1 /* 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 2022 Tomohiro Kusumi <tkusumi@netbsd.org> 5 * Copyright (c) 2011-2022 The DragonFly Project. All rights reserved. 6 * 7 * This code is derived from software contributed to The DragonFly Project 8 * by Matthew Dillon <dillon@dragonflybsd.org> 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in 18 * the documentation and/or other materials provided with the 19 * distribution. 20 * 3. Neither the name of The DragonFly Project nor the names of its 21 * contributors may be used to endorse or promote products derived 22 * from this software without specific, prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 26 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 27 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 28 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 29 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 30 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 31 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 32 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 33 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 34 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 35 * SUCH DAMAGE. 36 */ 37 /* 38 #include <sys/param.h> 39 #include <sys/systm.h> 40 #include <sys/kernel.h> 41 #include <sys/nlookup.h> 42 #include <sys/vnode.h> 43 #include <sys/mount.h> 44 #include <sys/fcntl.h> 45 #include <sys/vfsops.h> 46 #include <sys/sysctl.h> 47 #include <sys/socket.h> 48 #include <sys/objcache.h> 49 #include <sys/proc.h> 50 #include <sys/lock.h> 51 #include <sys/file.h> 52 */ 53 54 #include "hammer2.h" 55 56 TAILQ_HEAD(hammer2_mntlist, hammer2_dev); 57 static struct hammer2_mntlist hammer2_mntlist; 58 59 struct hammer2_pfslist hammer2_pfslist; 60 struct hammer2_pfslist hammer2_spmplist; 61 struct lock hammer2_mntlk; 62 63 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT; 64 int hammer2_debug; 65 int hammer2_aux_flags; 66 int hammer2_xop_nthreads; 67 int hammer2_xop_sgroups; 68 int hammer2_xop_xgroups; 69 int hammer2_xop_xbase; 70 int hammer2_xop_mod; 71 long hammer2_debug_inode; 72 int hammer2_cluster_meta_read = 1; /* physical read-ahead */ 73 int hammer2_cluster_data_read = 4; /* physical read-ahead */ 74 int hammer2_cluster_write = 0; /* physical write clustering */ 75 int hammer2_dedup_enable = 1; 76 int hammer2_always_compress = 0; /* always try to compress */ 77 int hammer2_flush_pipe = 100; 78 int hammer2_dio_count; 79 int hammer2_dio_limit = 256; 80 int hammer2_bulkfree_tps = 5000; 81 int hammer2_spread_workers; 82 int hammer2_limit_saved_depth; 83 long hammer2_chain_allocs; 84 long hammer2_limit_saved_chains; 85 long hammer2_limit_dirty_chains; 86 long hammer2_limit_dirty_inodes; 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 static long hammer2_iod_inode_creates; 102 static long hammer2_iod_inode_deletes; 103 104 long hammer2_process_icrc32; 105 long hammer2_process_xxhash64; 106 107 int hz; 108 int ticks; 109 int64_t vnode_count; 110 111 MALLOC_DECLARE(M_HAMMER2_CBUFFER); 112 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer", 113 "Buffer used for compression."); 114 115 MALLOC_DECLARE(M_HAMMER2_DEBUFFER); 116 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer", 117 "Buffer used for decompression."); 118 119 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem"); 120 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD, 122 &hammer2_supported_version, 0, ""); 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, aux_flags, CTLFLAG_RW, 124 &hammer2_aux_flags, 0, ""); 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW, 126 &hammer2_debug, 0, ""); 127 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW, 128 &hammer2_debug_inode, 0, ""); 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, spread_workers, CTLFLAG_RW, 130 &hammer2_spread_workers, 0, ""); 131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW, 132 &hammer2_cluster_meta_read, 0, ""); 133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW, 134 &hammer2_cluster_data_read, 0, ""); 135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW, 136 &hammer2_cluster_write, 0, ""); 137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW, 138 &hammer2_dedup_enable, 0, ""); 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW, 140 &hammer2_always_compress, 0, ""); 141 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW, 142 &hammer2_flush_pipe, 0, ""); 143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW, 144 &hammer2_bulkfree_tps, 0, ""); 145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RD, 146 &hammer2_chain_allocs, 0, ""); 147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_saved_chains, CTLFLAG_RW, 148 &hammer2_limit_saved_chains, 0, ""); 149 SYSCTL_INT(_vfs_hammer2, OID_AUTO, limit_saved_depth, CTLFLAG_RW, 150 &hammer2_limit_saved_depth, 0, ""); 151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW, 152 &hammer2_limit_dirty_chains, 0, ""); 153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW, 154 &hammer2_limit_dirty_inodes, 0, ""); 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RD, 156 &hammer2_count_modified_chains, 0, ""); 157 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD, 158 &hammer2_dio_count, 0, ""); 159 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW, 160 &hammer2_dio_limit, 0, ""); 161 162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RD, 163 &hammer2_iod_file_read, 0, ""); 164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RD, 165 &hammer2_iod_meta_read, 0, ""); 166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RD, 167 &hammer2_iod_indr_read, 0, ""); 168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RD, 169 &hammer2_iod_fmap_read, 0, ""); 170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RD, 171 &hammer2_iod_volu_read, 0, ""); 172 173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RD, 174 &hammer2_iod_file_write, 0, ""); 175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RD, 176 &hammer2_iod_file_wembed, 0, ""); 177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RD, 178 &hammer2_iod_file_wzero, 0, ""); 179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RD, 180 &hammer2_iod_file_wdedup, 0, ""); 181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RD, 182 &hammer2_iod_meta_write, 0, ""); 183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RD, 184 &hammer2_iod_indr_write, 0, ""); 185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RD, 186 &hammer2_iod_fmap_write, 0, ""); 187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RD, 188 &hammer2_iod_volu_write, 0, ""); 189 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RD, 190 &hammer2_iod_inode_creates, 0, ""); 191 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RD, 192 &hammer2_iod_inode_deletes, 0, ""); 193 194 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RD, 195 &hammer2_process_icrc32, 0, ""); 196 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RD, 197 &hammer2_process_xxhash64, 0, ""); 198 199 /* 200 static int hammer2_vfs_init(struct vfsconf *conf); 201 static int hammer2_vfs_uninit(struct vfsconf *vfsp); 202 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data, 203 struct ucred *cred); 204 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *, 205 struct ucred *); 206 */ 207 static int hammer2_recovery(hammer2_dev_t *hmp); 208 /* 209 static int hammer2_vfs_unmount(struct mount *mp, int mntflags); 210 static int hammer2_vfs_root(struct mount *mp, struct m_vnode **vpp); 211 */ 212 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, 213 struct ucred *cred); 214 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, 215 struct ucred *cred); 216 /* 217 static int hammer2_vfs_fhtovp(struct mount *mp, struct m_vnode *rootvp, 218 struct fid *fhp, struct m_vnode **vpp); 219 static int hammer2_vfs_vptofh(struct m_vnode *vp, struct fid *fhp); 220 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 221 int *exflagsp, struct ucred **credanonp); 222 static int hammer2_vfs_modifying(struct mount *mp); 223 */ 224 225 static void hammer2_update_pmps(hammer2_dev_t *hmp); 226 227 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp); 228 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, 229 hammer2_dev_t *hmp); 230 static int hammer2_fixup_pfses(hammer2_dev_t *hmp); 231 232 /* 233 * HAMMER2 vfs operations. 234 */ 235 /* 236 static struct vfsops hammer2_vfsops = { 237 .vfs_flags = 0, 238 .vfs_init = hammer2_vfs_init, 239 .vfs_uninit = hammer2_vfs_uninit, 240 .vfs_sync = hammer2_vfs_sync, 241 .vfs_mount = hammer2_vfs_mount, 242 .vfs_unmount = hammer2_vfs_unmount, 243 .vfs_root = hammer2_vfs_root, 244 .vfs_statfs = hammer2_vfs_statfs, 245 .vfs_statvfs = hammer2_vfs_statvfs, 246 .vfs_vget = hammer2_vfs_vget, 247 .vfs_vptofh = hammer2_vfs_vptofh, 248 .vfs_fhtovp = hammer2_vfs_fhtovp, 249 .vfs_checkexp = hammer2_vfs_checkexp, 250 .vfs_modifying = hammer2_vfs_modifying 251 }; 252 */ 253 254 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", ""); 255 256 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE); 257 MODULE_VERSION(hammer2, 1); 258 259 int 260 hammer2_vfs_init(void) 261 { 262 /* 263 static struct objcache_malloc_args margs_read; 264 static struct objcache_malloc_args margs_write; 265 static struct objcache_malloc_args margs_vop; 266 */ 267 268 int error; 269 int mod; 270 271 error = 0; 272 kmalloc_raise_limit(M_HAMMER2, 0); /* unlimited */ 273 274 /* 275 * hammer2_xop_nthreads must be a multiple of ncpus, 276 * minimum 2 * ncpus. 277 */ 278 const int ncpus = 1; 279 mod = ncpus; 280 hammer2_xop_mod = mod; 281 hammer2_xop_nthreads = mod * 2; 282 /* 283 while (hammer2_xop_nthreads / mod < HAMMER2_XOPGROUPS_MIN || 284 hammer2_xop_nthreads < HAMMER2_XOPTHREADS_MIN) 285 { 286 hammer2_xop_nthreads += mod; 287 } 288 hammer2_xop_sgroups = hammer2_xop_nthreads / mod / 2; 289 hammer2_xop_xgroups = hammer2_xop_nthreads / mod - hammer2_xop_sgroups; 290 hammer2_xop_xbase = hammer2_xop_sgroups * mod; 291 */ 292 293 /* 294 * A large DIO cache is needed to retain dedup enablement masks. 295 * The bulkfree code clears related masks as part of the disk block 296 * recycling algorithm, preventing it from being used for a later 297 * dedup. 298 * 299 * NOTE: A large buffer cache can actually interfere with dedup 300 * operation because we dedup based on media physical buffers 301 * and not logical buffers. Try to make the DIO case large 302 * enough to avoid this problem, but also cap it. 303 */ 304 const long nbuf = 100000; /* XXX */ 305 hammer2_dio_limit = nbuf * 2; 306 if (hammer2_dio_limit > 100000) 307 hammer2_dio_limit = 100000; 308 309 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref)) 310 error = EINVAL; 311 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data)) 312 error = EINVAL; 313 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data)) 314 error = EINVAL; 315 316 if (error) { 317 kprintf("HAMMER2 structure size mismatch; cannot continue.\n"); 318 return (error); 319 } 320 321 #if 0 322 margs_read.objsize = 65536; 323 margs_read.mtype = M_HAMMER2_DEBUFFER; 324 325 margs_write.objsize = 32768; 326 margs_write.mtype = M_HAMMER2_CBUFFER; 327 328 margs_vop.objsize = sizeof(hammer2_xop_t); 329 margs_vop.mtype = M_HAMMER2; 330 331 /* 332 * Note thaht for the XOPS cache we want backing store allocations 333 * to use M_ZERO. This is not allowed in objcache_get() (to avoid 334 * confusion), so use the backing store function that does it. This 335 * means that initial XOPS objects are zerod but REUSED objects are 336 * not. So we are responsible for cleaning the object up sufficiently 337 * for our needs before objcache_put()ing it back (typically just the 338 * FIFO indices). 