2 * Copyright (c) 2011-2018 The DragonFly Project. All rights reserved.
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)
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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
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.
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
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>
44 #include <sys/vfsops.h>
45 #include <sys/sysctl.h>
46 #include <sys/socket.h>
47 #include <sys/objcache.h>
50 #include <sys/namei.h>
51 #include <sys/mountctl.h>
52 #include <sys/dirent.h>
55 #include <sys/mutex.h>
56 #include <sys/mutex2.h>
59 #include "hammer2_disk.h"
60 #include "hammer2_mount.h"
61 #include "hammer2_lz4.h"
63 #include "zlib/hammer2_zlib.h"
65 #define REPORT_REFS_ERRORS 1 /* XXX remove me */
67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
69 struct hammer2_sync_info {
75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
76 static struct hammer2_mntlist hammer2_mntlist;
78 struct hammer2_pfslist hammer2_pfslist;
79 struct hammer2_pfslist hammer2_spmplist;
80 struct lock hammer2_mntlk;
82 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
84 int hammer2_cluster_meta_read = 1; /* physical read-ahead */
85 int hammer2_cluster_data_read = 4; /* physical read-ahead */
86 int hammer2_dedup_enable = 1;
87 int hammer2_always_compress = 0; /* always try to compress */
88 int hammer2_inval_enable = 0;
89 int hammer2_flush_pipe = 100;
90 int hammer2_dio_count;
91 int hammer2_dio_limit = 256;
92 int hammer2_bulkfree_tps = 5000;
93 long hammer2_chain_allocs;
94 long hammer2_chain_frees;
95 long hammer2_limit_dirty_chains;
96 long hammer2_limit_dirty_inodes;
97 long hammer2_count_modified_chains;
98 long hammer2_iod_invals;
99 long hammer2_iod_file_read;
100 long hammer2_iod_meta_read;
101 long hammer2_iod_indr_read;
102 long hammer2_iod_fmap_read;
103 long hammer2_iod_volu_read;
104 long hammer2_iod_file_write;
105 long hammer2_iod_file_wembed;
106 long hammer2_iod_file_wzero;
107 long hammer2_iod_file_wdedup;
108 long hammer2_iod_meta_write;
109 long hammer2_iod_indr_write;
110 long hammer2_iod_fmap_write;
111 long hammer2_iod_volu_write;
113 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
114 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
115 "Buffer used for compression.");
117 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
118 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
119 "Buffer used for decompression.");
121 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
124 &hammer2_supported_version, 0, "");
125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
126 &hammer2_debug, 0, "");
127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
128 &hammer2_cluster_meta_read, 0, "");
129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
130 &hammer2_cluster_data_read, 0, "");
131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
132 &hammer2_dedup_enable, 0, "");
133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
134 &hammer2_always_compress, 0, "");
135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
136 &hammer2_inval_enable, 0, "");
137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
138 &hammer2_flush_pipe, 0, "");
139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
140 &hammer2_bulkfree_tps, 0, "");
141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
142 &hammer2_chain_allocs, 0, "");
143 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
144 &hammer2_chain_frees, 0, "");
145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
146 &hammer2_limit_dirty_chains, 0, "");
147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
148 &hammer2_limit_dirty_inodes, 0, "");
149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
150 &hammer2_count_modified_chains, 0, "");
151 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
152 &hammer2_dio_count, 0, "");
153 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
154 &hammer2_dio_limit, 0, "");
156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
157 &hammer2_iod_invals, 0, "");
158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
159 &hammer2_iod_file_read, 0, "");
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
161 &hammer2_iod_meta_read, 0, "");
162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
163 &hammer2_iod_indr_read, 0, "");
164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
165 &hammer2_iod_fmap_read, 0, "");
166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
167 &hammer2_iod_volu_read, 0, "");
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
170 &hammer2_iod_file_write, 0, "");
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
172 &hammer2_iod_file_wembed, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
174 &hammer2_iod_file_wzero, 0, "");
175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
176 &hammer2_iod_file_wdedup, 0, "");
177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
178 &hammer2_iod_meta_write, 0, "");
179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
180 &hammer2_iod_indr_write, 0, "");
181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
182 &hammer2_iod_fmap_write, 0, "");
183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
184 &hammer2_iod_volu_write, 0, "");
186 long hammer2_process_icrc32;
187 long hammer2_process_xxhash64;
188 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
189 &hammer2_process_icrc32, 0, "");
190 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
191 &hammer2_process_xxhash64, 0, "");
193 static int hammer2_vfs_init(struct vfsconf *conf);
194 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
195 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
197 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
198 struct vnode *, struct ucred *);
199 static int hammer2_recovery(hammer2_dev_t *hmp);
200 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
201 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
202 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
204 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
206 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
207 struct fid *fhp, struct vnode **vpp);
208 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
209 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
210 int *exflagsp, struct ucred **credanonp);
212 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
213 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
215 static void hammer2_update_pmps(hammer2_dev_t *hmp);
217 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
218 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
220 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
223 * HAMMER2 vfs operations.
225 static struct vfsops hammer2_vfsops = {
226 .vfs_init = hammer2_vfs_init,
227 .vfs_uninit = hammer2_vfs_uninit,
228 .vfs_sync = hammer2_vfs_sync,
229 .vfs_mount = hammer2_vfs_mount,
230 .vfs_unmount = hammer2_vfs_unmount,
231 .vfs_root = hammer2_vfs_root,
232 .vfs_statfs = hammer2_vfs_statfs,
233 .vfs_statvfs = hammer2_vfs_statvfs,
234 .vfs_vget = hammer2_vfs_vget,
235 .vfs_vptofh = hammer2_vfs_vptofh,
236 .vfs_fhtovp = hammer2_vfs_fhtovp,
237 .vfs_checkexp = hammer2_vfs_checkexp
240 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
242 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
243 MODULE_VERSION(hammer2, 1);
247 hammer2_vfs_init(struct vfsconf *conf)
249 static struct objcache_malloc_args margs_read;
250 static struct objcache_malloc_args margs_write;
251 static struct objcache_malloc_args margs_vop;
258 * A large DIO cache is needed to retain dedup enablement masks.
259 * The bulkfree code clears related masks as part of the disk block
260 * recycling algorithm, preventing it from being used for a later
263 * NOTE: A large buffer cache can actually interfere with dedup
264 * operation because we dedup based on media physical buffers
265 * and not logical buffers. Try to make the DIO case large
266 * enough to avoid this problem, but also cap it.
268 hammer2_dio_limit = nbuf * 2;
269 if (hammer2_dio_limit > 100000)
270 hammer2_dio_limit = 100000;
272 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
274 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
276 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
280 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
282 margs_read.objsize = 65536;
283 margs_read.mtype = M_HAMMER2_DEBUFFER;
285 margs_write.objsize = 32768;
286 margs_write.mtype = M_HAMMER2_CBUFFER;
288 margs_vop.objsize = sizeof(hammer2_xop_t);
289 margs_vop.mtype = M_HAMMER2;
292 * Note thaht for the XOPS cache we want backing store allocations
293 * to use M_ZERO. This is not allowed in objcache_get() (to avoid
294 * confusion), so use the backing store function that does it. This
295 * means that initial XOPS objects are zerod but REUSED objects are
296 * not. So we are responsible for cleaning the object up sufficiently
297 * for our needs before objcache_put()ing it back (typically just the
300 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
301 0, 1, NULL, NULL, NULL,
302 objcache_malloc_alloc,
303 objcache_malloc_free,
305 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
306 0, 1, NULL, NULL, NULL,
307 objcache_malloc_alloc,
308 objcache_malloc_free,
310 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
311 0, 1, NULL, NULL, NULL,
312 objcache_malloc_alloc_zero,
313 objcache_malloc_free,
317 lockinit(&hammer2_mntlk, "mntlk", 0, 0);
318 TAILQ_INIT(&hammer2_mntlist);
319 TAILQ_INIT(&hammer2_pfslist);
320 TAILQ_INIT(&hammer2_spmplist);
322 hammer2_limit_dirty_chains = maxvnodes / 10;
323 if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
324 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
326 hammer2_limit_dirty_inodes = maxvnodes / 100;
327 if (hammer2_limit_dirty_inodes < 100)
328 hammer2_limit_dirty_inodes = 100;
329 if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
330 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
337 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
339 objcache_destroy(cache_buffer_read);
340 objcache_destroy(cache_buffer_write);
341 objcache_destroy(cache_xops);
346 * Core PFS allocator. Used to allocate or reference the pmp structure
347 * for PFS cluster mounts and the spmp structure for media (hmp) structures.
