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, 0);
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 const hammer2_inode_data_t *ripdata;
913 hammer2_blockref_t bref;
915 char devstr[MNAMELEN];
934 bzero(&info, sizeof(info));
935 info.cluster_fd = -1;
936 ksnprintf(devstr, sizeof(devstr), "%s",
937 mp->mnt_stat.f_mntfromname);
938 kprintf("hammer2_mount: root '%s'\n", devstr);
941 * Non-root mount or updating a mount
943 error = copyin(data, &info, sizeof(info));
947 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
950 kprintf("hammer2_mount: '%s'\n", devstr);
954 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
955 * if no label specified, based on the partition id. Error out if no
956 * label or device (with partition id) is specified. This is strictly
957 * a convenience to match the default label created by newfs_hammer2,
958 * our preference is that a label always be specified.
960 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
961 * that does not specify a device, as long as some H2 label
962 * has already been mounted from that device. This makes
963 * mounting snapshots a lot easier.
966 label = strchr(devstr, '@');
967 if (label && ((label + 1) - dev) > done) {
968 kprintf("hammer2: mount: bad label %s/%zd\n",
972 if (label == NULL || label[1] == 0) {
976 label = devstr + strlen(devstr);
978 *label = '\0'; /* clean up trailing @ */
997 kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
998 dev, label, (mp->mnt_flag & MNT_RDONLY));
1000 if (mp->mnt_flag & MNT_UPDATE) {
1002 * Update mount. Note that pmp->iroot->cluster is
1003 * an inode-embedded cluster and thus cannot be
1006 * XXX HAMMER2 needs to implement NFS export via
1009 hammer2_cluster_t *cluster;
1012 pmp->hflags = info.hflags;
1013 cluster = &pmp->iroot->cluster;
1014 for (i = 0; i < cluster->nchains; ++i) {
1015 if (cluster->array[i].chain == NULL)
1017 hmp = cluster->array[i].chain->hmp;
1019 error = hammer2_remount(hmp, mp, path,
1031 * If a path is specified and dev is not an empty string, lookup the
1032 * name and verify that it referes to a block device.
1034 * If a path is specified and dev is an empty string we fall through
1035 * and locate the label in the hmp search.
1037 if (path && *dev != 0) {
1038 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1040 error = nlookup(&nd);
1042 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1044 } else if (path == NULL) {
1046 cdev_t cdev = kgetdiskbyname(dev);
1047 error = bdevvp(cdev, &devvp);
1049 kprintf("hammer2: cannot find '%s'\n", dev);
1052 * We will locate the hmp using the label in the hmp loop.
1058 * Make sure its a block device. Do not check to see if it is
1059 * already mounted until we determine that its a fresh H2 device.
1061 if (error == 0 && devvp) {
1062 vn_isdisk(devvp, &error);
1066 * Determine if the device has already been mounted. After this
1067 * check hmp will be non-NULL if we are doing the second or more
1068 * hammer2 mounts from the same device.
1070 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1073 * Match the device. Due to the way devfs works,
1074 * we may not be able to directly match the vnode pointer,
1075 * so also check to see if the underlying device matches.
1077 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1078 if (hmp->devvp == devvp)
1080 if (devvp->v_rdev &&
1081 hmp->devvp->v_rdev == devvp->v_rdev) {
1087 * If no match this may be a fresh H2 mount, make sure
1088 * the device is not mounted on anything else.
1091 error = vfs_mountedon(devvp);
1092 } else if (error == 0) {
1094 * Match the label to a pmp already probed.
1096 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1097 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1098 if (pmp->pfs_names[i] &&
1099 strcmp(pmp->pfs_names[i], label) == 0) {
1100 hmp = pmp->pfs_hmps[i];
1112 * Open the device if this isn't a secondary mount and construct
1113 * the H2 device mount (hmp).
1116 hammer2_chain_t *schain;
1118 hammer2_xop_head_t xop;
1120 if (error == 0 && vcount(devvp) > 0) {
1121 kprintf("Primary device already has references\n");
1126 * Now open the device
1129 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1130 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1131 error = vinvalbuf(devvp, V_SAVE, 0, 0);
1133 error = VOP_OPEN(devvp,
1134 (ronly ? FREAD : FREAD | FWRITE),
1139 if (error && devvp) {
1144 lockmgr(&hammer2_mntlk, LK_RELEASE);
1147 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1148 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1151 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1152 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1153 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1154 RB_INIT(&hmp->iotree);
1155 spin_init(&hmp->io_spin, "hm2mount_io");
1156 spin_init(&hmp->list_spin, "hm2mount_list");
1157 TAILQ_INIT(&hmp->flushq);
1159 lockinit(&hmp->vollk, "h2vol", 0, 0);
1160 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1161 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1164 * vchain setup. vchain.data is embedded.
1165 * vchain.refs is initialized and will never drop to 0.
1167 * NOTE! voldata is not yet loaded.
1169 hmp->vchain.hmp = hmp;
1170 hmp->vchain.refs = 1;
1171 hmp->vchain.data = (void *)&hmp->voldata;
1172 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1173 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1174 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1176 hammer2_chain_core_init(&hmp->vchain);
1177 /* hmp->vchain.u.xxx is left NULL */
1180 * fchain setup. fchain.data is embedded.
1181 * fchain.refs is initialized and will never drop to 0.