339 */ 340 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc, 341 0, 1, NULL, NULL, NULL, 342 objcache_malloc_alloc, 343 objcache_malloc_free, 344 &margs_read); 345 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc, 346 0, 1, NULL, NULL, NULL, 347 objcache_malloc_alloc, 348 objcache_malloc_free, 349 &margs_write); 350 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc, 351 0, 1, NULL, NULL, NULL, 352 objcache_malloc_alloc_zero, 353 objcache_malloc_free, 354 &margs_vop); 355 #endif 356 357 358 lockinit(&hammer2_mntlk, "mntlk", 0, 0); 359 TAILQ_INIT(&hammer2_mntlist); 360 TAILQ_INIT(&hammer2_pfslist); 361 TAILQ_INIT(&hammer2_spmplist); 362 363 const int maxvnodes = 100000; /* XXX */ 364 hammer2_limit_dirty_chains = maxvnodes / 10; 365 if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS) 366 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS; 367 if (hammer2_limit_dirty_chains < 1000) 368 hammer2_limit_dirty_chains = 1000; 369 370 hammer2_limit_dirty_inodes = maxvnodes / 25; 371 if (hammer2_limit_dirty_inodes < 100) 372 hammer2_limit_dirty_inodes = 100; 373 if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES) 374 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES; 375 376 hammer2_limit_saved_chains = hammer2_limit_dirty_chains * 5; 377 378 return (error); 379 } 380 381 int 382 hammer2_vfs_uninit(void) 383 { 384 /* 385 objcache_destroy(cache_buffer_read); 386 objcache_destroy(cache_buffer_write); 387 objcache_destroy(cache_xops); 388 */ 389 return 0; 390 } 391 392 /* 393 * Core PFS allocator. Used to allocate or reference the pmp structure 394 * for PFS cluster mounts and the spmp structure for media (hmp) structures. 395 * The pmp can be passed in or loaded by this function using the chain and 396 * inode data. 397 * 398 * pmp->modify_tid tracks new modify_tid transaction ids for front-end 399 * transactions. Note that synchronization does not use this field. 400 * (typically frontend operations and synchronization cannot run on the 401 * same PFS node at the same time). 402 * 403 * XXX check locking 404 */ 405 hammer2_pfs_t * 406 hammer2_pfsalloc(hammer2_chain_t *chain, 407 const hammer2_inode_data_t *ripdata, 408 hammer2_dev_t *force_local) 409 { 410 hammer2_pfs_t *pmp; 411 hammer2_inode_t *iroot; 412 int count; 413 int i; 414 int j; 415 416 pmp = NULL; 417 418 /* 419 * Locate or create the PFS based on the cluster id. If ripdata 420 * is NULL this is a spmp which is unique and is always allocated. 421 * 422 * If the device is mounted in local mode all PFSs are considered 423 * independent and not part of any cluster (for debugging only). 424 */ 425 if (ripdata) { 426 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 427 if (force_local != pmp->force_local) 428 continue; 429 if (force_local == NULL && 430 bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid, 431 sizeof(pmp->pfs_clid)) == 0) { 432 break; 433 } else if (force_local && pmp->pfs_names[0] && 434 strcmp(pmp->pfs_names[0], (const char *)ripdata->filename) == 0) { 435 break; 436 } 437 } 438 } 439 440 if (pmp == NULL) { 441 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO); 442 pmp->force_local = force_local; 443 hammer2_trans_manage_init(pmp); 444 kmalloc_create_obj(&pmp->minode, "HAMMER2-inodes", 445 sizeof(struct hammer2_inode)); 446 lockinit(&pmp->lock, "pfslk", 0, 0); 447 hammer2_spin_init(&pmp->inum_spin, "hm2pfsalloc_inum"); 448 hammer2_spin_init(&pmp->xop_spin, "h2xop"); 449 hammer2_spin_init(&pmp->lru_spin, "h2lru"); 450 RB_INIT(&pmp->inum_tree); 451 TAILQ_INIT(&pmp->syncq); 452 TAILQ_INIT(&pmp->depq); 453 TAILQ_INIT(&pmp->lru_list); 454 hammer2_spin_init(&pmp->list_spin, "h2pfsalloc_list"); 455 456 /* 457 * Save the last media transaction id for the flusher. Set 458 * initial 459 */ 460 if (ripdata) { 461 pmp->pfs_clid = ripdata->meta.pfs_clid; 462 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry); 463 } else { 464 pmp->flags |= HAMMER2_PMPF_SPMP; 465 TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry); 466 } 467 468 /* 469 * The synchronization thread may start too early, make 470 * sure it stays frozen until we are ready to let it go. 471 * XXX 472 */ 473 /* 474 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN | 475 HAMMER2_THREAD_REMASTER; 476 */ 477 } 478 479 /* 480 * Create the PFS's root inode and any missing XOP helper threads. 481 */ 482 if ((iroot = pmp->iroot) == NULL) { 483 iroot = hammer2_inode_get(pmp, NULL, 1, -1); 484 if (ripdata) 485 iroot->meta = ripdata->meta; 486 pmp->iroot = iroot; 487 hammer2_inode_ref(iroot); 488 hammer2_inode_unlock(iroot); 489 } 490 491 /* 492 * Stop here if no chain is passed in. 493 */ 494 if (chain == NULL) 495 goto done; 496 497 /* 498 * When a chain is passed in we must add it to the PFS's root 499 * inode, update pmp->pfs_types[], and update the syncronization 500 * threads. 501 * 502 * When forcing local mode, mark the PFS as a MASTER regardless. 503 * 504 * At the moment empty spots can develop due to removals or failures. 505 * Ultimately we want to re-fill these spots but doing so might 506 * confused running code. XXX 507 */ 508 hammer2_inode_ref(iroot); 509 hammer2_mtx_ex(&iroot->lock); 510 j = iroot->cluster.nchains; 511 512 if (j == HAMMER2_MAXCLUSTER) { 513 kprintf("hammer2_pfsalloc: cluster full!\n"); 514 /* XXX fatal error? */ 515 } else { 516 KKASSERT(chain->pmp == NULL); 517 chain->pmp = pmp; 518 hammer2_chain_ref(chain); 519 iroot->cluster.array[j].chain = chain; 520 if (force_local) 521 pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER; 522 else 523 pmp->pfs_types[j] = ripdata->meta.pfs_type; 524 pmp->pfs_names[j] = kstrdup((const char *)ripdata->filename, M_HAMMER2); 525 pmp->pfs_hmps[j] = chain->hmp; 526 hammer2_spin_ex(&pmp->inum_spin); 527 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset; 528 hammer2_spin_unex(&pmp->inum_spin); 529 530 /* 531 * If the PFS is already mounted we must account 532 * for the mount_count here. 533 */ 534 if (pmp->mp) 535 ++chain->hmp->mount_count; 536 537 /* 538 * May have to fixup dirty chain tracking. Previous 539 * pmp was NULL so nothing to undo. 540 */ 541 if (chain->flags & HAMMER2_CHAIN_MODIFIED) 542 hammer2_pfs_memory_inc(pmp); 543 ++j; 544 } 545 iroot->cluster.nchains = j; 546 547 /* 548 * Update nmasters from any PFS inode which is part of the cluster. 549 * It is possible that this will result in a value which is too 550 * high. MASTER PFSs are authoritative for pfs_nmasters and will 551 * override this value later on. 552 * 553 * (This informs us of masters that might not currently be 554 * discoverable by this mount). 555 */ 556 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) { 557 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters; 558 } 559 560 /* 561 * Count visible masters. Masters are usually added with 562 * ripdata->meta.pfs_nmasters set to 1. This detects when there 563 * are more (XXX and must update the master inodes). 564 */ 565 count = 0; 566 for (i = 0; i < iroot->cluster.nchains; ++i) { 567 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) 568 ++count; 569 } 570 if (pmp->pfs_nmasters < count) 571 pmp->pfs_nmasters = count; 572 573 /* 574 * Create missing synchronization and support threads. 575 * 576 * Single-node masters (including snapshots) have nothing to 577 * synchronize and do not require this thread. 578 * 579 * Multi-node masters or any number of soft masters, slaves, copy, 580 * or other PFS types need the thread. 581 * 582 * Each thread is responsible for its particular cluster index. 583 * We use independent threads so stalls or mismatches related to 584 * any given target do not affect other targets. 585 */ 586 for (i = 0; i < iroot->cluster.nchains; ++i) { 587 /* 588 * Single-node masters (including snapshots) have nothing 589 * to synchronize and will make direct xops support calls, 590 * thus they do not require this thread. 591 * 592 * Note that there can be thousands of snapshots. We do not 593 * want to create thousands of threads. 594 */ 595 if (pmp->pfs_nmasters <= 1 && 596 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) { 597 continue; 598 } 599 600 /* 601 * Sync support thread 602 */ 603 /* 604 if (pmp->sync_thrs[i].td == NULL) { 605 hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL, 606 "h2nod", i, -1, 607 hammer2_primary_sync_thread); 608 } 609 */ 610 } 611 612 /* 613 * Create missing Xop threads 614 * 615 * NOTE: We create helper threads for all mounted PFSs or any 616 * PFSs with 2+ nodes (so the sync thread can update them, 617 * even if not mounted). 618 */ 619 if (pmp->mp || iroot->cluster.nchains >= 2) 620 hammer2_xop_helper_create(pmp); 621 622 hammer2_mtx_unlock(&iroot->lock); 623 hammer2_inode_drop(iroot); 624 done: 625 return pmp; 626 } 627 628 /* 629 * Deallocate an element of a probed PFS. If destroying and this is a 630 * MASTER, adjust nmasters. 631 * 632 * This function does not physically destroy the PFS element in its device 633 * under the super-root (see hammer2_ioctl_pfs_delete()). 634 */ 635 void 636 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying) 637 { 638 hammer2_inode_t *iroot; 639 hammer2_chain_t *chain; 640 int j; 641 642 /* 643 * Cleanup our reference on iroot. iroot is (should) not be needed 644 * by the flush code. 645 */ 646 iroot = pmp->iroot; 647 if (iroot) { 648 /* 649 * Stop synchronizing 650 * 651 * XXX flush after acquiring the iroot lock. 652 * XXX clean out the cluster index from all inode structures. 653 */ 654 hammer2_thr_delete(&pmp->sync_thrs[clindex]); 655 656 /* 657 * Remove the cluster index from the group. If destroying 658 * the PFS and this is a master, adjust pfs_nmasters. 659 */ 660 hammer2_mtx_ex(&iroot->lock); 661 chain = iroot->cluster.array[clindex].chain; 662 iroot->cluster.array[clindex].chain = NULL; 663 664 switch(pmp->pfs_types[clindex]) { 665 case HAMMER2_PFSTYPE_MASTER: 666 if (destroying && pmp->pfs_nmasters > 0) 667 --pmp->pfs_nmasters; 668 /* XXX adjust ripdata->meta.pfs_nmasters */ 669 break; 670 default: 671 break; 672 } 673 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE; 674 675 hammer2_mtx_unlock(&iroot->lock); 676 677 /* 678 * Release the chain. 679 */ 680 if (chain) { 681 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 682 hammer2_chain_drop(chain); 683 } 684 685 /* 686 * Terminate all XOP threads for the cluster index. 