348 * The pmp can be passed in or loaded by this function using the chain and
351 * pmp->modify_tid tracks new modify_tid transaction ids for front-end
352 * transactions. Note that synchronization does not use this field.
353 * (typically frontend operations and synchronization cannot run on the
354 * same PFS node at the same time).
359 hammer2_pfsalloc(hammer2_chain_t *chain,
360 const hammer2_inode_data_t *ripdata,
361 hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
364 hammer2_inode_t *iroot;
372 * Locate or create the PFS based on the cluster id. If ripdata
373 * is NULL this is a spmp which is unique and is always allocated.
375 * If the device is mounted in local mode all PFSs are considered
376 * independent and not part of any cluster (for debugging only).
379 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
380 if (force_local != pmp->force_local)
382 if (force_local == NULL &&
383 bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
384 sizeof(pmp->pfs_clid)) == 0) {
386 } else if (force_local && pmp->pfs_names[0] &&
387 strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
394 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
395 pmp->force_local = force_local;
396 hammer2_trans_manage_init(pmp);
397 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
398 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
399 lockinit(&pmp->lock, "pfslk", 0, 0);
400 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
401 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
402 spin_init(&pmp->xop_spin, "h2xop");
403 spin_init(&pmp->lru_spin, "h2lru");
404 RB_INIT(&pmp->inum_tree);
405 TAILQ_INIT(&pmp->sideq);
406 TAILQ_INIT(&pmp->lru_list);
407 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
410 * Distribute backend operations to threads
412 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
413 hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
416 * Save the last media transaction id for the flusher. Set
420 pmp->pfs_clid = ripdata->meta.pfs_clid;
421 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
423 pmp->flags |= HAMMER2_PMPF_SPMP;
424 TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
428 * The synchronization thread may start too early, make
429 * sure it stays frozen until we are ready to let it go.
433 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
434 HAMMER2_THREAD_REMASTER;
439 * Create the PFS's root inode and any missing XOP helper threads.
441 if ((iroot = pmp->iroot) == NULL) {
442 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
444 iroot->meta = ripdata->meta;
446 hammer2_inode_ref(iroot);
447 hammer2_inode_unlock(iroot);
451 * Stop here if no chain is passed in.
457 * When a chain is passed in we must add it to the PFS's root
458 * inode, update pmp->pfs_types[], and update the syncronization
461 * When forcing local mode, mark the PFS as a MASTER regardless.
463 * At the moment empty spots can develop due to removals or failures.
464 * Ultimately we want to re-fill these spots but doing so might
465 * confused running code. XXX
467 hammer2_inode_ref(iroot);
468 hammer2_mtx_ex(&iroot->lock);
469 j = iroot->cluster.nchains;
471 if (j == HAMMER2_MAXCLUSTER) {
472 kprintf("hammer2_mount: cluster full!\n");
473 /* XXX fatal error? */
475 KKASSERT(chain->pmp == NULL);
477 hammer2_chain_ref(chain);
478 iroot->cluster.array[j].chain = chain;
480 pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
482 pmp->pfs_types[j] = ripdata->meta.pfs_type;
483 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
484 pmp->pfs_hmps[j] = chain->hmp;
487 * If the PFS is already mounted we must account
488 * for the mount_count here.
491 ++chain->hmp->mount_count;
494 * May have to fixup dirty chain tracking. Previous
495 * pmp was NULL so nothing to undo.
497 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
498 hammer2_pfs_memory_inc(pmp);
501 iroot->cluster.nchains = j;
504 * Update nmasters from any PFS inode which is part of the cluster.
505 * It is possible that this will result in a value which is too
506 * high. MASTER PFSs are authoritative for pfs_nmasters and will
507 * override this value later on.
509 * (This informs us of masters that might not currently be
510 * discoverable by this mount).
512 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
513 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
517 * Count visible masters. Masters are usually added with
518 * ripdata->meta.pfs_nmasters set to 1. This detects when there
519 * are more (XXX and must update the master inodes).
522 for (i = 0; i < iroot->cluster.nchains; ++i) {
523 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
526 if (pmp->pfs_nmasters < count)
527 pmp->pfs_nmasters = count;
530 * Create missing synchronization and support threads.
532 * Single-node masters (including snapshots) have nothing to
533 * synchronize and do not require this thread.
535 * Multi-node masters or any number of soft masters, slaves, copy,
536 * or other PFS types need the thread.
538 * Each thread is responsible for its particular cluster index.
539 * We use independent threads so stalls or mismatches related to
540 * any given target do not affect other targets.
542 for (i = 0; i < iroot->cluster.nchains; ++i) {
544 * Single-node masters (including snapshots) have nothing
545 * to synchronize and will make direct xops support calls,
546 * thus they do not require this thread.
548 * Note that there can be thousands of snapshots. We do not
549 * want to create thousands of threads.
551 if (pmp->pfs_nmasters <= 1 &&
552 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
557 * Sync support thread
559 if (pmp->sync_thrs[i].td == NULL) {
560 hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
562 hammer2_primary_sync_thread);
567 * Create missing Xop threads
569 * NOTE: We create helper threads for all mounted PFSs or any
570 * PFSs with 2+ nodes (so the sync thread can update them,
571 * even if not mounted).
573 if (pmp->mp || iroot->cluster.nchains >= 2)
574 hammer2_xop_helper_create(pmp);
576 hammer2_mtx_unlock(&iroot->lock);
577 hammer2_inode_drop(iroot);
583 * Deallocate an element of a probed PFS. If destroying and this is a
584 * MASTER, adjust nmasters.
586 * This function does not physically destroy the PFS element in its device
587 * under the super-root (see hammer2_ioctl_pfs_delete()).
590 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
592 hammer2_inode_t *iroot;
593 hammer2_chain_t *chain;
597 * Cleanup our reference on iroot. iroot is (should) not be needed
605 * XXX flush after acquiring the iroot lock.
606 * XXX clean out the cluster index from all inode structures.
608 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
611 * Remove the cluster index from the group. If destroying
612 * the PFS and this is a master, adjust pfs_nmasters.
614 hammer2_mtx_ex(&iroot->lock);
615 chain = iroot->cluster.array[clindex].chain;
616 iroot->cluster.array[clindex].chain = NULL;
618 switch(pmp->pfs_types[clindex]) {
619 case HAMMER2_PFSTYPE_MASTER:
620 if (destroying && pmp->pfs_nmasters > 0)
622 /* XXX adjust ripdata->meta.pfs_nmasters */
627 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
629 hammer2_mtx_unlock(&iroot->lock);
635 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
636 hammer2_chain_drop(chain);
640 * Terminate all XOP threads for the cluster index.
642 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
643 hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
648 * Destroy a PFS, typically only occurs after the last mount on a device
652 hammer2_pfsfree(hammer2_pfs_t *pmp)
654 hammer2_inode_t *iroot;
655 hammer2_chain_t *chain;
660 * Cleanup our reference on iroot. iroot is (should) not be needed
663 if (pmp->flags & HAMMER2_PMPF_SPMP)
664 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
666 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
670 for (i = 0; i < iroot->cluster.nchains; ++i) {
671 hammer2_thr_delete(&pmp->sync_thrs[i]);
672 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
673 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
675 #if REPORT_REFS_ERRORS
676 if (pmp->iroot->refs != 1)
677 kprintf("PMP->IROOT %p REFS WRONG %d\n",
678 pmp->iroot, pmp->iroot->refs);
680 KKASSERT(pmp->iroot->refs == 1);
682 /* ref for pmp->iroot */
683 hammer2_inode_drop(pmp->iroot);
688 * Cleanup chains remaining on LRU list.
690 hammer2_spin_ex(&pmp->lru_spin);
691 while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
692 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
693 atomic_add_int(&pmp->lru_count, -1);
694 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
695 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
696 hammer2_chain_ref(chain);
697 hammer2_spin_unex(&pmp->lru_spin);
698 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
699 hammer2_chain_drop(chain);
700 hammer2_spin_ex(&pmp->lru_spin);
702 hammer2_spin_unex(&pmp->lru_spin);
705 * Free remaining pmp resources
707 kmalloc_destroy(&pmp->mmsg);
708 kmalloc_destroy(&pmp->minode);
710 kfree(pmp, M_HAMMER2);
714 * Remove all references to hmp from the pfs list. Any PFS which becomes
715 * empty is terminated and freed.
