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_count_modified_chains;
97 long hammer2_iod_invals;
98 long hammer2_iod_file_read;
99 long hammer2_iod_meta_read;
100 long hammer2_iod_indr_read;
101 long hammer2_iod_fmap_read;
102 long hammer2_iod_volu_read;
103 long hammer2_iod_file_write;
104 long hammer2_iod_file_wembed;
105 long hammer2_iod_file_wzero;
106 long hammer2_iod_file_wdedup;
107 long hammer2_iod_meta_write;
108 long hammer2_iod_indr_write;
109 long hammer2_iod_fmap_write;
110 long hammer2_iod_volu_write;
112 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
113 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
114 "Buffer used for compression.");
116 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
117 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
118 "Buffer used for decompression.");
120 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
122 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
123 &hammer2_supported_version, 0, "");
124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
125 &hammer2_debug, 0, "");
126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
127 &hammer2_cluster_meta_read, 0, "");
128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
129 &hammer2_cluster_data_read, 0, "");
130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
131 &hammer2_dedup_enable, 0, "");
132 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
133 &hammer2_always_compress, 0, "");
134 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
135 &hammer2_inval_enable, 0, "");
136 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
137 &hammer2_flush_pipe, 0, "");
138 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
139 &hammer2_bulkfree_tps, 0, "");
140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
141 &hammer2_chain_allocs, 0, "");
142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
143 &hammer2_chain_frees, 0, "");
144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
145 &hammer2_limit_dirty_chains, 0, "");
146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
147 &hammer2_count_modified_chains, 0, "");
148 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
149 &hammer2_dio_count, 0, "");
150 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
151 &hammer2_dio_limit, 0, "");
153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
154 &hammer2_iod_invals, 0, "");
155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
156 &hammer2_iod_file_read, 0, "");
157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
158 &hammer2_iod_meta_read, 0, "");
159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
160 &hammer2_iod_indr_read, 0, "");
161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
162 &hammer2_iod_fmap_read, 0, "");
163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
164 &hammer2_iod_volu_read, 0, "");
166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
167 &hammer2_iod_file_write, 0, "");
168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
169 &hammer2_iod_file_wembed, 0, "");
170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
171 &hammer2_iod_file_wzero, 0, "");
172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
173 &hammer2_iod_file_wdedup, 0, "");
174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
175 &hammer2_iod_meta_write, 0, "");
176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
177 &hammer2_iod_indr_write, 0, "");
178 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
179 &hammer2_iod_fmap_write, 0, "");
180 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
181 &hammer2_iod_volu_write, 0, "");
183 long hammer2_process_icrc32;
184 long hammer2_process_xxhash64;
185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
186 &hammer2_process_icrc32, 0, "");
187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
188 &hammer2_process_xxhash64, 0, "");
190 static int hammer2_vfs_init(struct vfsconf *conf);
191 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
192 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
194 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
195 struct vnode *, struct ucred *);
196 static int hammer2_recovery(hammer2_dev_t *hmp);
197 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
198 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
199 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
201 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
203 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
204 struct fid *fhp, struct vnode **vpp);
205 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
206 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
207 int *exflagsp, struct ucred **credanonp);
209 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
210 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
212 static void hammer2_update_pmps(hammer2_dev_t *hmp);
214 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
215 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
217 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
220 * HAMMER2 vfs operations.
222 static struct vfsops hammer2_vfsops = {
223 .vfs_init = hammer2_vfs_init,
224 .vfs_uninit = hammer2_vfs_uninit,
225 .vfs_sync = hammer2_vfs_sync,
226 .vfs_mount = hammer2_vfs_mount,
227 .vfs_unmount = hammer2_vfs_unmount,
228 .vfs_root = hammer2_vfs_root,
229 .vfs_statfs = hammer2_vfs_statfs,
230 .vfs_statvfs = hammer2_vfs_statvfs,
231 .vfs_vget = hammer2_vfs_vget,
232 .vfs_vptofh = hammer2_vfs_vptofh,
233 .vfs_fhtovp = hammer2_vfs_fhtovp,
234 .vfs_checkexp = hammer2_vfs_checkexp
237 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
239 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
240 MODULE_VERSION(hammer2, 1);
244 hammer2_vfs_init(struct vfsconf *conf)
246 static struct objcache_malloc_args margs_read;
247 static struct objcache_malloc_args margs_write;
248 static struct objcache_malloc_args margs_vop;
255 * A large DIO cache is needed to retain dedup enablement masks.
256 * The bulkfree code clears related masks as part of the disk block
257 * recycling algorithm, preventing it from being used for a later
260 * NOTE: A large buffer cache can actually interfere with dedup
261 * operation because we dedup based on media physical buffers
262 * and not logical buffers. Try to make the DIO case large
263 * enough to avoid this problem, but also cap it.
265 hammer2_dio_limit = nbuf * 2;
266 if (hammer2_dio_limit > 100000)
267 hammer2_dio_limit = 100000;
269 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
271 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
273 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
277 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
279 margs_read.objsize = 65536;
280 margs_read.mtype = M_HAMMER2_DEBUFFER;
282 margs_write.objsize = 32768;
283 margs_write.mtype = M_HAMMER2_CBUFFER;
285 margs_vop.objsize = sizeof(hammer2_xop_t);
286 margs_vop.mtype = M_HAMMER2;
289 * Note thaht for the XOPS cache we want backing store allocations
290 * to use M_ZERO. This is not allowed in objcache_get() (to avoid
291 * confusion), so use the backing store function that does it. This
292 * means that initial XOPS objects are zerod but REUSED objects are
293 * not. So we are responsible for cleaning the object up sufficiently
294 * for our needs before objcache_put()ing it back (typically just the
297 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
298 0, 1, NULL, NULL, NULL,
299 objcache_malloc_alloc,
300 objcache_malloc_free,
302 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
303 0, 1, NULL, NULL, NULL,
304 objcache_malloc_alloc,
305 objcache_malloc_free,
307 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
308 0, 1, NULL, NULL, NULL,
309 objcache_malloc_alloc_zero,
310 objcache_malloc_free,
314 lockinit(&hammer2_mntlk, "mntlk", 0, 0);
315 TAILQ_INIT(&hammer2_mntlist);
316 TAILQ_INIT(&hammer2_pfslist);
317 TAILQ_INIT(&hammer2_spmplist);
319 hammer2_limit_dirty_chains = maxvnodes / 10;
320 if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
321 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
328 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
330 objcache_destroy(cache_buffer_read);
331 objcache_destroy(cache_buffer_write);
332 objcache_destroy(cache_xops);
337 * Core PFS allocator. Used to allocate or reference the pmp structure
338 * for PFS cluster mounts and the spmp structure for media (hmp) structures.
339 * The pmp can be passed in or loaded by this function using the chain and
342 * pmp->modify_tid tracks new modify_tid transaction ids for front-end
343 * transactions. Note that synchronization does not use this field.
344 * (typically frontend operations and synchronization cannot run on the
345 * same PFS node at the same time).
350 hammer2_pfsalloc(hammer2_chain_t *chain,
351 const hammer2_inode_data_t *ripdata,
352 hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
355 hammer2_inode_t *iroot;
363 * Locate or create the PFS based on the cluster id. If ripdata
364 * is NULL this is a spmp which is unique and is always allocated.
366 * If the device is mounted in local mode all PFSs are considered
367 * independent and not part of any cluster (for debugging only).