687 */ 688 if (pmp->xop_groups) { 689 for (j = 0; j < hammer2_xop_nthreads; ++j) { 690 hammer2_thr_delete( 691 &pmp->xop_groups[j].thrs[clindex]); 692 } 693 } 694 } 695 } 696 697 /* 698 * Destroy a PFS, typically only occurs after the last mount on a device 699 * has gone away. 700 */ 701 static void 702 hammer2_pfsfree(hammer2_pfs_t *pmp) 703 { 704 hammer2_inode_t *iroot; 705 hammer2_chain_t *chain; 706 int chains_still_present = 0; 707 int i; 708 //int j; 709 710 /* 711 * Cleanup our reference on iroot. iroot is (should) not be needed 712 * by the flush code. 713 */ 714 if (pmp->flags & HAMMER2_PMPF_SPMP) 715 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry); 716 else 717 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry); 718 719 /* 720 * Cleanup chains remaining on LRU list. 721 */ 722 hammer2_spin_ex(&pmp->lru_spin); 723 while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) { 724 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU); 725 atomic_add_int(&pmp->lru_count, -1); 726 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU); 727 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node); 728 hammer2_chain_ref(chain); 729 hammer2_spin_unex(&pmp->lru_spin); 730 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 731 hammer2_chain_drop(chain); 732 hammer2_spin_ex(&pmp->lru_spin); 733 } 734 hammer2_spin_unex(&pmp->lru_spin); 735 736 /* 737 * Clean up iroot 738 */ 739 iroot = pmp->iroot; 740 if (iroot) { 741 for (i = 0; i < iroot->cluster.nchains; ++i) { 742 /* 743 hammer2_thr_delete(&pmp->sync_thrs[i]); 744 if (pmp->xop_groups) { 745 for (j = 0; j < hammer2_xop_nthreads; ++j) 746 hammer2_thr_delete( 747 &pmp->xop_groups[j].thrs[i]); 748 } 749 */ 750 chain = iroot->cluster.array[i].chain; 751 if (chain && !RB_EMPTY(&chain->core.rbtree)) { 752 kprintf("hammer2: Warning pmp %p still " 753 "has active chains\n", pmp); 754 chains_still_present = 1; 755 } 756 } 757 KASSERT(iroot->refs == 1, 758 ("PMP->IROOT %p REFS WRONG %d", iroot, iroot->refs)); 759 760 /* ref for iroot */ 761 hammer2_inode_drop(iroot); 762 pmp->iroot = NULL; 763 } 764 765 /* 766 * Free remaining pmp resources 767 */ 768 if (chains_still_present) { 769 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp); 770 } else { 771 /* 772 * In makefs HAMMER2, all inodes must be gone at this point. 773 * XXX vnode_count may not be 0 at this point. 774 */ 775 assert(hammer2_pfs_inode_count(pmp) == 0); 776 777 kmalloc_destroy_obj(&pmp->minode); 778 kfree(pmp, M_HAMMER2); 779 } 780 } 781 782 /* 783 * Remove all references to hmp from the pfs list. Any PFS which becomes 784 * empty is terminated and freed. 785 * 786 * XXX inefficient. 787 */ 788 static void 789 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which) 790 { 791 hammer2_pfs_t *pmp; 792 hammer2_inode_t *iroot; 793 hammer2_chain_t *rchain; 794 int i; 795 //int j; 796 struct hammer2_pfslist *wlist; 797 798 if (which == 0) 799 wlist = &hammer2_pfslist; 800 else 801 wlist = &hammer2_spmplist; 802 again: 803 TAILQ_FOREACH(pmp, wlist, mntentry) { 804 if ((iroot = pmp->iroot) == NULL) 805 continue; 806 807 /* 808 * Determine if this PFS is affected. If it is we must 809 * freeze all management threads and lock its iroot. 810 * 811 * Freezing a management thread forces it idle, operations 812 * in-progress will be aborted and it will have to start 813 * over again when unfrozen, or exit if told to exit. 814 */ 815 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 816 if (pmp->pfs_hmps[i] == hmp) 817 break; 818 } 819 if (i == HAMMER2_MAXCLUSTER) 820 continue; 821 822 hammer2_vfs_sync_pmp(pmp, MNT_WAIT); 823 824 /* 825 * Make sure all synchronization threads are locked 826 * down. 827 */ 828 /* 829 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 830 if (pmp->pfs_hmps[i] == NULL) 831 continue; 832 hammer2_thr_freeze_async(&pmp->sync_thrs[i]); 833 if (pmp->xop_groups) { 834 for (j = 0; j < hammer2_xop_nthreads; ++j) { 835 hammer2_thr_freeze_async( 836 &pmp->xop_groups[j].thrs[i]); 837 } 838 } 839 } 840 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 841 if (pmp->pfs_hmps[i] == NULL) 842 continue; 843 hammer2_thr_freeze(&pmp->sync_thrs[i]); 844 if (pmp->xop_groups) { 845 for (j = 0; j < hammer2_xop_nthreads; ++j) { 846 hammer2_thr_freeze( 847 &pmp->xop_groups[j].thrs[i]); 848 } 849 } 850 } 851 */ 852 853 /* 854 * Lock the inode and clean out matching chains. 855 * Note that we cannot use hammer2_inode_lock_*() 856 * here because that would attempt to validate the 857 * cluster that we are in the middle of ripping 858 * apart. 859 * 860 * WARNING! We are working directly on the inodes 861 * embedded cluster. 862 */ 863 hammer2_mtx_ex(&iroot->lock); 864 865 /* 866 * Remove the chain from matching elements of the PFS. 867 */ 868 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 869 if (pmp->pfs_hmps[i] != hmp) 870 continue; 871 /* 872 hammer2_thr_delete(&pmp->sync_thrs[i]); 873 if (pmp->xop_groups) { 874 for (j = 0; j < hammer2_xop_nthreads; ++j) { 875 hammer2_thr_delete( 876 &pmp->xop_groups[j].thrs[i]); 877 } 878 } 879 */ 880 rchain = iroot->cluster.array[i].chain; 881 iroot->cluster.array[i].chain = NULL; 882 pmp->pfs_types[i] = HAMMER2_PFSTYPE_NONE; 883 if (pmp->pfs_names[i]) { 884 kfree(pmp->pfs_names[i], M_HAMMER2); 885 pmp->pfs_names[i] = NULL; 886 } 887 if (rchain) { 888 hammer2_chain_drop(rchain); 889 /* focus hint */ 890 if (iroot->cluster.focus == rchain) 891 iroot->cluster.focus = NULL; 892 } 893 pmp->pfs_hmps[i] = NULL; 894 } 895 hammer2_mtx_unlock(&iroot->lock); 896 897 /* 898 * Cleanup trailing chains. Gaps may remain. 899 */ 900 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) { 901 if (pmp->pfs_hmps[i]) 902 break; 903 } 904 iroot->cluster.nchains = i + 1; 905 906 /* 907 * If the PMP has no elements remaining we can destroy it. 908 * (this will transition management threads from frozen->exit). 909 */ 910 if (iroot->cluster.nchains == 0) { 911 /* 912 * If this was the hmp's spmp, we need to clean 913 * a little more stuff out. 914 */ 915 if (hmp->spmp == pmp) { 916 hmp->spmp = NULL; 917 hmp->vchain.pmp = NULL; 918 hmp->fchain.pmp = NULL; 919 } 920 921 /* 922 * Free the pmp and restart the loop 923 */ 924 KKASSERT(TAILQ_EMPTY(&pmp->syncq)); 925 KKASSERT(TAILQ_EMPTY(&pmp->depq)); 926 hammer2_pfsfree(pmp); 927 goto again; 928 } 929 930 /* 931 * If elements still remain we need to set the REMASTER 932 * flag and unfreeze it. 933 */ 934 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 935 if (pmp->pfs_hmps[i] == NULL) 936 continue; 937 /* 938 hammer2_thr_remaster(&pmp->sync_thrs[i]); 939 hammer2_thr_unfreeze(&pmp->sync_thrs[i]); 940 if (pmp->xop_groups) { 941 for (j = 0; j < hammer2_xop_nthreads; ++j) { 942 hammer2_thr_remaster( 943 &pmp->xop_groups[j].thrs[i]); 944 hammer2_thr_unfreeze( 945 &pmp->xop_groups[j].thrs[i]); 946 } 947 } 948 */ 949 } 950 } 951 } 952 953 /* 954 * Mount or remount HAMMER2 fileystem from physical media 955 * 956 * mountroot 957 * mp mount point structure 958 * path NULL 959 * data <unused> 960 * cred <unused> 961 * 962 * mount 963 * mp mount point structure 964 * path path to mount point 965 * data pointer to argument structure in user space 966 * volume volume path (device@LABEL form) 967 * hflags user mount flags 968 * cred user credentials 969 * 970 * RETURNS: 0 Success 971 * !0 error number 972 */ 973 int 974 hammer2_vfs_mount(struct m_vnode *makefs_devvp, struct mount *mp, 975 const char *label, const struct hammer2_mount_info *mi) 976 { 977 struct hammer2_mount_info info = *mi; 978 hammer2_pfs_t *pmp; 979 hammer2_pfs_t *spmp; 980 hammer2_dev_t *hmp, *hmp_tmp; 981 hammer2_dev_t *force_local; 982 hammer2_key_t key_next; 983 hammer2_key_t key_dummy; 984 hammer2_key_t lhc; 985 hammer2_chain_t *parent; 986 hammer2_chain_t *chain; 987 const hammer2_inode_data_t *ripdata; 988 hammer2_devvp_list_t devvpl; 989 hammer2_devvp_t *e, *e_tmp; 990 char *devstr; 991 int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0); 992 int error; 993 int i; 994 995 hmp = NULL; 996 pmp = NULL; 997 devstr = NULL; 998 999 kprintf("hammer2_mount: device=\"%s\" label=\"%s\" rdonly=%d\n", 1000 devstr, label, ronly); 1001 1002 /* 1003 * Initialize all device vnodes. 1004 */ 1005 TAILQ_INIT(&devvpl); 1006 error = hammer2_init_devvp(makefs_devvp, &devvpl); 1007 if (error) { 1008 kprintf("hammer2: failed to initialize devvp in %s\n", devstr); 1009 hammer2_cleanup_devvp(&devvpl); 1010 return error; 1011 } 1012 1013 /* 1014 * Determine if the device has already been mounted. After this 1015 * check hmp will be non-NULL if we are doing the second or more 1016 * hammer2 mounts from the same device. 1017 */ 1018 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1019 if (!TAILQ_EMPTY(&devvpl)) { 1020 /* 1021 * Match the device. Due to the way devfs works, 1022 * we may not be able to directly match the vnode pointer, 1023 * so also check to see if the underlying device matches. 1024 */ 1025 TAILQ_FOREACH(hmp_tmp, &hammer2_mntlist, mntentry) { 1026 TAILQ_FOREACH(e_tmp, &hmp_tmp->devvpl, entry) { 1027 int devvp_found = 0; 1028 TAILQ_FOREACH(e, &devvpl, entry) { 1029 KKASSERT(e->devvp); 1030 if (e_tmp->devvp == e->devvp) 1031 devvp_found = 1; 1032 /* 1033 if (e_tmp->devvp->v_rdev && 1034 e_tmp->devvp->v_rdev == e->devvp->v_rdev) 1035 devvp_found = 1; 1036 */ 1037 } 1038 if (!devvp_found) 1039 goto next_hmp; 1040 } 1041 hmp = hmp_tmp; 1042 kprintf("hammer2_mount: hmp=%p matched\n", hmp); 1043 break; 1044 next_hmp: 1045 continue; 1046 } 1047 1048 /* 1049 * If no match this may be a fresh H2 mount, make sure 1050 * the device is not mounted on anything else. 1051 */ 1052 if (hmp == NULL) { 1053 TAILQ_FOREACH(e, &devvpl, entry) { 1054 struct m_vnode *devvp = e->devvp; 1055 KKASSERT(devvp); 1056 error = vfs_mountedon(devvp); 1057 if (error) { 1058 kprintf("hammer2_mount: %s mounted %d\n", 1059 e->path, error); 1060 hammer2_cleanup_devvp(&devvpl); 1061 lockmgr(&hammer2_mntlk, LK_RELEASE); 1062 return error; 1063 } 1064 } 1065 } 1066 } else { 1067 /* 1068 * Match the label to a pmp already probed. 1069 */ 1070 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) { 1071 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 1072 if (pmp->pfs_names[i] && 1073 strcmp(pmp->pfs_names[i], label) == 0) { 1074 hmp = pmp->pfs_hmps[i]; 1075 break; 1076 } 1077 } 1078 if (hmp) 1079 break; 1080 } 1081 if (hmp == NULL) { 1082 kprintf("hammer2_mount: PFS label \"%s\" not found\n", 1083 label); 1084 hammer2_cleanup_devvp(&devvpl); 1085 lockmgr(&hammer2_mntlk, LK_RELEASE); 1086 return ENOENT; 1087 } 1088 } 1089 1090 /* 1091 * Open the device if this isn't a secondary mount and construct 1092 * the H2 device mount (hmp). 1093 */ 1094 if (hmp == NULL) { 1095 hammer2_chain_t *schain; 1096 hammer2_xop_head_t xop; 1097 1098 /* 1099 * Now open the device 1100 */ 1101 KKASSERT(!TAILQ_EMPTY(&devvpl)); 1102 error = hammer2_open_devvp(&devvpl, ronly); 1103 if (error) { 1104 hammer2_close_devvp(&devvpl, ronly); 1105 hammer2_cleanup_devvp(&devvpl); 1106 lockmgr(&hammer2_mntlk, LK_RELEASE); 1107 return error; 1108 } 1109 1110 /* 1111 * Construct volumes and link with device vnodes. 1112 */ 1113 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO); 1114 hmp->devvp = NULL; 1115 error = hammer2_init_vfsvolumes(mp, &devvpl, hmp->volumes, 1116 &hmp->voldata, &hmp->volhdrno, 1117 &hmp->devvp); 1118 if (error) { 1119 hammer2_close_devvp(&devvpl, ronly); 1120 hammer2_cleanup_devvp(&devvpl); 1121 lockmgr(&hammer2_mntlk, LK_RELEASE); 1122 kfree(hmp, M_HAMMER2); 1123 return error; 1124 } 1125 if (!hmp->devvp) { 1126 kprintf("hammer2: failed to initialize root volume\n"); 1127 hammer2_unmount_helper(mp, NULL, hmp); 1128 lockmgr(&hammer2_mntlk, LK_RELEASE); 1129 hammer2_vfs_unmount(mp, MNT_FORCE); 1130 return EINVAL; 1131 } 1132 1133 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", devstr); 1134 hmp->ronly = ronly; 1135 hmp->hflags = info.hflags & HMNT2_DEVFLAGS; 1136 kmalloc_create_obj(&hmp->mchain, "HAMMER2-chains", 1137 sizeof(struct hammer2_chain)); 1138 kmalloc_create_obj(&hmp->mio, "HAMMER2-dio", 1139 sizeof(struct hammer2_io)); 1140 kmalloc_create(&hmp->mmsg, "HAMMER2-msg"); 1141 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry); 1142 RB_INIT(&hmp->iotree); 1143 hammer2_spin_init(&hmp->io_spin, "h2mount_io"); 1144 hammer2_spin_init(&hmp->list_spin, "h2mount_list"); 1145 1146 lockinit(&hmp->vollk, "h2vol", 0, 0); 1147 lockinit(&hmp->bulklk, "h2bulk", 0, 0); 1148 lockinit(&hmp->bflock, "h2bflk", 0, 0); 1149 1150 /* 1151 * vchain setup. vchain.data is embedded. 