty() in that 2107 * situation. 2108 */ 2109 ip = VTOI(vp); 2110 if (ip == NULL) { 2111 return(0); 2112 } 2113 if (vp->v_type == VNON || vp->v_type == VBAD) { 2114 vclrisdirty(vp); 2115 return(0); 2116 } 2117 2118 /* 2119 * VOP_FSYNC will start a new transaction so replicate some code 2120 * here to do it inline (see hammer2_vop_fsync()). 2121 * 2122 * WARNING: The vfsync interacts with the buffer cache and might 2123 * block, we can't hold the inode lock at that time. 2124 * However, we MUST ref ip before blocking to ensure that 2125 * it isn't ripped out from under us (since we do not 2126 * hold a lock on the vnode). 2127 */ 2128 hammer2_inode_ref(ip); 2129 if ((ip->flags & HAMMER2_INODE_MODIFIED) || 2130 !RB_EMPTY(&vp->v_rbdirty_tree)) { 2131 vfsync(vp, info->waitfor, 1, NULL, NULL); 2132 if (ip->flags & (HAMMER2_INODE_RESIZED | 2133 HAMMER2_INODE_MODIFIED)) { 2134 hammer2_inode_lock(ip, 0); 2135 hammer2_inode_chain_sync(ip); 2136 hammer2_inode_unlock(ip); 2137 } 2138 } 2139 if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 && 2140 RB_EMPTY(&vp->v_rbdirty_tree)) { 2141 vclrisdirty(vp); 2142 } 2143 2144 hammer2_inode_drop(ip); 2145 #if 1 2146 error = 0; 2147 if (error) 2148 info->error = error; 2149 #endif 2150 return(0); 2151 } 2152 2153 static 2154 int 2155 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp) 2156 { 2157 return (0); 2158 } 2159 2160 static 2161 int 2162 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp, 2163 struct fid *fhp, struct vnode **vpp) 2164 { 2165 return (0); 2166 } 2167 2168 static 2169 int 2170 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 2171 int *exflagsp, struct ucred **credanonp) 2172 { 2173 return (0); 2174 } 2175 2176 /* 2177 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume 2178 * header into the HMP 2179 * 2180 * XXX read four volhdrs and use the one with the highest TID whos CRC 2181 * matches. 2182 * 2183 * XXX check iCRCs. 2184 * 2185 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to 2186 * nonexistant locations. 2187 * 2188 * XXX Record selected volhdr and ring updates to each of 4 volhdrs 2189 */ 2190 static 2191 int 2192 hammer2_install_volume_header(hammer2_dev_t *hmp) 2193 { 2194 hammer2_volume_data_t *vd; 2195 struct buf *bp; 2196 hammer2_crc32_t crc0, crc, bcrc0, bcrc; 2197 int error_reported; 2198 int error; 2199 int valid; 2200 int i; 2201 2202 error_reported = 0; 2203 error = 0; 2204 valid = 0; 2205 bp = NULL; 2206 2207 /* 2208 * There are up to 4 copies of the volume header (syncs iterate 2209 * between them so there is no single master). We don't trust the 2210 * volu_size field so we don't know precisely how large the filesystem 2211 * is, so depend on the OS to return an error if we go beyond the 2212 * block device's EOF. 2213 */ 2214 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) { 2215 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64, 2216 HAMMER2_VOLUME_BYTES, &bp); 2217 if (error) { 2218 brelse(bp); 2219 bp = NULL; 2220 continue; 2221 } 2222 2223 vd = (struct hammer2_volume_data *) bp->b_data; 2224 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) && 2225 (vd->magic != HAMMER2_VOLUME_ID_ABO)) { 2226 brelse(bp); 2227 bp = NULL; 2228 continue; 2229 } 2230 2231 if (vd->magic == HAMMER2_VOLUME_ID_ABO) { 2232 /* XXX: Reversed-endianness filesystem */ 2233 kprintf("hammer2: reverse-endian filesystem detected"); 2234 brelse(bp); 2235 bp = NULL; 2236 continue; 2237 } 2238 2239 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0]; 2240 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF, 2241 HAMMER2_VOLUME_ICRC0_SIZE); 2242 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1]; 2243 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF, 2244 HAMMER2_VOLUME_ICRC1_SIZE); 2245 if ((crc0 != crc) || (bcrc0 != bcrc)) { 2246 kprintf("hammer2 volume header crc " 2247 "mismatch copy #%d %08x/%08x\n", 2248 i, crc0, crc); 2249 error_reported = 1; 2250 brelse(bp); 2251 bp = NULL; 2252 continue; 2253 } 2254 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) { 2255 valid = 1; 2256 hmp->voldata = *vd; 2257 hmp->volhdrno = i; 2258 } 2259 brelse(bp); 2260 bp = NULL; 2261 } 2262 if (valid) { 2263 hmp->volsync = hmp->voldata; 2264 error = 0; 2265 if (error_reported || bootverbose || 1) { /* 1/DEBUG */ 2266 kprintf("hammer2: using volume header #%d\n", 2267 hmp->volhdrno); 2268 } 2269 } else { 2270 error = EINVAL; 2271 kprintf("hammer2: no valid volume headers found!\n"); 2272 } 2273 return (error); 2274 } 2275 2276 /* 2277 * This handles hysteresis on regular file flushes. Because the BIOs are 2278 * routed to a thread it is possible for an excessive number to build up 2279 * and cause long front-end stalls long before the runningbuffspace limit 2280 * is hit, so we implement hammer2_flush_pipe to control the 2281 * hysteresis. 