370 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
371 if (force_local != pmp->force_local)
373 if (force_local == NULL &&
374 bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
375 sizeof(pmp->pfs_clid)) == 0) {
377 } else if (force_local && pmp->pfs_names[0] &&
378 strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
385 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
386 pmp->force_local = force_local;
387 hammer2_trans_manage_init(pmp);
388 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
389 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
390 lockinit(&pmp->lock, "pfslk", 0, 0);
391 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
392 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
393 spin_init(&pmp->xop_spin, "h2xop");
394 spin_init(&pmp->lru_spin, "h2lru");
395 RB_INIT(&pmp->inum_tree);
396 TAILQ_INIT(&pmp->sideq);
397 TAILQ_INIT(&pmp->lru_list);
398 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
401 * Distribute backend operations to threads
403 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
404 hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
407 * Save the last media transaction id for the flusher. Set
411 pmp->pfs_clid = ripdata->meta.pfs_clid;
412 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
414 pmp->flags |= HAMMER2_PMPF_SPMP;
415 TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
419 * The synchronization thread may start too early, make
420 * sure it stays frozen until we are ready to let it go.
424 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
425 HAMMER2_THREAD_REMASTER;
430 * Create the PFS's root inode and any missing XOP helper threads.
432 if ((iroot = pmp->iroot) == NULL) {
433 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
435 iroot->meta = ripdata->meta;
437 hammer2_inode_ref(iroot);
438 hammer2_inode_unlock(iroot);
442 * Stop here if no chain is passed in.
448 * When a chain is passed in we must add it to the PFS's root
449 * inode, update pmp->pfs_types[], and update the syncronization
452 * When forcing local mode, mark the PFS as a MASTER regardless.
454 * At the moment empty spots can develop due to removals or failures.
455 * Ultimately we want to re-fill these spots but doing so might
456 * confused running code. XXX
458 hammer2_inode_ref(iroot);
459 hammer2_mtx_ex(&iroot->lock);
460 j = iroot->cluster.nchains;
462 if (j == HAMMER2_MAXCLUSTER) {
463 kprintf("hammer2_mount: cluster full!\n");
464 /* XXX fatal error? */
466 KKASSERT(chain->pmp == NULL);
468 hammer2_chain_ref(chain);
469 iroot->cluster.array[j].chain = chain;
471 pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
473 pmp->pfs_types[j] = ripdata->meta.pfs_type;
474 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
475 pmp->pfs_hmps[j] = chain->hmp;
478 * If the PFS is already mounted we must account
479 * for the mount_count here.
482 ++chain->hmp->mount_count;
485 * May have to fixup dirty chain tracking. Previous
486 * pmp was NULL so nothing to undo.
488 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
489 hammer2_pfs_memory_inc(pmp);
492 iroot->cluster.nchains = j;
495 * Update nmasters from any PFS inode which is part of the cluster.
496 * It is possible that this will result in a value which is too
497 * high. MASTER PFSs are authoritative for pfs_nmasters and will
498 * override this value later on.
500 * (This informs us of masters that might not currently be
501 * discoverable by this mount).
503 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
504 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
508 * Count visible masters. Masters are usually added with
509 * ripdata->meta.pfs_nmasters set to 1. This detects when there
510 * are more (XXX and must update the master inodes).
513 for (i = 0; i < iroot->cluster.nchains; ++i) {
514 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
517 if (pmp->pfs_nmasters < count)
518 pmp->pfs_nmasters = count;
521 * Create missing synchronization and support threads.
523 * Single-node masters (including snapshots) have nothing to
524 * synchronize and do not require this thread.
526 * Multi-node masters or any number of soft masters, slaves, copy,
527 * or other PFS types need the thread.
529 * Each thread is responsible for its particular cluster index.
530 * We use independent threads so stalls or mismatches related to
531 * any given target do not affect other targets.
533 for (i = 0; i < iroot->cluster.nchains; ++i) {
535 * Single-node masters (including snapshots) have nothing
536 * to synchronize and will make direct xops support calls,
537 * thus they do not require this thread.
539 * Note that there can be thousands of snapshots. We do not
540 * want to create thousands of threads.
542 if (pmp->pfs_nmasters <= 1 &&
543 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
548 * Sync support thread
550 if (pmp->sync_thrs[i].td == NULL) {
551 hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
553 hammer2_primary_sync_thread);
558 * Create missing Xop threads
560 * NOTE: We create helper threads for all mounted PFSs or any
561 * PFSs with 2+ nodes (so the sync thread can update them,
562 * even if not mounted).
564 if (pmp->mp || iroot->cluster.nchains >= 2)
565 hammer2_xop_helper_create(pmp);
567 hammer2_mtx_unlock(&iroot->lock);
568 hammer2_inode_drop(iroot);
574 * Deallocate an element of a probed PFS. If destroying and this is a
575 * MASTER, adjust nmasters.
577 * This function does not physically destroy the PFS element in its device
578 * under the super-root (see hammer2_ioctl_pfs_delete()).
581 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
583 hammer2_inode_t *iroot;
584 hammer2_chain_t *chain;
588 * Cleanup our reference on iroot. iroot is (should) not be needed
596 * XXX flush after acquiring the iroot lock.
597 * XXX clean out the cluster index from all inode structures.
599 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
602 * Remove the cluster index from the group. If destroying
603 * the PFS and this is a master, adjust pfs_nmasters.
605 hammer2_mtx_ex(&iroot->lock);
606 chain = iroot->cluster.array[clindex].chain;
607 iroot->cluster.array[clindex].chain = NULL;
609 switch(pmp->pfs_types[clindex]) {
610 case HAMMER2_PFSTYPE_MASTER:
611 if (destroying && pmp->pfs_nmasters > 0)
613 /* XXX adjust ripdata->meta.pfs_nmasters */
618 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
620 hammer2_mtx_unlock(&iroot->lock);
626 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
627 hammer2_chain_drop(chain);
631 * Terminate all XOP threads for the cluster index.
633 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
634 hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
639 * Destroy a PFS, typically only occurs after the last mount on a device
643 hammer2_pfsfree(hammer2_pfs_t *pmp)
645 hammer2_inode_t *iroot;
646 hammer2_chain_t *chain;
651 * Cleanup our reference on iroot. iroot is (should) not be needed
654 if (pmp->flags & HAMMER2_PMPF_SPMP)
655 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
657 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
661 for (i = 0; i < iroot->cluster.nchains; ++i) {
662 hammer2_thr_delete(&pmp->sync_thrs[i]);
663 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
664 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
666 #if REPORT_REFS_ERRORS
667 if (pmp->iroot->refs != 1)
668 kprintf("PMP->IROOT %p REFS WRONG %d\n",
669 pmp->iroot, pmp->iroot->refs);
671 KKASSERT(pmp->iroot->refs == 1);
673 /* ref for pmp->iroot */
674 hammer2_inode_drop(pmp->iroot);
679 * Cleanup chains remaining on LRU list.
681 hammer2_spin_ex(&pmp->lru_spin);
682 while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
683 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
684 atomic_add_int(&pmp->lru_count, -1);
685 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
686 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
687 hammer2_chain_ref(chain);
688 hammer2_spin_unex(&pmp->lru_spin);
689 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
690 hammer2_chain_drop(chain);
691 hammer2_spin_ex(&pmp->lru_spin);
693 hammer2_spin_unex(&pmp->lru_spin);
696 * Free remaining pmp resources
698 kmalloc_destroy(&pmp->mmsg);
699 kmalloc_destroy(&pmp->minode);
701 kfree(pmp, M_HAMMER2);
705 * Remove all references to hmp from the pfs list. Any PFS which becomes
706 * empty is terminated and freed.
711 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
714 hammer2_inode_t *iroot;
715 hammer2_chain_t *rchain;
719 struct hammer2_pfslist *wlist;
722 wlist = &hammer2_pfslist;
724 wlist = &hammer2_spmplist;
726 TAILQ_FOREACH(pmp, wlist, mntentry) {
727 if ((iroot = pmp->iroot) == NULL)
729 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
730 hammer2_inode_run_sideq(pmp, 1);
731 hammer2_bioq_sync(pmp);
732 hammer2_trans_done(pmp);
735 * Determine if this PFS is affected. If it is we must
736 * freeze all management threads and lock its iroot.
738 * Freezing a management thread forces it idle, operations
739 * in-progress will be aborted and it will have to start
740 * over again when unfrozen, or exit if told to exit.