720 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
723 hammer2_inode_t *iroot;
724 hammer2_chain_t *rchain;
728 struct hammer2_pfslist *wlist;
731 wlist = &hammer2_pfslist;
733 wlist = &hammer2_spmplist;
735 TAILQ_FOREACH(pmp, wlist, mntentry) {
736 if ((iroot = pmp->iroot) == NULL)
738 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
739 hammer2_inode_run_sideq(pmp, 1);
740 hammer2_bioq_sync(pmp);
741 hammer2_trans_done(pmp);
744 * Determine if this PFS is affected. If it is we must
745 * freeze all management threads and lock its iroot.
747 * Freezing a management thread forces it idle, operations
748 * in-progress will be aborted and it will have to start
749 * over again when unfrozen, or exit if told to exit.
751 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
752 if (pmp->pfs_hmps[i] == hmp)
755 if (i != HAMMER2_MAXCLUSTER) {
757 * Make sure all synchronization threads are locked
760 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
761 if (pmp->pfs_hmps[i] == NULL)
763 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
764 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
765 hammer2_thr_freeze_async(
766 &pmp->xop_groups[j].thrs[i]);
769 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
770 if (pmp->pfs_hmps[i] == NULL)
772 hammer2_thr_freeze(&pmp->sync_thrs[i]);
773 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
775 &pmp->xop_groups[j].thrs[i]);
780 * Lock the inode and clean out matching chains.
781 * Note that we cannot use hammer2_inode_lock_*()
782 * here because that would attempt to validate the
783 * cluster that we are in the middle of ripping
786 * WARNING! We are working directly on the inodes
789 hammer2_mtx_ex(&iroot->lock);
792 * Remove the chain from matching elements of the PFS.
794 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
795 if (pmp->pfs_hmps[i] != hmp)
797 hammer2_thr_delete(&pmp->sync_thrs[i]);
798 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
800 &pmp->xop_groups[j].thrs[i]);
802 rchain = iroot->cluster.array[i].chain;
803 iroot->cluster.array[i].chain = NULL;
804 pmp->pfs_types[i] = 0;
805 if (pmp->pfs_names[i]) {
806 kfree(pmp->pfs_names[i], M_HAMMER2);
807 pmp->pfs_names[i] = NULL;
810 hammer2_chain_drop(rchain);
812 if (iroot->cluster.focus == rchain)
813 iroot->cluster.focus = NULL;
815 pmp->pfs_hmps[i] = NULL;
817 hammer2_mtx_unlock(&iroot->lock);
818 didfreeze = 1; /* remaster, unfreeze down below */
824 * Cleanup trailing chains. Gaps may remain.
826 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
827 if (pmp->pfs_hmps[i])
830 iroot->cluster.nchains = i + 1;
833 * If the PMP has no elements remaining we can destroy it.
834 * (this will transition management threads from frozen->exit).
836 if (iroot->cluster.nchains == 0) {
838 * If this was the hmp's spmp, we need to clean
839 * a little more stuff out.
841 if (hmp->spmp == pmp) {
843 hmp->vchain.pmp = NULL;
844 hmp->fchain.pmp = NULL;
848 * Free the pmp and restart the loop
850 hammer2_pfsfree(pmp);
855 * If elements still remain we need to set the REMASTER
856 * flag and unfreeze it.
859 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
860 if (pmp->pfs_hmps[i] == NULL)
862 hammer2_thr_remaster(&pmp->sync_thrs[i]);
863 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
864 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
865 hammer2_thr_remaster(
866 &pmp->xop_groups[j].thrs[i]);
867 hammer2_thr_unfreeze(
868 &pmp->xop_groups[j].thrs[i]);
876 * Mount or remount HAMMER2 fileystem from physical media
879 * mp mount point structure
885 * mp mount point structure
886 * path path to mount point
887 * data pointer to argument structure in user space
888 * volume volume path (device@LABEL form)
889 * hflags user mount flags
890 * cred user credentials
897 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
900 struct hammer2_mount_info info;
904 hammer2_dev_t *force_local;
905 hammer2_key_t key_next;
906 hammer2_key_t key_dummy;
909 struct nlookupdata nd;
910 hammer2_chain_t *parent;
911 hammer2_chain_t *chain;
912 hammer2_cluster_t *cluster;
913 const hammer2_inode_data_t *ripdata;
914 hammer2_blockref_t bref;
916 char devstr[MNAMELEN];
931 kprintf("hammer2_mount\n");
937 bzero(&info, sizeof(info));
938 info.cluster_fd = -1;
939 ksnprintf(devstr, sizeof(devstr), "%s",
940 mp->mnt_stat.f_mntfromname);
941 kprintf("hammer2_mount: root '%s'\n", devstr);
944 * Non-root mount or updating a mount
946 error = copyin(data, &info, sizeof(info));
950 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
956 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
957 * if no label specified, based on the partition id. Error out if no
958 * label or device (with partition id) is specified. This is strictly
959 * a convenience to match the default label created by newfs_hammer2,
960 * our preference is that a label always be specified.
962 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
963 * that does not specify a device, as long as some H2 label
964 * has already been mounted from that device. This makes
965 * mounting snapshots a lot easier.
968 label = strchr(devstr, '@');
969 if (label && ((label + 1) - dev) > done)
971 if (label == NULL || label[1] == 0) {
975 label = devstr + strlen(devstr);
993 kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
994 dev, label, (mp->mnt_flag & MNT_RDONLY));
996 if (mp->mnt_flag & MNT_UPDATE) {
998 * Update mount. Note that pmp->iroot->cluster is
999 * an inode-embedded cluster and thus cannot be
1002 * XXX HAMMER2 needs to implement NFS export via
1006 pmp->hflags = info.hflags;
1007 cluster = &pmp->iroot->cluster;
1008 for (i = 0; i < cluster->nchains; ++i) {
1009 if (cluster->array[i].chain == NULL)
1011 hmp = cluster->array[i].chain->hmp;
1013 error = hammer2_remount(hmp, mp, path,
1025 * If a path is specified and dev is not an empty string, lookup the
1026 * name and verify that it referes to a block device.
1028 * If a path is specified and dev is an empty string we fall through
1029 * and locate the label in the hmp search.
1031 if (path && *dev != 0) {
1032 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1034 error = nlookup(&nd);
1036 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1038 } else if (path == NULL) {
1040 cdev_t cdev = kgetdiskbyname(dev);
1041 error = bdevvp(cdev, &devvp);
1043 kprintf("hammer2: cannot find '%s'\n", dev);
1046 * We will locate the hmp using the label in the hmp loop.
1052 * Make sure its a block device. Do not check to see if it is
1053 * already mounted until we determine that its a fresh H2 device.
1055 if (error == 0 && devvp) {
1056 vn_isdisk(devvp, &error);
1060 * Determine if the device has already been mounted. After this
1061 * check hmp will be non-NULL if we are doing the second or more
1062 * hammer2 mounts from the same device.
1064 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1067 * Match the device. Due to the way devfs works,
1068 * we may not be able to directly match the vnode pointer,
1069 * so also check to see if the underlying device matches.
1071 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1072 if (hmp->devvp == devvp)
1074 if (devvp->v_rdev &&
1075 hmp->devvp->v_rdev == devvp->v_rdev) {
1081 * If no match this may be a fresh H2 mount, make sure
1082 * the device is not mounted on anything else.
1085 error = vfs_mountedon(devvp);
1086 } else if (error == 0) {
1088 * Match the label to a pmp already probed.
1090 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1091 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1092 if (pmp->pfs_names[i] &&
1093 strcmp(pmp->pfs_names[i], label) == 0) {
1094 hmp = pmp->pfs_hmps[i];
1106 * Open the device if this isn't a secondary mount and construct
1107 * the H2 device mount (hmp).
