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 static struct hammer2_mntlist hammer2_mntlist; 76 77 struct hammer2_pfslist hammer2_pfslist; 78 struct lock hammer2_mntlk; 79 80 int hammer2_debug; 81 int hammer2_cluster_read = 4; /* physical read-ahead */ 82 int hammer2_cluster_write = 0; /* bdwrite() so later inval works */ 83 int hammer2_dedup_enable = 1; 84 int hammer2_inval_enable = 0; 85 int hammer2_flush_pipe = 100; 86 int hammer2_synchronous_flush = 1; 87 int hammer2_dio_count; 88 long hammer2_chain_allocs; 89 long hammer2_chain_frees; 90 long hammer2_limit_dirty_chains; 91 long hammer2_count_modified_chains; 92 long hammer2_iod_invals; 93 long hammer2_iod_file_read; 94 long hammer2_iod_meta_read; 95 long hammer2_iod_indr_read; 96 long hammer2_iod_fmap_read; 97 long hammer2_iod_volu_read; 98 long hammer2_iod_file_write; 99 long hammer2_iod_file_wembed; 100 long hammer2_iod_file_wzero; 101 long hammer2_iod_file_wdedup; 102 long hammer2_iod_meta_write; 103 long hammer2_iod_indr_write; 104 long hammer2_iod_fmap_write; 105 long hammer2_iod_volu_write; 106 107 MALLOC_DECLARE(M_HAMMER2_CBUFFER); 108 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer", 109 "Buffer used for compression."); 110 111 MALLOC_DECLARE(M_HAMMER2_DEBUFFER); 112 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer", 113 "Buffer used for decompression."); 114 115 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem"); 116 117 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW, 118 &hammer2_debug, 0, ""); 119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW, 120 &hammer2_cluster_read, 0, ""); 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW, 122 &hammer2_cluster_write, 0, ""); 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW, 124 &hammer2_dedup_enable, 0, ""); 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW, 126 &hammer2_inval_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, chain_allocs, CTLFLAG_RW, 132 &hammer2_chain_allocs, 0, ""); 133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW, 134 &hammer2_chain_frees, 0, ""); 135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW, 136 &hammer2_limit_dirty_chains, 0, ""); 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW, 138 &hammer2_count_modified_chains, 0, ""); 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD, 140 &hammer2_dio_count, 0, ""); 141 142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW, 143 &hammer2_iod_invals, 0, ""); 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW, 145 &hammer2_iod_file_read, 0, ""); 146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW, 147 &hammer2_iod_meta_read, 0, ""); 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW, 149 &hammer2_iod_indr_read, 0, ""); 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW, 151 &hammer2_iod_fmap_read, 0, ""); 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW, 153 &hammer2_iod_volu_read, 0, ""); 154 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW, 156 &hammer2_iod_file_write, 0, ""); 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW, 158 &hammer2_iod_file_wembed, 0, ""); 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW, 160 &hammer2_iod_file_wzero, 0, ""); 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW, 162 &hammer2_iod_file_wdedup, 0, ""); 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW, 164 &hammer2_iod_meta_write, 0, ""); 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW, 166 &hammer2_iod_indr_write, 0, ""); 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW, 168 &hammer2_iod_fmap_write, 0, ""); 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW, 170 &hammer2_iod_volu_write, 0, ""); 171 172 long hammer2_check_icrc32; 173 long hammer2_check_xxhash64; 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW, 175 &hammer2_check_icrc32, 0, ""); 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW, 177 &hammer2_check_xxhash64, 0, ""); 178 179 static int hammer2_vfs_init(struct vfsconf *conf); 180 static int hammer2_vfs_uninit(struct vfsconf *vfsp); 181 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data, 182 struct ucred *cred); 183 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *, 184 struct vnode *, struct ucred *); 185 static int hammer2_recovery(hammer2_dev_t *hmp); 186 static int hammer2_vfs_unmount(struct mount *mp, int mntflags); 187 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp); 188 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, 189 struct ucred *cred); 190 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, 191 struct ucred *cred); 192 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp, 193 struct fid *fhp, struct vnode **vpp); 194 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp); 195 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 196 int *exflagsp, struct ucred **credanonp); 197 198 static int hammer2_install_volume_header(hammer2_dev_t *hmp); 199 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data); 200 201 static void hammer2_update_pmps(hammer2_dev_t *hmp); 202 203 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp); 204 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, 205 hammer2_dev_t *hmp); 206 207 /* 208 * HAMMER2 vfs operations. 209 */ 210 static struct vfsops hammer2_vfsops = { 211 .vfs_init = hammer2_vfs_init, 212 .vfs_uninit = hammer2_vfs_uninit, 213 .vfs_sync = hammer2_vfs_sync, 214 .vfs_mount = hammer2_vfs_mount, 215 .vfs_unmount = hammer2_vfs_unmount, 216 .vfs_root = hammer2_vfs_root, 217 .vfs_statfs = hammer2_vfs_statfs, 218 .vfs_statvfs = hammer2_vfs_statvfs, 219 .vfs_vget = hammer2_vfs_vget, 220 .vfs_vptofh = hammer2_vfs_vptofh, 221 .vfs_fhtovp = hammer2_vfs_fhtovp, 222 .vfs_checkexp = hammer2_vfs_checkexp 223 }; 224 225 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", ""); 226 227 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE); 228 MODULE_VERSION(hammer2, 1); 229 230 static 231 int 232 hammer2_vfs_init(struct vfsconf *conf) 233 { 234 static struct objcache_malloc_args margs_read; 235 static struct objcache_malloc_args margs_write; 236 static struct objcache_malloc_args margs_vop; 237 238 int error; 239 240 error = 0; 241 242 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref)) 243 error = EINVAL; 244 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data)) 245 error = EINVAL; 246 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data)) 247 error = EINVAL; 248 249 if (error) 250 kprintf("HAMMER2 structure size mismatch; cannot continue.\n"); 251 252 margs_read.objsize = 65536; 253 margs_read.mtype = M_HAMMER2_DEBUFFER; 254 255 margs_write.objsize = 32768; 256 margs_write.mtype = M_HAMMER2_CBUFFER; 257 258 margs_vop.objsize = sizeof(hammer2_xop_t); 259 margs_vop.mtype = M_HAMMER2; 260 261 /* 262 * Note thaht for the XOPS cache we want backing store allocations 263 * to use M_ZERO. This is not allowed in objcache_get() (to avoid 264 * confusion), so use the backing store function that does it. This 265 * means that initial XOPS objects are zerod but REUSED objects are 266 * not. So we are responsible for cleaning the object up sufficiently 267 * for our needs before objcache_put()ing it back (typically just the 268 * FIFO indices). 269 */ 270 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc, 271 0, 1, NULL, NULL, NULL, 272 objcache_malloc_alloc, 273 objcache_malloc_free, 274 &margs_read); 275 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc, 276 0, 1, NULL, NULL, NULL, 277 objcache_malloc_alloc, 278 objcache_malloc_free, 279 &margs_write); 280 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc, 281 0, 1, NULL, NULL, NULL, 282 objcache_malloc_alloc_zero, 283 objcache_malloc_free, 284 &margs_vop); 285 286 287 lockinit(&hammer2_mntlk, "mntlk", 0, 0); 288 TAILQ_INIT(&hammer2_mntlist); 289 TAILQ_INIT(&hammer2_pfslist); 290 291 hammer2_limit_dirty_chains = maxvnodes / 10; 292 if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS) 293 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS; 294 295 return (error); 296 } 297 298 static 299 int 300 hammer2_vfs_uninit(struct vfsconf *vfsp __unused) 301 { 302 objcache_destroy(cache_buffer_read); 303 objcache_destroy(cache_buffer_write); 304 objcache_destroy(cache_xops); 305 return 0; 306 } 307 308 /* 309 * Core PFS allocator. Used to allocate or reference the pmp structure 310 * for PFS cluster mounts and the spmp structure for media (hmp) structures. 311 * The pmp can be passed in or loaded by this function using the chain and 312 * inode data. 313 * 314 * pmp->modify_tid tracks new modify_tid transaction ids for front-end 315 * transactions. Note that synchronization does not use this field. 316 * (typically frontend operations and synchronization cannot run on the 317 * same PFS node at the same time). 318 * 319 * XXX check locking 320 */ 321 hammer2_pfs_t * 322 hammer2_pfsalloc(hammer2_chain_t *chain, 323 const hammer2_inode_data_t *ripdata, 324 hammer2_tid_t modify_tid, hammer2_dev_t *force_local) 325 { 326 hammer2_pfs_t *pmp; 327 hammer2_inode_t *iroot; 328 int count; 329 int i; 330 int j; 331 332 pmp = NULL; 333 334 /* 335 * Locate or create the PFS based on the cluster id. If ripdata 336 * is NULL this is a spmp which is unique and is always allocated. 337 * 338 * If the device is mounted in local mode all PFSs are considered 339 * independent and not part of any cluster (for debugging only). 