742 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
743 if (pmp->pfs_hmps[i] == hmp)
746 if (i != HAMMER2_MAXCLUSTER) {
748 * Make sure all synchronization threads are locked
751 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
752 if (pmp->pfs_hmps[i] == NULL)
754 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
755 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
756 hammer2_thr_freeze_async(
757 &pmp->xop_groups[j].thrs[i]);
760 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
761 if (pmp->pfs_hmps[i] == NULL)
763 hammer2_thr_freeze(&pmp->sync_thrs[i]);
764 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
766 &pmp->xop_groups[j].thrs[i]);
771 * Lock the inode and clean out matching chains.
772 * Note that we cannot use hammer2_inode_lock_*()
773 * here because that would attempt to validate the
774 * cluster that we are in the middle of ripping
777 * WARNING! We are working directly on the inodes
780 hammer2_mtx_ex(&iroot->lock);
783 * Remove the chain from matching elements of the PFS.
785 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
786 if (pmp->pfs_hmps[i] != hmp)
788 hammer2_thr_delete(&pmp->sync_thrs[i]);
789 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
791 &pmp->xop_groups[j].thrs[i]);
793 rchain = iroot->cluster.array[i].chain;
794 iroot->cluster.array[i].chain = NULL;
795 pmp->pfs_types[i] = 0;
796 if (pmp->pfs_names[i]) {
797 kfree(pmp->pfs_names[i], M_HAMMER2);
798 pmp->pfs_names[i] = NULL;
801 hammer2_chain_drop(rchain);
803 if (iroot->cluster.focus == rchain)
804 iroot->cluster.focus = NULL;
806 pmp->pfs_hmps[i] = NULL;
808 hammer2_mtx_unlock(&iroot->lock);
809 didfreeze = 1; /* remaster, unfreeze down below */
815 * Cleanup trailing chains. Gaps may remain.
817 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
818 if (pmp->pfs_hmps[i])
821 iroot->cluster.nchains = i + 1;
824 * If the PMP has no elements remaining we can destroy it.
825 * (this will transition management threads from frozen->exit).
827 if (iroot->cluster.nchains == 0) {
829 * If this was the hmp's spmp, we need to clean
830 * a little more stuff out.
832 if (hmp->spmp == pmp) {
834 hmp->vchain.pmp = NULL;
835 hmp->fchain.pmp = NULL;
839 * Free the pmp and restart the loop
841 hammer2_pfsfree(pmp);
846 * If elements still remain we need to set the REMASTER
847 * flag and unfreeze it.
850 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
851 if (pmp->pfs_hmps[i] == NULL)
853 hammer2_thr_remaster(&pmp->sync_thrs[i]);
854 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
855 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
856 hammer2_thr_remaster(
857 &pmp->xop_groups[j].thrs[i]);
858 hammer2_thr_unfreeze(
859 &pmp->xop_groups[j].thrs[i]);
867 * Mount or remount HAMMER2 fileystem from physical media
870 * mp mount point structure
876 * mp mount point structure
877 * path path to mount point
878 * data pointer to argument structure in user space
879 * volume volume path (device@LABEL form)
880 * hflags user mount flags
881 * cred user credentials
888 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
891 struct hammer2_mount_info info;
895 hammer2_dev_t *force_local;
896 hammer2_key_t key_next;
897 hammer2_key_t key_dummy;
900 struct nlookupdata nd;
901 hammer2_chain_t *parent;
902 hammer2_chain_t *chain;
903 hammer2_cluster_t *cluster;
904 const hammer2_inode_data_t *ripdata;
905 hammer2_blockref_t bref;
907 char devstr[MNAMELEN];
922 kprintf("hammer2_mount\n");
928 bzero(&info, sizeof(info));
929 info.cluster_fd = -1;
930 ksnprintf(devstr, sizeof(devstr), "%s",
931 mp->mnt_stat.f_mntfromname);
932 kprintf("hammer2_mount: root '%s'\n", devstr);
935 * Non-root mount or updating a mount
937 error = copyin(data, &info, sizeof(info));
941 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
947 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
948 * if no label specified, based on the partition id. Error out if no
949 * label or device (with partition id) is specified. This is strictly
950 * a convenience to match the default label created by newfs_hammer2,
951 * our preference is that a label always be specified.
953 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
954 * that does not specify a device, as long as some H2 label
955 * has already been mounted from that device. This makes
956 * mounting snapshots a lot easier.
959 label = strchr(devstr, '@');
960 if (label && ((label + 1) - dev) > done)
962 if (label == NULL || label[1] == 0) {
966 label = devstr + strlen(devstr);
984 kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
985 dev, label, (mp->mnt_flag & MNT_RDONLY));
987 if (mp->mnt_flag & MNT_UPDATE) {
989 * Update mount. Note that pmp->iroot->cluster is
990 * an inode-embedded cluster and thus cannot be
993 * XXX HAMMER2 needs to implement NFS export via
997 pmp->hflags = info.hflags;
998 cluster = &pmp->iroot->cluster;
999 for (i = 0; i < cluster->nchains; ++i) {
1000 if (cluster->array[i].chain == NULL)
1002 hmp = cluster->array[i].chain->hmp;
1004 error = hammer2_remount(hmp, mp, path,
1016 * If a path is specified and dev is not an empty string, lookup the
1017 * name and verify that it referes to a block device.
1019 * If a path is specified and dev is an empty string we fall through
1020 * and locate the label in the hmp search.
1022 if (path && *dev != 0) {
1023 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1025 error = nlookup(&nd);
1027 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1029 } else if (path == NULL) {
1031 cdev_t cdev = kgetdiskbyname(dev);
1032 error = bdevvp(cdev, &devvp);
1034 kprintf("hammer2: cannot find '%s'\n", dev);
1037 * We will locate the hmp using the label in the hmp loop.
1043 * Make sure its a block device. Do not check to see if it is
1044 * already mounted until we determine that its a fresh H2 device.
1046 if (error == 0 && devvp) {
1047 vn_isdisk(devvp, &error);
1051 * Determine if the device has already been mounted. After this
1052 * check hmp will be non-NULL if we are doing the second or more
1053 * hammer2 mounts from the same device.
1055 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1058 * Match the device. Due to the way devfs works,
1059 * we may not be able to directly match the vnode pointer,
1060 * so also check to see if the underlying device matches.
1062 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1063 if (hmp->devvp == devvp)
1065 if (devvp->v_rdev &&
1066 hmp->devvp->v_rdev == devvp->v_rdev) {
1072 * If no match this may be a fresh H2 mount, make sure
1073 * the device is not mounted on anything else.
1076 error = vfs_mountedon(devvp);
1077 } else if (error == 0) {
1079 * Match the label to a pmp already probed.
1081 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1082 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1083 if (pmp->pfs_names[i] &&
1084 strcmp(pmp->pfs_names[i], label) == 0) {
1085 hmp = pmp->pfs_hmps[i];
1097 * Open the device if this isn't a secondary mount and construct
1098 * the H2 device mount (hmp).
1101 hammer2_chain_t *schain;
1104 if (error == 0 && vcount(devvp) > 0) {
1105 kprintf("Primary device already has references\n");
1110 * Now open the device
1113 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1114 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1115 error = vinvalbuf(devvp, V_SAVE, 0, 0);
1117 error = VOP_OPEN(devvp,
1118 (ronly ? FREAD : FREAD | FWRITE),
1123 if (error && devvp) {
1128 lockmgr(&hammer2_mntlk, LK_RELEASE);
1131 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1132 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1135 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1136 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1137 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1138 RB_INIT(&hmp->iotree);
1139 spin_init(&hmp->io_spin, "hm2mount_io");
1140 spin_init(&hmp->list_spin, "hm2mount_list");
1141 TAILQ_INIT(&hmp->flushq);
1143 lockinit(&hmp->vollk, "h2vol", 0, 0);
1144 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1145 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1148 * vchain setup. vchain.data is embedded.