1110 hammer2_chain_t *schain;
1113 if (error == 0 && vcount(devvp) > 0) {
1114 kprintf("Primary device already has references\n");
1119 * Now open the device
1122 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1123 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1124 error = vinvalbuf(devvp, V_SAVE, 0, 0);
1126 error = VOP_OPEN(devvp,
1127 (ronly ? FREAD : FREAD | FWRITE),
1132 if (error && devvp) {
1137 lockmgr(&hammer2_mntlk, LK_RELEASE);
1140 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1141 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1144 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1145 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1146 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1147 RB_INIT(&hmp->iotree);
1148 spin_init(&hmp->io_spin, "hm2mount_io");
1149 spin_init(&hmp->list_spin, "hm2mount_list");
1150 TAILQ_INIT(&hmp->flushq);
1152 lockinit(&hmp->vollk, "h2vol", 0, 0);
1153 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1154 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1157 * vchain setup. vchain.data is embedded.
1158 * vchain.refs is initialized and will never drop to 0.
1160 * NOTE! voldata is not yet loaded.
1162 hmp->vchain.hmp = hmp;
1163 hmp->vchain.refs = 1;
1164 hmp->vchain.data = (void *)&hmp->voldata;
1165 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1166 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1167 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1169 hammer2_chain_core_init(&hmp->vchain);
1170 /* hmp->vchain.u.xxx is left NULL */
1173 * fchain setup. fchain.data is embedded.
1174 * fchain.refs is initialized and will never drop to 0.
1176 * The data is not used but needs to be initialized to
1177 * pass assertion muster. We use this chain primarily
1178 * as a placeholder for the freemap's top-level RBTREE
1179 * so it does not interfere with the volume's topology
1182 hmp->fchain.hmp = hmp;
1183 hmp->fchain.refs = 1;
1184 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1185 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1186 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1187 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1188 hmp->fchain.bref.methods =
1189 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1190 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1192 hammer2_chain_core_init(&hmp->fchain);
1193 /* hmp->fchain.u.xxx is left NULL */
1196 * Install the volume header and initialize fields from
1199 error = hammer2_install_volume_header(hmp);
1201 hammer2_unmount_helper(mp, NULL, hmp);
1202 lockmgr(&hammer2_mntlk, LK_RELEASE);
1203 hammer2_vfs_unmount(mp, MNT_FORCE);
1208 * Really important to get these right or flush will get
1211 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1215 * Dummy-up vchain and fchain's modify_tid. mirror_tid
1216 * is inherited from the volume header.
1219 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1220 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1221 hmp->vchain.pmp = spmp;
1222 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1223 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1224 hmp->fchain.pmp = spmp;
1227 * First locate the super-root inode, which is key 0
1228 * relative to the volume header's blockset.
1230 * Then locate the root inode by scanning the directory keyspace
1231 * represented by the label.
1233 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1234 schain = hammer2_chain_lookup(&parent, &key_dummy,
1235 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1237 hammer2_chain_lookup_done(parent);
1238 if (schain == NULL) {
1239 kprintf("hammer2_mount: invalid super-root\n");
1240 hammer2_unmount_helper(mp, NULL, hmp);
1241 lockmgr(&hammer2_mntlk, LK_RELEASE);
1242 hammer2_vfs_unmount(mp, MNT_FORCE);
1245 if (schain->error) {
1246 kprintf("hammer2_mount: error %s reading super-root\n",
1247 hammer2_error_str(schain->error));
1248 hammer2_chain_unlock(schain);
1249 hammer2_chain_drop(schain);
1251 hammer2_unmount_helper(mp, NULL, hmp);
1252 lockmgr(&hammer2_mntlk, LK_RELEASE);
1253 hammer2_vfs_unmount(mp, MNT_FORCE);
1258 * The super-root always uses an inode_tid of 1 when
1261 spmp->inode_tid = 1;
1262 spmp->modify_tid = schain->bref.modify_tid + 1;
1265 * Sanity-check schain's pmp and finish initialization.
1266 * Any chain belonging to the super-root topology should
1267 * have a NULL pmp (not even set to spmp).
1269 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1270 KKASSERT(schain->pmp == NULL);
1271 spmp->pfs_clid = ripdata->meta.pfs_clid;
1274 * Replace the dummy spmp->iroot with a real one. It's
1275 * easier to just do a wholesale replacement than to try
1276 * to update the chain and fixup the iroot fields.
1278 * The returned inode is locked with the supplied cluster.
1280 cluster = hammer2_cluster_from_chain(schain);
1281 hammer2_inode_drop(spmp->iroot);
1283 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1284 spmp->spmp_hmp = hmp;
1285 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1286 spmp->pfs_hmps[0] = hmp;
1287 hammer2_inode_ref(spmp->iroot);
1288 hammer2_inode_unlock(spmp->iroot);
1289 hammer2_cluster_unlock(cluster);
1290 hammer2_cluster_drop(cluster);
1292 /* leave spmp->iroot with one ref */
1294 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1295 error = hammer2_recovery(hmp);
1297 error |= hammer2_fixup_pfses(hmp);
1298 /* XXX do something with error */
1300 hammer2_update_pmps(hmp);
1301 hammer2_iocom_init(hmp);
1302 hammer2_bulkfree_init(hmp);
1305 * Ref the cluster management messaging descriptor. The mount
1306 * program deals with the other end of the communications pipe.
1308 * Root mounts typically do not supply one.
1310 if (info.cluster_fd >= 0) {
1311 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1313 hammer2_cluster_reconnect(hmp, fp);
1315 kprintf("hammer2_mount: bad cluster_fd!\n");
1320 if (info.hflags & HMNT2_DEVFLAGS) {
1321 kprintf("hammer2: Warning: mount flags pertaining "
1322 "to the whole device may only be specified "
1323 "on the first mount of the device: %08x\n",
1324 info.hflags & HMNT2_DEVFLAGS);
1329 * Force local mount (disassociate all PFSs from their clusters).
1330 * Used primarily for debugging.
1332 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1335 * Lookup the mount point under the media-localized super-root.
1336 * Scanning hammer2_pfslist doesn't help us because it represents
1337 * PFS cluster ids which can aggregate several named PFSs together.
1339 * cluster->pmp will incorrectly point to spmp and must be fixed
1342 hammer2_inode_lock(spmp->iroot, 0);
1343 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1344 lhc = hammer2_dirhash(label, strlen(label));
1345 chain = hammer2_chain_lookup(&parent, &key_next,
1346 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1349 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1350 strcmp(label, chain->data->ipdata.filename) == 0) {
1353 chain = hammer2_chain_next(&parent, chain, &key_next,
1355 lhc + HAMMER2_DIRHASH_LOMASK,
1359 hammer2_chain_unlock(parent);
1360 hammer2_chain_drop(parent);
1362 hammer2_inode_unlock(spmp->iroot);
1365 * PFS could not be found?
1367 if (chain == NULL) {
1369 kprintf("hammer2_mount: PFS label I/O error\n");
1371 kprintf("hammer2_mount: PFS label not found\n");
1372 hammer2_unmount_helper(mp, NULL, hmp);
1373 lockmgr(&hammer2_mntlk, LK_RELEASE);
1374 hammer2_vfs_unmount(mp, MNT_FORCE);
1380 * Acquire the pmp structure (it should have already been allocated
1381 * via hammer2_update_pmps() so do not pass cluster in to add to
1382 * available chains).
1384 * Check if the cluster has already been mounted. A cluster can
1385 * only be mounted once, use null mounts to mount additional copies.
1388 kprintf("hammer2_mount: PFS label I/O error\n");
1390 ripdata = &chain->data->ipdata;
1392 pmp = hammer2_pfsalloc(NULL, ripdata,
1393 bref.modify_tid, force_local);
1395 hammer2_chain_unlock(chain);
1396 hammer2_chain_drop(chain);
1401 kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1404 kprintf("hammer2_mount: PFS already mounted!\n");
1405 hammer2_unmount_helper(mp, NULL, hmp);
1406 lockmgr(&hammer2_mntlk, LK_RELEASE);
1407 hammer2_vfs_unmount(mp, MNT_FORCE);
1412 pmp->hflags = info.hflags;
1413 mp->mnt_flag |= MNT_LOCAL;
1414 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */
1415 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */
1418 * required mount structure initializations
1420 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1421 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1423 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1424 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1429 mp->mnt_iosize_max = MAXPHYS;
1432 * Connect up mount pointers.