2282 * 2283 * This is a particular problem when compression is used. 2284 */ 2285 void 2286 hammer2_lwinprog_ref(hammer2_pfs_t *pmp) 2287 { 2288 atomic_add_int(&pmp->count_lwinprog, 1); 2289 } 2290 2291 void 2292 hammer2_lwinprog_drop(hammer2_pfs_t *pmp) 2293 { 2294 int lwinprog; 2295 2296 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1); 2297 if ((lwinprog & HAMMER2_LWINPROG_WAITING) && 2298 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) { 2299 atomic_clear_int(&pmp->count_lwinprog, 2300 HAMMER2_LWINPROG_WAITING); 2301 wakeup(&pmp->count_lwinprog); 2302 } 2303 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) && 2304 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) { 2305 atomic_clear_int(&pmp->count_lwinprog, 2306 HAMMER2_LWINPROG_WAITING0); 2307 wakeup(&pmp->count_lwinprog); 2308 } 2309 } 2310 2311 void 2312 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe) 2313 { 2314 int lwinprog; 2315 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING : 2316 HAMMER2_LWINPROG_WAITING0; 2317 2318 for (;;) { 2319 lwinprog = pmp->count_lwinprog; 2320 cpu_ccfence(); 2321 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 2322 break; 2323 tsleep_interlock(&pmp->count_lwinprog, 0); 2324 atomic_set_int(&pmp->count_lwinprog, lwflag); 2325 lwinprog = pmp->count_lwinprog; 2326 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 2327 break; 2328 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz); 2329 } 2330 } 2331 2332 /* 2333 * Manage excessive memory resource use for chain and related 2334 * structures. 2335 */ 2336 void 2337 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp) 2338 { 2339 uint32_t waiting; 2340 uint32_t count; 2341 uint32_t limit; 2342 #if 0 2343 static int zzticks; 2344 #endif 2345 2346 /* 2347 * Atomic check condition and wait. Also do an early speedup of 2348 * the syncer to try to avoid hitting the wait. 2349 */ 2350 for (;;) { 2351 waiting = pmp->inmem_dirty_chains; 2352 cpu_ccfence(); 2353 count = waiting & HAMMER2_DIRTYCHAIN_MASK; 2354 2355 limit = pmp->mp->mnt_nvnodelistsize / 10; 2356 if (limit < hammer2_limit_dirty_chains) 2357 limit = hammer2_limit_dirty_chains; 2358 if (limit < 1000) 2359 limit = 1000; 2360 2361 #if 0 2362 if ((int)(ticks - zzticks) > hz) { 2363 zzticks = ticks; 2364 kprintf("count %ld %ld\n", count, limit); 2365 } 2366 #endif 2367 2368 /* 2369 * Block if there are too many dirty chains present, wait 2370 * for the flush to clean some out. 2371 */ 2372 if (count > limit) { 2373 tsleep_interlock(&pmp->inmem_dirty_chains, 0); 2374 if (atomic_cmpset_int(&pmp->inmem_dirty_chains, 2375 waiting, 2376 waiting | HAMMER2_DIRTYCHAIN_WAITING)) { 2377 speedup_syncer(pmp->mp); 2378 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED, 2379 "chnmem", hz); 2380 } 2381 continue; /* loop on success or fail */ 2382 } 2383 2384 /* 2385 * Try to start an early flush before we are forced to block. 2386 */ 2387 if (count > limit * 7 / 10) 2388 speedup_syncer(pmp->mp); 2389 break; 2390 } 2391 } 2392 2393 void 2394 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp) 2395 { 2396 if (pmp) { 2397 atomic_add_int(&pmp->inmem_dirty_chains, 1); 2398 } 2399 } 2400 2401 void 2402 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp) 2403 { 2404 uint32_t waiting; 2405 2406 if (pmp == NULL) 2407 return; 2408 2409 for (;;) { 2410 waiting = pmp->inmem_dirty_chains; 2411 cpu_ccfence(); 2412 if (atomic_cmpset_int(&pmp->inmem_dirty_chains, 2413 waiting, 2414 (waiting - 1) & 2415 ~HAMMER2_DIRTYCHAIN_WAITING)) { 2416 break; 2417 } 2418 } 2419 2420 if (waiting & HAMMER2_DIRTYCHAIN_WAITING) 2421 wakeup(&pmp->inmem_dirty_chains); 2422 } 2423 2424 /* 2425 * Debugging 2426 */ 2427 void 2428 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx) 2429 { 2430 hammer2_chain_t *scan; 2431 hammer2_chain_t *parent; 2432 2433 --*countp; 2434 if (*countp == 0) { 2435 kprintf("%*.*s...\n", tab, tab, ""); 2436 return; 2437 } 2438 if (*countp < 0) 2439 return; 2440 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n", 2441 tab, tab, "", pfx, 2442 chain, chain->bref.type, 2443 chain->bref.key, chain->bref.keybits, 2444 chain->bref.mirror_tid); 2445 2446 kprintf("%*.*s [%08x] (%s) refs=%d", 2447 tab, tab, "", 2448 chain->flags, 2449 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE && 2450 chain->data) ? (char *)chain->data->ipdata.filename : "?"), 2451 chain->refs); 2452 2453 parent = chain->parent; 2454 if (parent) 2455 kprintf("\n%*.*s p=%p [pflags %08x prefs %d", 2456 tab, tab, "", 2457 parent, parent->flags, parent->refs); 2458 if (RB_EMPTY(&chain->core.rbtree)) { 2459 kprintf("\n"); 2460 } else { 2461 kprintf(" {\n"); 2462 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) 2463 hammer2_dump_chain(scan, tab + 4, countp, 'a'); 2464 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data) 2465 kprintf("%*.*s}(%s)\n", tab, tab, "", 2466 chain->data->ipdata.filename); 2467 else 2468 kprintf("%*.*s}\n", tab, tab, ""); 2469 } 2470 } 2471