340 */ 341 if (ripdata) { 342 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 343 if (force_local != pmp->force_local) 344 continue; 345 if (force_local == NULL && 346 bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid, 347 sizeof(pmp->pfs_clid)) == 0) { 348 break; 349 } else if (force_local && pmp->pfs_names[0] && 350 strcmp(pmp->pfs_names[0], ripdata->filename) == 0) { 351 break; 352 } 353 } 354 } 355 356 if (pmp == NULL) { 357 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO); 358 pmp->force_local = force_local; 359 hammer2_trans_manage_init(pmp); 360 kmalloc_create(&pmp->minode, "HAMMER2-inodes"); 361 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg"); 362 lockinit(&pmp->lock, "pfslk", 0, 0); 363 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0); 364 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum"); 365 spin_init(&pmp->xop_spin, "h2xop"); 366 spin_init(&pmp->lru_spin, "h2lru"); 367 RB_INIT(&pmp->inum_tree); 368 TAILQ_INIT(&pmp->sideq); 369 TAILQ_INIT(&pmp->lru_list); 370 spin_init(&pmp->list_spin, "hm2pfsalloc_list"); 371 372 /* 373 * Distribute backend operations to threads 374 */ 375 for (i = 0; i < HAMMER2_XOPGROUPS; ++i) 376 hammer2_xop_group_init(pmp, &pmp->xop_groups[i]); 377 378 /* 379 * Save the last media transaction id for the flusher. Set 380 * initial 381 */ 382 if (ripdata) 383 pmp->pfs_clid = ripdata->meta.pfs_clid; 384 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry); 385 386 /* 387 * The synchronization thread may start too early, make 388 * sure it stays frozen until we are ready to let it go. 389 * XXX 390 */ 391 /* 392 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN | 393 HAMMER2_THREAD_REMASTER; 394 */ 395 } 396 397 /* 398 * Create the PFS's root inode and any missing XOP helper threads. 399 */ 400 if ((iroot = pmp->iroot) == NULL) { 401 iroot = hammer2_inode_get(pmp, NULL, NULL, -1); 402 if (ripdata) 403 iroot->meta = ripdata->meta; 404 pmp->iroot = iroot; 405 hammer2_inode_ref(iroot); 406 hammer2_inode_unlock(iroot); 407 } 408 409 /* 410 * Stop here if no chain is passed in. 411 */ 412 if (chain == NULL) 413 goto done; 414 415 /* 416 * When a chain is passed in we must add it to the PFS's root 417 * inode, update pmp->pfs_types[], and update the syncronization 418 * threads. 419 * 420 * When forcing local mode, mark the PFS as a MASTER regardless. 421 * 422 * At the moment empty spots can develop due to removals or failures. 423 * Ultimately we want to re-fill these spots but doing so might 424 * confused running code. XXX 425 */ 426 hammer2_inode_ref(iroot); 427 hammer2_mtx_ex(&iroot->lock); 428 j = iroot->cluster.nchains; 429 430 kprintf("add PFS to pmp %p[%d]\n", pmp, j); 431 432 if (j == HAMMER2_MAXCLUSTER) { 433 kprintf("hammer2_mount: cluster full!\n"); 434 /* XXX fatal error? */ 435 } else { 436 KKASSERT(chain->pmp == NULL); 437 chain->pmp = pmp; 438 hammer2_chain_ref(chain); 439 iroot->cluster.array[j].chain = chain; 440 if (force_local) 441 pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER; 442 else 443 pmp->pfs_types[j] = ripdata->meta.pfs_type; 444 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2); 445 pmp->pfs_hmps[j] = chain->hmp; 446 447 /* 448 * If the PFS is already mounted we must account 449 * for the mount_count here. 450 */ 451 if (pmp->mp) 452 ++chain->hmp->mount_count; 453 454 /* 455 * May have to fixup dirty chain tracking. Previous 456 * pmp was NULL so nothing to undo. 457 */ 458 if (chain->flags & HAMMER2_CHAIN_MODIFIED) 459 hammer2_pfs_memory_inc(pmp); 460 ++j; 461 } 462 iroot->cluster.nchains = j; 463 464 /* 465 * Update nmasters from any PFS inode which is part of the cluster. 466 * It is possible that this will result in a value which is too 467 * high. MASTER PFSs are authoritative for pfs_nmasters and will 468 * override this value later on. 469 * 470 * (This informs us of masters that might not currently be 471 * discoverable by this mount). 472 */ 473 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) { 474 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters; 475 } 476 477 /* 478 * Count visible masters. Masters are usually added with 479 * ripdata->meta.pfs_nmasters set to 1. This detects when there 480 * are more (XXX and must update the master inodes). 481 */ 482 count = 0; 483 for (i = 0; i < iroot->cluster.nchains; ++i) { 484 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) 485 ++count; 486 } 487 if (pmp->pfs_nmasters < count) 488 pmp->pfs_nmasters = count; 489 490 /* 491 * Create missing synchronization and support threads. 492 * 493 * Single-node masters (including snapshots) have nothing to 494 * synchronize and do not require this thread. 495 * 496 * Multi-node masters or any number of soft masters, slaves, copy, 497 * or other PFS types need the thread. 498 * 499 * Each thread is responsible for its particular cluster index. 500 * We use independent threads so stalls or mismatches related to 501 * any given target do not affect other targets. 502 */ 503 for (i = 0; i < iroot->cluster.nchains; ++i) { 504 /* 505 * Single-node masters (including snapshots) have nothing 506 * to synchronize and will make direct xops support calls, 507 * thus they do not require this thread. 508 * 509 * Note that there can be thousands of snapshots. We do not 510 * want to create thousands of threads. 511 */ 512 if (pmp->pfs_nmasters <= 1 && 513 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) { 514 continue; 515 } 516 517 /* 518 * Sync support thread 519 */ 520 if (pmp->sync_thrs[i].td == NULL) { 521 hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL, 522 "h2nod", i, -1, 523 hammer2_primary_sync_thread); 524 } 525 } 526 527 /* 528 * Create missing Xop threads 529 * 530 * NOTE: We create helper threads for all mounted PFSs or any 531 * PFSs with 2+ nodes (so the sync thread can update them, 532 * even if not mounted). 533 */ 534 if (pmp->mp || iroot->cluster.nchains >= 2) 535 hammer2_xop_helper_create(pmp); 536 537 hammer2_mtx_unlock(&iroot->lock); 538 hammer2_inode_drop(iroot); 539 done: 540 return pmp; 541 } 542 543 /* 544 * Deallocate an element of a probed PFS. If destroying and this is a 545 * MASTER, adjust nmasters. 546 * 547 * This function does not physically destroy the PFS element in its device 548 * under the super-root (see hammer2_ioctl_pfs_delete()). 549 */ 550 void 551 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying) 552 { 553 hammer2_inode_t *iroot; 554 hammer2_chain_t *chain; 555 int j; 556 557 /* 558 * Cleanup our reference on iroot. iroot is (should) not be needed 559 * by the flush code. 560 */ 561 iroot = pmp->iroot; 562 if (iroot) { 563 /* 564 * Stop synchronizing 565 * 566 * XXX flush after acquiring the iroot lock. 567 * XXX clean out the cluster index from all inode structures. 568 */ 569 hammer2_thr_delete(&pmp->sync_thrs[clindex]); 570 571 /* 572 * Remove the cluster index from the group. If destroying 573 * the PFS and this is a master, adjust pfs_nmasters. 574 */ 575 hammer2_mtx_ex(&iroot->lock); 576 chain = iroot->cluster.array[clindex].chain; 577 iroot->cluster.array[clindex].chain = NULL; 578 579 switch(pmp->pfs_types[clindex]) { 580 case HAMMER2_PFSTYPE_MASTER: 581 if (destroying && pmp->pfs_nmasters > 0) 582 --pmp->pfs_nmasters; 583 /* XXX adjust ripdata->meta.pfs_nmasters */ 584 break; 585 default: 586 break; 587 } 588 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE; 589 590 hammer2_mtx_unlock(&iroot->lock); 591 592 /* 593 * Release the chain. 594 */ 595 if (chain) { 596 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 597 hammer2_chain_drop(chain); 598 } 599 600 /* 601 * Terminate all XOP threads for the cluster index. 602 */ 603 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) 604 hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]); 605 } 606 } 607 608 /* 609 * Destroy a PFS, typically only occurs after the last mount on a device 610 * has gone away. 611 */ 612 static void 613 hammer2_pfsfree(hammer2_pfs_t *pmp) 614 { 615 hammer2_inode_t *iroot; 616 hammer2_chain_t *chain; 617 int i; 618 int j; 619 620 /* 621 * Cleanup our reference on iroot. iroot is (should) not be needed 622 * by the flush code. 623 */ 624 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry); 625 626 iroot = pmp->iroot; 627 if (iroot) { 628 for (i = 0; i < iroot->cluster.nchains; ++i) { 629 hammer2_thr_delete(&pmp->sync_thrs[i]); 630 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) 631 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]); 632 } 633 #if REPORT_REFS_ERRORS 634 if (pmp->iroot->refs != 1) 635 kprintf("PMP->IROOT %p REFS WRONG %d\n", 636 pmp->iroot, pmp->iroot->refs); 637 #else 638 KKASSERT(pmp->iroot->refs == 1); 639 #endif 640 /* ref for pmp->iroot */ 641 hammer2_inode_drop(pmp->iroot); 642 pmp->iroot = NULL; 643 } 644 645 /* 646 * Cleanup chains remaining on LRU list. 647 */ 648 kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count); 649 while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) { 650 hammer2_chain_ref(chain); 651 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 652 hammer2_chain_drop(chain); 653 } 654 655 /* 656 * Free remaining pmp resources 657 */ 658 kmalloc_destroy(&pmp->mmsg); 659 kmalloc_destroy(&pmp->minode); 660 661 kfree(pmp, M_HAMMER2); 662 } 663 664 /* 665 * Remove all references to hmp from the pfs list. Any PFS which becomes 666 * empty is terminated and freed. 