1152 * vchain.refs is initialized and will never drop to 0. 1153 */ 1154 hmp->vchain.hmp = hmp; 1155 hmp->vchain.refs = 1; 1156 hmp->vchain.data = (void *)&hmp->voldata; 1157 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME; 1158 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1159 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1160 hammer2_chain_init(&hmp->vchain); 1161 1162 /* 1163 * fchain setup. fchain.data is embedded. 1164 * fchain.refs is initialized and will never drop to 0. 1165 * 1166 * The data is not used but needs to be initialized to 1167 * pass assertion muster. We use this chain primarily 1168 * as a placeholder for the freemap's top-level radix tree 1169 * so it does not interfere with the volume's topology 1170 * radix tree. 1171 */ 1172 hmp->fchain.hmp = hmp; 1173 hmp->fchain.refs = 1; 1174 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset; 1175 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP; 1176 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX; 1177 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1178 hmp->fchain.bref.methods = 1179 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) | 1180 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE); 1181 hammer2_chain_init(&hmp->fchain); 1182 1183 /* 1184 * Initialize volume header related fields. 1185 */ 1186 KKASSERT(hmp->voldata.magic == HAMMER2_VOLUME_ID_HBO || 1187 hmp->voldata.magic == HAMMER2_VOLUME_ID_ABO); 1188 hmp->volsync = hmp->voldata; 1189 hmp->free_reserved = hmp->voldata.allocator_size / 20; 1190 /* 1191 * Must use hmp instead of volume header for these two 1192 * in order to handle volume versions transparently. 1193 */ 1194 if (hmp->voldata.version >= HAMMER2_VOL_VERSION_MULTI_VOLUMES) { 1195 hmp->nvolumes = hmp->voldata.nvolumes; 1196 hmp->total_size = hmp->voldata.total_size; 1197 } else { 1198 hmp->nvolumes = 1; 1199 hmp->total_size = hmp->voldata.volu_size; 1200 } 1201 KKASSERT(hmp->nvolumes > 0); 1202 1203 /* 1204 * Move devvpl entries to hmp. 1205 */ 1206 TAILQ_INIT(&hmp->devvpl); 1207 while ((e = TAILQ_FIRST(&devvpl)) != NULL) { 1208 TAILQ_REMOVE(&devvpl, e, entry); 1209 TAILQ_INSERT_TAIL(&hmp->devvpl, e, entry); 1210 } 1211 KKASSERT(TAILQ_EMPTY(&devvpl)); 1212 KKASSERT(!TAILQ_EMPTY(&hmp->devvpl)); 1213 1214 /* 1215 * Really important to get these right or the flush and 1216 * teardown code will get confused. 1217 */ 1218 hmp->spmp = hammer2_pfsalloc(NULL, NULL, NULL); 1219 spmp = hmp->spmp; 1220 spmp->pfs_hmps[0] = hmp; 1221 1222 /* 1223 * Dummy-up vchain and fchain's modify_tid. mirror_tid 1224 * is inherited from the volume header. 1225 */ 1226 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid; 1227 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid; 1228 hmp->vchain.pmp = spmp; 1229 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid; 1230 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid; 1231 hmp->fchain.pmp = spmp; 1232 1233 /* 1234 * First locate the super-root inode, which is key 0 1235 * relative to the volume header's blockset. 1236 * 1237 * Then locate the root inode by scanning the directory keyspace 1238 * represented by the label. 1239 */ 1240 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 1241 schain = hammer2_chain_lookup(&parent, &key_dummy, 1242 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY, 1243 &error, 0); 1244 hammer2_chain_lookup_done(parent); 1245 if (schain == NULL) { 1246 kprintf("hammer2_mount: invalid super-root\n"); 1247 hammer2_unmount_helper(mp, NULL, hmp); 1248 lockmgr(&hammer2_mntlk, LK_RELEASE); 1249 hammer2_vfs_unmount(mp, MNT_FORCE); 1250 return EINVAL; 1251 } 1252 if (schain->error) { 1253 kprintf("hammer2_mount: error %s reading super-root\n", 1254 hammer2_error_str(schain->error)); 1255 hammer2_chain_unlock(schain); 1256 hammer2_chain_drop(schain); 1257 schain = NULL; 1258 hammer2_unmount_helper(mp, NULL, hmp); 1259 lockmgr(&hammer2_mntlk, LK_RELEASE); 1260 hammer2_vfs_unmount(mp, MNT_FORCE); 1261 return EINVAL; 1262 } 1263 1264 /* 1265 * The super-root always uses an inode_tid of 1 when 1266 * creating PFSs. 1267 */ 1268 spmp->inode_tid = 1; 1269 spmp->modify_tid = schain->bref.modify_tid + 1; 1270 1271 /* 1272 * Sanity-check schain's pmp and finish initialization. 1273 * Any chain belonging to the super-root topology should 1274 * have a NULL pmp (not even set to spmp). 1275 */ 1276 ripdata = &schain->data->ipdata; 1277 KKASSERT(schain->pmp == NULL); 1278 spmp->pfs_clid = ripdata->meta.pfs_clid; 1279 1280 /* 1281 * Replace the dummy spmp->iroot with a real one. It's 1282 * easier to just do a wholesale replacement than to try 1283 * to update the chain and fixup the iroot fields. 1284 * 1285 * The returned inode is locked with the supplied cluster. 1286 */ 1287 hammer2_dummy_xop_from_chain(&xop, schain); 1288 hammer2_inode_drop(spmp->iroot); 1289 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1); 1290 spmp->spmp_hmp = hmp; 1291 spmp->pfs_types[0] = ripdata->meta.pfs_type; 1292 spmp->pfs_hmps[0] = hmp; 1293 hammer2_inode_ref(spmp->iroot); 1294 hammer2_inode_unlock(spmp->iroot); 1295 hammer2_cluster_unlock(&xop.cluster); 1296 hammer2_chain_drop(schain); 1297 /* do not call hammer2_cluster_drop() on an embedded cluster */ 1298 schain = NULL; /* now invalid */ 1299 /* leave spmp->iroot with one ref */ 1300 1301 if (!hmp->ronly) { 1302 error = hammer2_recovery(hmp); 1303 if (error == 0) 1304 error |= hammer2_fixup_pfses(hmp); 1305 /* XXX do something with error */ 1306 } 1307 hammer2_update_pmps(hmp); 1308 hammer2_iocom_init(hmp); 1309 hammer2_bulkfree_init(hmp); 1310 1311 /* 1312 * Ref the cluster management messaging descriptor. The mount 1313 * program deals with the other end of the communications pipe. 1314 * 1315 * Root mounts typically do not supply one. 1316 */ 1317 /* 1318 if (info.cluster_fd >= 0) { 1319 fp = holdfp(curthread, info.cluster_fd, -1); 1320 if (fp) { 1321 hammer2_cluster_reconnect(hmp, fp); 1322 } else { 1323 kprintf("hammer2_mount: bad cluster_fd!\n"); 1324 } 1325 } 1326 */ 1327 } else { 1328 /* hmp->devvp_list is already constructed. */ 1329 hammer2_cleanup_devvp(&devvpl); 1330 spmp = hmp->spmp; 1331 if (info.hflags & HMNT2_DEVFLAGS) { 1332 kprintf("hammer2_mount: Warning: mount flags pertaining " 1333 "to the whole device may only be specified " 1334 "on the first mount of the device: %08x\n", 1335 info.hflags & HMNT2_DEVFLAGS); 1336 } 1337 } 1338 1339 /* 1340 * Force local mount (disassociate all PFSs from their clusters). 1341 * Used primarily for debugging. 1342 */ 1343 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1344 1345 /* 1346 * Lookup the mount point under the media-localized super-root. 1347 * Scanning hammer2_pfslist doesn't help us because it represents 1348 * PFS cluster ids which can aggregate several named PFSs together. 1349 * 1350 * cluster->pmp will incorrectly point to spmp and must be fixed 1351 * up later on. 1352 */ 1353 hammer2_inode_lock(spmp->iroot, 0); 1354 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1355 lhc = hammer2_dirhash(label, strlen(label)); 1356 chain = hammer2_chain_lookup(&parent, &key_next, 1357 lhc, lhc + HAMMER2_DIRHASH_LOMASK, 1358 &error, 0); 1359 while (chain) { 1360 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && 1361 strcmp(label, (char *)chain->data->ipdata.filename) == 0) { 1362 break; 1363 } 1364 chain = hammer2_chain_next(&parent, chain, &key_next, 1365 key_next, 1366 lhc + HAMMER2_DIRHASH_LOMASK, 1367 &error, 0); 1368 } 1369 if (parent) { 1370 hammer2_chain_unlock(parent); 1371 hammer2_chain_drop(parent); 1372 } 1373 hammer2_inode_unlock(spmp->iroot); 1374 1375 /* 1376 * PFS could not be found? 1377 */ 1378 if (chain == NULL) { 1379 hammer2_unmount_helper(mp, NULL, hmp); 1380 lockmgr(&hammer2_mntlk, LK_RELEASE); 1381 hammer2_vfs_unmount(mp, MNT_FORCE); 1382 1383 if (error) { 1384 kprintf("hammer2_mount: PFS label I/O error\n"); 1385 return EINVAL; 1386 } else { 1387 kprintf("hammer2_mount: PFS label \"%s\" not found\n", 1388 label); 1389 return ENOENT; 1390 } 1391 } 1392 1393 /* 1394 * Acquire the pmp structure (it should have already been allocated 1395 * via hammer2_update_pmps()). 1396 */ 1397 if (chain->error) { 1398 kprintf("hammer2_mount: PFS label I/O error\n"); 1399 } else { 1400 ripdata = &chain->data->ipdata; 1401 pmp = hammer2_pfsalloc(NULL, ripdata, force_local); 1402 } 1403 hammer2_chain_unlock(chain); 1404 hammer2_chain_drop(chain); 1405 1406 /* 1407 * PFS to mount must exist at this point. 1408 */ 1409 if (pmp == NULL) { 1410 kprintf("hammer2_mount: Failed to acquire PFS structure\n"); 1411 hammer2_unmount_helper(mp, NULL, hmp); 1412 lockmgr(&hammer2_mntlk, LK_RELEASE); 1413 hammer2_vfs_unmount(mp, MNT_FORCE); 1414 return EINVAL; 1415 } 1416 1417 /* 1418 * Finish the mount 1419 */ 1420 kprintf("hammer2_mount: hmp=%p pmp=%p\n", hmp, pmp); 1421 1422 /* Check if the pmp has already been mounted. */ 1423 if (pmp->mp) { 1424 kprintf("hammer2_mount: PFS already mounted!\n"); 1425 hammer2_unmount_helper(mp, NULL, hmp); 1426 lockmgr(&hammer2_mntlk, LK_RELEASE); 1427 hammer2_vfs_unmount(mp, MNT_FORCE); 1428 return EBUSY; 1429 } 1430 1431 pmp->hflags = info.hflags; 1432 mp->mnt_flag |= MNT_LOCAL; 1433 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */ 1434 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */ 1435 1436 /* 1437 * required mount structure initializations 1438 */ 1439 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE; 1440 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE; 1441 1442 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE; 1443 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE; 1444 1445 /* 1446 * Optional fields 1447 */ 1448 mp->mnt_iosize_max = MAXPHYS; 1449 1450 /* 1451 * Connect up mount pointers. 1452 */ 1453 hammer2_mount_helper(mp, pmp); 1454 lockmgr(&hammer2_mntlk, LK_RELEASE); 1455 1456 #if 0 1457 /* 1458 * Finish setup 1459 */ 1460 vfs_getnewfsid(mp); 1461 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops); 1462 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops); 1463 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops); 1464 1465 if (path) { 1466 copyinstr(info.volume, mp->mnt_stat.f_mntfromname, 1467 MNAMELEN - 1, &size); 1468 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size); 1469 } /* else root mount, already in there */ 1470 1471 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname)); 1472 if (path) { 1473 copyinstr(path, mp->mnt_stat.f_mntonname, 1474 sizeof(mp->mnt_stat.f_mntonname) - 1, 1475 &size); 1476 } else { 1477 /* root mount */ 1478 mp->mnt_stat.f_mntonname[0] = '/'; 1479 } 1480 #endif 1481 1482 /* 1483 * Initial statfs to prime mnt_stat. 1484 */ 1485 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL); 1486 hammer2_vfs_statvfs(mp, &mp->mnt_vstat, NULL); 1487 1488 return 0; 1489 } 1490 1491 /* 1492 * Scan PFSs under the super-root and create hammer2_pfs structures. 