1183 * The data is not used but needs to be initialized to
1184 * pass assertion muster. We use this chain primarily
1185 * as a placeholder for the freemap's top-level RBTREE
1186 * so it does not interfere with the volume's topology
1189 hmp->fchain.hmp = hmp;
1190 hmp->fchain.refs = 1;
1191 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1192 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1193 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1194 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1195 hmp->fchain.bref.methods =
1196 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1197 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1199 hammer2_chain_core_init(&hmp->fchain);
1200 /* hmp->fchain.u.xxx is left NULL */
1203 * Install the volume header and initialize fields from
1206 error = hammer2_install_volume_header(hmp);
1208 hammer2_unmount_helper(mp, NULL, hmp);
1209 lockmgr(&hammer2_mntlk, LK_RELEASE);
1210 hammer2_vfs_unmount(mp, MNT_FORCE);
1215 * Really important to get these right or flush will get
1218 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1222 * Dummy-up vchain and fchain's modify_tid. mirror_tid
1223 * is inherited from the volume header.
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;
1234 * First locate the super-root inode, which is key 0
1235 * relative to the volume header's blockset.
1237 * Then locate the root inode by scanning the directory keyspace
1238 * represented by the label.
1240 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1241 schain = hammer2_chain_lookup(&parent, &key_dummy,
1242 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
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);
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);
1258 hammer2_unmount_helper(mp, NULL, hmp);
1259 lockmgr(&hammer2_mntlk, LK_RELEASE);
1260 hammer2_vfs_unmount(mp, MNT_FORCE);
1265 * The super-root always uses an inode_tid of 1 when
1268 spmp->inode_tid = 1;
1269 spmp->modify_tid = schain->bref.modify_tid + 1;
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).
1276 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1277 KKASSERT(schain->pmp == NULL);
1278 spmp->pfs_clid = ripdata->meta.pfs_clid;
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.
1285 * The returned inode is locked with the supplied cluster.
1287 hammer2_dummy_xop_from_chain(&xop, schain);
1288 hammer2_inode_drop(spmp->iroot);
1290 spmp->iroot = hammer2_inode_get(spmp, NULL, &xop, -1);
1291 spmp->spmp_hmp = hmp;
1292 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1293 spmp->pfs_hmps[0] = hmp;
1294 hammer2_inode_ref(spmp->iroot);
1295 hammer2_inode_unlock(spmp->iroot);
1296 hammer2_cluster_unlock(&xop.cluster);
1297 hammer2_chain_drop(schain);
1298 /* do not call hammer2_cluster_drop() on an embedded cluster */
1299 schain = NULL; /* now invalid */
1300 /* leave spmp->iroot with one ref */
1302 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1303 error = hammer2_recovery(hmp);
1305 error |= hammer2_fixup_pfses(hmp);
1306 /* XXX do something with error */
1308 hammer2_update_pmps(hmp);
1309 hammer2_iocom_init(hmp);
1310 hammer2_bulkfree_init(hmp);
1313 * Ref the cluster management messaging descriptor. The mount
1314 * program deals with the other end of the communications pipe.
1316 * Root mounts typically do not supply one.
1318 if (info.cluster_fd >= 0) {
1319 fp = holdfp(curthread, info.cluster_fd, -1);
1321 hammer2_cluster_reconnect(hmp, fp);
1323 kprintf("hammer2_mount: bad cluster_fd!\n");
1328 if (info.hflags & HMNT2_DEVFLAGS) {
1329 kprintf("hammer2: Warning: mount flags pertaining "
1330 "to the whole device may only be specified "
1331 "on the first mount of the device: %08x\n",
1332 info.hflags & HMNT2_DEVFLAGS);
1337 * Force local mount (disassociate all PFSs from their clusters).
1338 * Used primarily for debugging.
1340 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1343 * Lookup the mount point under the media-localized super-root.
1344 * Scanning hammer2_pfslist doesn't help us because it represents
1345 * PFS cluster ids which can aggregate several named PFSs together.
1347 * cluster->pmp will incorrectly point to spmp and must be fixed
1350 hammer2_inode_lock(spmp->iroot, 0);
1351 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1352 lhc = hammer2_dirhash(label, strlen(label));
1353 chain = hammer2_chain_lookup(&parent, &key_next,
1354 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1357 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1358 strcmp(label, chain->data->ipdata.filename) == 0) {
1361 chain = hammer2_chain_next(&parent, chain, &key_next,
1363 lhc + HAMMER2_DIRHASH_LOMASK,
1367 hammer2_chain_unlock(parent);
1368 hammer2_chain_drop(parent);
1370 hammer2_inode_unlock(spmp->iroot);
1373 * PFS could not be found?
1375 if (chain == NULL) {
1377 kprintf("hammer2_mount: PFS label I/O error\n");
1379 kprintf("hammer2_mount: PFS label not found\n");
1380 hammer2_unmount_helper(mp, NULL, hmp);
1381 lockmgr(&hammer2_mntlk, LK_RELEASE);
1382 hammer2_vfs_unmount(mp, MNT_FORCE);
1388 * Acquire the pmp structure (it should have already been allocated
1389 * via hammer2_update_pmps() so do not pass cluster in to add to
1390 * available chains).
1392 * Check if the cluster has already been mounted. A cluster can
1393 * only be mounted once, use null mounts to mount additional copies.
1396 kprintf("hammer2_mount: PFS label I/O error\n");
1398 ripdata = &chain->data->ipdata;
1400 pmp = hammer2_pfsalloc(NULL, ripdata,
1401 bref.modify_tid, force_local);
1403 hammer2_chain_unlock(chain);
1404 hammer2_chain_drop(chain);
1409 kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1412 kprintf("hammer2_mount: PFS already mounted!\n");
1413 hammer2_unmount_helper(mp, NULL, hmp);
1414 lockmgr(&hammer2_mntlk, LK_RELEASE);
1415 hammer2_vfs_unmount(mp, MNT_FORCE);
1420 pmp->hflags = info.hflags;
1421 mp->mnt_flag |= MNT_LOCAL;
1422 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */
1423 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */
1426 * required mount structure initializations
1428 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1429 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1431 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1432 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1437 mp->mnt_iosize_max = MAXPHYS;
1440 * Connect up mount pointers.