1149 * vchain.refs is initialized and will never drop to 0.
1151 * NOTE! voldata is not yet loaded.
1153 hmp->vchain.hmp = hmp;
1154 hmp->vchain.refs = 1;
1155 hmp->vchain.data = (void *)&hmp->voldata;
1156 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1157 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1158 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1160 hammer2_chain_core_init(&hmp->vchain);
1161 /* hmp->vchain.u.xxx is left NULL */
1164 * fchain setup. fchain.data is embedded.
1165 * fchain.refs is initialized and will never drop to 0.
1167 * The data is not used but needs to be initialized to
1168 * pass assertion muster. We use this chain primarily
1169 * as a placeholder for the freemap's top-level RBTREE
1170 * so it does not interfere with the volume's topology
1173 hmp->fchain.hmp = hmp;
1174 hmp->fchain.refs = 1;
1175 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1176 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1177 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1178 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1179 hmp->fchain.bref.methods =
1180 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1181 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1183 hammer2_chain_core_init(&hmp->fchain);
1184 /* hmp->fchain.u.xxx is left NULL */
1187 * Install the volume header and initialize fields from
1190 error = hammer2_install_volume_header(hmp);
1192 hammer2_unmount_helper(mp, NULL, hmp);
1193 lockmgr(&hammer2_mntlk, LK_RELEASE);
1194 hammer2_vfs_unmount(mp, MNT_FORCE);
1199 * Really important to get these right or flush will get
1202 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1206 * Dummy-up vchain and fchain's modify_tid. mirror_tid
1207 * is inherited from the volume header.
1210 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1211 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1212 hmp->vchain.pmp = spmp;
1213 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1214 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1215 hmp->fchain.pmp = spmp;
1218 * First locate the super-root inode, which is key 0
1219 * relative to the volume header's blockset.
1221 * Then locate the root inode by scanning the directory keyspace
1222 * represented by the label.
1224 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1225 schain = hammer2_chain_lookup(&parent, &key_dummy,
1226 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1228 hammer2_chain_lookup_done(parent);
1229 if (schain == NULL) {
1230 kprintf("hammer2_mount: invalid super-root\n");
1231 hammer2_unmount_helper(mp, NULL, hmp);
1232 lockmgr(&hammer2_mntlk, LK_RELEASE);
1233 hammer2_vfs_unmount(mp, MNT_FORCE);
1236 if (schain->error) {
1237 kprintf("hammer2_mount: error %s reading super-root\n",
1238 hammer2_error_str(schain->error));
1239 hammer2_chain_unlock(schain);
1240 hammer2_chain_drop(schain);
1242 hammer2_unmount_helper(mp, NULL, hmp);
1243 lockmgr(&hammer2_mntlk, LK_RELEASE);
1244 hammer2_vfs_unmount(mp, MNT_FORCE);
1249 * The super-root always uses an inode_tid of 1 when
1252 spmp->inode_tid = 1;
1253 spmp->modify_tid = schain->bref.modify_tid + 1;
1256 * Sanity-check schain's pmp and finish initialization.
1257 * Any chain belonging to the super-root topology should
1258 * have a NULL pmp (not even set to spmp).
1260 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1261 KKASSERT(schain->pmp == NULL);
1262 spmp->pfs_clid = ripdata->meta.pfs_clid;
1265 * Replace the dummy spmp->iroot with a real one. It's
1266 * easier to just do a wholesale replacement than to try
1267 * to update the chain and fixup the iroot fields.
1269 * The returned inode is locked with the supplied cluster.
1271 cluster = hammer2_cluster_from_chain(schain);
1272 hammer2_inode_drop(spmp->iroot);
1274 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1275 spmp->spmp_hmp = hmp;
1276 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1277 spmp->pfs_hmps[0] = hmp;
1278 hammer2_inode_ref(spmp->iroot);
1279 hammer2_inode_unlock(spmp->iroot);
1280 hammer2_cluster_unlock(cluster);
1281 hammer2_cluster_drop(cluster);
1283 /* leave spmp->iroot with one ref */
1285 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1286 error = hammer2_recovery(hmp);
1288 error |= hammer2_fixup_pfses(hmp);
1289 /* XXX do something with error */
1291 hammer2_update_pmps(hmp);
1292 hammer2_iocom_init(hmp);
1293 hammer2_bulkfree_init(hmp);
1296 * Ref the cluster management messaging descriptor. The mount
1297 * program deals with the other end of the communications pipe.
1299 * Root mounts typically do not supply one.
1301 if (info.cluster_fd >= 0) {
1302 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1304 hammer2_cluster_reconnect(hmp, fp);
1306 kprintf("hammer2_mount: bad cluster_fd!\n");
1311 if (info.hflags & HMNT2_DEVFLAGS) {
1312 kprintf("hammer2: Warning: mount flags pertaining "
1313 "to the whole device may only be specified "
1314 "on the first mount of the device: %08x\n",
1315 info.hflags & HMNT2_DEVFLAGS);
1320 * Force local mount (disassociate all PFSs from their clusters).
1321 * Used primarily for debugging.
1323 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1326 * Lookup the mount point under the media-localized super-root.
1327 * Scanning hammer2_pfslist doesn't help us because it represents
1328 * PFS cluster ids which can aggregate several named PFSs together.
1330 * cluster->pmp will incorrectly point to spmp and must be fixed
1333 hammer2_inode_lock(spmp->iroot, 0);
1334 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1335 lhc = hammer2_dirhash(label, strlen(label));
1336 chain = hammer2_chain_lookup(&parent, &key_next,
1337 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1340 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1341 strcmp(label, chain->data->ipdata.filename) == 0) {
1344 chain = hammer2_chain_next(&parent, chain, &key_next,
1346 lhc + HAMMER2_DIRHASH_LOMASK,
1350 hammer2_chain_unlock(parent);
1351 hammer2_chain_drop(parent);
1353 hammer2_inode_unlock(spmp->iroot);
1356 * PFS could not be found?
1358 if (chain == NULL) {
1360 kprintf("hammer2_mount: PFS label I/O error\n");
1362 kprintf("hammer2_mount: PFS label not found\n");
1363 hammer2_unmount_helper(mp, NULL, hmp);
1364 lockmgr(&hammer2_mntlk, LK_RELEASE);
1365 hammer2_vfs_unmount(mp, MNT_FORCE);
1371 * Acquire the pmp structure (it should have already been allocated
1372 * via hammer2_update_pmps() so do not pass cluster in to add to
1373 * available chains).
1375 * Check if the cluster has already been mounted. A cluster can
1376 * only be mounted once, use null mounts to mount additional copies.
1379 kprintf("hammer2_mount: PFS label I/O error\n");
1381 ripdata = &chain->data->ipdata;
1383 pmp = hammer2_pfsalloc(NULL, ripdata,
1384 bref.modify_tid, force_local);
1386 hammer2_chain_unlock(chain);
1387 hammer2_chain_drop(chain);
1392 kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1395 kprintf("hammer2_mount: PFS already mounted!\n");
1396 hammer2_unmount_helper(mp, NULL, hmp);
1397 lockmgr(&hammer2_mntlk, LK_RELEASE);
1398 hammer2_vfs_unmount(mp, MNT_FORCE);
1403 pmp->hflags = info.hflags;
1404 mp->mnt_flag |= MNT_LOCAL;
1405 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */
1406 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */
1409 * required mount structure initializations
1411 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1412 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1414 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1415 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1420 mp->mnt_iosize_max = MAXPHYS;
1423 * Connect up mount pointers.
1425 hammer2_mount_helper(mp, pmp);
1427 lockmgr(&hammer2_mntlk, LK_RELEASE);
1433 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1434 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1435 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1438 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1439 MNAMELEN - 1, &size);
1440 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1441 } /* else root mount, already in there */
1443 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1445 copyinstr(path, mp->mnt_stat.f_mntonname,
1446 sizeof(mp->mnt_stat.f_mntonname) - 1,
1450 mp->mnt_stat.f_mntonname[0] = '/';
1454 * Initial statfs to prime mnt_stat.
