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