667 * 668 * XXX inefficient. 669 */ 670 static void 671 hammer2_pfsfree_scan(hammer2_dev_t *hmp) 672 { 673 hammer2_pfs_t *pmp; 674 hammer2_inode_t *iroot; 675 hammer2_chain_t *rchain; 676 int didfreeze; 677 int i; 678 int j; 679 680 again: 681 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 682 if ((iroot = pmp->iroot) == NULL) 683 continue; 684 if (hmp->spmp == pmp) { 685 kprintf("unmount hmp %p remove spmp %p\n", 686 hmp, pmp); 687 hmp->spmp = NULL; 688 } 689 690 /* 691 * Determine if this PFS is affected. If it is we must 692 * freeze all management threads and lock its iroot. 693 * 694 * Freezing a management thread forces it idle, operations 695 * in-progress will be aborted and it will have to start 696 * over again when unfrozen, or exit if told to exit. 697 */ 698 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 699 if (pmp->pfs_hmps[i] == hmp) 700 break; 701 } 702 if (i != HAMMER2_MAXCLUSTER) { 703 /* 704 * Make sure all synchronization threads are locked 705 * down. 706 */ 707 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 708 if (pmp->pfs_hmps[i] == NULL) 709 continue; 710 hammer2_thr_freeze_async(&pmp->sync_thrs[i]); 711 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) { 712 hammer2_thr_freeze_async( 713 &pmp->xop_groups[j].thrs[i]); 714 } 715 } 716 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 717 if (pmp->pfs_hmps[i] == NULL) 718 continue; 719 hammer2_thr_freeze(&pmp->sync_thrs[i]); 720 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) { 721 hammer2_thr_freeze( 722 &pmp->xop_groups[j].thrs[i]); 723 } 724 } 725 726 /* 727 * Lock the inode and clean out matching chains. 728 * Note that we cannot use hammer2_inode_lock_*() 729 * here because that would attempt to validate the 730 * cluster that we are in the middle of ripping 731 * apart. 732 * 733 * WARNING! We are working directly on the inodes 734 * embedded cluster. 735 */ 736 hammer2_mtx_ex(&iroot->lock); 737 738 /* 739 * Remove the chain from matching elements of the PFS. 740 */ 741 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 742 if (pmp->pfs_hmps[i] != hmp) 743 continue; 744 hammer2_thr_delete(&pmp->sync_thrs[i]); 745 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) { 746 hammer2_thr_delete( 747 &pmp->xop_groups[j].thrs[i]); 748 } 749 rchain = iroot->cluster.array[i].chain; 750 iroot->cluster.array[i].chain = NULL; 751 pmp->pfs_types[i] = 0; 752 if (pmp->pfs_names[i]) { 753 kfree(pmp->pfs_names[i], M_HAMMER2); 754 pmp->pfs_names[i] = NULL; 755 } 756 if (rchain) { 757 hammer2_chain_drop(rchain); 758 /* focus hint */ 759 if (iroot->cluster.focus == rchain) 760 iroot->cluster.focus = NULL; 761 } 762 pmp->pfs_hmps[i] = NULL; 763 } 764 hammer2_mtx_unlock(&iroot->lock); 765 didfreeze = 1; /* remaster, unfreeze down below */ 766 } else { 767 didfreeze = 0; 768 } 769 770 /* 771 * Cleanup trailing chains. Gaps may remain. 772 */ 773 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) { 774 if (pmp->pfs_hmps[i]) 775 break; 776 } 777 iroot->cluster.nchains = i + 1; 778 779 /* 780 * If the PMP has no elements remaining we can destroy it. 781 * (this will transition management threads from frozen->exit). 782 */ 783 if (iroot->cluster.nchains == 0) { 784 kprintf("unmount hmp %p last ref to PMP=%p\n", 785 hmp, pmp); 786 hammer2_pfsfree(pmp); 787 goto again; 788 } 789 790 /* 791 * If elements still remain we need to set the REMASTER 792 * flag and unfreeze it. 793 */ 794 if (didfreeze) { 795 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 796 if (pmp->pfs_hmps[i] == NULL) 797 continue; 798 hammer2_thr_remaster(&pmp->sync_thrs[i]); 799 hammer2_thr_unfreeze(&pmp->sync_thrs[i]); 800 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) { 801 hammer2_thr_remaster( 802 &pmp->xop_groups[j].thrs[i]); 803 hammer2_thr_unfreeze( 804 &pmp->xop_groups[j].thrs[i]); 805 } 806 } 807 } 808 } 809 } 810 811 /* 812 * Mount or remount HAMMER2 fileystem from physical media 813 * 814 * mountroot 815 * mp mount point structure 816 * path NULL 817 * data <unused> 818 * cred <unused> 819 * 820 * mount 821 * mp mount point structure 822 * path path to mount point 823 * data pointer to argument structure in user space 824 * volume volume path (device@LABEL form) 825 * hflags user mount flags 826 * cred user credentials 827 * 828 * RETURNS: 0 Success 829 * !0 error number 830 */ 831 static 832 int 833 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data, 834 struct ucred *cred) 835 { 836 struct hammer2_mount_info info; 837 hammer2_pfs_t *pmp; 838 hammer2_pfs_t *spmp; 839 hammer2_dev_t *hmp; 840 hammer2_dev_t *force_local; 841 hammer2_key_t key_next; 842 hammer2_key_t key_dummy; 843 hammer2_key_t lhc; 844 struct vnode *devvp; 845 struct nlookupdata nd; 846 hammer2_chain_t *parent; 847 hammer2_chain_t *chain; 848 hammer2_cluster_t *cluster; 849 const hammer2_inode_data_t *ripdata; 850 hammer2_blockref_t bref; 851 struct file *fp; 852 char devstr[MNAMELEN]; 853 size_t size; 854 size_t done; 855 char *dev; 856 char *label; 857 int ronly = 1; 858 int error; 859 int cache_index; 860 int i; 861 862 hmp = NULL; 863 pmp = NULL; 864 dev = NULL; 865 label = NULL; 866 devvp = NULL; 867 cache_index = -1; 868 869 kprintf("hammer2_mount\n"); 870 871 if (path == NULL) { 872 /* 873 * Root mount 874 */ 875 bzero(&info, sizeof(info)); 876 info.cluster_fd = -1; 877 ksnprintf(devstr, sizeof(devstr), "%s", 878 mp->mnt_stat.f_mntfromname); 879 kprintf("hammer2_mount: root '%s'\n", devstr); 880 } else { 881 /* 882 * Non-root mount or updating a mount 883 */ 884 error = copyin(data, &info, sizeof(info)); 885 if (error) 886 return (error); 887 888 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done); 889 if (error) 890 return (error); 891 } 892 893 /* 894 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA 895 * if no label specified, based on the partition id. Error out if no 896 * partition id. This is strictly a convenience to match the 897 * default label created by newfs_hammer2, our preference is 898 * that a label always be specified. 899 */ 900 dev = devstr; 901 label = strchr(devstr, '@'); 902 if (label && ((label + 1) - dev) > done) 903 return (EINVAL); 904 if (label && label == devstr) 905 return (EINVAL); 906 if (label == NULL || label[1] == 0) { 907 char slice; 908 909 if (label == NULL) 910 label = devstr + strlen(devstr); 911 slice = label[-1]; 912 switch(slice) { 913 case 'a': 914 label = "BOOT"; 915 break; 916 case 'd': 917 label = "ROOT"; 918 break; 919 default: 920 label = "DATA"; 921 break; 922 } 923 } else { 924 *label = '\0'; 925 label++; 926 } 927 928 kprintf("hammer2_mount: dev=\"%s\" label=\"%s\"\n", 929 dev, label); 930 931 if (mp->mnt_flag & MNT_UPDATE) { 932 /* 933 * Update mount. Note that pmp->iroot->cluster is 934 * an inode-embedded cluster and thus cannot be 935 * directly locked. 936 * 937 * XXX HAMMER2 needs to implement NFS export via 938 * mountctl. 939 */ 940 pmp = MPTOPMP(mp); 941 pmp->hflags = info.hflags; 942 cluster = &pmp->iroot->cluster; 943 for (i = 0; i < cluster->nchains; ++i) { 944 if (cluster->array[i].chain == NULL) 945 continue; 946 hmp = cluster->array[i].chain->hmp; 947 devvp = hmp->devvp; 948 error = hammer2_remount(hmp, mp, path, 949 devvp, cred); 950 if (error) 951 break; 952 } 953 954 return error; 955 } 956 957 /* 958 * HMP device mount 959 * 960 * If a path is specified and dev is not an empty string, lookup the 961 * name and verify that it referes to a block device. 962 * 963 * If a path is specified and dev is an empty string we fall through 964 * and locate the label in the hmp search. 965 */ 966 if (path && *dev != 0) { 967 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW); 968 if (error == 0) 969 error = nlookup(&nd); 970 if (error == 0) 971 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp); 972 nlookup_done(&nd); 973 } else if (path == NULL) { 974 /* root mount */ 975 cdev_t cdev = kgetdiskbyname(dev); 976 error = bdevvp(cdev, &devvp); 977 if (error) 978 kprintf("hammer2: cannot find '%s'\n", dev); 979 } else { 980 /* 981 * We will locate the hmp using the label in the hmp loop. 982 */ 983 error = 0; 984 } 985 986 if (error == 0 && devvp) { 987 if (vn_isdisk(devvp, &error)) 988 error = vfs_mountedon(devvp); 989 } 990 991 /* 992 * Determine if the device has already been mounted. After this 993 * check hmp will be non-NULL if we are doing the second or more 994 * hammer2 mounts from the same device. 995 */ 996 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 997 if (devvp) { 998 /* 999 * Match the device 1000 */ 1001 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1002 if (hmp->devvp == devvp) 1003 break; 1004 } 1005 } else if (error == 0) { 1006 /* 1007 * Match the label to a pmp already probed. 1008 */ 1009 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 1010 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 1011 if (pmp->pfs_names[i] && 1012 strcmp(pmp->pfs_names[i], label) == 0) { 1013 hmp = pmp->pfs_hmps[i]; 1014 break; 1015 } 1016 } 1017 if (hmp) 1018 break; 1019 } 1020 if (hmp == NULL) 1021 error = ENOENT; 1022 } 1023 1024 /* 1025 * Open the device if this isn't a secondary mount and construct 1026 * the H2 device mount (hmp). 