1456 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1462 * Scan PFSs under the super-root and create hammer2_pfs structures.
1466 hammer2_update_pmps(hammer2_dev_t *hmp)
1468 const hammer2_inode_data_t *ripdata;
1469 hammer2_chain_t *parent;
1470 hammer2_chain_t *chain;
1471 hammer2_blockref_t bref;
1472 hammer2_dev_t *force_local;
1473 hammer2_pfs_t *spmp;
1475 hammer2_key_t key_next;
1479 * Force local mount (disassociate all PFSs from their clusters).
1480 * Used primarily for debugging.
1482 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1485 * Lookup mount point under the media-localized super-root.
1487 * cluster->pmp will incorrectly point to spmp and must be fixed
1491 hammer2_inode_lock(spmp->iroot, 0);
1492 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1493 chain = hammer2_chain_lookup(&parent, &key_next,
1494 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1497 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1500 kprintf("I/O error scanning PFS labels\n");
1502 ripdata = &chain->data->ipdata;
1505 pmp = hammer2_pfsalloc(chain, ripdata,
1506 bref.modify_tid, force_local);
1508 chain = hammer2_chain_next(&parent, chain, &key_next,
1509 key_next, HAMMER2_KEY_MAX,
1513 hammer2_chain_unlock(parent);
1514 hammer2_chain_drop(parent);
1516 hammer2_inode_unlock(spmp->iroot);
1521 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1522 struct vnode *devvp, struct ucred *cred)
1526 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1527 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1528 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1530 error = hammer2_recovery(hmp);
1531 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1533 VOP_CLOSE(devvp, FREAD, NULL);
1536 VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1547 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1558 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1561 * If mount initialization proceeded far enough we must flush
1562 * its vnodes and sync the underlying mount points. Three syncs
1563 * are required to fully flush the filesystem (freemap updates lag
1564 * by one flush, and one extra for safety).
1566 if (mntflags & MNT_FORCE)
1571 error = vflush(mp, 0, flags);
1574 hammer2_vfs_sync(mp, MNT_WAIT);
1575 hammer2_vfs_sync(mp, MNT_WAIT);
1576 hammer2_vfs_sync(mp, MNT_WAIT);
1580 * Cleanup the frontend support XOPS threads
1582 hammer2_xop_helper_cleanup(pmp);
1585 hammer2_unmount_helper(mp, pmp, NULL);
1589 lockmgr(&hammer2_mntlk, LK_RELEASE);
1595 * Mount helper, hook the system mount into our PFS.
1596 * The mount lock is held.
1598 * We must bump the mount_count on related devices for any
1603 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1605 hammer2_cluster_t *cluster;
1606 hammer2_chain_t *rchain;
1609 mp->mnt_data = (qaddr_t)pmp;
1613 * After pmp->mp is set we have to adjust hmp->mount_count.
1615 cluster = &pmp->iroot->cluster;
1616 for (i = 0; i < cluster->nchains; ++i) {
1617 rchain = cluster->array[i].chain;
1620 ++rchain->hmp->mount_count;
1624 * Create missing Xop threads
1626 hammer2_xop_helper_create(pmp);
1630 * Mount helper, unhook the system mount from our PFS.
1631 * The mount lock is held.
1633 * If hmp is supplied a mount responsible for being the first to open
1634 * the block device failed and the block device and all PFSs using the
1635 * block device must be cleaned up.
1637 * If pmp is supplied multiple devices might be backing the PFS and each
1638 * must be disconnected. This might not be the last PFS using some of the
1639 * underlying devices. Also, we have to adjust our hmp->mount_count
1640 * accounting for the devices backing the pmp which is now undergoing an
1645 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1647 hammer2_cluster_t *cluster;
1648 hammer2_chain_t *rchain;
1649 struct vnode *devvp;
1655 * If no device supplied this is a high-level unmount and we have to
1656 * to disconnect the mount, adjust mount_count, and locate devices
1657 * that might now have no mounts.
1660 KKASSERT(hmp == NULL);
1661 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1663 mp->mnt_data = NULL;
1666 * After pmp->mp is cleared we have to account for
1669 cluster = &pmp->iroot->cluster;
1670 for (i = 0; i < cluster->nchains; ++i) {
1671 rchain = cluster->array[i].chain;
1674 --rchain->hmp->mount_count;
1675 /* scrapping hmp now may invalidate the pmp */
1678 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1679 if (hmp->mount_count == 0) {
1680 hammer2_unmount_helper(NULL, NULL, hmp);
1688 * Try to terminate the block device. We can't terminate it if
1689 * there are still PFSs referencing it.
1691 if (hmp->mount_count)
1695 * Decomission the network before we start messing with the
1698 hammer2_iocom_uninit(hmp);
1700 hammer2_bulkfree_uninit(hmp);
1701 hammer2_pfsfree_scan(hmp, 0);
1702 hammer2_dev_exlock(hmp); /* XXX order */
1705 * Cycle the volume data lock as a safety (probably not needed any
1706 * more). To ensure everything is out we need to flush at least
1707 * three times. (1) The running of the sideq can dirty the
1708 * filesystem, (2) A normal flush can dirty the freemap, and
1709 * (3) ensure that the freemap is fully synchronized.
1711 * The next mount's recovery scan can clean everything up but we want
1712 * to leave the filesystem in a 100% clean state on a normal unmount.
1715 hammer2_voldata_lock(hmp);
1716 hammer2_voldata_unlock(hmp);
1720 * Flush whatever is left. Unmounted but modified PFS's might still
1721 * have some dirty chains on them.
1723 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1724 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1726 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1727 hammer2_voldata_modify(hmp);
1728 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1731 hammer2_chain_unlock(&hmp->fchain);
1733 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1734 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1737 hammer2_chain_unlock(&hmp->vchain);
1739 if ((hmp->vchain.flags | hmp->fchain.flags) &
1740 HAMMER2_CHAIN_FLUSH_MASK) {
1741 kprintf("hammer2_unmount: chains left over "
1742 "after final sync\n");
1743 kprintf(" vchain %08x\n", hmp->vchain.flags);
1744 kprintf(" fchain %08x\n", hmp->fchain.flags);
1746 if (hammer2_debug & 0x0010)
1747 Debugger("entered debugger");
1750 hammer2_pfsfree_scan(hmp, 1);
1752 KKASSERT(hmp->spmp == NULL);
1755 * Finish up with the device vnode
1757 if ((devvp = hmp->devvp) != NULL) {
1759 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1760 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1761 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1763 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1764 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1765 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1767 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1774 * Clear vchain/fchain flags that might prevent final cleanup
1777 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1778 atomic_add_long(&hammer2_count_modified_chains, -1);
1779 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1780 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1782 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1783 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1786 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1787 atomic_add_long(&hammer2_count_modified_chains, -1);
1788 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1789 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1791 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1792 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1796 * Final drop of embedded freemap root chain to
1797 * clean up fchain.core (fchain structure is not
1798 * flagged ALLOCATED so it is cleaned out and then
1801 hammer2_chain_drop(&hmp->fchain);
1804 * Final drop of embedded volume root chain to clean
1805 * up vchain.core (vchain structure is not flagged
1806 * ALLOCATED so it is cleaned out and then left to
1810 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1812 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1813 hammer2_dev_unlock(hmp);
1814 hammer2_chain_drop(&hmp->vchain);
1816 hammer2_io_cleanup(hmp, &hmp->iotree);
1817 if (hmp->iofree_count) {
1818 kprintf("io_cleanup: %d I/O's left hanging\n",
1822 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1823 kmalloc_destroy(&hmp->mchain);
1824 kfree(hmp, M_HAMMER2);
1828 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1829 ino_t ino, struct vnode **vpp)
1831 hammer2_xop_lookup_t *xop;
1833 hammer2_inode_t *ip;
1837 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1843 * Easy if we already have it cached
1845 ip = hammer2_inode_lookup(pmp, inum);
1847 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1848 *vpp = hammer2_igetv(ip, &error);
1849 hammer2_inode_unlock(ip);
1850 hammer2_inode_drop(ip); /* from lookup */
1856 * Otherwise we have to find the inode
1858 xop = hammer2_xop_alloc(pmp->iroot, 0);
1860 hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1861 error = hammer2_xop_collect(&xop->head, 0);
1864 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1865 kprintf("vget: no collect error but also no rdata\n");
1866 kprintf("xop %p\n", xop);
1867 while ((hammer2_debug & 0x80000) == 0) {
1868 tsleep(xop, PCATCH, "wait", hz * 10);
1872 ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1875 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1878 *vpp = hammer2_igetv(ip, &error);
1879 hammer2_inode_unlock(ip);
1889 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1896 if (pmp->iroot == NULL) {
1902 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1904 while (pmp->inode_tid == 0) {
1905 hammer2_xop_ipcluster_t *xop;
1906 hammer2_inode_meta_t *meta;
1908 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1909 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1910 error = hammer2_xop_collect(&xop->head, 0);
1913 meta = &xop->head.cluster.focus->data->ipdata.meta;
1914 pmp->iroot->meta = *meta;
1915 pmp->inode_tid = meta->pfs_inum + 1;
1916 if (pmp->inode_tid < HAMMER2_INODE_START)
1917 pmp->inode_tid = HAMMER2_INODE_START;
1919 xop->head.cluster.focus->bref.modify_tid + 1;
1921 kprintf("PFS: Starting inode %jd\n",
1922 (intmax_t)pmp->inode_tid);
1923 kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1924 pmp->inode_tid, pmp->modify_tid);
1926 wakeup(&pmp->iroot);
1928 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1931 * Prime the mount info.