1442 hammer2_mount_helper(mp, pmp);
1444 lockmgr(&hammer2_mntlk, LK_RELEASE);
1450 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1451 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1452 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1455 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1456 MNAMELEN - 1, &size);
1457 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1458 } /* else root mount, already in there */
1460 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1462 copyinstr(path, mp->mnt_stat.f_mntonname,
1463 sizeof(mp->mnt_stat.f_mntonname) - 1,
1467 mp->mnt_stat.f_mntonname[0] = '/';
1471 * Initial statfs to prime mnt_stat.
1473 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1479 * Scan PFSs under the super-root and create hammer2_pfs structures.
1483 hammer2_update_pmps(hammer2_dev_t *hmp)
1485 const hammer2_inode_data_t *ripdata;
1486 hammer2_chain_t *parent;
1487 hammer2_chain_t *chain;
1488 hammer2_blockref_t bref;
1489 hammer2_dev_t *force_local;
1490 hammer2_pfs_t *spmp;
1492 hammer2_key_t key_next;
1496 * Force local mount (disassociate all PFSs from their clusters).
1497 * Used primarily for debugging.
1499 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1502 * Lookup mount point under the media-localized super-root.
1504 * cluster->pmp will incorrectly point to spmp and must be fixed
1508 hammer2_inode_lock(spmp->iroot, 0);
1509 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1510 chain = hammer2_chain_lookup(&parent, &key_next,
1511 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1514 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1517 kprintf("I/O error scanning PFS labels\n");
1519 ripdata = &chain->data->ipdata;
1522 pmp = hammer2_pfsalloc(chain, ripdata,
1523 bref.modify_tid, force_local);
1525 chain = hammer2_chain_next(&parent, chain, &key_next,
1526 key_next, HAMMER2_KEY_MAX,
1530 hammer2_chain_unlock(parent);
1531 hammer2_chain_drop(parent);
1533 hammer2_inode_unlock(spmp->iroot);
1538 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1539 struct vnode *devvp, struct ucred *cred)
1543 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1544 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1545 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1547 error = hammer2_recovery(hmp);
1549 error |= hammer2_fixup_pfses(hmp);
1550 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1552 VOP_CLOSE(devvp, FREAD, NULL);
1555 VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1566 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1577 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1580 * If mount initialization proceeded far enough we must flush
1581 * its vnodes and sync the underlying mount points. Three syncs
1582 * are required to fully flush the filesystem (freemap updates lag
1583 * by one flush, and one extra for safety).
1585 if (mntflags & MNT_FORCE)
1590 error = vflush(mp, 0, flags);
1593 hammer2_vfs_sync(mp, MNT_WAIT);
1594 hammer2_vfs_sync(mp, MNT_WAIT);
1595 hammer2_vfs_sync(mp, MNT_WAIT);
1599 * Cleanup the frontend support XOPS threads
1601 hammer2_xop_helper_cleanup(pmp);
1604 hammer2_unmount_helper(mp, pmp, NULL);
1608 lockmgr(&hammer2_mntlk, LK_RELEASE);
1614 * Mount helper, hook the system mount into our PFS.
1615 * The mount lock is held.
1617 * We must bump the mount_count on related devices for any
1622 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1624 hammer2_cluster_t *cluster;
1625 hammer2_chain_t *rchain;
1628 mp->mnt_data = (qaddr_t)pmp;
1632 * After pmp->mp is set we have to adjust hmp->mount_count.
1634 cluster = &pmp->iroot->cluster;
1635 for (i = 0; i < cluster->nchains; ++i) {
1636 rchain = cluster->array[i].chain;
1639 ++rchain->hmp->mount_count;
1643 * Create missing Xop threads
1645 hammer2_xop_helper_create(pmp);
1649 * Mount helper, unhook the system mount from our PFS.
1650 * The mount lock is held.
1652 * If hmp is supplied a mount responsible for being the first to open
1653 * the block device failed and the block device and all PFSs using the
1654 * block device must be cleaned up.
1656 * If pmp is supplied multiple devices might be backing the PFS and each
1657 * must be disconnected. This might not be the last PFS using some of the
1658 * underlying devices. Also, we have to adjust our hmp->mount_count
1659 * accounting for the devices backing the pmp which is now undergoing an
1664 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1666 hammer2_cluster_t *cluster;
1667 hammer2_chain_t *rchain;
1668 struct vnode *devvp;
1674 * If no device supplied this is a high-level unmount and we have to
1675 * to disconnect the mount, adjust mount_count, and locate devices
1676 * that might now have no mounts.
1679 KKASSERT(hmp == NULL);
1680 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1682 mp->mnt_data = NULL;
1685 * After pmp->mp is cleared we have to account for
1688 cluster = &pmp->iroot->cluster;
1689 for (i = 0; i < cluster->nchains; ++i) {
1690 rchain = cluster->array[i].chain;
1693 --rchain->hmp->mount_count;
1694 /* scrapping hmp now may invalidate the pmp */
1697 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1698 if (hmp->mount_count == 0) {
1699 hammer2_unmount_helper(NULL, NULL, hmp);
1707 * Try to terminate the block device. We can't terminate it if
1708 * there are still PFSs referencing it.
