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