1434 hammer2_mount_helper(mp, pmp);
1436 lockmgr(&hammer2_mntlk, LK_RELEASE);
1442 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1443 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1444 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1447 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1448 MNAMELEN - 1, &size);
1449 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1450 } /* else root mount, already in there */
1452 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1454 copyinstr(path, mp->mnt_stat.f_mntonname,
1455 sizeof(mp->mnt_stat.f_mntonname) - 1,
1459 mp->mnt_stat.f_mntonname[0] = '/';
1463 * Initial statfs to prime mnt_stat.
1465 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1471 * Scan PFSs under the super-root and create hammer2_pfs structures.
1475 hammer2_update_pmps(hammer2_dev_t *hmp)
1477 const hammer2_inode_data_t *ripdata;
1478 hammer2_chain_t *parent;
1479 hammer2_chain_t *chain;
1480 hammer2_blockref_t bref;
1481 hammer2_dev_t *force_local;
1482 hammer2_pfs_t *spmp;
1484 hammer2_key_t key_next;
1488 * Force local mount (disassociate all PFSs from their clusters).
1489 * Used primarily for debugging.
1491 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1494 * Lookup mount point under the media-localized super-root.
1496 * cluster->pmp will incorrectly point to spmp and must be fixed
1500 hammer2_inode_lock(spmp->iroot, 0);
1501 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1502 chain = hammer2_chain_lookup(&parent, &key_next,
1503 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1506 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1509 kprintf("I/O error scanning PFS labels\n");
1511 ripdata = &chain->data->ipdata;
1514 pmp = hammer2_pfsalloc(chain, ripdata,
1515 bref.modify_tid, force_local);
1517 chain = hammer2_chain_next(&parent, chain, &key_next,
1518 key_next, HAMMER2_KEY_MAX,
1522 hammer2_chain_unlock(parent);
1523 hammer2_chain_drop(parent);
1525 hammer2_inode_unlock(spmp->iroot);
1530 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1531 struct vnode *devvp, struct ucred *cred)
1535 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1536 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1537 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1539 error = hammer2_recovery(hmp);
1541 error |= hammer2_fixup_pfses(hmp);
1542 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1544 VOP_CLOSE(devvp, FREAD, NULL);
1547 VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1558 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1569 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1572 * If mount initialization proceeded far enough we must flush
1573 * its vnodes and sync the underlying mount points. Three syncs
1574 * are required to fully flush the filesystem (freemap updates lag
1575 * by one flush, and one extra for safety).
1577 if (mntflags & MNT_FORCE)
1582 error = vflush(mp, 0, flags);
1585 hammer2_vfs_sync(mp, MNT_WAIT);
1586 hammer2_vfs_sync(mp, MNT_WAIT);
1587 hammer2_vfs_sync(mp, MNT_WAIT);
1591 * Cleanup the frontend support XOPS threads
1593 hammer2_xop_helper_cleanup(pmp);
1596 hammer2_unmount_helper(mp, pmp, NULL);
1600 lockmgr(&hammer2_mntlk, LK_RELEASE);
1606 * Mount helper, hook the system mount into our PFS.
1607 * The mount lock is held.
1609 * We must bump the mount_count on related devices for any
1614 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1616 hammer2_cluster_t *cluster;
1617 hammer2_chain_t *rchain;
1620 mp->mnt_data = (qaddr_t)pmp;
1624 * After pmp->mp is set we have to adjust hmp->mount_count.
1626 cluster = &pmp->iroot->cluster;
1627 for (i = 0; i < cluster->nchains; ++i) {
1628 rchain = cluster->array[i].chain;
1631 ++rchain->hmp->mount_count;
1635 * Create missing Xop threads
1637 hammer2_xop_helper_create(pmp);
1641 * Mount helper, unhook the system mount from our PFS.
1642 * The mount lock is held.
1644 * If hmp is supplied a mount responsible for being the first to open
1645 * the block device failed and the block device and all PFSs using the
1646 * block device must be cleaned up.
1648 * If pmp is supplied multiple devices might be backing the PFS and each
1649 * must be disconnected. This might not be the last PFS using some of the
1650 * underlying devices. Also, we have to adjust our hmp->mount_count
1651 * accounting for the devices backing the pmp which is now undergoing an
1656 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1658 hammer2_cluster_t *cluster;
1659 hammer2_chain_t *rchain;
1660 struct vnode *devvp;
1666 * If no device supplied this is a high-level unmount and we have to
1667 * to disconnect the mount, adjust mount_count, and locate devices
1668 * that might now have no mounts.
1671 KKASSERT(hmp == NULL);
1672 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1674 mp->mnt_data = NULL;
1677 * After pmp->mp is cleared we have to account for
1680 cluster = &pmp->iroot->cluster;
1681 for (i = 0; i < cluster->nchains; ++i) {
1682 rchain = cluster->array[i].chain;
1685 --rchain->hmp->mount_count;
1686 /* scrapping hmp now may invalidate the pmp */
1689 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1690 if (hmp->mount_count == 0) {
1691 hammer2_unmount_helper(NULL, NULL, hmp);
1699 * Try to terminate the block device. We can't terminate it if
1700 * there are still PFSs referencing it.
1702 if (hmp->mount_count)
1706 * Decomission the network before we start messing with the
1709 hammer2_iocom_uninit(hmp);
1711 hammer2_bulkfree_uninit(hmp);
1712 hammer2_pfsfree_scan(hmp, 0);
1713 hammer2_dev_exlock(hmp); /* XXX order */
1716 * Cycle the volume data lock as a safety (probably not needed any
1717 * more). To ensure everything is out we need to flush at least
1718 * three times. (1) The running of the sideq can dirty the
1719 * filesystem, (2) A normal flush can dirty the freemap, and
1720 * (3) ensure that the freemap is fully synchronized.
1722 * The next mount's recovery scan can clean everything up but we want
1723 * to leave the filesystem in a 100% clean state on a normal unmount.
1726 hammer2_voldata_lock(hmp);
1727 hammer2_voldata_unlock(hmp);
1731 * Flush whatever is left. Unmounted but modified PFS's might still
1732 * have some dirty chains on them.
1734 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1735 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1737 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1738 hammer2_voldata_modify(hmp);
1739 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1742 hammer2_chain_unlock(&hmp->fchain);
1744 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1745 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1748 hammer2_chain_unlock(&hmp->vchain);
1750 if ((hmp->vchain.flags | hmp->fchain.flags) &
1751 HAMMER2_CHAIN_FLUSH_MASK) {
1752 kprintf("hammer2_unmount: chains left over "
1753 "after final sync\n");
1754 kprintf(" vchain %08x\n", hmp->vchain.flags);
1755 kprintf(" fchain %08x\n", hmp->fchain.flags);
1757 if (hammer2_debug & 0x0010)
1758 Debugger("entered debugger");
1761 hammer2_pfsfree_scan(hmp, 1);
1763 KKASSERT(hmp->spmp == NULL);
1766 * Finish up with the device vnode
1768 if ((devvp = hmp->devvp) != NULL) {
1770 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1771 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1772 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1774 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1775 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1776 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1778 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1785 * Clear vchain/fchain flags that might prevent final cleanup
1788 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1789 atomic_add_long(&hammer2_count_modified_chains, -1);
1790 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1791 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1793 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1794 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1797 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1798 atomic_add_long(&hammer2_count_modified_chains, -1);
1799 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1800 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1802 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1803 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1807 * Final drop of embedded freemap root chain to
1808 * clean up fchain.core (fchain structure is not
1809 * flagged ALLOCATED so it is cleaned out and then
1812 hammer2_chain_drop(&hmp->fchain);
1815 * Final drop of embedded volume root chain to clean
1816 * up vchain.core (vchain structure is not flagged
1817 * ALLOCATED so it is cleaned out and then left to
1821 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1823 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1824 hammer2_dev_unlock(hmp);
1825 hammer2_chain_drop(&hmp->vchain);
1827 hammer2_io_cleanup(hmp, &hmp->iotree);
1828 if (hmp->iofree_count) {
1829 kprintf("io_cleanup: %d I/O's left hanging\n",
1833 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1834 kmalloc_destroy(&hmp->mchain);
1835 kfree(hmp, M_HAMMER2);
1839 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1840 ino_t ino, struct vnode **vpp)
1842 hammer2_xop_lookup_t *xop;
1844 hammer2_inode_t *ip;
1848 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1854 * Easy if we already have it cached
1856 ip = hammer2_inode_lookup(pmp, inum);
1858 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1859 *vpp = hammer2_igetv(ip, &error);
1860 hammer2_inode_unlock(ip);
1861 hammer2_inode_drop(ip); /* from lookup */
1867 * Otherwise we have to find the inode
1869 xop = hammer2_xop_alloc(pmp->iroot, 0);
1871 hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1872 error = hammer2_xop_collect(&xop->head, 0);
1875 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1876 kprintf("vget: no collect error but also no rdata\n");
1877 kprintf("xop %p\n", xop);
1878 while ((hammer2_debug & 0x80000) == 0) {
1879 tsleep(xop, PCATCH, "wait", hz * 10);
1883 ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1886 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1889 *vpp = hammer2_igetv(ip, &error);
1890 hammer2_inode_unlock(ip);
1900 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1907 if (pmp->iroot == NULL) {
1913 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1915 while (pmp->inode_tid == 0) {
1916 hammer2_xop_ipcluster_t *xop;
1917 hammer2_inode_meta_t *meta;
1919 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1920 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1921 error = hammer2_xop_collect(&xop->head, 0);
1924 meta = &xop->head.cluster.focus->data->ipdata.meta;
1925 pmp->iroot->meta = *meta;
1926 pmp->inode_tid = meta->pfs_inum + 1;
1927 if (pmp->inode_tid < HAMMER2_INODE_START)
1928 pmp->inode_tid = HAMMER2_INODE_START;
1930 xop->head.cluster.focus->bref.modify_tid + 1;
1932 kprintf("PFS: Starting inode %jd\n",
1933 (intmax_t)pmp->inode_tid);
1934 kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1935 pmp->inode_tid, pmp->modify_tid);
1937 wakeup(&pmp->iroot);
1939 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1942 * Prime the mount info.