1710 if (hmp->mount_count)
1714 * Decomission the network before we start messing with the
1717 hammer2_iocom_uninit(hmp);
1719 hammer2_bulkfree_uninit(hmp);
1720 hammer2_pfsfree_scan(hmp, 0);
1721 hammer2_dev_exlock(hmp); /* XXX order */
1724 * Cycle the volume data lock as a safety (probably not needed any
1725 * more). To ensure everything is out we need to flush at least
1726 * three times. (1) The running of the sideq can dirty the
1727 * filesystem, (2) A normal flush can dirty the freemap, and
1728 * (3) ensure that the freemap is fully synchronized.
1730 * The next mount's recovery scan can clean everything up but we want
1731 * to leave the filesystem in a 100% clean state on a normal unmount.
1734 hammer2_voldata_lock(hmp);
1735 hammer2_voldata_unlock(hmp);
1739 * Flush whatever is left. Unmounted but modified PFS's might still
1740 * have some dirty chains on them.
1742 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1743 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1745 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1746 hammer2_voldata_modify(hmp);
1747 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1750 hammer2_chain_unlock(&hmp->fchain);
1752 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1753 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1756 hammer2_chain_unlock(&hmp->vchain);
1758 if ((hmp->vchain.flags | hmp->fchain.flags) &
1759 HAMMER2_CHAIN_FLUSH_MASK) {
1760 kprintf("hammer2_unmount: chains left over "
1761 "after final sync\n");
1762 kprintf(" vchain %08x\n", hmp->vchain.flags);
1763 kprintf(" fchain %08x\n", hmp->fchain.flags);
1765 if (hammer2_debug & 0x0010)
1766 Debugger("entered debugger");
1769 hammer2_pfsfree_scan(hmp, 1);
1771 KKASSERT(hmp->spmp == NULL);
1774 * Finish up with the device vnode
1776 if ((devvp = hmp->devvp) != NULL) {
1778 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1779 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1780 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1782 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1783 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1784 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1786 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1793 * Clear vchain/fchain flags that might prevent final cleanup
1796 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1797 atomic_add_long(&hammer2_count_modified_chains, -1);
1798 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1799 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1801 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1802 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1805 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1806 atomic_add_long(&hammer2_count_modified_chains, -1);
1807 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1808 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1810 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1811 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1815 * Final drop of embedded freemap root chain to
1816 * clean up fchain.core (fchain structure is not
1817 * flagged ALLOCATED so it is cleaned out and then
1820 hammer2_chain_drop(&hmp->fchain);
1823 * Final drop of embedded volume root chain to clean
1824 * up vchain.core (vchain structure is not flagged
1825 * ALLOCATED so it is cleaned out and then left to
1829 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1831 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1832 hammer2_dev_unlock(hmp);
1833 hammer2_chain_drop(&hmp->vchain);
1835 hammer2_io_cleanup(hmp, &hmp->iotree);
1836 if (hmp->iofree_count) {
1837 kprintf("io_cleanup: %d I/O's left hanging\n",
1841 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1842 kmalloc_destroy(&hmp->mchain);
1843 kfree(hmp, M_HAMMER2);
1847 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1848 ino_t ino, struct vnode **vpp)
1850 hammer2_xop_lookup_t *xop;
1852 hammer2_inode_t *ip;
1856 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1862 * Easy if we already have it cached
1864 ip = hammer2_inode_lookup(pmp, inum);
1866 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1867 *vpp = hammer2_igetv(ip, &error);
1868 hammer2_inode_unlock(ip);
1869 hammer2_inode_drop(ip); /* from lookup */
1875 * Otherwise we have to find the inode
1877 xop = hammer2_xop_alloc(pmp->iroot, 0);
1879 hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1880 error = hammer2_xop_collect(&xop->head, 0);
1883 ip = hammer2_inode_get(pmp, NULL, &xop->head, -1);
1884 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1887 *vpp = hammer2_igetv(ip, &error);
1888 hammer2_inode_unlock(ip);
1898 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1905 if (pmp->iroot == NULL) {
1906 kprintf("hammer2 (%s): no root inode\n",
1907 mp->mnt_stat.f_mntfromname);
1913 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1915 while (pmp->inode_tid == 0) {
1916 hammer2_xop_ipcluster_t *xop;
1917 const 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 = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1925 pmp->iroot->meta = *meta;
1926 pmp->inode_tid = meta->pfs_inum + 1;
1927 hammer2_xop_pdata(&xop->head);
1930 if (pmp->inode_tid < HAMMER2_INODE_START)
1931 pmp->inode_tid = HAMMER2_INODE_START;
1933 xop->head.cluster.focus->bref.modify_tid + 1;
1935 kprintf("PFS: Starting inode %jd\n",
1936 (intmax_t)pmp->inode_tid);
1937 kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1938 pmp->inode_tid, pmp->modify_tid);
1940 wakeup(&pmp->iroot);
1942 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1945 * Prime the mount info.
1947 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1954 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1955 hammer2_inode_unlock(pmp->iroot);
1956 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1957 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1963 hammer2_inode_unlock(pmp->iroot);
1966 vp = hammer2_igetv(pmp->iroot, &error);
1967 hammer2_inode_unlock(pmp->iroot);
1977 * XXX incorporate ipdata->meta.inode_quota and data_quota
1981 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1985 hammer2_blockref_t bref;
1990 * NOTE: iroot might not have validated the cluster yet.