1027 */ 1028 if (hmp == NULL) { 1029 hammer2_chain_t *schain; 1030 hammer2_xid_t xid; 1031 1032 if (error == 0 && vcount(devvp) > 0) 1033 error = EBUSY; 1034 1035 /* 1036 * Now open the device 1037 */ 1038 if (error == 0) { 1039 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0); 1040 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1041 error = vinvalbuf(devvp, V_SAVE, 0, 0); 1042 if (error == 0) { 1043 error = VOP_OPEN(devvp, 1044 ronly ? FREAD : FREAD | FWRITE, 1045 FSCRED, NULL); 1046 } 1047 vn_unlock(devvp); 1048 } 1049 if (error && devvp) { 1050 vrele(devvp); 1051 devvp = NULL; 1052 } 1053 if (error) { 1054 lockmgr(&hammer2_mntlk, LK_RELEASE); 1055 return error; 1056 } 1057 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO); 1058 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev); 1059 hmp->ronly = ronly; 1060 hmp->devvp = devvp; 1061 hmp->hflags = info.hflags & HMNT2_DEVFLAGS; 1062 kmalloc_create(&hmp->mchain, "HAMMER2-chains"); 1063 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry); 1064 RB_INIT(&hmp->iotree); 1065 spin_init(&hmp->io_spin, "hm2mount_io"); 1066 spin_init(&hmp->list_spin, "hm2mount_list"); 1067 TAILQ_INIT(&hmp->flushq); 1068 1069 lockinit(&hmp->vollk, "h2vol", 0, 0); 1070 lockinit(&hmp->bulklk, "h2bulk", 0, 0); 1071 lockinit(&hmp->bflock, "h2bflk", 0, 0); 1072 1073 /* 1074 * vchain setup. vchain.data is embedded. 1075 * vchain.refs is initialized and will never drop to 0. 1076 * 1077 * NOTE! voldata is not yet loaded. 1078 */ 1079 hmp->vchain.hmp = hmp; 1080 hmp->vchain.refs = 1; 1081 hmp->vchain.data = (void *)&hmp->voldata; 1082 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME; 1083 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1084 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1085 1086 hammer2_chain_core_init(&hmp->vchain); 1087 /* hmp->vchain.u.xxx is left NULL */ 1088 1089 /* 1090 * fchain setup. fchain.data is embedded. 1091 * fchain.refs is initialized and will never drop to 0. 1092 * 1093 * The data is not used but needs to be initialized to 1094 * pass assertion muster. We use this chain primarily 1095 * as a placeholder for the freemap's top-level RBTREE 1096 * so it does not interfere with the volume's topology 1097 * RBTREE. 1098 */ 1099 hmp->fchain.hmp = hmp; 1100 hmp->fchain.refs = 1; 1101 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset; 1102 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP; 1103 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1104 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1105 hmp->fchain.bref.methods = 1106 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) | 1107 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE); 1108 1109 hammer2_chain_core_init(&hmp->fchain); 1110 /* hmp->fchain.u.xxx is left NULL */ 1111 1112 /* 1113 * Install the volume header and initialize fields from 1114 * voldata. 1115 */ 1116 error = hammer2_install_volume_header(hmp); 1117 if (error) { 1118 hammer2_unmount_helper(mp, NULL, hmp); 1119 lockmgr(&hammer2_mntlk, LK_RELEASE); 1120 hammer2_vfs_unmount(mp, MNT_FORCE); 1121 return error; 1122 } 1123 1124 /* 1125 * Really important to get these right or flush will get 1126 * confused. 1127 */ 1128 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL); 1129 kprintf("alloc spmp %p tid %016jx\n", 1130 hmp->spmp, hmp->voldata.mirror_tid); 1131 spmp = hmp->spmp; 1132 1133 /* 1134 * Dummy-up vchain and fchain's modify_tid. mirror_tid 1135 * is inherited from the volume header. 1136 */ 1137 xid = 0; 1138 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1139 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid; 1140 hmp->vchain.pmp = spmp; 1141 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1142 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid; 1143 hmp->fchain.pmp = spmp; 1144 1145 /* 1146 * First locate the super-root inode, which is key 0 1147 * relative to the volume header's blockset. 1148 * 1149 * Then locate the root inode by scanning the directory keyspace 1150 * represented by the label. 1151 */ 1152 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 1153 schain = hammer2_chain_lookup(&parent, &key_dummy, 1154 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY, 1155 &cache_index, 0); 1156 hammer2_chain_lookup_done(parent); 1157 if (schain == NULL) { 1158 kprintf("hammer2_mount: invalid super-root\n"); 1159 hammer2_unmount_helper(mp, NULL, hmp); 1160 lockmgr(&hammer2_mntlk, LK_RELEASE); 1161 hammer2_vfs_unmount(mp, MNT_FORCE); 1162 return EINVAL; 1163 } 1164 if (schain->error) { 1165 kprintf("hammer2_mount: error %s reading super-root\n", 1166 hammer2_error_str(schain->error)); 1167 hammer2_chain_unlock(schain); 1168 hammer2_chain_drop(schain); 1169 schain = NULL; 1170 hammer2_unmount_helper(mp, NULL, hmp); 1171 lockmgr(&hammer2_mntlk, LK_RELEASE); 1172 hammer2_vfs_unmount(mp, MNT_FORCE); 1173 return EINVAL; 1174 } 1175 1176 /* 1177 * The super-root always uses an inode_tid of 1 when 1178 * creating PFSs. 1179 */ 1180 spmp->inode_tid = 1; 1181 spmp->modify_tid = schain->bref.modify_tid + 1; 1182 1183 /* 1184 * Sanity-check schain's pmp and finish initialization. 1185 * Any chain belonging to the super-root topology should 1186 * have a NULL pmp (not even set to spmp). 1187 */ 1188 ripdata = &hammer2_chain_rdata(schain)->ipdata; 1189 KKASSERT(schain->pmp == NULL); 1190 spmp->pfs_clid = ripdata->meta.pfs_clid; 1191 1192 /* 1193 * Replace the dummy spmp->iroot with a real one. It's 1194 * easier to just do a wholesale replacement than to try 1195 * to update the chain and fixup the iroot fields. 1196 * 1197 * The returned inode is locked with the supplied cluster. 1198 */ 1199 cluster = hammer2_cluster_from_chain(schain); 1200 hammer2_inode_drop(spmp->iroot); 1201 spmp->iroot = NULL; 1202 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1); 1203 spmp->spmp_hmp = hmp; 1204 spmp->pfs_types[0] = ripdata->meta.pfs_type; 1205 spmp->pfs_hmps[0] = hmp; 1206 hammer2_inode_ref(spmp->iroot); 1207 hammer2_inode_unlock(spmp->iroot); 1208 hammer2_cluster_unlock(cluster); 1209 hammer2_cluster_drop(cluster); 1210 schain = NULL; 1211 /* leave spmp->iroot with one ref */ 1212 1213 if ((mp->mnt_flag & MNT_RDONLY) == 0) { 1214 error = hammer2_recovery(hmp); 1215 /* XXX do something with error */ 1216 } 1217 hammer2_update_pmps(hmp); 1218 hammer2_iocom_init(hmp); 1219 hammer2_bulkfree_init(hmp); 1220 1221 /* 1222 * Ref the cluster management messaging descriptor. The mount 1223 * program deals with the other end of the communications pipe. 1224 * 1225 * Root mounts typically do not supply one. 1226 */ 1227 if (info.cluster_fd >= 0) { 1228 fp = holdfp(curproc->p_fd, info.cluster_fd, -1); 1229 if (fp) { 1230 hammer2_cluster_reconnect(hmp, fp); 1231 } else { 1232 kprintf("hammer2_mount: bad cluster_fd!\n"); 1233 } 1234 } 1235 } else { 1236 spmp = hmp->spmp; 1237 if (info.hflags & HMNT2_DEVFLAGS) { 1238 kprintf("hammer2: Warning: mount flags pertaining " 1239 "to the whole device may only be specified " 1240 "on the first mount of the device: %08x\n", 1241 info.hflags & HMNT2_DEVFLAGS); 1242 } 1243 } 1244 1245 /* 1246 * Force local mount (disassociate all PFSs from their clusters). 1247 * Used primarily for debugging. 1248 */ 1249 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1250 1251 /* 1252 * Lookup the mount point under the media-localized super-root. 1253 * Scanning hammer2_pfslist doesn't help us because it represents 1254 * PFS cluster ids which can aggregate several named PFSs together. 1255 * 1256 * cluster->pmp will incorrectly point to spmp and must be fixed 1257 * up later on. 1258 */ 1259 hammer2_inode_lock(spmp->iroot, 0); 1260 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1261 lhc = hammer2_dirhash(label, strlen(label)); 1262 chain = hammer2_chain_lookup(&parent, &key_next, 1263 lhc, lhc + HAMMER2_DIRHASH_LOMASK, 1264 &cache_index, 0); 1265 while (chain) { 1266 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && 1267 strcmp(label, chain->data->ipdata.filename) == 0) { 1268 break; 1269 } 1270 chain = hammer2_chain_next(&parent, chain, &key_next, 1271 key_next, 1272 lhc + HAMMER2_DIRHASH_LOMASK, 1273 &cache_index, 0); 1274 } 1275 if (parent) { 1276 hammer2_chain_unlock(parent); 1277 hammer2_chain_drop(parent); 1278 } 1279 hammer2_inode_unlock(spmp->iroot); 1280 1281 /* 1282 * PFS could not be found? 1283 */ 1284 if (chain == NULL) { 1285 kprintf("hammer2_mount: PFS label not found\n"); 1286 hammer2_unmount_helper(mp, NULL, hmp); 1287 lockmgr(&hammer2_mntlk, LK_RELEASE); 1288 hammer2_vfs_unmount(mp, MNT_FORCE); 1289 1290 return EINVAL; 1291 } 1292 1293 /* 1294 * Acquire the pmp structure (it should have already been allocated 1295 * via hammer2_update_pmps() so do not pass cluster in to add to 1296 * available chains). 1297 * 1298 * Check if the cluster has already been mounted. A cluster can 1299 * only be mounted once, use null mounts to mount additional copies. 1300 */ 1301 ripdata = &chain->data->ipdata; 1302 bref = chain->bref; 1303 pmp = hammer2_pfsalloc(NULL, ripdata, 1304 bref.modify_tid, force_local); 1305 hammer2_chain_unlock(chain); 1306 hammer2_chain_drop(chain); 1307 1308 /* 1309 * Finish the mount 1310 */ 1311 kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp); 1312 1313 if (pmp->mp) { 1314 kprintf("hammer2_mount: PFS already mounted!