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