1994 bzero(&tmp, sizeof(tmp));
1996 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1997 hmp = pmp->pfs_hmps[i];
2000 if (pmp->iroot->cluster.array[i].chain)
2001 bref = pmp->iroot->cluster.array[i].chain->bref;
2003 bzero(&bref, sizeof(bref));
2005 tmp.f_files = bref.embed.stats.inode_count;
2007 tmp.f_blocks = hmp->voldata.allocator_size /
2008 mp->mnt_vstat.f_bsize;
2009 tmp.f_bfree = hmp->voldata.allocator_free /
2010 mp->mnt_vstat.f_bsize;
2011 tmp.f_bavail = tmp.f_bfree;
2013 if (cred && cred->cr_uid != 0) {
2017 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2018 tmp.f_blocks -= adj;
2020 tmp.f_bavail -= adj;
2023 mp->mnt_stat.f_blocks = tmp.f_blocks;
2024 mp->mnt_stat.f_bfree = tmp.f_bfree;
2025 mp->mnt_stat.f_bavail = tmp.f_bavail;
2026 mp->mnt_stat.f_files = tmp.f_files;
2027 mp->mnt_stat.f_ffree = tmp.f_ffree;
2029 *sbp = mp->mnt_stat;
2036 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2040 hammer2_blockref_t bref;
2045 * NOTE: iroot might not have validated the cluster yet.
2048 bzero(&tmp, sizeof(tmp));
2050 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2051 hmp = pmp->pfs_hmps[i];
2054 if (pmp->iroot->cluster.array[i].chain)
2055 bref = pmp->iroot->cluster.array[i].chain->bref;
2057 bzero(&bref, sizeof(bref));
2059 tmp.f_files = bref.embed.stats.inode_count;
2061 tmp.f_blocks = hmp->voldata.allocator_size /
2062 mp->mnt_vstat.f_bsize;
2063 tmp.f_bfree = hmp->voldata.allocator_free /
2064 mp->mnt_vstat.f_bsize;
2065 tmp.f_bavail = tmp.f_bfree;
2067 if (cred && cred->cr_uid != 0) {
2071 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2072 tmp.f_blocks -= adj;
2074 tmp.f_bavail -= adj;
2077 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2078 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2079 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2080 mp->mnt_vstat.f_files = tmp.f_files;
2081 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2083 *sbp = mp->mnt_vstat;
2089 * Mount-time recovery (RW mounts)
2091 * Updates to the free block table are allowed to lag flushes by one
2092 * transaction. In case of a crash, then on a fresh mount we must do an
2093 * incremental scan of the last committed transaction id and make sure that
2094 * all related blocks have been marked allocated.
2096 * The super-root topology and each PFS has its own transaction id domain,
2097 * so we must track PFS boundary transitions.
2099 struct hammer2_recovery_elm {
2100 TAILQ_ENTRY(hammer2_recovery_elm) entry;
2101 hammer2_chain_t *chain;
2102 hammer2_tid_t sync_tid;
2105 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2107 struct hammer2_recovery_info {
2108 struct hammer2_recovery_list list;
2113 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2114 hammer2_chain_t *parent,
2115 struct hammer2_recovery_info *info,
2116 hammer2_tid_t sync_tid);
2118 #define HAMMER2_RECOVERY_MAXDEPTH 10
2122 hammer2_recovery(hammer2_dev_t *hmp)
2124 struct hammer2_recovery_info info;
2125 struct hammer2_recovery_elm *elm;
2126 hammer2_chain_t *parent;
2127 hammer2_tid_t sync_tid;
2128 hammer2_tid_t mirror_tid;
2131 hammer2_trans_init(hmp->spmp, 0);
2133 sync_tid = hmp->voldata.freemap_tid;
2134 mirror_tid = hmp->voldata.mirror_tid;
2136 kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2137 if (sync_tid >= mirror_tid) {
2138 kprintf(" no recovery needed\n");
2140 kprintf(" freemap recovery %016jx-%016jx\n",
2141 sync_tid + 1, mirror_tid);
2144 TAILQ_INIT(&info.list);
2146 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2147 error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2148 hammer2_chain_lookup_done(parent);
2150 while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2151 TAILQ_REMOVE(&info.list, elm, entry);
2152 parent = elm->chain;
2153 sync_tid = elm->sync_tid;
2154 kfree(elm, M_HAMMER2);
2156 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2157 error |= hammer2_recovery_scan(hmp, parent, &info,
2158 hmp->voldata.freemap_tid);
2159 hammer2_chain_unlock(parent);
2160 hammer2_chain_drop(parent); /* drop elm->chain ref */
2163 hammer2_trans_done(hmp->spmp, 0);
2170 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2171 struct hammer2_recovery_info *info,
2172 hammer2_tid_t sync_tid)
2174 const hammer2_inode_data_t *ripdata;
2175 hammer2_chain_t *chain;
2176 hammer2_blockref_t bref;
2183 * Adjust freemap to ensure that the block(s) are marked allocated.
2185 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2186 hammer2_freemap_adjust(hmp, &parent->bref,
2187 HAMMER2_FREEMAP_DORECOVER);
2191 * Check type for recursive scan
2193 switch(parent->bref.type) {
2194 case HAMMER2_BREF_TYPE_VOLUME:
2195 /* data already instantiated */
2197 case HAMMER2_BREF_TYPE_INODE:
2199 * Must instantiate data for DIRECTDATA test and also
2202 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2203 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2204 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2205 /* not applicable to recovery scan */
2206 hammer2_chain_unlock(parent);
2209 hammer2_chain_unlock(parent);
2211 case HAMMER2_BREF_TYPE_INDIRECT:
2213 * Must instantiate data for recursion
2215 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2216 hammer2_chain_unlock(parent);
2218 case HAMMER2_BREF_TYPE_DIRENT:
2219 case HAMMER2_BREF_TYPE_DATA:
2220 case HAMMER2_BREF_TYPE_FREEMAP:
2221 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2222 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2223 /* not applicable to recovery scan */
2227 return HAMMER2_ERROR_BADBREF;
2231 * Defer operation if depth limit reached or if we are crossing a
2234 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2235 struct hammer2_recovery_elm *elm;
2237 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2238 elm->chain = parent;
2239 elm->sync_tid = sync_tid;
2240 hammer2_chain_ref(parent);
2241 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2242 /* unlocked by caller */
2249 * Recursive scan of the last flushed transaction only. We are
2250 * doing this without pmp assignments so don't leave the chains
2251 * hanging around after we are done with them.