\n"); 1315 hammer2_unmount_helper(mp, NULL, hmp); 1316 lockmgr(&hammer2_mntlk, LK_RELEASE); 1317 hammer2_vfs_unmount(mp, MNT_FORCE); 1318 1319 return EBUSY; 1320 } 1321 1322 pmp->hflags = info.hflags; 1323 mp->mnt_flag = MNT_LOCAL; 1324 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */ 1325 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */ 1326 1327 /* 1328 * required mount structure initializations 1329 */ 1330 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE; 1331 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE; 1332 1333 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE; 1334 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE; 1335 1336 /* 1337 * Optional fields 1338 */ 1339 mp->mnt_iosize_max = MAXPHYS; 1340 1341 /* 1342 * Connect up mount pointers. 1343 */ 1344 hammer2_mount_helper(mp, pmp); 1345 1346 lockmgr(&hammer2_mntlk, LK_RELEASE); 1347 1348 /* 1349 * Finish setup 1350 */ 1351 vfs_getnewfsid(mp); 1352 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops); 1353 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops); 1354 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops); 1355 1356 if (path) { 1357 copyinstr(info.volume, mp->mnt_stat.f_mntfromname, 1358 MNAMELEN - 1, &size); 1359 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size); 1360 } /* else root mount, already in there */ 1361 1362 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname)); 1363 if (path) { 1364 copyinstr(path, mp->mnt_stat.f_mntonname, 1365 sizeof(mp->mnt_stat.f_mntonname) - 1, 1366 &size); 1367 } else { 1368 /* root mount */ 1369 mp->mnt_stat.f_mntonname[0] = '/'; 1370 } 1371 1372 /* 1373 * Initial statfs to prime mnt_stat. 1374 */ 1375 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred); 1376 1377 return 0; 1378 } 1379 1380 /* 1381 * Scan PFSs under the super-root and create hammer2_pfs structures. 1382 */ 1383 static 1384 void 1385 hammer2_update_pmps(hammer2_dev_t *hmp) 1386 { 1387 const hammer2_inode_data_t *ripdata; 1388 hammer2_chain_t *parent; 1389 hammer2_chain_t *chain; 1390 hammer2_blockref_t bref; 1391 hammer2_dev_t *force_local; 1392 hammer2_pfs_t *spmp; 1393 hammer2_pfs_t *pmp; 1394 hammer2_key_t key_next; 1395 int cache_index = -1; 1396 1397 /* 1398 * Force local mount (disassociate all PFSs from their clusters). 1399 * Used primarily for debugging. 1400 */ 1401 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1402 1403 /* 1404 * Lookup mount point under the media-localized super-root. 1405 * 1406 * cluster->pmp will incorrectly point to spmp and must be fixed 1407 * up later on. 1408 */ 1409 spmp = hmp->spmp; 1410 hammer2_inode_lock(spmp->iroot, 0); 1411 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1412 chain = hammer2_chain_lookup(&parent, &key_next, 1413 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 1414 &cache_index, 0); 1415 while (chain) { 1416 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) 1417 continue; 1418 ripdata = &chain->data->ipdata; 1419 bref = chain->bref; 1420 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename); 1421 1422 pmp = hammer2_pfsalloc(chain, ripdata, 1423 bref.modify_tid, force_local); 1424 chain = hammer2_chain_next(&parent, chain, &key_next, 1425 key_next, HAMMER2_KEY_MAX, 1426 &cache_index, 0); 1427 } 1428 if (parent) { 1429 hammer2_chain_unlock(parent); 1430 hammer2_chain_drop(parent); 1431 } 1432 hammer2_inode_unlock(spmp->iroot); 1433 } 1434 1435 static 1436 int 1437 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused, 1438 struct vnode *devvp, struct ucred *cred) 1439 { 1440 int error; 1441 1442 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) { 1443 error = hammer2_recovery(hmp); 1444 } else { 1445 error = 0; 1446 } 1447 return error; 1448 } 1449 1450 static 1451 int 1452 hammer2_vfs_unmount(struct mount *mp, int mntflags) 1453 { 1454 hammer2_pfs_t *pmp; 1455 int flags; 1456 int error = 0; 1457 1458 pmp = MPTOPMP(mp); 1459 1460 if (pmp == NULL) 1461 return(0); 1462 1463 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1464 1465 /* 1466 * If mount initialization proceeded far enough we must flush 1467 * its vnodes and sync the underlying mount points. Three syncs 1468 * are required to fully flush the filesystem (freemap updates lag 1469 * by one flush, and one extra for safety). 1470 */ 1471 if (mntflags & MNT_FORCE) 1472 flags = FORCECLOSE; 1473 else 1474 flags = 0; 1475 if (pmp->iroot) { 1476 error = vflush(mp, 0, flags); 1477 if (error) 1478 goto failed; 1479 hammer2_vfs_sync(mp, MNT_WAIT); 1480 hammer2_vfs_sync(mp, MNT_WAIT); 1481 hammer2_vfs_sync(mp, MNT_WAIT); 1482 } 1483 1484 /* 1485 * Cleanup the frontend support XOPS threads 1486 */ 1487 hammer2_xop_helper_cleanup(pmp); 1488 1489 if (pmp->mp) 1490 hammer2_unmount_helper(mp, pmp, NULL); 1491 1492 error = 0; 1493 failed: 1494 lockmgr(&hammer2_mntlk, LK_RELEASE); 1495 1496 return (error); 1497 } 1498 1499 /* 1500 * Mount helper, hook the system mount into our PFS. 1501 * The mount lock is held. 1502 * 1503 * We must bump the mount_count on related devices for any 1504 * mounted PFSs. 1505 */ 1506 static 1507 void 1508 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp) 1509 { 1510 hammer2_cluster_t *cluster; 1511 hammer2_chain_t *rchain; 1512 int i; 1513 1514 mp->mnt_data = (qaddr_t)pmp; 1515 pmp->mp = mp; 1516 1517 /* 1518 * After pmp->mp is set we have to adjust hmp->mount_count. 1519 */ 1520 cluster = &pmp->iroot->cluster; 1521 for (i = 0; i < cluster->nchains; ++i) { 1522 rchain = cluster->array[i].chain; 1523 if (rchain == NULL) 1524 continue; 1525 ++rchain->hmp->mount_count; 1526 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n", 1527 rchain->hmp, rchain->hmp->mount_count); 1528 } 1529 1530 /* 1531 * Create missing Xop threads 1532 */ 1533 hammer2_xop_helper_create(pmp); 1534 } 1535 1536 /* 1537 * Mount helper, unhook the system mount from our PFS. 1538 * The mount lock is held. 1539 * 1540 * If hmp is supplied a mount responsible for being the first to open 1541 * the block device failed and the block device and all PFSs using the 1542 * block device must be cleaned up. 1543 * 1544 * If pmp is supplied multiple devices might be backing the PFS and each 1545 * must be disconnected. This might not be the last PFS using some of the 1546 * underlying devices. Also, we have to adjust our hmp->mount_count 1547 * accounting for the devices backing the pmp which is now undergoing an 1548 * unmount. 1549 */ 1550 static 1551 void 1552 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp) 1553 { 1554 hammer2_cluster_t *cluster; 1555 hammer2_chain_t *rchain; 1556 struct vnode *devvp; 1557 int dumpcnt; 1558 int ronly = 0; 1559 int i; 1560 1561 /* 1562 * If no device supplied this is a high-level unmount and we have to 1563 * to disconnect the mount, adjust mount_count, and locate devices 1564 * that might now have no mounts. 1565 */ 1566 if (pmp) { 1567 KKASSERT(hmp == NULL); 1568 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp); 1569 pmp->mp = NULL; 1570 mp->mnt_data = NULL; 1571 1572 /* 1573 * After pmp->mp is cleared we have to account for 1574 * mount_count. 1575 */ 1576 cluster = &pmp->iroot->cluster; 1577 for (i = 0; i < cluster->nchains; ++i) { 1578 rchain = cluster->array[i].chain; 1579 if (rchain == NULL) 1580 continue; 1581 --rchain->hmp->mount_count; 1582 kprintf("hammer2_unmount hmp=%p --mount_count=%d\n", 1583 rchain->hmp, rchain->hmp->mount_count); 1584 /* scrapping hmp now may invalidate the pmp */ 1585 } 1586 again: 1587 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1588 if (hmp->mount_count == 0) { 1589 hammer2_unmount_helper(NULL, NULL, hmp); 1590 goto again; 1591 } 1592 } 1593 return; 1594 } 1595 1596 /* 1597 * Try to terminate the block device. We can't terminate it if 1598 * there are still PFSs referencing it. 1599 */ 1600 kprintf("hammer2_unmount hmp=%p mount_count=%d\n", 1601 hmp, hmp->mount_count); 1602 if (hmp->mount_count) 1603 return; 1604 1605 hammer2_bulkfree_uninit(hmp); 1606 hammer2_pfsfree_scan(hmp); 1607 hammer2_dev_exlock(hmp); /* XXX order */ 1608 1609 /* 1610 * Cycle the volume data lock as a safety (probably not needed any 1611 * more). To ensure everything is out we need to flush at least 1612 * three times. (1) The running of the sideq can dirty the 1613 * filesystem, (2) A normal flush can dirty the freemap, and 1614 * (3) ensure that the freemap is fully synchronized. 1615 * 1616 * The next mount's recovery scan can clean everything up but we want 1617 * to leave the filesystem in a 100% clean state on a normal unmount. 1618 */ 1619 #if 0 1620 hammer2_voldata_lock(hmp); 1621 hammer2_voldata_unlock(hmp); 1622 #endif 1623 hammer2_iocom_uninit(hmp); 1624 1625 if ((hmp->vchain.flags | hmp->fchain.flags) & 1626 HAMMER2_CHAIN_FLUSH_MASK) { 1627 kprintf("hammer2_unmount: chains left over " 1628 "after final sync\n"); 1629 kprintf(" vchain %08x\n", hmp->vchain.flags); 1630 kprintf(" fchain %08x\n", hmp->fchain.flags); 1631 1632 if (hammer2_debug & 0x0010) 1633 Debugger("entered debugger"); 1634 } 1635 1636 KKASSERT(hmp->spmp == NULL); 1637 1638 /* 1639 * Finish up with the device vnode 1640 */ 1641 if ((devvp = hmp->devvp) != NULL) { 1642 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1643 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0); 1644 hmp->devvp = NULL; 1645 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL); 1646 vn_unlock(devvp); 1647 vrele(devvp); 1648 devvp = NULL; 1649 } 1650 1651 /* 1652 * Clear vchain/fchain flags that might prevent final cleanup 1653 * of these chains. 