1493 */ 1494 static 1495 void 1496 hammer2_update_pmps(hammer2_dev_t *hmp) 1497 { 1498 const hammer2_inode_data_t *ripdata; 1499 hammer2_chain_t *parent; 1500 hammer2_chain_t *chain; 1501 hammer2_dev_t *force_local; 1502 hammer2_pfs_t *spmp; 1503 hammer2_key_t key_next; 1504 int error; 1505 1506 /* 1507 * Force local mount (disassociate all PFSs from their clusters). 1508 * Used primarily for debugging. 1509 */ 1510 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL; 1511 1512 /* 1513 * Lookup mount point under the media-localized super-root. 1514 * 1515 * cluster->pmp will incorrectly point to spmp and must be fixed 1516 * up later on. 1517 */ 1518 spmp = hmp->spmp; 1519 hammer2_inode_lock(spmp->iroot, 0); 1520 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 1521 chain = hammer2_chain_lookup(&parent, &key_next, 1522 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 1523 &error, 0); 1524 while (chain) { 1525 if (chain->error) { 1526 kprintf("I/O error scanning PFS labels\n"); 1527 } else if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) { 1528 kprintf("Non inode chain type %d under super-root\n", 1529 chain->bref.type); 1530 } else { 1531 ripdata = &chain->data->ipdata; 1532 hammer2_pfsalloc(chain, ripdata, force_local); 1533 } 1534 chain = hammer2_chain_next(&parent, chain, &key_next, 1535 key_next, HAMMER2_KEY_MAX, 1536 &error, 0); 1537 } 1538 if (parent) { 1539 hammer2_chain_unlock(parent); 1540 hammer2_chain_drop(parent); 1541 } 1542 hammer2_inode_unlock(spmp->iroot); 1543 } 1544 1545 #if 0 1546 static 1547 int 1548 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused, 1549 struct ucred *cred) 1550 { 1551 hammer2_volume_t *vol; 1552 struct m_vnode *devvp; 1553 int i, error, result = 0; 1554 1555 if (!(hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR))) 1556 return 0; 1557 1558 for (i = 0; i < hmp->nvolumes; ++i) { 1559 vol = &hmp->volumes[i]; 1560 devvp = vol->dev->devvp; 1561 KKASSERT(devvp); 1562 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1563 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL); 1564 vn_unlock(devvp); 1565 error = 0; 1566 if (vol->id == HAMMER2_ROOT_VOLUME) { 1567 error = hammer2_recovery(hmp); 1568 if (error == 0) 1569 error |= hammer2_fixup_pfses(hmp); 1570 } 1571 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1572 if (error == 0) { 1573 VOP_CLOSE(devvp, FREAD, NULL); 1574 } else { 1575 VOP_CLOSE(devvp, FREAD | FWRITE, NULL); 1576 } 1577 vn_unlock(devvp); 1578 result |= error; 1579 } 1580 if (result == 0) { 1581 kprintf("hammer2: enable read/write\n"); 1582 hmp->ronly = 0; 1583 } 1584 1585 return result; 1586 } 1587 #endif 1588 1589 int 1590 hammer2_vfs_unmount(struct mount *mp, int mntflags) 1591 { 1592 hammer2_pfs_t *pmp; 1593 int flags; 1594 int error = 0; 1595 1596 pmp = MPTOPMP(mp); 1597 1598 if (pmp == NULL) 1599 return(0); 1600 1601 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE); 1602 1603 /* 1604 * If mount initialization proceeded far enough we must flush 1605 * its vnodes and sync the underlying mount points. Three syncs 1606 * are required to fully flush the filesystem (freemap updates lag 1607 * by one flush, and one extra for safety). 1608 */ 1609 if (mntflags & MNT_FORCE) 1610 flags = FORCECLOSE; 1611 else 1612 flags = 0; 1613 if (pmp->iroot) { 1614 error = vflush(mp, 0, flags); 1615 if (error) 1616 goto failed; 1617 hammer2_vfs_sync(mp, MNT_WAIT); 1618 hammer2_vfs_sync(mp, MNT_WAIT); 1619 hammer2_vfs_sync(mp, MNT_WAIT); 1620 } 1621 1622 /* 1623 * Cleanup the frontend support XOPS threads 1624 */ 1625 hammer2_xop_helper_cleanup(pmp); 1626 1627 if (pmp->mp) 1628 hammer2_unmount_helper(mp, pmp, NULL); 1629 1630 error = 0; 1631 failed: 1632 lockmgr(&hammer2_mntlk, LK_RELEASE); 1633 1634 return (error); 1635 } 1636 1637 /* 1638 * Mount helper, hook the system mount into our PFS. 1639 * The mount lock is held. 1640 * 1641 * We must bump the mount_count on related devices for any 1642 * mounted PFSs. 1643 */ 1644 static 1645 void 1646 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp) 1647 { 1648 hammer2_cluster_t *cluster; 1649 hammer2_chain_t *rchain; 1650 int i; 1651 1652 mp->mnt_data = (qaddr_t)pmp; 1653 pmp->mp = mp; 1654 1655 /* 1656 * After pmp->mp is set we have to adjust hmp->mount_count. 1657 */ 1658 cluster = &pmp->iroot->cluster; 1659 for (i = 0; i < cluster->nchains; ++i) { 1660 rchain = cluster->array[i].chain; 1661 if (rchain == NULL) 1662 continue; 1663 ++rchain->hmp->mount_count; 1664 } 1665 1666 /* 1667 * Create missing Xop threads 1668 */ 1669 hammer2_xop_helper_create(pmp); 1670 } 1671 1672 /* 1673 * Unmount helper, unhook the system mount from our PFS. 1674 * The mount lock is held. 1675 * 1676 * If hmp is supplied a mount responsible for being the first to open 1677 * the block device failed and the block device and all PFSs using the 1678 * block device must be cleaned up. 1679 * 1680 * If pmp is supplied multiple devices might be backing the PFS and each 1681 * must be disconnected. This might not be the last PFS using some of the 1682 * underlying devices. Also, we have to adjust our hmp->mount_count 1683 * accounting for the devices backing the pmp which is now undergoing an 1684 * unmount. 1685 */ 1686 static 1687 void 1688 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp) 1689 { 1690 hammer2_cluster_t *cluster; 1691 hammer2_chain_t *rchain; 1692 int dumpcnt; 1693 int i; 1694 1695 /* 1696 * If no device supplied this is a high-level unmount and we have to 1697 * to disconnect the mount, adjust mount_count, and locate devices 1698 * that might now have no mounts. 1699 */ 1700 if (pmp) { 1701 KKASSERT(hmp == NULL); 1702 KKASSERT(MPTOPMP(mp) == pmp); 1703 pmp->mp = NULL; 1704 mp->mnt_data = NULL; 1705 1706 /* 1707 * After pmp->mp is cleared we have to account for 1708 * mount_count. 1709 */ 1710 cluster = &pmp->iroot->cluster; 1711 for (i = 0; i < cluster->nchains; ++i) { 1712 rchain = cluster->array[i].chain; 1713 if (rchain == NULL) 1714 continue; 1715 --rchain->hmp->mount_count; 1716 /* scrapping hmp now may invalidate the pmp */ 1717 } 1718 again: 1719 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) { 1720 if (hmp->mount_count == 0) { 1721 hammer2_unmount_helper(NULL, NULL, hmp); 1722 goto again; 1723 } 1724 } 1725 return; 1726 } 1727 1728 /* 1729 * Try to terminate the block device. We can't terminate it if 1730 * there are still PFSs referencing it. 1731 */ 1732 if (hmp->mount_count) 1733 return; 1734 1735 /* 1736 * Decomission the network before we start messing with the 1737 * device and PFS. 1738 */ 1739 hammer2_iocom_uninit(hmp); 1740 1741 hammer2_bulkfree_uninit(hmp); 1742 hammer2_pfsfree_scan(hmp, 0); 1743 1744 /* 1745 * Cycle the volume data lock as a safety (probably not needed any 1746 * more). To ensure everything is out we need to flush at least 1747 * three times. (1) The running of the sideq can dirty the 1748 * filesystem, (2) A normal flush can dirty the freemap, and 1749 * (3) ensure that the freemap is fully synchronized. 1750 * 1751 * The next mount's recovery scan can clean everything up but we want 1752 * to leave the filesystem in a 100% clean state on a normal unmount. 1753 */ 1754 #if 0 1755 hammer2_voldata_lock(hmp); 1756 hammer2_voldata_unlock(hmp); 1757 #endif 1758 1759 /* 1760 * Flush whatever is left. Unmounted but modified PFS's might still 1761 * have some dirty chains on them. 1762 */ 1763 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS); 1764 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS); 1765 1766 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) { 1767 hammer2_voldata_modify(hmp); 1768 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP | 1769 HAMMER2_FLUSH_ALL); 1770 } 1771 hammer2_chain_unlock(&hmp->fchain); 1772 1773 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) { 1774 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP | 1775 HAMMER2_FLUSH_ALL); 1776 } 1777 hammer2_chain_unlock(&hmp->vchain); 1778 1779 if ((hmp->vchain.flags | hmp->fchain.flags) & 1780 HAMMER2_CHAIN_FLUSH_MASK) { 1781 kprintf("hammer2_unmount: chains left over after final sync\n"); 1782 kprintf(" vchain %08x\n", hmp->vchain.flags); 1783 kprintf(" fchain %08x\n", hmp->fchain.flags); 1784 1785 if (hammer2_debug & 0x0010) 1786 Debugger("entered debugger"); 1787 } 1788 1789 hammer2_pfsfree_scan(hmp, 1); 1790 1791 KKASSERT(hmp->spmp == NULL); 1792 1793 /* 1794 * Finish up with the device vnode 1795 */ 1796 if (!TAILQ_EMPTY(&hmp->devvpl)) { 1797 hammer2_close_devvp(&hmp->devvpl, hmp->ronly); 1798 hammer2_cleanup_devvp(&hmp->devvpl); 1799 } 1800 KKASSERT(TAILQ_EMPTY(&hmp->devvpl)); 1801 1802 /* 1803 * Clear vchain/fchain flags that might prevent final cleanup 1804 * of these chains. 1805 */ 1806 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) { 1807 atomic_add_long(&hammer2_count_modified_chains, -1); 1808 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 1809 hammer2_pfs_memory_wakeup(hmp->vchain.pmp, -1); 1810 } 1811 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) { 1812 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE); 1813 } 1814 1815 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) { 1816 atomic_add_long(&hammer2_count_modified_chains, -1); 1817 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED); 1818 hammer2_pfs_memory_wakeup(hmp->fchain.pmp, -1); 1819 } 1820 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) { 1821 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE); 1822 } 1823 1824 dumpcnt = 50; 1825 hammer2_dump_chain(&hmp->vchain, 0, 0, &dumpcnt, 'v', (u_int)-1); 1826 dumpcnt = 50; 1827 hammer2_dump_chain(&hmp->fchain, 0, 0, &dumpcnt, 'f', (u_int)-1); 1828 1829 /* 1830 * Final drop of embedded freemap root chain to 1831 * clean up fchain.core (fchain structure is not 1832 * flagged ALLOCATED so it is cleaned out and then 1833 * left to rot). 1834 */ 1835 hammer2_chain_drop(&hmp->fchain); 1836 1837 /* 1838 * Final drop of embedded volume root chain to clean 1839 * up vchain.core (vchain structure is not flagged 1840 * ALLOCATED so it is cleaned out and then left to 1841 * rot). 