2253 * error Cumulative error this level only
2254 * rup_error Cumulative error for recursion
2255 * tmp_error Specific non-cumulative recursion error
2263 error |= hammer2_chain_scan(parent, &chain, &bref,
2265 HAMMER2_LOOKUP_NODATA);
2268 * Problem during scan or EOF
2276 if (chain == NULL) {
2277 if (bref.mirror_tid > sync_tid) {
2278 hammer2_freemap_adjust(hmp, &bref,
2279 HAMMER2_FREEMAP_DORECOVER);
2285 * This may or may not be a recursive node.
2287 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2288 if (bref.mirror_tid > sync_tid) {
2290 tmp_error = hammer2_recovery_scan(hmp, chain,
2298 * Flush the recovery at the PFS boundary to stage it for
2299 * the final flush of the super-root topology.
2301 if (tmp_error == 0 &&
2302 (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2303 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2304 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2307 rup_error |= tmp_error;
2309 return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2313 * This fixes up an error introduced in earlier H2 implementations where
2314 * moving a PFS inode into an indirect block wound up causing the
2315 * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2319 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2321 const hammer2_inode_data_t *ripdata;
2322 hammer2_chain_t *parent;
2323 hammer2_chain_t *chain;
2324 hammer2_key_t key_next;
2325 hammer2_pfs_t *spmp;
2331 * Lookup mount point under the media-localized super-root.
2333 * cluster->pmp will incorrectly point to spmp and must be fixed
2337 hammer2_inode_lock(spmp->iroot, 0);
2338 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2339 chain = hammer2_chain_lookup(&parent, &key_next,
2340 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2343 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2346 kprintf("I/O error scanning PFS labels\n");
2347 error |= chain->error;
2348 } else if ((chain->bref.flags &
2349 HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2352 ripdata = &chain->data->ipdata;
2353 hammer2_trans_init(hmp->spmp, 0);
2354 error2 = hammer2_chain_modify(chain,
2355 chain->bref.modify_tid,
2358 kprintf("hammer2: Correct mis-flagged PFS %s\n",
2360 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2364 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2366 hammer2_trans_done(hmp->spmp, 0);
2368 chain = hammer2_chain_next(&parent, chain, &key_next,
2369 key_next, HAMMER2_KEY_MAX,
2373 hammer2_chain_unlock(parent);
2374 hammer2_chain_drop(parent);
2376 hammer2_inode_unlock(spmp->iroot);
2382 * Sync a mount point; this is called periodically on a per-mount basis from
2383 * the filesystem syncer, and whenever a user issues a sync.
2386 hammer2_vfs_sync(struct mount *mp, int waitfor)
2388 hammer2_xop_flush_t *xop;
2389 struct hammer2_sync_info info;
2390 hammer2_inode_t *iroot;
2398 KKASSERT(iroot->pmp == pmp);
2401 * We can't acquire locks on existing vnodes while in a transaction
2402 * without risking a deadlock. This assumes that vfsync() can be
2403 * called without the vnode locked (which it can in DragonFly).
2404 * Otherwise we'd have to implement a multi-pass or flag the lock
2405 * failures and retry.
2407 * The reclamation code interlocks with the sync list's token
2408 * (by removing the vnode from the scan list) before unlocking
2409 * the inode, giving us time to ref the inode.
2411 /*flags = VMSC_GETVP;*/
2413 if (waitfor & MNT_LAZY)
2414 flags |= VMSC_ONEPASS;
2417 * Flush vnodes individually using a normal transaction to avoid
2418 * stalling any concurrent operations. This will flush the related
2419 * buffer cache buffers and inodes to the media.
2421 * For efficiency do an async pass before making sure with a
2422 * synchronous pass on all related buffer cache buffers.
2424 hammer2_trans_init(pmp, 0);
2428 info.waitfor = MNT_NOWAIT;
2430 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2433 * Now do two passes making sure we get everything. The first pass
2434 * vfsync()s dirty vnodes. The second pass waits for their I/O's
2435 * to finish and cleans up the dirty flag on the vnode.
2438 info.waitfor = MNT_WAIT;
2439 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2442 info.waitfor = MNT_WAIT;
2443 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2446 * We must also run the sideq to handle any disconnected inodes
2447 * as the vnode scan will not see these.
2449 hammer2_inode_run_sideq(pmp, 1);
2450 hammer2_trans_done(pmp, 0);
2453 * Start our flush transaction and flush the root topology down to
2454 * the inodes, but not the inodes themselves (which we already flushed
2455 * above). Any concurrent activity effecting inode contents will not
2457 * The flush sequence will
2459 * NOTE! It is still possible for the paging code to push pages
2460 * out via a UIO_NOCOPY hammer2_vop_write() during the main
2463 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2466 * sync dirty vnodes again while in the flush transaction. This is
2467 * currently an expensive shim to makre sure the logical topology is
2468 * completely consistent before we flush the volume header.
2471 info.waitfor = MNT_WAIT;
2472 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2475 info.waitfor = MNT_WAIT;
2476 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2479 * Use the XOP interface to concurrently flush all nodes to
2480 * synchronize the PFSROOT subtopology to the media. A standard
2481 * end-of-scan ENOENT error indicates cluster sufficiency.