1944 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1951 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1952 hammer2_inode_unlock(pmp->iroot);
1953 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1954 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1960 hammer2_inode_unlock(pmp->iroot);
1963 vp = hammer2_igetv(pmp->iroot, &error);
1964 hammer2_inode_unlock(pmp->iroot);
1974 * XXX incorporate ipdata->meta.inode_quota and data_quota
1978 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1982 hammer2_blockref_t bref;
1987 * NOTE: iroot might not have validated the cluster yet.
1991 bzero(&tmp, sizeof(tmp));
1993 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1994 hmp = pmp->pfs_hmps[i];
1997 if (pmp->iroot->cluster.array[i].chain)
1998 bref = pmp->iroot->cluster.array[i].chain->bref;
2000 bzero(&bref, sizeof(bref));
2002 tmp.f_files = bref.embed.stats.inode_count;
2004 tmp.f_blocks = hmp->voldata.allocator_size /
2005 mp->mnt_vstat.f_bsize;
2006 tmp.f_bfree = hmp->voldata.allocator_free /
2007 mp->mnt_vstat.f_bsize;
2008 tmp.f_bavail = tmp.f_bfree;
2010 if (cred && cred->cr_uid != 0) {
2014 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2015 tmp.f_blocks -= adj;
2017 tmp.f_bavail -= adj;
2020 mp->mnt_stat.f_blocks = tmp.f_blocks;
2021 mp->mnt_stat.f_bfree = tmp.f_bfree;
2022 mp->mnt_stat.f_bavail = tmp.f_bavail;
2023 mp->mnt_stat.f_files = tmp.f_files;
2024 mp->mnt_stat.f_ffree = tmp.f_ffree;
2026 *sbp = mp->mnt_stat;
2033 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2037 hammer2_blockref_t bref;
2042 * NOTE: iroot might not have validated the cluster yet.
2045 bzero(&tmp, sizeof(tmp));
2047 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2048 hmp = pmp->pfs_hmps[i];
2051 if (pmp->iroot->cluster.array[i].chain)
2052 bref = pmp->iroot->cluster.array[i].chain->bref;
2054 bzero(&bref, sizeof(bref));
2056 tmp.f_files = bref.embed.stats.inode_count;
2058 tmp.f_blocks = hmp->voldata.allocator_size /
2059 mp->mnt_vstat.f_bsize;
2060 tmp.f_bfree = hmp->voldata.allocator_free /
2061 mp->mnt_vstat.f_bsize;
2062 tmp.f_bavail = tmp.f_bfree;
2064 if (cred && cred->cr_uid != 0) {
2068 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2069 tmp.f_blocks -= adj;
2071 tmp.f_bavail -= adj;
2074 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2075 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2076 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2077 mp->mnt_vstat.f_files = tmp.f_files;
2078 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2080 *sbp = mp->mnt_vstat;
2086 * Mount-time recovery (RW mounts)
2088 * Updates to the free block table are allowed to lag flushes by one
2089 * transaction. In case of a crash, then on a fresh mount we must do an
2090 * incremental scan of the last committed transaction id and make sure that
2091 * all related blocks have been marked allocated.
2093 * The super-root topology and each PFS has its own transaction id domain,
2094 * so we must track PFS boundary transitions.
2096 struct hammer2_recovery_elm {
2097 TAILQ_ENTRY(hammer2_recovery_elm) entry;
2098 hammer2_chain_t *chain;
2099 hammer2_tid_t sync_tid;
2102 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2104 struct hammer2_recovery_info {
2105 struct hammer2_recovery_list list;
2110 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2111 hammer2_chain_t *parent,
2112 struct hammer2_recovery_info *info,
2113 hammer2_tid_t sync_tid);
2115 #define HAMMER2_RECOVERY_MAXDEPTH 10
2119 hammer2_recovery(hammer2_dev_t *hmp)
2121 struct hammer2_recovery_info info;
2122 struct hammer2_recovery_elm *elm;
2123 hammer2_chain_t *parent;
2124 hammer2_tid_t sync_tid;
2125 hammer2_tid_t mirror_tid;
2128 hammer2_trans_init(hmp->spmp, 0);
2130 sync_tid = hmp->voldata.freemap_tid;
2131 mirror_tid = hmp->voldata.mirror_tid;
2133 kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2134 if (sync_tid >= mirror_tid) {
2135 kprintf(" no recovery needed\n");
2137 kprintf(" freemap recovery %016jx-%016jx\n",
2138 sync_tid + 1, mirror_tid);
2141 TAILQ_INIT(&info.list);
2143 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2144 error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2145 hammer2_chain_lookup_done(parent);
2147 while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2148 TAILQ_REMOVE(&info.list, elm, entry);
2149 parent = elm->chain;
2150 sync_tid = elm->sync_tid;
2151 kfree(elm, M_HAMMER2);
2153 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2154 error |= hammer2_recovery_scan(hmp, parent, &info,
2155 hmp->voldata.freemap_tid);
2156 hammer2_chain_unlock(parent);
2157 hammer2_chain_drop(parent); /* drop elm->chain ref */
2160 hammer2_trans_done(hmp->spmp);
2167 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2168 struct hammer2_recovery_info *info,
2169 hammer2_tid_t sync_tid)
2171 const hammer2_inode_data_t *ripdata;
2172 hammer2_chain_t *chain;
2173 hammer2_blockref_t bref;
2180 * Adjust freemap to ensure that the block(s) are marked allocated.
2182 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2183 hammer2_freemap_adjust(hmp, &parent->bref,
2184 HAMMER2_FREEMAP_DORECOVER);
2188 * Check type for recursive scan
2190 switch(parent->bref.type) {
2191 case HAMMER2_BREF_TYPE_VOLUME:
2192 /* data already instantiated */
2194 case HAMMER2_BREF_TYPE_INODE:
2196 * Must instantiate data for DIRECTDATA test and also
2199 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2200 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2201 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2202 /* not applicable to recovery scan */
2203 hammer2_chain_unlock(parent);
2206 hammer2_chain_unlock(parent);
2208 case HAMMER2_BREF_TYPE_INDIRECT:
2210 * Must instantiate data for recursion
2212 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2213 hammer2_chain_unlock(parent);
2215 case HAMMER2_BREF_TYPE_DIRENT:
2216 case HAMMER2_BREF_TYPE_DATA:
2217 case HAMMER2_BREF_TYPE_FREEMAP:
2218 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2219 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2220 /* not applicable to recovery scan */
2224 return HAMMER2_ERROR_BADBREF;
2228 * Defer operation if depth limit reached or if we are crossing a
2231 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2232 struct hammer2_recovery_elm *elm;
2234 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2235 elm->chain = parent;
2236 elm->sync_tid = sync_tid;
2237 hammer2_chain_ref(parent);
2238 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2239 /* unlocked by caller */
2246 * Recursive scan of the last flushed transaction only. We are
2247 * doing this without pmp assignments so don't leave the chains
2248 * hanging around after we are done with them.