1842 */ 1843 hammer2_chain_drop(&hmp->vchain); 1844 1845 hammer2_io_cleanup(hmp, &hmp->iotree); 1846 if (hmp->iofree_count) { 1847 kprintf("io_cleanup: %d I/O's left hanging\n", 1848 hmp->iofree_count); 1849 } 1850 1851 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry); 1852 kmalloc_destroy_obj(&hmp->mchain); 1853 kmalloc_destroy_obj(&hmp->mio); 1854 kmalloc_destroy(&hmp->mmsg); 1855 kfree(hmp, M_HAMMER2); 1856 } 1857 1858 int 1859 hammer2_vfs_vget(struct mount *mp, struct m_vnode *dvp, 1860 ino_t ino, struct m_vnode **vpp) 1861 { 1862 hammer2_xop_lookup_t *xop; 1863 hammer2_pfs_t *pmp; 1864 hammer2_inode_t *ip; 1865 hammer2_tid_t inum; 1866 int error; 1867 1868 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK; 1869 1870 error = 0; 1871 pmp = MPTOPMP(mp); 1872 1873 /* 1874 * Easy if we already have it cached 1875 */ 1876 ip = hammer2_inode_lookup(pmp, inum); 1877 if (ip) { 1878 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED); 1879 *vpp = hammer2_igetv(ip, &error); 1880 hammer2_inode_unlock(ip); 1881 hammer2_inode_drop(ip); /* from lookup */ 1882 1883 return error; 1884 } 1885 1886 /* 1887 * Otherwise we have to find the inode 1888 */ 1889 xop = hammer2_xop_alloc(pmp->iroot, 0); 1890 xop->lhc = inum; 1891 hammer2_xop_start(&xop->head, &hammer2_lookup_desc); 1892 error = hammer2_xop_collect(&xop->head, 0); 1893 1894 if (error == 0) 1895 ip = hammer2_inode_get(pmp, &xop->head, -1, -1); 1896 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1897 1898 if (ip) { 1899 *vpp = hammer2_igetv(ip, &error); 1900 hammer2_inode_unlock(ip); 1901 } else { 1902 *vpp = NULL; 1903 error = ENOENT; 1904 } 1905 return (error); 1906 } 1907 1908 int 1909 hammer2_vfs_root(struct mount *mp, struct m_vnode **vpp) 1910 { 1911 hammer2_pfs_t *pmp; 1912 struct m_vnode *vp; 1913 int error; 1914 1915 pmp = MPTOPMP(mp); 1916 if (pmp->iroot == NULL) { 1917 kprintf("hammer2 (%s): no root inode\n", 1918 mp->mnt_stat.f_mntfromname); 1919 *vpp = NULL; 1920 return EINVAL; 1921 } 1922 1923 error = 0; 1924 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1925 1926 while (pmp->inode_tid == 0) { 1927 hammer2_xop_ipcluster_t *xop; 1928 const hammer2_inode_meta_t *meta; 1929 1930 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING); 1931 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc); 1932 error = hammer2_xop_collect(&xop->head, 0); 1933 1934 if (error == 0) { 1935 meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta; 1936 pmp->iroot->meta = *meta; 1937 pmp->inode_tid = meta->pfs_inum + 1; 1938 hammer2_xop_pdata(&xop->head); 1939 /* meta invalid */ 1940 1941 if (pmp->inode_tid < HAMMER2_INODE_START) 1942 pmp->inode_tid = HAMMER2_INODE_START; 1943 pmp->modify_tid = 1944 xop->head.cluster.focus->bref.modify_tid + 1; 1945 #if 0 1946 kprintf("PFS: Starting inode %jd\n", 1947 (intmax_t)pmp->inode_tid); 1948 kprintf("PMP focus good set nextino=%ld mod=%016jx\n", 1949 pmp->inode_tid, pmp->modify_tid); 1950 #endif 1951 //wakeup(&pmp->iroot); XXX 1952 1953 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1954 1955 /* 1956 * Prime the mount info. 1957 */ 1958 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL); 1959 break; 1960 } 1961 1962 /* 1963 * Loop, try again 1964 */ 1965 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1966 hammer2_inode_unlock(pmp->iroot); 1967 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz); 1968 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED); 1969 if (error == EINTR) 1970 break; 1971 } 1972 1973 if (error) { 1974 hammer2_inode_unlock(pmp->iroot); 1975 *vpp = NULL; 1976 } else { 1977 vp = hammer2_igetv(pmp->iroot, &error); 1978 hammer2_inode_unlock(pmp->iroot); 1979 *vpp = vp; 1980 } 1981 1982 return (error); 1983 } 1984 1985 /* 1986 * Filesystem status 1987 * 1988 * XXX incorporate ipdata->meta.inode_quota and data_quota 1989 */ 1990 static 1991 int 1992 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred) 1993 { 1994 hammer2_pfs_t *pmp; 1995 hammer2_dev_t *hmp; 1996 hammer2_blockref_t bref; 1997 struct statfs tmp; 1998 int i; 1999 2000 /* 2001 * NOTE: iroot might not have validated the cluster yet. 2002 */ 2003 pmp = MPTOPMP(mp); 2004 2005 bzero(&tmp, sizeof(tmp)); 2006 2007 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 2008 hmp = pmp->pfs_hmps[i]; 2009 if (hmp == NULL) 2010 continue; 2011 if (pmp->iroot->cluster.array[i].chain) 2012 bref = pmp->iroot->cluster.array[i].chain->bref; 2013 else 2014 bzero(&bref, sizeof(bref)); 2015 2016 tmp.f_files = bref.embed.stats.inode_count; 2017 tmp.f_ffree = 0; 2018 tmp.f_blocks = hmp->voldata.allocator_size / 2019 mp->mnt_vstat.f_bsize; 2020 tmp.f_bfree = hmp->voldata.allocator_free / 2021 mp->mnt_vstat.f_bsize; 2022 tmp.f_bavail = tmp.f_bfree; 2023 2024 if (cred && cred->cr_uid != 0) { 2025 uint64_t adj; 2026 2027 /* 5% */ 2028 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize; 2029 tmp.f_blocks -= adj; 2030 tmp.f_bfree -= adj; 2031 tmp.f_bavail -= adj; 2032 } 2033 2034 mp->mnt_stat.f_blocks = tmp.f_blocks; 2035 mp->mnt_stat.f_bfree = tmp.f_bfree; 2036 mp->mnt_stat.f_bavail = tmp.f_bavail; 2037 mp->mnt_stat.f_files = tmp.f_files; 2038 mp->mnt_stat.f_ffree = tmp.f_ffree; 2039 2040 *sbp = mp->mnt_stat; 2041 } 2042 return (0); 2043 } 2044 2045 static 2046 int 2047 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred) 2048 { 2049 hammer2_pfs_t *pmp; 2050 hammer2_dev_t *hmp; 2051 hammer2_blockref_t bref; 2052 struct statvfs tmp; 2053 int i; 2054 2055 /* 2056 * NOTE: iroot might not have validated the cluster yet. 2057 */ 2058 pmp = MPTOPMP(mp); 2059 bzero(&tmp, sizeof(tmp)); 2060 2061 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 2062 hmp = pmp->pfs_hmps[i]; 2063 if (hmp == NULL) 2064 continue; 2065 if (pmp->iroot->cluster.array[i].chain) 2066 bref = pmp->iroot->cluster.array[i].chain->bref; 2067 else 2068 bzero(&bref, sizeof(bref)); 2069 2070 tmp.f_files = bref.embed.stats.inode_count; 2071 tmp.f_ffree = 0; 2072 tmp.f_blocks = hmp->voldata.allocator_size / 2073 mp->mnt_vstat.f_bsize; 2074 tmp.f_bfree = hmp->voldata.allocator_free / 2075 mp->mnt_vstat.f_bsize; 2076 tmp.f_bavail = tmp.f_bfree; 2077 2078 if (cred && cred->cr_uid != 0) { 2079 uint64_t adj; 2080 2081 /* 5% */ 2082 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize; 2083 tmp.f_blocks -= adj; 2084 tmp.f_bfree -= adj; 2085 tmp.f_bavail -= adj; 2086 } 2087 2088 mp->mnt_vstat.f_blocks = tmp.f_blocks; 2089 mp->mnt_vstat.f_bfree = tmp.f_bfree; 2090 mp->mnt_vstat.f_bavail = tmp.f_bavail; 2091 mp->mnt_vstat.f_files = tmp.f_files; 2092 mp->mnt_vstat.f_ffree = tmp.f_ffree; 2093 2094 *sbp = mp->mnt_vstat; 2095 } 2096 return (0); 2097 } 2098 2099 /* 2100 * Mount-time recovery (RW mounts) 2101 * 2102 * Updates to the free block table are allowed to lag flushes by one 2103 * transaction. In case of a crash, then on a fresh mount we must do an 2104 * incremental scan of the last committed transaction id and make sure that 2105 * all related blocks have been marked allocated. 2106 */ 2107 struct hammer2_recovery_elm { 2108 TAILQ_ENTRY(hammer2_recovery_elm) entry; 2109 hammer2_chain_t *chain; 2110 hammer2_tid_t sync_tid; 2111 }; 2112 2113 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm); 2114 2115 struct hammer2_recovery_info { 2116 struct hammer2_recovery_list list; 2117 hammer2_tid_t mtid; 2118 int depth; 2119 }; 2120 2121 static int hammer2_recovery_scan(hammer2_dev_t *hmp, 2122 hammer2_chain_t *parent, 2123 struct hammer2_recovery_info *info, 2124 hammer2_tid_t sync_tid); 2125 2126 #define HAMMER2_RECOVERY_MAXDEPTH 10 2127 2128 static 2129 int 2130 hammer2_recovery(hammer2_dev_t *hmp) 2131 { 2132 struct hammer2_recovery_info info; 2133 struct hammer2_recovery_elm *elm; 2134 hammer2_chain_t *parent; 2135 hammer2_tid_t sync_tid; 2136 hammer2_tid_t mirror_tid; 2137 int error; 2138 2139 hammer2_trans_init(hmp->spmp, 0); 2140 2141 sync_tid = hmp->voldata.freemap_tid; 2142 mirror_tid = hmp->voldata.mirror_tid; 2143 2144 kprintf("hammer2_mount: \"%s\": ", hmp->devrepname); 2145 if (sync_tid >= mirror_tid) { 2146 kprintf("no recovery needed\n"); 2147 } else { 2148 kprintf("freemap recovery %016jx-%016jx\n", 2149 sync_tid + 1, mirror_tid); 2150 } 2151 2152 TAILQ_INIT(&info.list); 2153 info.depth = 0; 2154 parent = hammer2_chain_lookup_init(&hmp->vchain, 0); 2155 error = hammer2_recovery_scan(hmp, parent, &info, sync_tid); 2156 hammer2_chain_lookup_done(parent); 2157 2158 while ((elm = TAILQ_FIRST(&info.list)) != NULL) { 2159 TAILQ_REMOVE(&info.list, elm, entry); 2160 parent = elm->chain; 2161 sync_tid = elm->sync_tid; 2162 kfree(elm, M_HAMMER2); 2163 2164 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2165 error |= hammer2_recovery_scan(hmp, parent, &info, 2166 hmp->voldata.freemap_tid); 2167 hammer2_chain_unlock(parent); 2168 hammer2_chain_drop(parent); /* drop elm->chain ref */ 2169 } 2170 2171 hammer2_trans_done(hmp->spmp, 0); 2172 2173 return error; 2174 } 2175 2176 static 2177 int 2178 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent, 2179 struct hammer2_recovery_info *info, 2180 hammer2_tid_t sync_tid) 2181 { 2182 const hammer2_inode_data_t *ripdata; 2183 hammer2_chain_t *chain; 2184 hammer2_blockref_t bref; 2185 int tmp_error; 2186 int rup_error; 2187 int error; 2188 int first; 2189 2190 /* 2191 * Adjust freemap to ensure that the block(s) are marked allocated. 2192 */ 2193 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) { 2194 hammer2_freemap_adjust(hmp, &parent->bref, 2195 HAMMER2_FREEMAP_DORECOVER); 2196 } 2197 2198 /* 2199 * Check type for recursive scan 2200 */ 2201 switch(parent->bref.type) { 2202 case HAMMER2_BREF_TYPE_VOLUME: 2203 /* data already instantiated */ 2204 break; 2205 case HAMMER2_BREF_TYPE_INODE: 2206 /* 2207 * Must instantiate data for DIRECTDATA test and also 2208 * for recursion. 2209 */ 2210 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2211 ripdata = &parent->data->ipdata; 2212 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) { 2213 /* not applicable to recovery scan */ 2214 hammer2_chain_unlock(parent); 2215 return 0; 2216 } 2217 hammer2_chain_unlock(parent); 2218 break; 2219 case HAMMER2_BREF_TYPE_INDIRECT: 2220 /* 2221 * Must instantiate data for recursion 2222 */ 2223 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2224 hammer2_chain_unlock(parent); 2225 break; 2226 case HAMMER2_BREF_TYPE_DIRENT: 2227 case HAMMER2_BREF_TYPE_DATA: 2228 case HAMMER2_BREF_TYPE_FREEMAP: 2229 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2230 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 2231 /* not applicable to recovery scan */ 2232 return 0; 2233 break; 2234 default: 2235 return HAMMER2_ERROR_BADBREF; 2236 } 2237 2238 /* 2239 * Defer operation if depth limit reached. 2240 */ 2241 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) { 2242 struct hammer2_recovery_elm *elm; 2243 2244 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK); 2245 elm->chain = parent; 2246 elm->sync_tid = sync_tid; 2247 hammer2_chain_ref(parent); 2248 TAILQ_INSERT_TAIL(&info->list, elm, entry); 2249 /* unlocked by caller */ 2250 2251 return(0); 2252 } 2253 2254 2255 /* 2256 * Recursive scan of the last flushed transaction only. We are 2257 * doing this without pmp assignments so don't leave the chains 2258 * hanging around after we are done with them. 2259 * 2260 * error Cumulative error this level only 2261 * rup_error Cumulative error for recursion 2262 * tmp_error Specific non-cumulative recursion error 2263 */ 2264 chain = NULL; 2265 first = 1; 2266 rup_error = 0; 2267 error = 0; 2268 2269 for (;;) { 2270 error |= hammer2_chain_scan(parent, &chain, &bref, 2271 &first, 2272 HAMMER2_LOOKUP_NODATA); 2273 2274 /* 2275 * Problem during scan or EOF 2276 */ 2277 if (error) 2278 break; 2279 2280 /* 2281 * If this is a leaf 2282 */ 2283 if (chain == NULL) { 2284 if (bref.mirror_tid > sync_tid) { 2285 hammer2_freemap_adjust(hmp, &bref, 2286 HAMMER2_FREEMAP_DORECOVER); 2287 } 2288 continue; 2289 } 2290 2291 /* 2292 * This may or may not be a recursive node. 2293 */ 2294 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 2295 if (bref.mirror_tid > sync_tid) { 2296 ++info->depth; 2297 tmp_error = hammer2_recovery_scan(hmp, chain, 2298 info, sync_tid); 2299 --info->depth; 2300 } else { 2301 tmp_error = 0; 2302 } 2303 2304 /* 2305 * Flush the recovery at the PFS boundary to stage it for 2306 * the final flush of the super-root topology. 