1933 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1940 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1941 hammer2_inode_unlock(pmp->iroot);
1942 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1943 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1949 hammer2_inode_unlock(pmp->iroot);
1952 vp = hammer2_igetv(pmp->iroot, &error);
1953 hammer2_inode_unlock(pmp->iroot);
1963 * XXX incorporate ipdata->meta.inode_quota and data_quota
1967 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1971 hammer2_blockref_t bref;
1976 * NOTE: iroot might not have validated the cluster yet.
1980 bzero(&tmp, sizeof(tmp));
1982 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1983 hmp = pmp->pfs_hmps[i];
1986 if (pmp->iroot->cluster.array[i].chain)
1987 bref = pmp->iroot->cluster.array[i].chain->bref;
1989 bzero(&bref, sizeof(bref));
1991 tmp.f_files = bref.embed.stats.inode_count;
1993 tmp.f_blocks = hmp->voldata.allocator_size /
1994 mp->mnt_vstat.f_bsize;
1995 tmp.f_bfree = hmp->voldata.allocator_free /
1996 mp->mnt_vstat.f_bsize;
1997 tmp.f_bavail = tmp.f_bfree;
1999 if (cred && cred->cr_uid != 0) {
2003 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2004 tmp.f_blocks -= adj;
2006 tmp.f_bavail -= adj;
2009 mp->mnt_stat.f_blocks = tmp.f_blocks;
2010 mp->mnt_stat.f_bfree = tmp.f_bfree;
2011 mp->mnt_stat.f_bavail = tmp.f_bavail;
2012 mp->mnt_stat.f_files = tmp.f_files;
2013 mp->mnt_stat.f_ffree = tmp.f_ffree;
2015 *sbp = mp->mnt_stat;
2022 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2026 hammer2_blockref_t bref;
2031 * NOTE: iroot might not have validated the cluster yet.
2034 bzero(&tmp, sizeof(tmp));
2036 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2037 hmp = pmp->pfs_hmps[i];
2040 if (pmp->iroot->cluster.array[i].chain)
2041 bref = pmp->iroot->cluster.array[i].chain->bref;
2043 bzero(&bref, sizeof(bref));
2045 tmp.f_files = bref.embed.stats.inode_count;
2047 tmp.f_blocks = hmp->voldata.allocator_size /
2048 mp->mnt_vstat.f_bsize;
2049 tmp.f_bfree = hmp->voldata.allocator_free /
2050 mp->mnt_vstat.f_bsize;
2051 tmp.f_bavail = tmp.f_bfree;
2053 if (cred && cred->cr_uid != 0) {
2057 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2058 tmp.f_blocks -= adj;
2060 tmp.f_bavail -= adj;
2063 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2064 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2065 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2066 mp->mnt_vstat.f_files = tmp.f_files;
2067 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2069 *sbp = mp->mnt_vstat;
2075 * Mount-time recovery (RW mounts)
2077 * Updates to the free block table are allowed to lag flushes by one
2078 * transaction. In case of a crash, then on a fresh mount we must do an
2079 * incremental scan of the last committed transaction id and make sure that
2080 * all related blocks have been marked allocated.
2082 * The super-root topology and each PFS has its own transaction id domain,
2083 * so we must track PFS boundary transitions.
2085 struct hammer2_recovery_elm {
2086 TAILQ_ENTRY(hammer2_recovery_elm) entry;
2087 hammer2_chain_t *chain;
2088 hammer2_tid_t sync_tid;
2091 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2093 struct hammer2_recovery_info {
2094 struct hammer2_recovery_list list;
2099 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2100 hammer2_chain_t *parent,
2101 struct hammer2_recovery_info *info,
2102 hammer2_tid_t sync_tid);
2104 #define HAMMER2_RECOVERY_MAXDEPTH 10
2108 hammer2_recovery(hammer2_dev_t *hmp)
2110 struct hammer2_recovery_info info;
2111 struct hammer2_recovery_elm *elm;
2112 hammer2_chain_t *parent;
2113 hammer2_tid_t sync_tid;
2114 hammer2_tid_t mirror_tid;
2117 hammer2_trans_init(hmp->spmp, 0);
2119 sync_tid = hmp->voldata.freemap_tid;
2120 mirror_tid = hmp->voldata.mirror_tid;
2122 kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2123 if (sync_tid >= mirror_tid) {
2124 kprintf(" no recovery needed\n");
2126 kprintf(" freemap recovery %016jx-%016jx\n",
2127 sync_tid + 1, mirror_tid);
2130 TAILQ_INIT(&info.list);
2132 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2133 error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2134 hammer2_chain_lookup_done(parent);
2136 while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2137 TAILQ_REMOVE(&info.list, elm, entry);
2138 parent = elm->chain;
2139 sync_tid = elm->sync_tid;
2140 kfree(elm, M_HAMMER2);
2142 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2143 error |= hammer2_recovery_scan(hmp, parent, &info,
2144 hmp->voldata.freemap_tid);
2145 hammer2_chain_unlock(parent);
2146 hammer2_chain_drop(parent); /* drop elm->chain ref */
2149 hammer2_trans_done(hmp->spmp);
2156 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2157 struct hammer2_recovery_info *info,
2158 hammer2_tid_t sync_tid)
2160 const hammer2_inode_data_t *ripdata;
2161 hammer2_chain_t *chain;
2162 hammer2_blockref_t bref;
2169 * Adjust freemap to ensure that the block(s) are marked allocated.