1654 */ 1655 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) { 1656 atomic_add_long(&hammer2_count_modified_chains, -1); 1657 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 1658 hammer2_pfs_memory_wakeup(hmp->vchain.pmp); 1659 } 1660 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) { 1661 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE); 1662 } 1663 1664 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) { 1665 atomic_add_long(&hammer2_count_modified_chains, -1); 1666 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED); 1667 hammer2_pfs_memory_wakeup(hmp->fchain.pmp); 1668 } 1669 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) { 1670 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE); 1671 } 1672 1673 /* 1674 * Final drop of embedded freemap root chain to 1675 * clean up fchain.core (fchain structure is not 1676 * flagged ALLOCATED so it is cleaned out and then 1677 * left to rot). 1678 */ 1679 hammer2_chain_drop(&hmp->fchain); 1680 1681 /* 1682 * Final drop of embedded volume root chain to clean 1683 * up vchain.core (vchain structure is not flagged 1684 * ALLOCATED so it is cleaned out and then left to 1685 * rot). 1686 */ 1687 dumpcnt = 50; 1688 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v'); 1689 dumpcnt = 50; 1690 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f'); 1691 hammer2_dev_unlock(hmp); 1692 hammer2_chain_drop(&hmp->vchain); 1693 1694 hammer2_io_cleanup(hmp, &hmp->iotree); 1695 if (hmp->iofree_count) { 1696 kprintf("io_cleanup: %d I/O's left hanging\n", 1697 hmp->iofree_count); 1698 } 1699 1700 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry); 1701 kmalloc_destroy(&hmp->mchain); 1702 kfree(hmp, M_HAMMER2); 1703 } 1704 1705 int 1706 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp, 1707 ino_t ino, struct vnode **vpp) 1708 { 1709 hammer2_xop_lookup_t *xop; 1710 hammer2_pfs_t *pmp; 1711 hammer2_inode_t *ip; 1712 hammer2_tid_t inum; 1713 int error; 1714 1715 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK; 1716 1717 error = 0; 1718 pmp = MPTOPMP(mp); 1719 1720 /* 1721 * Easy if we already have it cached 1722 */ 1723 ip = hammer2_inode_lookup(pmp, inum); 1724 if (ip) { 1725 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED); 1726 *vpp = hammer2_igetv(ip, &error); 1727 hammer2_inode_unlock(ip); 1728 hammer2_inode_drop(ip); /* from lookup */ 1729 1730 return error; 1731 } 1732 1733 /* 1734 * Otherwise we have to find the inode 1735 */ 1736 xop = hammer2_xop_alloc(pmp->iroot, 0); 1737 xop->lhc = inum; 1738 hammer2_xop_start(&xop->head, hammer2_xop_lookup); 1739 error = hammer2_xop_collect(&xop->head, 0); 1740 1741 if (error == 0) { 1742 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) { 1743 kprintf("vget: no collect error but also no rdata\n"); 1744 kprintf("xop %p\n", xop); 1745 while ((hammer2_debug & 0x80000) == 0) { 1746 tsleep(xop, PCATCH, "wait", hz * 10); 1747 } 1748 ip = NULL; 1749 } else { 1750 ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1); 1751 } 1752 } 1753 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1754 1755 if (ip) { 1756 *vpp = hammer2_igetv(ip, &error); 1757 hammer2_inode_unlock(ip); 1758 } else { 1759 *vpp = NULL; 1760 error = ENOENT; 1761 } 1762 return (error); 1763 } 1764 1765 static 1766 int 1767 hammer2_vfs_root(struct mount *mp, struct vnode **vpp) 1768 { 1769 hammer2_pfs_t *pmp; 1770 struct vnode *vp; 1771 int error; 1772 1773 pmp = MPTOPMP(mp); 1774 if (pmp->iroot == NULL) { 1775 *vpp = NULL; 1776 return EINVAL; 1777 } 1778 1779 error = 0; 1780 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1781 1782 while (pmp->inode_tid == 0) { 1783 hammer2_xop_ipcluster_t *xop; 1784 hammer2_inode_meta_t *meta; 1785 1786 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING); 1787 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster); 1788 error = hammer2_xop_collect(&xop->head, 0); 1789 1790 if (error == 0) { 1791 meta = &xop->head.cluster.focus->data->ipdata.meta; 1792 pmp->iroot->meta = *meta; 1793 pmp->inode_tid = meta->pfs_inum + 1; 1794 if (pmp->inode_tid < HAMMER2_INODE_START) 1795 pmp->inode_tid = HAMMER2_INODE_START; 1796 pmp->modify_tid = 1797 xop->head.cluster.focus->bref.modify_tid + 1; 1798 kprintf("PFS: Starting inode %jd\n", 1799 (intmax_t)pmp->inode_tid); 1800 kprintf("PMP focus good set nextino=%ld mod=%016jx\n", 1801 pmp->inode_tid, pmp->modify_tid); 1802 wakeup(&pmp->iroot); 1803 1804 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1805 1806 /* 1807 * Prime the mount info. 1808 */ 1809 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL); 1810 break; 1811 } 1812 1813 /* 1814 * Loop, try again 1815 */ 1816 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1817 hammer2_inode_unlock(pmp->iroot); 1818 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz); 1819 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1820 if (error == EINTR) 1821 break; 1822 } 1823 1824 if (error) { 1825 hammer2_inode_unlock(pmp->iroot); 1826 *vpp = NULL; 1827 } else { 1828 vp = hammer2_igetv(pmp->iroot, &error); 1829 hammer2_inode_unlock(pmp->iroot); 1830 *vpp = vp; 1831 } 1832 1833 return (error); 1834 } 1835 1836 /* 1837 * Filesystem status 1838 * 1839 * XXX incorporate ipdata->meta.inode_quota and data_quota 1840 */ 1841 static 1842 int 1843 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred) 1844 { 1845 hammer2_pfs_t *pmp; 1846 hammer2_dev_t *hmp; 1847 hammer2_blockref_t bref; 1848 int i; 1849 1850 /* 1851 * NOTE: iroot might not have validated the cluster yet. 1852 */ 1853 pmp = MPTOPMP(mp); 1854 1855 mp->mnt_stat.f_files = 0; 1856 mp->mnt_stat.f_ffree = 0; 1857 mp->mnt_stat.f_blocks = 0; 1858 mp->mnt_stat.f_bfree = 0; 1859 mp->mnt_stat.f_bavail = 0; 1860 1861 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 1862 hmp = pmp->pfs_hmps[i]; 1863 if (hmp == NULL) 1864 continue; 1865 if (pmp->iroot->cluster.array[i].chain) 1866 bref = pmp->iroot->cluster.array[i].chain->bref; 1867 else 1868 bzero(&bref, sizeof(bref)); 1869 1870 mp->mnt_stat.f_files = bref.embed.stats.inode_count; 1871 mp->mnt_stat.f_ffree = 0; 1872 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size / 1873 mp->mnt_vstat.f_bsize; 1874 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free / 1875 mp->mnt_vstat.f_bsize; 1876 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree; 1877 1878 *sbp = mp->mnt_stat; 1879 } 1880 return (0); 1881 } 1882 1883 static 1884 int 1885 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred) 1886 { 1887 hammer2_pfs_t *pmp; 1888 hammer2_dev_t *hmp; 1889 hammer2_blockref_t bref; 1890 int i; 1891 1892 /* 1893 * NOTE: iroot might not have validated the cluster yet. 1894 */ 1895 pmp = MPTOPMP(mp); 1896 1897 mp->mnt_vstat.f_bsize = 0; 1898 mp->mnt_vstat.f_files = 0; 1899 mp->mnt_vstat.f_ffree = 0; 1900 mp->mnt_vstat.f_blocks = 0; 1901 mp->mnt_vstat.f_bfree = 0; 1902 mp->mnt_vstat.f_bavail = 0; 1903 1904 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 1905 hmp = pmp->pfs_hmps[i]; 1906 if (hmp == NULL) 1907 continue; 1908 if (pmp->iroot->cluster.array[i].chain) 1909 bref = pmp->iroot->cluster.array[i].chain->bref; 1910 else 1911 bzero(&bref, sizeof(bref)); 1912 1913 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE; 1914 mp->mnt_vstat.f_files = bref.embed.stats.inode_count; 1915 mp->mnt_vstat.f_ffree = 0; 1916 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size / 1917 mp->mnt_vstat.f_bsize; 1918 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free / 1919 mp->mnt_vstat.f_bsize; 1920 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree; 1921 1922 *sbp = mp->mnt_vstat; 1923 } 1924 return (0); 1925 } 1926 1927 /* 1928 * Mount-time recovery (RW mounts) 1929 * 1930 * Updates to the free block table are allowed to lag flushes by one 1931 * transaction. In case of a crash, then on a fresh mount we must do an 1932 * incremental scan of the last committed transaction id and make sure that 1933 * all related blocks have been marked allocated. 1934 * 1935 * The super-root topology and each PFS has its own transaction id domain, 1936 * so we must track PFS boundary transitions. 1937 */ 1938 struct hammer2_recovery_elm { 1939 TAILQ_ENTRY(hammer2_recovery_elm) entry; 1940 hammer2_chain_t *chain; 1941 hammer2_tid_t sync_tid; 1942 }; 1943 1944 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm); 1945 1946 struct hammer2_recovery_info { 1947 struct hammer2_recovery_list list; 1948 hammer2_tid_t mtid; 1949 int depth; 1950 }; 1951 1952 static int hammer2_recovery_scan(hammer2_dev_t *hmp, 1953 hammer2_chain_t *parent, 1954 struct hammer2_recovery_info *info, 1955 hammer2_tid_t sync_tid); 1956 1957 #define HAMMER2_RECOVERY_MAXDEPTH 10 1958 1959 static 1960 int 1961 hammer2_recovery(hammer2_dev_t *hmp) 1962 { 1963 struct hammer2_recovery_info info; 1964 struct hammer2_recovery_elm *elm; 1965 hammer2_chain_t *parent; 1966 hammer2_tid_t sync_tid; 1967 hammer2_tid_t mirror_tid; 1968 int error; 1969 int cumulative_error = 0; 1970 1971 hammer2_trans_init(hmp->spmp, 0); 1972 1973 sync_tid = hmp->voldata.freemap_tid; 1974 mirror_tid = hmp->voldata.mirror_tid; 1975 1976 kprintf("hammer2 mount \"%s\": ", hmp->devrepname); 1977 if (sync_tid >= mirror_tid) { 1978 kprintf(" no recovery needed\n"); 1979 } else { 1980 kprintf(" freemap recovery %016jx-%016jx\n", 1981 sync_tid + 1, mirror_tid); 1982 } 1983 1984 TAILQ_INIT(&info.list); 1985 info.depth = 0; 1986 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 1987 cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid); 1988 hammer2_chain_lookup_done(parent); 1989 1990 while ((elm = TAILQ_FIRST(&info.list)) != NULL) { 1991 TAILQ_REMOVE(&info.list, elm, entry); 1992 parent = elm->chain; 1993 sync_tid = elm->sync_tid; 1994 kfree(elm, M_HAMMER2); 1995 1996 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 1997 error = hammer2_recovery_scan(hmp, parent, &info, 1998 hmp->voldata.freemap_tid); 1999 hammer2_chain_unlock(parent); 2000 hammer2_chain_drop(parent); /* drop elm->chain ref */ 2001 if (error) 2002 cumulative_error = error; 2003 } 2004 hammer2_trans_done(hmp->spmp); 2005 2006 return cumulative_error; 2007 } 2008 2009 static 2010 int 2011 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent, 2012 struct hammer2_recovery_info *info, 2013 hammer2_tid_t sync_tid) 2014 { 2015 const hammer2_inode_data_t *ripdata; 2016 hammer2_chain_t *chain; 2017 hammer2_blockref_t bref; 2018 int cache_index; 2019 int cumulative_error = 0; 2020 int error; 2021 int first; 2022 2023 /* 2024 * Adjust freemap to ensure that the block(s) are marked allocated. 