2250 * error Cumulative error this level only
2251 * rup_error Cumulative error for recursion
2252 * tmp_error Specific non-cumulative recursion error
2260 error |= hammer2_chain_scan(parent, &chain, &bref,
2262 HAMMER2_LOOKUP_NODATA);
2265 * Problem during scan or EOF
2273 if (chain == NULL) {
2274 if (bref.mirror_tid > sync_tid) {
2275 hammer2_freemap_adjust(hmp, &bref,
2276 HAMMER2_FREEMAP_DORECOVER);
2282 * This may or may not be a recursive node.
2284 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2285 if (bref.mirror_tid > sync_tid) {
2287 tmp_error = hammer2_recovery_scan(hmp, chain,
2295 * Flush the recovery at the PFS boundary to stage it for
2296 * the final flush of the super-root topology.
2298 if (tmp_error == 0 &&
2299 (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2300 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2301 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2304 rup_error |= tmp_error;
2306 return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2310 * This fixes up an error introduced in earlier H2 implementations where
2311 * moving a PFS inode into an indirect block wound up causing the
2312 * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2316 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2318 const hammer2_inode_data_t *ripdata;
2319 hammer2_chain_t *parent;
2320 hammer2_chain_t *chain;
2321 hammer2_key_t key_next;
2322 hammer2_pfs_t *spmp;
2328 * Lookup mount point under the media-localized super-root.
2330 * cluster->pmp will incorrectly point to spmp and must be fixed
2334 hammer2_inode_lock(spmp->iroot, 0);
2335 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2336 chain = hammer2_chain_lookup(&parent, &key_next,
2337 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2340 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2343 kprintf("I/O error scanning PFS labels\n");
2344 error |= chain->error;
2345 } else if ((chain->bref.flags &
2346 HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2349 ripdata = &chain->data->ipdata;
2350 hammer2_trans_init(hmp->spmp, 0);
2351 error2 = hammer2_chain_modify(chain,
2352 chain->bref.modify_tid,
2355 kprintf("hammer2: Correct mis-flagged PFS %s\n",
2357 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2361 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2363 hammer2_trans_done(hmp->spmp);
2365 chain = hammer2_chain_next(&parent, chain, &key_next,
2366 key_next, HAMMER2_KEY_MAX,
2370 hammer2_chain_unlock(parent);
2371 hammer2_chain_drop(parent);
2373 hammer2_inode_unlock(spmp->iroot);
2379 * Sync a mount point; this is called periodically on a per-mount basis from
2380 * the filesystem syncer, and whenever a user issues a sync.
2383 hammer2_vfs_sync(struct mount *mp, int waitfor)
2385 hammer2_xop_flush_t *xop;
2386 struct hammer2_sync_info info;
2387 hammer2_inode_t *iroot;
2395 KKASSERT(iroot->pmp == pmp);
2398 * We can't acquire locks on existing vnodes while in a transaction
2399 * without risking a deadlock. This assumes that vfsync() can be
2400 * called without the vnode locked (which it can in DragonFly).
2401 * Otherwise we'd have to implement a multi-pass or flag the lock
2402 * failures and retry.
2404 * The reclamation code interlocks with the sync list's token
2405 * (by removing the vnode from the scan list) before unlocking
2406 * the inode, giving us time to ref the inode.
2408 /*flags = VMSC_GETVP;*/
2410 if (waitfor & MNT_LAZY)
2411 flags |= VMSC_ONEPASS;
2414 * Flush vnodes individually using a normal transaction to avoid
2415 * stalling any concurrent operations. This will flush the related
2416 * buffer cache buffers and inodes to the media.
2418 * For efficiency do an async pass before making sure with a
2419 * synchronous pass on all related buffer cache buffers.
2421 hammer2_trans_init(pmp, 0);
2425 info.waitfor = MNT_NOWAIT;
2427 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2430 * Now do two passes making sure we get everything. The first pass
2431 * vfsync()s dirty vnodes. The second pass waits for their I/O's
2432 * to finish and cleans up the dirty flag on the vnode.
2435 info.waitfor = MNT_WAIT;
2436 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2439 info.waitfor = MNT_WAIT;
2440 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2443 * We must also run the sideq to handle any disconnected inodes
2444 * as the vnode scan will not see these.
2446 hammer2_inode_run_sideq(pmp, 1);
2447 hammer2_trans_done(pmp);
2450 * Start our flush transaction and flush the root topology down to
2451 * the inodes, but not the inodes themselves (which we already flushed
2452 * above). Any concurrent activity effecting inode contents will not
2454 * The flush sequence will
2456 * NOTE! It is still possible for the paging code to push pages
2457 * out via a UIO_NOCOPY hammer2_vop_write() during the main
2460 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2463 * sync dirty vnodes again while in the flush transaction. This is
2464 * currently an expensive shim to makre sure the logical topology is
2465 * completely consistent before we flush the volume header.
2468 info.waitfor = MNT_WAIT;
2469 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2472 info.waitfor = MNT_WAIT;
2473 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2476 * Use the XOP interface to concurrently flush all nodes to
2477 * synchronize the PFSROOT subtopology to the media. A standard
2478 * end-of-scan ENOENT error indicates cluster sufficiency.
2480 * Note that this flush will not be visible on crash recovery until
2481 * we flush the super-root topology in the next loop.
2483 * XXX For now wait for all flushes to complete.
2487 * If unmounting try to flush everything including any
2488 * sub-trees under inodes, just in case there is dangling
2489 * modified data, as a safety. Otherwise just flush up to
2490 * the inodes in this stage.
2492 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2493 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2494 HAMMER2_XOP_VOLHDR);
2496 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2497 HAMMER2_XOP_INODE_STOP |
2498 HAMMER2_XOP_VOLHDR);
2500 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2501 error = hammer2_xop_collect(&xop->head,
2502 HAMMER2_XOP_COLLECT_WAITALL);
2503 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2504 if (error == HAMMER2_ERROR_ENOENT)
2507 error = hammer2_error_to_errno(error);
2511 hammer2_trans_done(pmp);
2519 * Note that we ignore the tranasction mtid we got above. Instead,
2520 * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2523 * WARNING! The frontend might be waiting on chnmem (limit_dirty_chains)
2524 * while holding a vnode locked. When this situation occurs we cannot
2525 * safely test whether it is ok to clear the dirty bit on the vnode.
2526 * However, we can still flush the inode's topology.
2529 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2531 struct hammer2_sync_info *info = data;
2532 hammer2_inode_t *ip;
2536 * Degenerate cases. Note that ip == NULL typically means the
2537 * syncer vnode itself and we don't want to vclrisdirty() in that
2544 if (vp->v_type == VNON || vp->v_type == VBAD) {
2550 * Synchronize the buffer cche and inode meta-data to the backing
2553 * vfsync is not necessarily synchronous, so it is best NOT to try
2554 * to flush the backing topology to media at this point.
2556 hammer2_inode_ref(ip);
2557 if ((ip->flags & (HAMMER2_INODE_RESIZED|HAMMER2_INODE_MODIFIED)) ||
2558 !RB_EMPTY(&vp->v_rbdirty_tree)) {
2559 if (info->pass == 1)
2560 vfsync(vp, info->waitfor, 1, NULL, NULL);
2562 bio_track_wait(&vp->v_track_write, 0, 0);
2564 if (info->pass == 2 && (vp->v_flag & VISDIRTY)) {
2566 * v_token is needed to interlock v_rbdirty_tree.