2307 */ 2308 if (tmp_error == 0 && 2309 (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) && 2310 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) { 2311 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 2312 HAMMER2_FLUSH_ALL); 2313 } 2314 rup_error |= tmp_error; 2315 } 2316 return ((error | rup_error) & ~HAMMER2_ERROR_EOF); 2317 } 2318 2319 /* 2320 * This fixes up an error introduced in earlier H2 implementations where 2321 * moving a PFS inode into an indirect block wound up causing the 2322 * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared. 2323 */ 2324 static 2325 int 2326 hammer2_fixup_pfses(hammer2_dev_t *hmp) 2327 { 2328 const hammer2_inode_data_t *ripdata; 2329 hammer2_chain_t *parent; 2330 hammer2_chain_t *chain; 2331 hammer2_key_t key_next; 2332 hammer2_pfs_t *spmp; 2333 int error; 2334 2335 error = 0; 2336 2337 /* 2338 * Lookup mount point under the media-localized super-root. 2339 * 2340 * cluster->pmp will incorrectly point to spmp and must be fixed 2341 * up later on. 2342 */ 2343 spmp = hmp->spmp; 2344 hammer2_inode_lock(spmp->iroot, 0); 2345 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS); 2346 chain = hammer2_chain_lookup(&parent, &key_next, 2347 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX, 2348 &error, 0); 2349 while (chain) { 2350 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) 2351 continue; 2352 if (chain->error) { 2353 kprintf("I/O error scanning PFS labels\n"); 2354 error |= chain->error; 2355 } else if ((chain->bref.flags & 2356 HAMMER2_BREF_FLAG_PFSROOT) == 0) { 2357 int error2; 2358 2359 ripdata = &chain->data->ipdata; 2360 hammer2_trans_init(hmp->spmp, 0); 2361 error2 = hammer2_chain_modify(chain, 2362 chain->bref.modify_tid, 2363 0, 0); 2364 if (error2 == 0) { 2365 kprintf("hammer2: Correct mis-flagged PFS %s\n", 2366 ripdata->filename); 2367 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT; 2368 } else { 2369 error |= error2; 2370 } 2371 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 2372 HAMMER2_FLUSH_ALL); 2373 hammer2_trans_done(hmp->spmp, 0); 2374 } 2375 chain = hammer2_chain_next(&parent, chain, &key_next, 2376 key_next, HAMMER2_KEY_MAX, 2377 &error, 0); 2378 } 2379 if (parent) { 2380 hammer2_chain_unlock(parent); 2381 hammer2_chain_drop(parent); 2382 } 2383 hammer2_inode_unlock(spmp->iroot); 2384 2385 return error; 2386 } 2387 2388 /* 2389 * Sync a mount point; this is called periodically on a per-mount basis from 2390 * the filesystem syncer, and whenever a user issues a sync. 2391 */ 2392 int 2393 hammer2_vfs_sync(struct mount *mp, int waitfor) 2394 { 2395 int error; 2396 2397 error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor); 2398 2399 return error; 2400 } 2401 2402 /* 2403 * Because frontend operations lock vnodes before we get a chance to 2404 * lock the related inode, we can't just acquire a vnode lock without 2405 * risking a deadlock. The frontend may be holding a vnode lock while 2406 * also blocked on our SYNCQ flag while trying to get the inode lock. 2407 * 2408 * To deal with this situation we can check the vnode lock situation 2409 * after locking the inode and perform a work-around. 2410 */ 2411 int 2412 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor) 2413 { 2414 hammer2_inode_t *ip; 2415 hammer2_depend_t *depend; 2416 hammer2_depend_t *depend_next; 2417 struct m_vnode *vp; 2418 uint32_t pass2; 2419 int error; 2420 int wakecount; 2421 int dorestart; 2422 2423 /* 2424 * Move all inodes on sideq to syncq. This will clear sideq. 2425 * This should represent all flushable inodes. These inodes 2426 * will already have refs due to being on syncq or sideq. We 2427 * must do this all at once with the spinlock held to ensure that 2428 * all inode dependencies are part of the same flush. 2429 * 2430 * We should be able to do this asynchronously from frontend 2431 * operations because we will be locking the inodes later on 2432 * to actually flush them, and that will partition any frontend 2433 * op using the same inode. Either it has already locked the 2434 * inode and we will block, or it has not yet locked the inode 2435 * and it will block until we are finished flushing that inode. 2436 * 2437 * When restarting, only move the inodes flagged as PASS2 from 2438 * SIDEQ to SYNCQ. PASS2 propagation by inode_lock4() and 2439 * inode_depend() are atomic with the spin-lock. 2440 */ 2441 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH); 2442 #ifdef HAMMER2_DEBUG_SYNC 2443 kprintf("FILESYSTEM SYNC BOUNDARY\n"); 2444 #endif 2445 dorestart = 0; 2446 2447 /* 2448 * Move inodes from depq to syncq, releasing the related 2449 * depend structures. 2450 */ 2451 restart: 2452 #ifdef HAMMER2_DEBUG_SYNC 2453 kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart); 2454 #endif 2455 hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/); 2456 hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN); 2457 2458 /* 2459 * Move inodes from depq to syncq. When restarting, only depq's 2460 * marked pass2 are moved. 2461 */ 2462 hammer2_spin_ex(&pmp->list_spin); 2463 depend_next = TAILQ_FIRST(&pmp->depq); 2464 wakecount = 0; 2465 2466 while ((depend = depend_next) != NULL) { 2467 depend_next = TAILQ_NEXT(depend, entry); 2468 if (dorestart && depend->pass2 == 0) 2469 continue; 2470 TAILQ_FOREACH(ip, &depend->sideq, entry) { 2471 KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ); 2472 atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ); 2473 atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ); 2474 ip->depend = NULL; 2475 } 2476 2477 /* 2478 * NOTE: pmp->sideq_count includes both sideq and syncq 2479 */ 2480 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry); 2481 2482 depend->count = 0; 2483 depend->pass2 = 0; 2484 TAILQ_REMOVE(&pmp->depq, depend, entry); 2485 } 2486 2487 hammer2_spin_unex(&pmp->list_spin); 2488 hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/ 2489 HAMMER2_TRANS_WAITING); 2490 dorestart = 0; 2491 2492 /* 2493 * sideq_count may have dropped enough to allow us to unstall 2494 * the frontend. 2495 */ 2496 hammer2_pfs_memory_wakeup(pmp, 0); 2497 2498 /* 2499 * Now run through all inodes on syncq. 2500 * 2501 * Flush transactions only interlock with other flush transactions. 2502 * Any conflicting frontend operations will block on the inode, but 2503 * may hold a vnode lock while doing so. 2504 */ 2505 hammer2_spin_ex(&pmp->list_spin); 2506 while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) { 2507 /* 2508 * Remove the inode from the SYNCQ, transfer the syncq ref 2509 * to us. We must clear SYNCQ to allow any potential 2510 * front-end deadlock to proceed. We must set PASS2 so 2511 * the dependency code knows what to do. 2512 */ 2513 pass2 = ip->flags; 2514 cpu_ccfence(); 2515 if (atomic_cmpset_int(&ip->flags, 2516 pass2, 2517 (pass2 & ~(HAMMER2_INODE_SYNCQ | 2518 HAMMER2_INODE_SYNCQ_WAKEUP)) | 2519 HAMMER2_INODE_SYNCQ_PASS2) == 0) 2520 { 2521 continue; 2522 } 2523 TAILQ_REMOVE(&pmp->syncq, ip, entry); 2524 --pmp->sideq_count; 2525 hammer2_spin_unex(&pmp->list_spin); 2526 2527 /* 2528 * Tickle anyone waiting on ip->flags or the hysteresis 2529 * on the dirty inode count. 2530 */ 2531 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP) 2532 wakeup(&ip->flags); 2533 if (++wakecount >= hammer2_limit_dirty_inodes / 20 + 1) { 2534 wakecount = 0; 2535 hammer2_pfs_memory_wakeup(pmp, 0); 2536 } 2537 2538 /* 2539 * Relock the inode, and we inherit a ref from the above. 2540 * We will check for a race after we acquire the vnode. 2541 */ 2542 hammer2_mtx_ex(&ip->lock); 2543 2544 /* 2545 * We need the vp in order to vfsync() dirty buffers, so if 2546 * one isn't attached we can skip it. 2547 * 2548 * Ordering the inode lock and then the vnode lock has the 2549 * potential to deadlock. If we had left SYNCQ set that could 2550 * also deadlock us against the frontend even if we don't hold 2551 * any locks, but the latter is not a problem now since we 2552 * cleared it. igetv will temporarily release the inode lock 2553 * in a safe manner to work-around the deadlock. 2554 * 2555 * Unfortunately it is still possible to deadlock when the 2556 * frontend obtains multiple inode locks, because all the 2557 * related vnodes are already locked (nor can the vnode locks 2558 * be released and reacquired without messing up RECLAIM and 2559 * INACTIVE sequencing). 2560 * 2561 * The solution for now is to move the vp back onto SIDEQ 2562 * and set dorestart, which will restart the flush after we 2563 * exhaust the current SYNCQ. Note that additional 2564 * dependencies may build up, so we definitely need to move 2565 * the whole SIDEQ back to SYNCQ when we restart. 2566 */ 2567 vp = ip->vp; 2568 if (vp) { 2569 if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) { 2570 /* 2571 * Failed to get the vnode, requeue the inode 2572 * (PASS2 is already set so it will be found 2573 * again on the restart). 2574 * 2575 * Then unlock, possibly sleep, and retry 2576 * later. We sleep if PASS2 was *previously* 2577 * set, before we set it again above. 2578 */ 2579 vp = NULL; 2580 dorestart = 1; 2581 #ifdef HAMMER2_DEBUG_SYNC 2582 kprintf("inum %ld (sync delayed by vnode)\n", 2583 (long)ip->meta.inum); 2584 #endif 2585 hammer2_inode_delayed_sideq(ip); 2586 2587 hammer2_mtx_unlock(&ip->lock); 2588 hammer2_inode_drop(ip); 2589 2590 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) { 2591 tsleep(&dorestart, 0, "h2syndel", 2); 2592 } 2593 hammer2_spin_ex(&pmp->list_spin); 2594 continue; 2595 } 2596 } else { 2597 vp = NULL; 2598 } 2599 2600 /* 2601 * If the inode wound up on a SIDEQ again it will already be 2602 * prepped for another PASS2. In this situation if we flush 2603 * it now we will just wind up flushing it again in the same 2604 * syncer run, so we might as well not flush it now. 2605 */ 2606 if (ip->flags & HAMMER2_INODE_SIDEQ) { 2607 hammer2_mtx_unlock(&ip->lock); 2608 hammer2_inode_drop(ip); 2609 if (vp) 2610 vput(vp); 2611 dorestart = 1; 2612 hammer2_spin_ex(&pmp->list_spin); 2613 continue; 2614 } 2615 2616 /* 2617 * Ok we have the inode exclusively locked and if vp is 2618 * not NULL that will also be exclusively locked. Do the 2619 * meat of the flush. 2620 * 2621 * vp token needed for v_rbdirty_tree check / vclrisdirty 2622 * sequencing. Though we hold the vnode exclusively so 2623 * we shouldn't need to hold the token also in this case. 2624 */ 2625 if (vp) { 2626 vfsync(vp, MNT_WAIT, 1, NULL, NULL); 2627 bio_track_wait(NULL, 0, 0); /* XXX */ 2628 } 2629 2630 /* 2631 * If the inode has not yet been inserted into the tree 2632 * we must do so. Then sync and flush it. The flush should 2633 * update the parent. 2634 */ 2635 if (ip->flags & HAMMER2_INODE_DELETING) { 2636 #ifdef HAMMER2_DEBUG_SYNC 2637 kprintf("inum %ld destroy\n", (long)ip->meta.inum); 2638 #endif 2639 hammer2_inode_chain_des(ip); 2640 atomic_add_long(&hammer2_iod_inode_deletes, 1); 2641 } else if (ip->flags & HAMMER2_INODE_CREATING) { 2642 #ifdef HAMMER2_DEBUG_SYNC 2643 kprintf("inum %ld insert\n", (long)ip->meta.inum); 2644 #endif 2645 hammer2_inode_chain_ins(ip); 2646 atomic_add_long(&hammer2_iod_inode_creates, 1); 2647 } 2648 #ifdef HAMMER2_DEBUG_SYNC 2649 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum); 2650 #endif 2651 2652 /* 2653 * Because I kinda messed up the design and index the inodes 2654 * under the root inode, along side the directory entries, 2655 * we can't flush the inode index under the iroot until the 2656 * end. If we do it now we might miss effects created by 2657 * other inodes on the SYNCQ. 2658 * 2659 * Do a normal (non-FSSYNC) flush instead, which allows the 2660 * vnode code to work the same. We don't want to force iroot 2661 * back onto the SIDEQ, and we also don't want the flush code 2662 * to update pfs_iroot_blocksets until the final flush later. 2663 * 2664 * XXX at the moment this will likely result in a double-flush 2665 * of the iroot chain. 