2483 * Note that this flush will not be visible on crash recovery until
2484 * we flush the super-root topology in the next loop.
2486 * XXX For now wait for all flushes to complete.
2490 * If unmounting try to flush everything including any
2491 * sub-trees under inodes, just in case there is dangling
2492 * modified data, as a safety. Otherwise just flush up to
2493 * the inodes in this stage.
2495 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2496 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2497 HAMMER2_XOP_VOLHDR);
2499 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2500 HAMMER2_XOP_INODE_STOP |
2501 HAMMER2_XOP_VOLHDR);
2503 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2504 error = hammer2_xop_collect(&xop->head,
2505 HAMMER2_XOP_COLLECT_WAITALL);
2506 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2507 if (error == HAMMER2_ERROR_ENOENT)
2510 error = hammer2_error_to_errno(error);
2514 hammer2_trans_done(pmp, 0);
2522 * Note that we ignore the tranasction mtid we got above. Instead,
2523 * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2526 * WARNING! The frontend might be waiting on chnmem (limit_dirty_chains)
2527 * while holding a vnode locked. When this situation occurs we cannot
2528 * safely test whether it is ok to clear the dirty bit on the vnode.
2529 * However, we can still flush the inode's topology.
2532 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2534 struct hammer2_sync_info *info = data;
2535 hammer2_inode_t *ip;
2539 * Degenerate cases. Note that ip == NULL typically means the
2540 * syncer vnode itself and we don't want to vclrisdirty() in that
2547 if (vp->v_type == VNON || vp->v_type == VBAD) {
2553 * Synchronize the buffer cche and inode meta-data to the backing
2556 * vfsync is not necessarily synchronous, so it is best NOT to try
2557 * to flush the backing topology to media at this point.
2559 hammer2_inode_ref(ip);
2560 if ((ip->flags & (HAMMER2_INODE_RESIZED|HAMMER2_INODE_MODIFIED)) ||
2561 !RB_EMPTY(&vp->v_rbdirty_tree)) {
2562 if (info->pass == 1)
2563 vfsync(vp, info->waitfor, 1, NULL, NULL);
2565 bio_track_wait(&vp->v_track_write, 0, 0);
2567 if (info->pass == 2 && (vp->v_flag & VISDIRTY)) {
2569 * v_token is needed to interlock v_rbdirty_tree.
2571 lwkt_gettoken(&vp->v_token);
2572 hammer2_inode_lock(ip, 0);
2573 hammer2_inode_chain_sync(ip);
2574 hammer2_inode_chain_flush(ip);
2575 if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2576 HAMMER2_INODE_RESIZED |
2577 HAMMER2_INODE_DIRTYDATA)) == 0 &&
2578 RB_EMPTY(&vp->v_rbdirty_tree) &&
2579 !bio_track_active(&vp->v_track_write)) {
2582 hammer2_inode_unlock(ip);
2583 lwkt_reltoken(&vp->v_token);
2585 hammer2_inode_drop(ip);
2589 info->error = error;
2596 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2598 hammer2_inode_t *ip;
2600 KKASSERT(MAXFIDSZ >= 16);
2602 fhp->fid_len = offsetof(struct fid, fid_data[16]);
2604 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2605 ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2612 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2613 struct fid *fhp, struct vnode **vpp)
2620 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2623 error = hammer2_vfs_root(mp, vpp);
2625 error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2630 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2636 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2637 int *exflagsp, struct ucred **credanonp)
2644 np = vfs_export_lookup(mp, &pmp->export, nam);
2646 *exflagsp = np->netc_exflags;
2647 *credanonp = &np->netc_anon;
2656 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume
2657 * header into the HMP
2659 * XXX read four volhdrs and use the one with the highest TID whos CRC
2664 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2665 * nonexistant locations.
2667 * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2671 hammer2_install_volume_header(hammer2_dev_t *hmp)
2673 hammer2_volume_data_t *vd;
2675 hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2687 * There are up to 4 copies of the volume header (syncs iterate
2688 * between them so there is no single master). We don't trust the
2689 * volu_size field so we don't know precisely how large the filesystem
2690 * is, so depend on the OS to return an error if we go beyond the
2691 * block device's EOF.
2693 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2694 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2695 HAMMER2_VOLUME_BYTES, &bp);
2702 vd = (struct hammer2_volume_data *) bp->b_data;
2703 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2704 (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2710 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2711 /* XXX: Reversed-endianness filesystem */
2712 kprintf("hammer2: reverse-endian filesystem detected");
2718 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2719 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2720 HAMMER2_VOLUME_ICRC0_SIZE);
2721 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2722 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2723 HAMMER2_VOLUME_ICRC1_SIZE);
2724 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2725 kprintf("hammer2 volume header crc "
2726 "mismatch copy #%d %08x/%08x\n",
2733 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2742 hmp->volsync = hmp->voldata;
2743 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2745 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2746 kprintf("hammer2: using volume header #%d\n",
2751 kprintf("hammer2: no valid volume headers found!\n");
2757 * This handles hysteresis on regular file flushes. Because the BIOs are
2758 * routed to a thread it is possible for an excessive number to build up
2759 * and cause long front-end stalls long before the runningbuffspace limit
2760 * is hit, so we implement hammer2_flush_pipe to control the
2763 * This is a particular problem when compression is used.