2171 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2172 hammer2_freemap_adjust(hmp, &parent->bref,
2173 HAMMER2_FREEMAP_DORECOVER);
2177 * Check type for recursive scan
2179 switch(parent->bref.type) {
2180 case HAMMER2_BREF_TYPE_VOLUME:
2181 /* data already instantiated */
2183 case HAMMER2_BREF_TYPE_INODE:
2185 * Must instantiate data for DIRECTDATA test and also
2188 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2189 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2190 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2191 /* not applicable to recovery scan */
2192 hammer2_chain_unlock(parent);
2195 hammer2_chain_unlock(parent);
2197 case HAMMER2_BREF_TYPE_INDIRECT:
2199 * Must instantiate data for recursion
2201 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2202 hammer2_chain_unlock(parent);
2204 case HAMMER2_BREF_TYPE_DIRENT:
2205 case HAMMER2_BREF_TYPE_DATA:
2206 case HAMMER2_BREF_TYPE_FREEMAP:
2207 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2208 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2209 /* not applicable to recovery scan */
2213 return HAMMER2_ERROR_BADBREF;
2217 * Defer operation if depth limit reached or if we are crossing a
2220 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2221 struct hammer2_recovery_elm *elm;
2223 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2224 elm->chain = parent;
2225 elm->sync_tid = sync_tid;
2226 hammer2_chain_ref(parent);
2227 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2228 /* unlocked by caller */
2235 * Recursive scan of the last flushed transaction only. We are
2236 * doing this without pmp assignments so don't leave the chains
2237 * hanging around after we are done with them.
2239 * error Cumulative error this level only
2240 * rup_error Cumulative error for recursion
2241 * tmp_error Specific non-cumulative recursion error
2249 error |= hammer2_chain_scan(parent, &chain, &bref,
2251 HAMMER2_LOOKUP_NODATA);
2254 * Problem during scan or EOF
2262 if (chain == NULL) {
2263 if (bref.mirror_tid > sync_tid) {
2264 hammer2_freemap_adjust(hmp, &bref,
2265 HAMMER2_FREEMAP_DORECOVER);
2271 * This may or may not be a recursive node.
2273 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2274 if (bref.mirror_tid > sync_tid) {
2276 tmp_error = hammer2_recovery_scan(hmp, chain,
2284 * Flush the recovery at the PFS boundary to stage it for
2285 * the final flush of the super-root topology.
2287 if (tmp_error == 0 &&
2288 (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2289 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2290 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2293 rup_error |= tmp_error;
2295 return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2299 * This fixes up an error introduced in earlier H2 implementations where
2300 * moving a PFS inode into an indirect block wound up causing the
2301 * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2305 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2307 const hammer2_inode_data_t *ripdata;
2308 hammer2_chain_t *parent;
2309 hammer2_chain_t *chain;
2310 hammer2_key_t key_next;
2311 hammer2_pfs_t *spmp;
2317 * Lookup mount point under the media-localized super-root.
2319 * cluster->pmp will incorrectly point to spmp and must be fixed
2323 hammer2_inode_lock(spmp->iroot, 0);
2324 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2325 chain = hammer2_chain_lookup(&parent, &key_next,
2326 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2329 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2332 kprintf("I/O error scanning PFS labels\n");
2333 error |= chain->error;
2334 } else if ((chain->bref.flags &
2335 HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2338 ripdata = &chain->data->ipdata;
2339 hammer2_trans_init(hmp->spmp, 0);
2340 error2 = hammer2_chain_modify(chain,
2341 chain->bref.modify_tid,
2344 kprintf("hammer2: Correct mis-flagged PFS %s\n",
2346 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2350 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2352 hammer2_trans_done(hmp->spmp);
2354 chain = hammer2_chain_next(&parent, chain, &key_next,
2355 key_next, HAMMER2_KEY_MAX,
2359 hammer2_chain_unlock(parent);
2360 hammer2_chain_drop(parent);
2362 hammer2_inode_unlock(spmp->iroot);
2368 * Sync a mount point; this is called periodically on a per-mount basis from
2369 * the filesystem syncer, and whenever a user issues a sync.
2372 hammer2_vfs_sync(struct mount *mp, int waitfor)
2374 hammer2_xop_flush_t *xop;
2375 struct hammer2_sync_info info;
2376 hammer2_inode_t *iroot;
2384 KKASSERT(iroot->pmp == pmp);
2387 * We can't acquire locks on existing vnodes while in a transaction
2388 * without risking a deadlock. This assumes that vfsync() can be
2389 * called without the vnode locked (which it can in DragonFly).
2390 * Otherwise we'd have to implement a multi-pass or flag the lock
2391 * failures and retry.
2393 * The reclamation code interlocks with the sync list's token
2394 * (by removing the vnode from the scan list) before unlocking
2395 * the inode, giving us time to ref the inode.
2397 /*flags = VMSC_GETVP;*/
2399 if (waitfor & MNT_LAZY)
2400 flags |= VMSC_ONEPASS;
2403 * Flush vnodes individually using a normal transaction to avoid
2404 * stalling any concurrent operations. This will flush the related
2405 * buffer cache buffers and inodes to the media.
2407 * For efficiency do an async pass before making sure with a
2408 * synchronous pass on all related buffer cache buffers.
2410 hammer2_trans_init(pmp, 0);
2414 info.waitfor = MNT_NOWAIT;
2416 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2419 * Now do two passes making sure we get everything. The first pass
2420 * vfsync()s dirty vnodes. The second pass waits for their I/O's
2421 * to finish and cleans up the dirty flag on the vnode.
2424 info.waitfor = MNT_WAIT;
2425 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2428 info.waitfor = MNT_WAIT;
2429 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2432 * We must also run the sideq to handle any disconnected inodes
2433 * as the vnode scan will not see these.
2435 hammer2_inode_run_sideq(pmp, 1);
2436 hammer2_trans_done(pmp);
2439 * Start our flush transaction and flush the root topology down to
2440 * the inodes, but not the inodes themselves (which we already flushed
2441 * above). Any concurrent activity effecting inode contents will not
2443 * The flush sequence will
2445 * NOTE! It is still possible for the paging code to push pages
2446 * out via a UIO_NOCOPY hammer2_vop_write() during the main
2449 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2452 * sync dirty vnodes again while in the flush transaction. This is
2453 * currently an expensive shim to makre sure the logical topology is
2454 * completely consistent before we flush the volume header.
2457 info.waitfor = MNT_WAIT;
2458 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2461 info.waitfor = MNT_WAIT;
2462 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2465 * Use the XOP interface to concurrently flush all nodes to
2466 * synchronize the PFSROOT subtopology to the media. A standard
2467 * end-of-scan ENOENT error indicates cluster sufficiency.
2469 * Note that this flush will not be visible on crash recovery until
2470 * we flush the super-root topology in the next loop.
2472 * XXX For now wait for all flushes to complete.
2476 * If unmounting try to flush everything including any
2477 * sub-trees under inodes, just in case there is dangling
2478 * modified data, as a safety. Otherwise just flush up to
2479 * the inodes in this stage.
2481 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2482 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2483 HAMMER2_XOP_VOLHDR);
2485 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2486 HAMMER2_XOP_INODE_STOP |
2487 HAMMER2_XOP_VOLHDR);
2489 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2490 error = hammer2_xop_collect(&xop->head,
2491 HAMMER2_XOP_COLLECT_WAITALL);
2492 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2493 if (error == HAMMER2_ERROR_ENOENT)
2496 error = hammer2_error_to_errno(error);
2500 hammer2_trans_done(pmp);
2508 * Note that we ignore the tranasction mtid we got above. Instead,
2509 * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2512 * WARNING! The frontend might be waiting on chnmem (limit_dirty_chains)
2513 * while holding a vnode locked. When this situation occurs we cannot
2514 * safely test whether it is ok to clear the dirty bit on the vnode.
2515 * However, we can still flush the inode's topology.
2518 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2520 struct hammer2_sync_info *info = data;
2521 hammer2_inode_t *ip;
2525 * Degenerate cases. Note that ip == NULL typically means the
2526 * syncer vnode itself and we don't want to vclrisdirty() in that
2533 if (vp->v_type == VNON || vp->v_type == VBAD) {
2539 * Synchronize the buffer cche and inode meta-data to the backing
2542 * vfsync is not necessarily synchronous, so it is best NOT to try
2543 * to flush the backing topology to media at this point.