2025 */ 2026 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) { 2027 hammer2_freemap_adjust(hmp, &parent->bref, 2028 HAMMER2_FREEMAP_DORECOVER); 2029 } 2030 2031 /* 2032 * Check type for recursive scan 2033 */ 2034 switch(parent->bref.type) { 2035 case HAMMER2_BREF_TYPE_VOLUME: 2036 /* data already instantiated */ 2037 break; 2038 case HAMMER2_BREF_TYPE_INODE: 2039 /* 2040 * Must instantiate data for DIRECTDATA test and also 2041 * for recursion. 2042 */ 2043 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2044 ripdata = &hammer2_chain_rdata(parent)->ipdata; 2045 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) { 2046 /* not applicable to recovery scan */ 2047 hammer2_chain_unlock(parent); 2048 return 0; 2049 } 2050 hammer2_chain_unlock(parent); 2051 break; 2052 case HAMMER2_BREF_TYPE_INDIRECT: 2053 /* 2054 * Must instantiate data for recursion 2055 */ 2056 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2057 hammer2_chain_unlock(parent); 2058 break; 2059 case HAMMER2_BREF_TYPE_DIRENT: 2060 case HAMMER2_BREF_TYPE_DATA: 2061 case HAMMER2_BREF_TYPE_FREEMAP: 2062 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2063 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 2064 /* not applicable to recovery scan */ 2065 return 0; 2066 break; 2067 default: 2068 return EDOM; 2069 } 2070 2071 /* 2072 * Defer operation if depth limit reached or if we are crossing a 2073 * PFS boundary. 2074 */ 2075 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) { 2076 struct hammer2_recovery_elm *elm; 2077 2078 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK); 2079 elm->chain = parent; 2080 elm->sync_tid = sync_tid; 2081 hammer2_chain_ref(parent); 2082 TAILQ_INSERT_TAIL(&info->list, elm, entry); 2083 /* unlocked by caller */ 2084 2085 return(0); 2086 } 2087 2088 2089 /* 2090 * Recursive scan of the last flushed transaction only. We are 2091 * doing this without pmp assignments so don't leave the chains 2092 * hanging around after we are done with them. 2093 */ 2094 cache_index = 0; 2095 chain = NULL; 2096 first = 1; 2097 2098 while (hammer2_chain_scan(parent, &chain, &bref, 2099 &first, &cache_index, 2100 HAMMER2_LOOKUP_NODATA) != NULL) { 2101 /* 2102 * If this is a leaf 2103 */ 2104 if (chain == NULL) { 2105 if (bref.mirror_tid > sync_tid) { 2106 hammer2_freemap_adjust(hmp, &bref, 2107 HAMMER2_FREEMAP_DORECOVER); 2108 } 2109 continue; 2110 } 2111 2112 /* 2113 * This may or may not be a recursive node. 2114 */ 2115 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 2116 if (bref.mirror_tid > sync_tid) { 2117 ++info->depth; 2118 error = hammer2_recovery_scan(hmp, chain, 2119 info, sync_tid); 2120 --info->depth; 2121 if (error) 2122 cumulative_error = error; 2123 } 2124 2125 /* 2126 * Flush the recovery at the PFS boundary to stage it for 2127 * the final flush of the super-root topology. 2128 */ 2129 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) && 2130 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) { 2131 hammer2_flush(chain, HAMMER2_FLUSH_TOP); 2132 } 2133 } 2134 2135 return cumulative_error; 2136 } 2137 2138 /* 2139 * Sync a mount point; this is called on a per-mount basis from the 2140 * filesystem syncer process periodically and whenever a user issues 2141 * a sync. 2142 */ 2143 int 2144 hammer2_vfs_sync(struct mount *mp, int waitfor) 2145 { 2146 hammer2_xop_flush_t *xop; 2147 struct hammer2_sync_info info; 2148 hammer2_inode_t *iroot; 2149 hammer2_pfs_t *pmp; 2150 int flags; 2151 int error; 2152 2153 pmp = MPTOPMP(mp); 2154 iroot = pmp->iroot; 2155 KKASSERT(iroot); 2156 KKASSERT(iroot->pmp == pmp); 2157 2158 /* 2159 * We can't acquire locks on existing vnodes while in a transaction 2160 * without risking a deadlock. This assumes that vfsync() can be 2161 * called without the vnode locked (which it can in DragonFly). 2162 * Otherwise we'd have to implement a multi-pass or flag the lock 2163 * failures and retry. 2164 * 2165 * The reclamation code interlocks with the sync list's token 2166 * (by removing the vnode from the scan list) before unlocking 2167 * the inode, giving us time to ref the inode. 2168 */ 2169 /*flags = VMSC_GETVP;*/ 2170 flags = 0; 2171 if (waitfor & MNT_LAZY) 2172 flags |= VMSC_ONEPASS; 2173 2174 /* 2175 * Preflush the vnodes using a normal transaction before interlocking 2176 * with a flush transaction. We do this to try to run as much of 2177 * the compression as possible outside the flush transaction. 2178 * 2179 * For efficiency do an async pass before making sure with a 2180 * synchronous pass on all related buffer cache buffers. 2181 */ 2182 hammer2_trans_init(pmp, 0); 2183 info.error = 0; 2184 info.waitfor = MNT_NOWAIT; 2185 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info); 2186 info.waitfor = MNT_WAIT; 2187 vsyncscan(mp, flags, hammer2_sync_scan2, &info); 2188 hammer2_trans_done(pmp); 2189 2190 /* 2191 * Start our flush transaction. This does not return until all 2192 * concurrent transactions have completed and will prevent any 2193 * new transactions from running concurrently, except for the 2194 * buffer cache transactions. 2195 * 2196 * NOTE! It is still possible for the paging code to push pages 2197 * out via a UIO_NOCOPY hammer2_vop_write() during the main 2198 * flush. 2199 */ 2200 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH); 2201 hammer2_inode_run_sideq(pmp); 2202 2203 info.error = 0; 2204 info.waitfor = MNT_NOWAIT; 2205 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info); 2206 info.waitfor = MNT_WAIT; 2207 vsyncscan(mp, flags, hammer2_sync_scan2, &info); 2208 hammer2_bioq_sync(pmp); 2209 2210 /* 2211 * Use the XOP interface to concurrently flush all nodes to 2212 * synchronize the PFSROOT subtopology to the media. A standard 2213 * end-of-scan ENOENT error indicates cluster sufficiency. 2214 * 2215 * Note that this flush will not be visible on crash recovery until 2216 * we flush the super-root topology in the next loop. 2217 * 2218 * XXX For now wait for all flushes to complete. 2219 */ 2220 if (iroot) { 2221 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING); 2222 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush); 2223 error = hammer2_xop_collect(&xop->head, 2224 HAMMER2_XOP_COLLECT_WAITALL); 2225 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 2226 if (error == ENOENT) 2227 error = 0; 2228 } else { 2229 error = 0; 2230 } 2231 hammer2_trans_done(pmp); 2232 2233 return (error); 2234 } 2235 2236 /* 2237 * Sync passes. 2238 * 2239 * Note that we ignore the tranasction mtid we got above. Instead, 2240 * each vfsync below will ultimately get its own via TRANS_BUFCACHE 2241 * transactions. 2242 */ 2243 static int 2244 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data) 2245 { 2246 struct hammer2_sync_info *info = data; 2247 hammer2_inode_t *ip; 2248 int error; 2249 2250 /* 2251 * Degenerate cases. Note that ip == NULL typically means the 2252 * syncer vnode itself and we don't want to vclrisdirty() in that 2253 * situation. 2254 */ 2255 ip = VTOI(vp); 2256 if (ip == NULL) { 2257 return(0); 2258 } 2259 if (vp->v_type == VNON || vp->v_type == VBAD) { 2260 vclrisdirty(vp); 2261 return(0); 2262 } 2263 2264 /* 2265 * VOP_FSYNC will start a new transaction so replicate some code 2266 * here to do it inline (see hammer2_vop_fsync()). 2267 * 2268 * WARNING: The vfsync interacts with the buffer cache and might 2269 * block, we can't hold the inode lock at that time. 2270 * However, we MUST ref ip before blocking to ensure that 2271 * it isn't ripped out from under us (since we do not 2272 * hold a lock on the vnode). 