2766 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2768 atomic_add_int(&pmp->count_lwinprog, 1);
2772 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2776 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2777 if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2778 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2779 atomic_clear_int(&pmp->count_lwinprog,
2780 HAMMER2_LWINPROG_WAITING);
2781 wakeup(&pmp->count_lwinprog);
2783 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2784 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2785 atomic_clear_int(&pmp->count_lwinprog,
2786 HAMMER2_LWINPROG_WAITING0);
2787 wakeup(&pmp->count_lwinprog);
2792 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2795 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2796 HAMMER2_LWINPROG_WAITING0;
2799 lwinprog = pmp->count_lwinprog;
2801 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2803 tsleep_interlock(&pmp->count_lwinprog, 0);
2804 atomic_set_int(&pmp->count_lwinprog, lwflag);
2805 lwinprog = pmp->count_lwinprog;
2806 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2808 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2813 * Attempt to proactively fsync dirty vnodes if we have too many. This
2814 * solves an issue where the kernel syncer thread can get seriously behind
2815 * when multiple user processes/threads are furiously modifying inodes.
2816 * This situation can occur on slow storage and is only limited by
2817 * kern.maxvnodes without the moderation code below. It is made worse
2818 * when the device buffers underlying the modified inodes (which are clean)
2819 * get evicted before the flush can occur, forcing a re-read.
2821 * We do not want sysads to feel that they have to torpedo kern.maxvnodes
2822 * to solve this problem, so we implement vfs.hammer2.limit_dirty_inodes
2823 * (per-mount-basis) and default it to something reasonable.
2826 hammer2_pfs_moderate(hammer2_inode_t *ip, int always_moderate)
2828 hammer2_pfs_t *pmp = ip->pmp;
2829 struct mount *mp = pmp->mp;
2831 if (mp && vn_syncer_count(mp) > hammer2_limit_dirty_inodes) {
2837 * Manage excessive memory resource use for chain and related
2840 * Called without any inode locks or transaction locks. VNodes
2841 * might be locked by the kernel in the call stack.
2844 hammer2_pfs_memory_wait(hammer2_inode_t *ip, int always_moderate)
2846 hammer2_pfs_t *pmp = ip->pmp;
2855 * Moderate the number of dirty inodes
2857 hammer2_pfs_moderate(ip, always_moderate);
2860 * Atomic check condition and wait. Also do an early speedup of
2861 * the syncer to try to avoid hitting the wait.
2864 waiting = pmp->inmem_dirty_chains;
2866 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2868 limit = pmp->mp->mnt_nvnodelistsize / 10;
2869 if (limit < hammer2_limit_dirty_chains)
2870 limit = hammer2_limit_dirty_chains;
2875 if ((int)(ticks - zzticks) > hz) {
2877 kprintf("count %ld %ld\n", count, limit);
2882 * Block if there are too many dirty chains present, wait
2883 * for the flush to clean some out.
2885 if (count > limit) {
2886 hammer2_pfs_moderate(ip, always_moderate);
2887 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2888 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2890 waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2891 if (ticks != pmp->speedup_ticks) {
2892 pmp->speedup_ticks = ticks;
2893 speedup_syncer(pmp->mp);
2895 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2898 continue; /* loop on success or fail */
2902 * Try to start an early flush before we are forced to block.
2904 if (count > limit * 5 / 10 &&
2905 ticks != pmp->speedup_ticks) {
2906 pmp->speedup_ticks = ticks;
2907 speedup_syncer(pmp->mp);
2914 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2917 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2922 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2927 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
2928 /* don't need --waiting to test flag */
2929 if (waiting & HAMMER2_DIRTYCHAIN_WAITING) {
2930 atomic_clear_int(&pmp->inmem_dirty_chains,
2931 HAMMER2_DIRTYCHAIN_WAITING);
2932 wakeup(&pmp->inmem_dirty_chains);
2938 * Returns 0 if the filesystem has tons of free space
2939 * Returns 1 if the filesystem has less than 10% remaining
2940 * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2943 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2947 hammer2_off_t free_reserved;
2948 hammer2_off_t free_nominal;
2953 if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2954 free_reserved = HAMMER2_SEGSIZE;
2955 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2956 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2957 hmp = pmp->pfs_hmps[i];
2960 if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2961 pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2964 if (free_nominal > hmp->voldata.allocator_free)
2965 free_nominal = hmp->voldata.allocator_free;
2966 if (free_reserved < hmp->free_reserved)
2967 free_reserved = hmp->free_reserved;
2973 pmp->free_reserved = free_reserved;
2974 pmp->free_nominal = free_nominal;
2975 pmp->free_ticks = ticks;
2977 free_reserved = pmp->free_reserved;
2978 free_nominal = pmp->free_nominal;
2980 if (cred && cred->cr_uid != 0) {
2981 if ((int64_t)(free_nominal - bytes) <
2982 (int64_t)free_reserved) {
2986 if ((int64_t)(free_nominal - bytes) <
2987 (int64_t)free_reserved / 2) {
2991 if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
3000 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
3003 hammer2_chain_t *scan;
3004 hammer2_chain_t *parent;
3008 kprintf("%*.*s...\n", tab, tab, "");
3013 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
3015 chain, chain->bref.type,
3016 chain->bref.key, chain->bref.keybits,
3017 chain->bref.mirror_tid);
3019 kprintf("%*.*s [%08x] (%s) refs=%d",
3022 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3023 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
3026 parent = chain->parent;
3028 kprintf("\n%*.*s p=%p [pflags %08x prefs %d",
3030 parent, parent->flags, parent->refs);
3031 if (RB_EMPTY(&chain->core.rbtree)) {
3035 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3036 if ((scan->flags & flags) || flags == (u_int)-1) {
3037 hammer2_dump_chain(scan, tab + 4, countp, 'a',
3041 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3042 kprintf("%*.*s}(%s)\n", tab, tab, "",
3043 chain->data->ipdata.filename);
3045 kprintf("%*.*s}\n", tab, tab, "");