2545 hammer2_inode_ref(ip);
2546 if ((ip->flags & (HAMMER2_INODE_RESIZED|HAMMER2_INODE_MODIFIED)) ||
2547 !RB_EMPTY(&vp->v_rbdirty_tree)) {
2548 if (info->pass == 1)
2549 vfsync(vp, info->waitfor, 1, NULL, NULL);
2551 bio_track_wait(&vp->v_track_write, 0, 0);
2553 if (info->pass == 2 && (vp->v_flag & VISDIRTY)) {
2555 * v_token is needed to interlock v_rbdirty_tree.
2557 lwkt_gettoken(&vp->v_token);
2558 hammer2_inode_lock(ip, 0);
2559 hammer2_inode_chain_sync(ip);
2560 hammer2_inode_chain_flush(ip);
2561 if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2562 HAMMER2_INODE_RESIZED |
2563 HAMMER2_INODE_DIRTYDATA)) == 0 &&
2564 RB_EMPTY(&vp->v_rbdirty_tree) &&
2565 !bio_track_active(&vp->v_track_write)) {
2568 hammer2_inode_unlock(ip);
2569 lwkt_reltoken(&vp->v_token);
2571 hammer2_inode_drop(ip);
2575 info->error = error;
2582 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2584 hammer2_inode_t *ip;
2586 KKASSERT(MAXFIDSZ >= 16);
2588 fhp->fid_len = offsetof(struct fid, fid_data[16]);
2590 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2591 ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2598 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2599 struct fid *fhp, struct vnode **vpp)
2606 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2609 error = hammer2_vfs_root(mp, vpp);
2611 error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2616 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2622 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2623 int *exflagsp, struct ucred **credanonp)
2630 np = vfs_export_lookup(mp, &pmp->export, nam);
2632 *exflagsp = np->netc_exflags;
2633 *credanonp = &np->netc_anon;
2642 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume
2643 * header into the HMP
2645 * XXX read four volhdrs and use the one with the highest TID whos CRC
2650 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2651 * nonexistant locations.
2653 * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2657 hammer2_install_volume_header(hammer2_dev_t *hmp)
2659 hammer2_volume_data_t *vd;
2661 hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2673 * There are up to 4 copies of the volume header (syncs iterate
2674 * between them so there is no single master). We don't trust the
2675 * volu_size field so we don't know precisely how large the filesystem
2676 * is, so depend on the OS to return an error if we go beyond the
2677 * block device's EOF.
2679 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2680 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2681 HAMMER2_VOLUME_BYTES, &bp);
2688 vd = (struct hammer2_volume_data *) bp->b_data;
2689 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2690 (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2696 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2697 /* XXX: Reversed-endianness filesystem */
2698 kprintf("hammer2: reverse-endian filesystem detected");
2704 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2705 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2706 HAMMER2_VOLUME_ICRC0_SIZE);
2707 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2708 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2709 HAMMER2_VOLUME_ICRC1_SIZE);
2710 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2711 kprintf("hammer2 volume header crc "
2712 "mismatch copy #%d %08x/%08x\n",
2719 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2728 hmp->volsync = hmp->voldata;
2729 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2731 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2732 kprintf("hammer2: using volume header #%d\n",
2737 kprintf("hammer2: no valid volume headers found!\n");
2743 * This handles hysteresis on regular file flushes. Because the BIOs are
2744 * routed to a thread it is possible for an excessive number to build up
2745 * and cause long front-end stalls long before the runningbuffspace limit
2746 * is hit, so we implement hammer2_flush_pipe to control the
2749 * This is a particular problem when compression is used.
2752 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2754 atomic_add_int(&pmp->count_lwinprog, 1);
2758 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2762 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2763 if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2764 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2765 atomic_clear_int(&pmp->count_lwinprog,
2766 HAMMER2_LWINPROG_WAITING);
2767 wakeup(&pmp->count_lwinprog);
2769 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2770 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2771 atomic_clear_int(&pmp->count_lwinprog,
2772 HAMMER2_LWINPROG_WAITING0);
2773 wakeup(&pmp->count_lwinprog);
2778 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2781 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2782 HAMMER2_LWINPROG_WAITING0;
2785 lwinprog = pmp->count_lwinprog;
2787 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2789 tsleep_interlock(&pmp->count_lwinprog, 0);
2790 atomic_set_int(&pmp->count_lwinprog, lwflag);
2791 lwinprog = pmp->count_lwinprog;
2792 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2794 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2799 * Manage excessive memory resource use for chain and related
2803 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2813 * Atomic check condition and wait. Also do an early speedup of
2814 * the syncer to try to avoid hitting the wait.
2817 waiting = pmp->inmem_dirty_chains;
2819 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2821 limit = pmp->mp->mnt_nvnodelistsize / 10;
2822 if (limit < hammer2_limit_dirty_chains)
2823 limit = hammer2_limit_dirty_chains;
2828 if ((int)(ticks - zzticks) > hz) {
2830 kprintf("count %ld %ld\n", count, limit);
2835 * Block if there are too many dirty chains present, wait
2836 * for the flush to clean some out.
2838 if (count > limit) {
2839 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2840 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2842 waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2843 speedup_syncer(pmp->mp);
2844 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2847 continue; /* loop on success or fail */
2851 * Try to start an early flush before we are forced to block.
2853 if (count > limit * 5 / 10)
2854 speedup_syncer(pmp->mp);
2860 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2863 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2868 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2873 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
2874 /* don't need --waiting to test flag */
2875 if (waiting & HAMMER2_DIRTYCHAIN_WAITING) {
2876 atomic_clear_int(&pmp->inmem_dirty_chains,
2877 HAMMER2_DIRTYCHAIN_WAITING);
2878 wakeup(&pmp->inmem_dirty_chains);
2884 * Returns 0 if the filesystem has tons of free space
2885 * Returns 1 if the filesystem has less than 10% remaining
2886 * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2889 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2893 hammer2_off_t free_reserved;
2894 hammer2_off_t free_nominal;
2899 if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2900 free_reserved = HAMMER2_SEGSIZE;
2901 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2902 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2903 hmp = pmp->pfs_hmps[i];
2906 if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2907 pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2910 if (free_nominal > hmp->voldata.allocator_free)
2911 free_nominal = hmp->voldata.allocator_free;
2912 if (free_reserved < hmp->free_reserved)
2913 free_reserved = hmp->free_reserved;
2919 pmp->free_reserved = free_reserved;
2920 pmp->free_nominal = free_nominal;
2921 pmp->free_ticks = ticks;
2923 free_reserved = pmp->free_reserved;
2924 free_nominal = pmp->free_nominal;
2926 if (cred && cred->cr_uid != 0) {
2927 if ((int64_t)(free_nominal - bytes) <
2928 (int64_t)free_reserved) {
2932 if ((int64_t)(free_nominal - bytes) <
2933 (int64_t)free_reserved / 2) {
2937 if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2946 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
2949 hammer2_chain_t *scan;
2950 hammer2_chain_t *parent;
2954 kprintf("%*.*s...\n", tab, tab, "");
2959 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2961 chain, chain->bref.type,
2962 chain->bref.key, chain->bref.keybits,
2963 chain->bref.mirror_tid);
2965 kprintf("%*.*s [%08x] (%s) refs=%d",
2968 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2969 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
2972 parent = chain->parent;
2974 kprintf("\n%*.*s p=%p [pflags %08x prefs %d",
2976 parent, parent->flags, parent->refs);
2977 if (RB_EMPTY(&chain->core.rbtree)) {
2981 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
2982 if ((scan->flags & flags) || flags == (u_int)-1) {
2983 hammer2_dump_chain(scan, tab + 4, countp, 'a',
2987 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2988 kprintf("%*.*s}(%s)\n", tab, tab, "",
2989 chain->data->ipdata.filename);
2991 kprintf("%*.*s}\n", tab, tab, "");