2273 */ 2274 hammer2_inode_ref(ip); 2275 if ((ip->flags & HAMMER2_INODE_MODIFIED) || 2276 !RB_EMPTY(&vp->v_rbdirty_tree)) { 2277 vfsync(vp, info->waitfor, 1, NULL, NULL); 2278 if (ip->flags & (HAMMER2_INODE_RESIZED | 2279 HAMMER2_INODE_MODIFIED)) { 2280 hammer2_inode_lock(ip, 0); 2281 hammer2_inode_chain_sync(ip); 2282 hammer2_inode_unlock(ip); 2283 } 2284 } 2285 if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 && 2286 RB_EMPTY(&vp->v_rbdirty_tree)) { 2287 vclrisdirty(vp); 2288 } 2289 2290 hammer2_inode_drop(ip); 2291 #if 1 2292 error = 0; 2293 if (error) 2294 info->error = error; 2295 #endif 2296 return(0); 2297 } 2298 2299 static 2300 int 2301 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp) 2302 { 2303 hammer2_inode_t *ip; 2304 2305 KKASSERT(MAXFIDSZ >= 16); 2306 ip = VTOI(vp); 2307 fhp->fid_len = offsetof(struct fid, fid_data[16]); 2308 fhp->fid_ext = 0; 2309 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum; 2310 ((hammer2_tid_t *)fhp->fid_data)[1] = 0; 2311 2312 return 0; 2313 } 2314 2315 static 2316 int 2317 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp, 2318 struct fid *fhp, struct vnode **vpp) 2319 { 2320 hammer2_pfs_t *pmp; 2321 hammer2_tid_t inum; 2322 int error; 2323 2324 pmp = MPTOPMP(mp); 2325 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK; 2326 if (vpp) { 2327 if (inum == 1) 2328 error = hammer2_vfs_root(mp, vpp); 2329 else 2330 error = hammer2_vfs_vget(mp, NULL, inum, vpp); 2331 } else { 2332 error = 0; 2333 } 2334 if (error) 2335 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error); 2336 return error; 2337 } 2338 2339 static 2340 int 2341 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 2342 int *exflagsp, struct ucred **credanonp) 2343 { 2344 hammer2_pfs_t *pmp; 2345 struct netcred *np; 2346 int error; 2347 2348 pmp = MPTOPMP(mp); 2349 np = vfs_export_lookup(mp, &pmp->export, nam); 2350 if (np) { 2351 *exflagsp = np->netc_exflags; 2352 *credanonp = &np->netc_anon; 2353 error = 0; 2354 } else { 2355 error = EACCES; 2356 } 2357 return error; 2358 } 2359 2360 /* 2361 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume 2362 * header into the HMP 2363 * 2364 * XXX read four volhdrs and use the one with the highest TID whos CRC 2365 * matches. 2366 * 2367 * XXX check iCRCs. 2368 * 2369 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to 2370 * nonexistant locations. 2371 * 2372 * XXX Record selected volhdr and ring updates to each of 4 volhdrs 2373 */ 2374 static 2375 int 2376 hammer2_install_volume_header(hammer2_dev_t *hmp) 2377 { 2378 hammer2_volume_data_t *vd; 2379 struct buf *bp; 2380 hammer2_crc32_t crc0, crc, bcrc0, bcrc; 2381 int error_reported; 2382 int error; 2383 int valid; 2384 int i; 2385 2386 error_reported = 0; 2387 error = 0; 2388 valid = 0; 2389 bp = NULL; 2390 2391 /* 2392 * There are up to 4 copies of the volume header (syncs iterate 2393 * between them so there is no single master). We don't trust the 2394 * volu_size field so we don't know precisely how large the filesystem 2395 * is, so depend on the OS to return an error if we go beyond the 2396 * block device's EOF. 2397 */ 2398 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) { 2399 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64, 2400 HAMMER2_VOLUME_BYTES, &bp); 2401 if (error) { 2402 brelse(bp); 2403 bp = NULL; 2404 continue; 2405 } 2406 2407 vd = (struct hammer2_volume_data *) bp->b_data; 2408 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) && 2409 (vd->magic != HAMMER2_VOLUME_ID_ABO)) { 2410 brelse(bp); 2411 bp = NULL; 2412 continue; 2413 } 2414 2415 if (vd->magic == HAMMER2_VOLUME_ID_ABO) { 2416 /* XXX: Reversed-endianness filesystem */ 2417 kprintf("hammer2: reverse-endian filesystem detected"); 2418 brelse(bp); 2419 bp = NULL; 2420 continue; 2421 } 2422 2423 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0]; 2424 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF, 2425 HAMMER2_VOLUME_ICRC0_SIZE); 2426 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1]; 2427 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF, 2428 HAMMER2_VOLUME_ICRC1_SIZE); 2429 if ((crc0 != crc) || (bcrc0 != bcrc)) { 2430 kprintf("hammer2 volume header crc " 2431 "mismatch copy #%d %08x/%08x\n", 2432 i, crc0, crc); 2433 error_reported = 1; 2434 brelse(bp); 2435 bp = NULL; 2436 continue; 2437 } 2438 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) { 2439 valid = 1; 2440 hmp->voldata = *vd; 2441 hmp->volhdrno = i; 2442 } 2443 brelse(bp); 2444 bp = NULL; 2445 } 2446 if (valid) { 2447 hmp->volsync = hmp->voldata; 2448 error = 0; 2449 if (error_reported || bootverbose || 1) { /* 1/DEBUG */ 2450 kprintf("hammer2: using volume header #%d\n", 2451 hmp->volhdrno); 2452 } 2453 } else { 2454 error = EINVAL; 2455 kprintf("hammer2: no valid volume headers found!\n"); 2456 } 2457 return (error); 2458 } 2459 2460 /* 2461 * This handles hysteresis on regular file flushes. Because the BIOs are 2462 * routed to a thread it is possible for an excessive number to build up 2463 * and cause long front-end stalls long before the runningbuffspace limit 2464 * is hit, so we implement hammer2_flush_pipe to control the 2465 * hysteresis. 2466 * 2467 * This is a particular problem when compression is used. 2468 */ 2469 void 2470 hammer2_lwinprog_ref(hammer2_pfs_t *pmp) 2471 { 2472 atomic_add_int(&pmp->count_lwinprog, 1); 2473 } 2474 2475 void 2476 hammer2_lwinprog_drop(hammer2_pfs_t *pmp) 2477 { 2478 int lwinprog; 2479 2480 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1); 2481 if ((lwinprog & HAMMER2_LWINPROG_WAITING) && 2482 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) { 2483 atomic_clear_int(&pmp->count_lwinprog, 2484 HAMMER2_LWINPROG_WAITING); 2485 wakeup(&pmp->count_lwinprog); 2486 } 2487 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) && 2488 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) { 2489 atomic_clear_int(&pmp->count_lwinprog, 2490 HAMMER2_LWINPROG_WAITING0); 2491 wakeup(&pmp->count_lwinprog); 2492 } 2493 } 2494 2495 void 2496 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe) 2497 { 2498 int lwinprog; 2499 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING : 2500 HAMMER2_LWINPROG_WAITING0; 2501 2502 for (;;) { 2503 lwinprog = pmp->count_lwinprog; 2504 cpu_ccfence(); 2505 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 2506 break; 2507 tsleep_interlock(&pmp->count_lwinprog, 0); 2508 atomic_set_int(&pmp->count_lwinprog, lwflag); 2509 lwinprog = pmp->count_lwinprog; 2510 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 2511 break; 2512 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz); 2513 } 2514 } 2515 2516 /* 2517 * Manage excessive memory resource use for chain and related 2518 * structures. 2519 */ 2520 void 2521 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp) 2522 { 2523 uint32_t waiting; 2524 uint32_t count; 2525 uint32_t limit; 2526 #if 0 2527 static int zzticks; 2528 #endif 2529 2530 /* 2531 * Atomic check condition and wait. Also do an early speedup of 2532 * the syncer to try to avoid hitting the wait. 2533 */ 2534 for (;;) { 2535 waiting = pmp->inmem_dirty_chains; 2536 cpu_ccfence(); 2537 count = waiting & HAMMER2_DIRTYCHAIN_MASK; 2538 2539 limit = pmp->mp->mnt_nvnodelistsize / 10; 2540 if (limit < hammer2_limit_dirty_chains) 2541 limit = hammer2_limit_dirty_chains; 2542 if (limit < 1000) 2543 limit = 1000; 2544 2545 #if 0 2546 if ((int)(ticks - zzticks) > hz) { 2547 zzticks = ticks; 2548 kprintf("count %ld %ld\n", count, limit); 2549 } 2550 #endif 2551 2552 /* 2553 * Block if there are too many dirty chains present, wait 2554 * for the flush to clean some out. 2555 */ 2556 if (count > limit) { 2557 tsleep_interlock(&pmp->inmem_dirty_chains, 0); 2558 if (atomic_cmpset_int(&pmp->inmem_dirty_chains, 2559 waiting, 2560 waiting | HAMMER2_DIRTYCHAIN_WAITING)) { 2561 speedup_syncer(pmp->mp); 2562 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED, 2563 "chnmem", hz); 2564 } 2565 continue; /* loop on success or fail */ 2566 } 2567 2568 /* 2569 * Try to start an early flush before we are forced to block. 2570 */ 2571 if (count > limit * 7 / 10) 2572 speedup_syncer(pmp->mp); 2573 break; 2574 } 2575 } 2576 2577 void 2578 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp) 2579 { 2580 if (pmp) { 2581 atomic_add_int(&pmp->inmem_dirty_chains, 1); 2582 } 2583 } 2584 2585 void 2586 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp) 2587 { 2588 uint32_t waiting; 2589 2590 if (pmp == NULL) 2591 return; 2592 2593 for (;;) { 2594 waiting = pmp->inmem_dirty_chains; 2595 cpu_ccfence(); 2596 if (atomic_cmpset_int(&pmp->inmem_dirty_chains, 2597 waiting, 2598 (waiting - 1) & 2599 ~HAMMER2_DIRTYCHAIN_WAITING)) { 2600 break; 2601 } 2602 } 2603 2604 if (waiting & HAMMER2_DIRTYCHAIN_WAITING) 2605 wakeup(&pmp->inmem_dirty_chains); 2606 } 2607 2608 /* 2609 * Debugging 2610 */ 2611 void 2612 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx) 2613 { 2614 hammer2_chain_t *scan; 2615 hammer2_chain_t *parent; 2616 2617 --*countp; 2618 if (*countp == 0) { 2619 kprintf("%*.*s...\n", tab, tab, ""); 2620 return; 2621 } 2622 if (*countp < 0) 2623 return; 2624 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n", 2625 tab, tab, "", pfx, 2626 chain, chain->bref.type, 2627 chain->bref.key, chain->bref.keybits, 2628 chain->bref.mirror_tid); 2629 2630 kprintf("%*.*s [%08x] (%s) refs=%d", 2631 tab, tab, "", 2632 chain->flags, 2633 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE && 2634 chain->data) ? (char *)chain->data->ipdata.filename : "?"), 2635 chain->refs); 2636 2637 parent = chain->parent; 2638 if (parent) 2639 kprintf("\n%*.*s p=%p [pflags %08x prefs %d", 2640 tab, tab, "", 2641 parent, parent->flags, parent->refs); 2642 if (RB_EMPTY(&chain->core.rbtree)) { 2643 kprintf("\n"); 2644 } else { 2645 kprintf(" {\n"); 2646 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) 2647 hammer2_dump_chain(scan, tab + 4, countp, 'a'); 2648 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data) 2649 kprintf("%*.*s}(%s)\n", tab, tab, "", 2650 chain->data->ipdata.filename); 2651 else 2652 kprintf("%*.*s}\n", tab, tab, ""); 2653 } 2654 } 2655