2666 */ 2667 hammer2_inode_chain_sync(ip); 2668 if (ip == pmp->iroot) { 2669 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP); 2670 } else { 2671 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP | 2672 HAMMER2_XOP_FSSYNC); 2673 } 2674 if (vp) { 2675 lwkt_gettoken(NULL); 2676 if ((ip->flags & (HAMMER2_INODE_MODIFIED | 2677 HAMMER2_INODE_RESIZED | 2678 HAMMER2_INODE_DIRTYDATA)) == 0) { 2679 //RB_EMPTY(&vp->v_rbdirty_tree) && 2680 //!bio_track_active(&vp->v_track_write)) { 2681 vclrisdirty(vp); 2682 } else { 2683 hammer2_inode_delayed_sideq(ip); 2684 } 2685 lwkt_reltoken(NULL); 2686 vput(vp); 2687 vp = NULL; /* safety */ 2688 } 2689 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2); 2690 hammer2_inode_unlock(ip); /* unlock+drop */ 2691 /* ip pointer invalid */ 2692 2693 /* 2694 * If the inode got dirted after we dropped our locks, 2695 * it will have already been moved back to the SIDEQ. 2696 */ 2697 hammer2_spin_ex(&pmp->list_spin); 2698 } 2699 hammer2_spin_unex(&pmp->list_spin); 2700 hammer2_pfs_memory_wakeup(pmp, 0); 2701 2702 if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) { 2703 #ifdef HAMMER2_DEBUG_SYNC 2704 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n"); 2705 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/ 2706 #endif 2707 dorestart = 1; 2708 goto restart; 2709 } 2710 #ifdef HAMMER2_DEBUG_SYNC 2711 kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n"); 2712 /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/ 2713 #endif 2714 2715 /* 2716 * We have to flush the PFS root last, even if it does not appear to 2717 * be dirty, because all the inodes in the PFS are indexed under it. 2718 * The normal flushing of iroot above would only occur if directory 2719 * entries under the root were changed. 2720 * 2721 * Specifying VOLHDR will cause an additionl flush of hmp->spmp 2722 * for the media making up the cluster. 2723 */ 2724 if ((ip = pmp->iroot) != NULL) { 2725 hammer2_inode_ref(ip); 2726 hammer2_mtx_ex(&ip->lock); 2727 hammer2_inode_chain_sync(ip); 2728 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP | 2729 HAMMER2_XOP_FSSYNC | 2730 HAMMER2_XOP_VOLHDR); 2731 hammer2_inode_unlock(ip); /* unlock+drop */ 2732 } 2733 #ifdef HAMMER2_DEBUG_SYNC 2734 kprintf("FILESYSTEM SYNC STAGE 2 DONE\n"); 2735 #endif 2736 2737 /* 2738 * device bioq sync 2739 */ 2740 hammer2_bioq_sync(pmp); 2741 2742 error = 0; /* XXX */ 2743 hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH); 2744 2745 return (error); 2746 } 2747 2748 #if 0 2749 static 2750 int 2751 hammer2_vfs_vptofh(struct m_vnode *vp, struct fid *fhp) 2752 { 2753 hammer2_inode_t *ip; 2754 2755 KKASSERT(MAXFIDSZ >= 16); 2756 ip = VTOI(vp); 2757 fhp->fid_len = offsetof(struct fid, fid_data[16]); 2758 fhp->fid_ext = 0; 2759 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum; 2760 ((hammer2_tid_t *)fhp->fid_data)[1] = 0; 2761 2762 return 0; 2763 } 2764 2765 static 2766 int 2767 hammer2_vfs_fhtovp(struct mount *mp, struct m_vnode *rootvp, 2768 struct fid *fhp, struct m_vnode **vpp) 2769 { 2770 hammer2_tid_t inum; 2771 int error; 2772 2773 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK; 2774 if (vpp) { 2775 if (inum == 1) 2776 error = hammer2_vfs_root(mp, vpp); 2777 else 2778 error = hammer2_vfs_vget(mp, NULL, inum, vpp); 2779 } else { 2780 error = 0; 2781 } 2782 return error; 2783 } 2784 2785 static 2786 int 2787 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam, 2788 int *exflagsp, struct ucred **credanonp) 2789 { 2790 hammer2_pfs_t *pmp; 2791 struct netcred *np; 2792 int error; 2793 2794 pmp = MPTOPMP(mp); 2795 np = vfs_export_lookup(mp, &pmp->export, nam); 2796 if (np) { 2797 *exflagsp = np->netc_exflags; 2798 *credanonp = &np->netc_anon; 2799 error = 0; 2800 } else { 2801 error = EACCES; 2802 } 2803 return error; 2804 } 2805 #endif 2806 2807 /* 2808 * This handles hysteresis on regular file flushes. Because the BIOs are 2809 * routed to a thread it is possible for an excessive number to build up 2810 * and cause long front-end stalls long before the runningbuffspace limit 2811 * is hit, so we implement hammer2_flush_pipe to control the 2812 * hysteresis. 2813 * 2814 * This is a particular problem when compression is used. 2815 */ 2816 void 2817 hammer2_lwinprog_ref(hammer2_pfs_t *pmp) 2818 { 2819 atomic_add_int(&pmp->count_lwinprog, 1); 2820 } 2821 2822 void 2823 hammer2_lwinprog_drop(hammer2_pfs_t *pmp) 2824 { 2825 #if 0 2826 int lwinprog; 2827 2828 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1); 2829 if ((lwinprog & HAMMER2_LWINPROG_WAITING) && 2830 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) { 2831 atomic_clear_int(&pmp->count_lwinprog, 2832 HAMMER2_LWINPROG_WAITING); 2833 wakeup(&pmp->count_lwinprog); 2834 } 2835 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) && 2836 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) { 2837 atomic_clear_int(&pmp->count_lwinprog, 2838 HAMMER2_LWINPROG_WAITING0); 2839 wakeup(&pmp->count_lwinprog); 2840 } 2841 #endif 2842 } 2843 2844 void 2845 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe) 2846 { 2847 #if 0 2848 int lwinprog; 2849 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING : 2850 HAMMER2_LWINPROG_WAITING0; 2851 2852 for (;;) { 2853 lwinprog = pmp->count_lwinprog; 2854 cpu_ccfence(); 2855 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 2856 break; 2857 tsleep_interlock(&pmp->count_lwinprog, 0); 2858 atomic_set_int(&pmp->count_lwinprog, lwflag); 2859 lwinprog = pmp->count_lwinprog; 2860 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe) 2861 break; 2862 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz); 2863 } 2864 #endif 2865 } 2866 2867 #if 0 2868 /* 2869 * It is possible for an excessive number of dirty chains or dirty inodes 2870 * to build up. When this occurs we start an asynchronous filesystem sync. 2871 * If the level continues to build up, we stall, waiting for it to drop, 2872 * with some hysteresis. 2873 * 2874 * This relies on the kernel calling hammer2_vfs_modifying() prior to 2875 * obtaining any vnode locks before making a modifying VOP call. 2876 */ 2877 static int 2878 hammer2_vfs_modifying(struct mount *mp) 2879 { 2880 if (mp->mnt_flag & MNT_RDONLY) 2881 return EROFS; 2882 hammer2_pfs_memory_wait(MPTOPMP(mp)); 2883 2884 return 0; 2885 } 2886 #endif 2887 2888 /* 2889 * Initiate an asynchronous filesystem sync and, with hysteresis, 2890 * stall if the internal data structure count becomes too bloated. 2891 */ 2892 void 2893 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp) 2894 { 2895 uint32_t waiting; 2896 int pcatch; 2897 int error; 2898 2899 if (pmp == NULL || pmp->mp == NULL) 2900 return; 2901 2902 for (;;) { 2903 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK; 2904 cpu_ccfence(); 2905 2906 /* 2907 * Start the syncer running at 1/2 the limit 2908 */ 2909 if (waiting > hammer2_limit_dirty_chains / 2 || 2910 pmp->sideq_count > hammer2_limit_dirty_inodes / 2) { 2911 trigger_syncer(pmp->mp); 2912 } 2913 2914 /* 2915 * Stall at the limit waiting for the counts to drop. 2916 * This code will typically be woken up once the count 2917 * drops below 3/4 the limit, or in one second. 2918 */ 2919 if (waiting < hammer2_limit_dirty_chains && 2920 pmp->sideq_count < hammer2_limit_dirty_inodes) { 2921 break; 2922 } 2923 2924 pcatch = curthread->td_proc ? PCATCH : 0; 2925 2926 tsleep_interlock(&pmp->inmem_dirty_chains, pcatch); 2927 atomic_set_int(&pmp->inmem_dirty_chains, 2928 HAMMER2_DIRTYCHAIN_WAITING); 2929 if (waiting < hammer2_limit_dirty_chains && 2930 pmp->sideq_count < hammer2_limit_dirty_inodes) { 2931 break; 2932 } 2933 trigger_syncer(pmp->mp); 2934 error = tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED | pcatch, 2935 "h2memw", hz); 2936 if (error == ERESTART) 2937 break; 2938 } 2939 } 2940 2941 /* 2942 * Wake up any stalled frontend ops waiting, with hysteresis, using 2943 * 2/3 of the limit. 2944 */ 2945 void 2946 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp, int count) 2947 { 2948 uint32_t waiting; 2949 2950 if (pmp) { 2951 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, count); 2952 /* don't need --waiting to test flag */ 2953 2954 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) && 2955 (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <= 2956 hammer2_limit_dirty_chains * 2 / 3 && 2957 pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) { 2958 atomic_clear_int(&pmp->inmem_dirty_chains, 2959 HAMMER2_DIRTYCHAIN_WAITING); 2960 wakeup(&pmp->inmem_dirty_chains); 2961 } 2962 } 2963 } 2964 2965 void 2966 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp) 2967 { 2968 if (pmp) { 2969 atomic_add_int(&pmp->inmem_dirty_chains, 1); 2970 } 2971 } 2972 2973 /* 2974 * Volume header data locks 2975 */ 2976 void 2977 hammer2_voldata_lock(hammer2_dev_t *hmp) 2978 { 2979 lockmgr(&hmp->vollk, LK_EXCLUSIVE); 2980 } 2981 2982 void 2983 hammer2_voldata_unlock(hammer2_dev_t *hmp) 2984 { 2985 lockmgr(&hmp->vollk, LK_RELEASE); 2986 } 2987 2988 /* 2989 * Caller indicates that the volume header is being modified. Flag 2990 * the related chain and adjust its transaction id. 2991 * 2992 * The transaction id is set to voldata.mirror_tid + 1, similar to 2993 * what hammer2_chain_modify() does. Be very careful here, volume 2994 * data can be updated independently of the rest of the filesystem. 2995 */ 2996 void 2997 hammer2_voldata_modify(hammer2_dev_t *hmp) 2998 { 2999 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) { 3000 atomic_add_long(&hammer2_count_modified_chains, 1); 3001 atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED); 3002 hammer2_pfs_memory_inc(hmp->vchain.pmp); 3003 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid + 1; 3004 } 3005 } 3006 3007 /* 3008 * Returns 0 if the filesystem has tons of free space 3009 * Returns 1 if the filesystem has less than 10% remaining 3010 * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining. 3011 */ 3012 int 3013 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred) 3014 { 3015 hammer2_pfs_t *pmp; 3016 hammer2_dev_t *hmp; 3017 hammer2_off_t free_reserved; 3018 hammer2_off_t free_nominal; 3019 int i; 3020 3021 pmp = ip->pmp; 3022 3023 if (/*XXX*/ 1 || pmp->free_ticks == 0 || pmp->free_ticks != ticks) { 3024 free_reserved = HAMMER2_SEGSIZE; 3025 free_nominal = 0x7FFFFFFFFFFFFFFFLLU; 3026 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 3027 hmp = pmp->pfs_hmps[i]; 3028 if (hmp == NULL) 3029 continue; 3030 if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER && 3031 pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER) 3032 continue; 3033 3034 if (free_nominal > hmp->voldata.allocator_free) 3035 free_nominal = hmp->voldata.allocator_free; 3036 if (free_reserved < hmp->free_reserved) 3037 free_reserved = hmp->free_reserved; 3038 } 3039 3040 /* 3041 * SMP races ok 3042 */ 3043 pmp->free_reserved = free_reserved; 3044 pmp->free_nominal = free_nominal; 3045 pmp->free_ticks = ticks; 3046 } else { 3047 free_reserved = pmp->free_reserved; 3048 free_nominal = pmp->free_nominal; 3049 } 3050 if (cred && cred->cr_uid != 0) { 3051 if ((int64_t)(free_nominal - bytes) < 3052 (int64_t)free_reserved) { 3053 return 2; 3054 } 3055 } else { 3056 if ((int64_t)(free_nominal - bytes) < 3057 (int64_t)free_reserved / 2) { 3058 return 2; 3059 } 3060 } 3061 if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2) 3062 return 1; 3063 return 0; 3064 } 3065