2568 lwkt_gettoken(&vp->v_token);
2569 hammer2_inode_lock(ip, 0);
2570 hammer2_inode_chain_sync(ip);
2571 hammer2_inode_chain_flush(ip);
2572 if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2573 HAMMER2_INODE_RESIZED |
2574 HAMMER2_INODE_DIRTYDATA)) == 0 &&
2575 RB_EMPTY(&vp->v_rbdirty_tree) &&
2576 !bio_track_active(&vp->v_track_write)) {
2579 hammer2_inode_unlock(ip);
2580 lwkt_reltoken(&vp->v_token);
2582 hammer2_inode_drop(ip);
2586 info->error = error;
2593 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2595 hammer2_inode_t *ip;
2597 KKASSERT(MAXFIDSZ >= 16);
2599 fhp->fid_len = offsetof(struct fid, fid_data[16]);
2601 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2602 ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2609 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2610 struct fid *fhp, struct vnode **vpp)
2617 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2620 error = hammer2_vfs_root(mp, vpp);
2622 error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2627 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2633 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2634 int *exflagsp, struct ucred **credanonp)
2641 np = vfs_export_lookup(mp, &pmp->export, nam);
2643 *exflagsp = np->netc_exflags;
2644 *credanonp = &np->netc_anon;
2653 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume
2654 * header into the HMP
2656 * XXX read four volhdrs and use the one with the highest TID whos CRC
2661 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2662 * nonexistant locations.
2664 * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2668 hammer2_install_volume_header(hammer2_dev_t *hmp)
2670 hammer2_volume_data_t *vd;
2672 hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2684 * There are up to 4 copies of the volume header (syncs iterate
2685 * between them so there is no single master). We don't trust the
2686 * volu_size field so we don't know precisely how large the filesystem
2687 * is, so depend on the OS to return an error if we go beyond the
2688 * block device's EOF.
2690 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2691 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2692 HAMMER2_VOLUME_BYTES, &bp);
2699 vd = (struct hammer2_volume_data *) bp->b_data;
2700 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2701 (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2707 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2708 /* XXX: Reversed-endianness filesystem */
2709 kprintf("hammer2: reverse-endian filesystem detected");
2715 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2716 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2717 HAMMER2_VOLUME_ICRC0_SIZE);
2718 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2719 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2720 HAMMER2_VOLUME_ICRC1_SIZE);
2721 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2722 kprintf("hammer2 volume header crc "
2723 "mismatch copy #%d %08x/%08x\n",
2730 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2739 hmp->volsync = hmp->voldata;
2740 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2742 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2743 kprintf("hammer2: using volume header #%d\n",
2748 kprintf("hammer2: no valid volume headers found!\n");
2754 * This handles hysteresis on regular file flushes. Because the BIOs are
2755 * routed to a thread it is possible for an excessive number to build up
2756 * and cause long front-end stalls long before the runningbuffspace limit
2757 * is hit, so we implement hammer2_flush_pipe to control the
2760 * This is a particular problem when compression is used.
2763 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2765 atomic_add_int(&pmp->count_lwinprog, 1);
2769 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2773 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2774 if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2775 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2776 atomic_clear_int(&pmp->count_lwinprog,
2777 HAMMER2_LWINPROG_WAITING);
2778 wakeup(&pmp->count_lwinprog);
2780 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2781 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2782 atomic_clear_int(&pmp->count_lwinprog,
2783 HAMMER2_LWINPROG_WAITING0);
2784 wakeup(&pmp->count_lwinprog);
2789 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2792 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2793 HAMMER2_LWINPROG_WAITING0;
2796 lwinprog = pmp->count_lwinprog;
2798 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2800 tsleep_interlock(&pmp->count_lwinprog, 0);
2801 atomic_set_int(&pmp->count_lwinprog, lwflag);
2802 lwinprog = pmp->count_lwinprog;
2803 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2805 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2810 * Attempt to proactively fsync dirty vnodes if we have too many. This
2811 * solves an issue where the kernel syncer thread can get seriously behind
2812 * when multiple user processes/threads are furiously modifying inodes.
2813 * This situation can occur on slow storage and is only limited by
2814 * kern.maxvnodes without the moderation code below. It is made worse
2815 * when the device buffers underlying the modified inodes (which are clean)
2816 * get evicted before the flush can occur, forcing a re-read.
2818 * We do not want sysads to feel that they have to torpedo kern.maxvnodes
2819 * to solve this problem, so we implement vfs.hammer2.limit_dirty_inodes
2820 * (per-mount-basis) and default it to something reasonable.
2823 hammer2_pfs_moderate(hammer2_inode_t *ip, int always_moderate)
2825 hammer2_pfs_t *pmp = ip->pmp;
2826 struct mount *mp = pmp->mp;
2828 if (mp && vn_syncer_count(mp) > hammer2_limit_dirty_inodes) {
2834 * Manage excessive memory resource use for chain and related
2837 * Called without any inode locks or transaction locks. VNodes
2838 * might be locked by the kernel in the call stack.
2841 hammer2_pfs_memory_wait(hammer2_inode_t *ip, int always_moderate)
2843 hammer2_pfs_t *pmp = ip->pmp;
2852 * Moderate the number of dirty inodes
2854 hammer2_pfs_moderate(ip, always_moderate);
2857 * Atomic check condition and wait. Also do an early speedup of
2858 * the syncer to try to avoid hitting the wait.
2861 waiting = pmp->inmem_dirty_chains;
2863 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2865 limit = pmp->mp->mnt_nvnodelistsize / 10;
2866 if (limit < hammer2_limit_dirty_chains)
2867 limit = hammer2_limit_dirty_chains;
2872 if ((int)(ticks - zzticks) > hz) {
2874 kprintf("count %ld %ld\n", count, limit);
2879 * Block if there are too many dirty chains present, wait
2880 * for the flush to clean some out.
2882 if (count > limit) {
2883 hammer2_pfs_moderate(ip, always_moderate);
2884 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2885 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2887 waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2888 if (ticks != pmp->speedup_ticks) {
2889 pmp->speedup_ticks = ticks;
2890 speedup_syncer(pmp->mp);
2892 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2895 continue; /* loop on success or fail */
2899 * Try to start an early flush before we are forced to block.
2901 if (count > limit * 5 / 10 &&
2902 ticks != pmp->speedup_ticks) {
2903 pmp->speedup_ticks = ticks;
2904 speedup_syncer(pmp->mp);
2911 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2914 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2919 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2924 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
2925 /* don't need --waiting to test flag */
2926 if (waiting & HAMMER2_DIRTYCHAIN_WAITING) {
2927 atomic_clear_int(&pmp->inmem_dirty_chains,
2928 HAMMER2_DIRTYCHAIN_WAITING);
2929 wakeup(&pmp->inmem_dirty_chains);
2935 * Returns 0 if the filesystem has tons of free space
2936 * Returns 1 if the filesystem has less than 10% remaining
2937 * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2940 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2944 hammer2_off_t free_reserved;
2945 hammer2_off_t free_nominal;
2950 if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2951 free_reserved = HAMMER2_SEGSIZE;
2952 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2953 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2954 hmp = pmp->pfs_hmps[i];
2957 if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2958 pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2961 if (free_nominal > hmp->voldata.allocator_free)
2962 free_nominal = hmp->voldata.allocator_free;
2963 if (free_reserved < hmp->free_reserved)
2964 free_reserved = hmp->free_reserved;
2970 pmp->free_reserved = free_reserved;
2971 pmp->free_nominal = free_nominal;
2972 pmp->free_ticks = ticks;
2974 free_reserved = pmp->free_reserved;
2975 free_nominal = pmp->free_nominal;
2977 if (cred && cred->cr_uid != 0) {
2978 if ((int64_t)(free_nominal - bytes) <
2979 (int64_t)free_reserved) {
2983 if ((int64_t)(free_nominal - bytes) <
2984 (int64_t)free_reserved / 2) {
2988 if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2997 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
3000 hammer2_chain_t *scan;
3001 hammer2_chain_t *parent;
3005 kprintf("%*.*s...\n", tab, tab, "");
3010 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
3012 chain, chain->bref.type,
3013 chain->bref.key, chain->bref.keybits,
3014 chain->bref.mirror_tid);
3016 kprintf("%*.*s [%08x] (%s) refs=%d",
3019 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3020 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
3023 parent = chain->parent;
3025 kprintf("\n%*.*s p=%p [pflags %08x prefs %d",
3027 parent, parent->flags, parent->refs);
3028 if (RB_EMPTY(&chain->core.rbtree)) {
3032 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3033 if ((scan->flags & flags) || flags == (u_int)-1) {
3034 hammer2_dump_chain(scan, tab + 4, countp, 'a',
3038 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3039 kprintf("%*.*s}(%s)\n", tab, tab, "",
3040 chain->data->ipdata.filename);
3042 kprintf("%*.*s}\n", tab, tab, "");