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 long hammer2_debug_inode;
85 int hammer2_cluster_meta_read = 1; /* physical read-ahead */
86 int hammer2_cluster_data_read = 4; /* physical read-ahead */
87 int hammer2_cluster_write = 0; /* physical write clustering */
88 int hammer2_dedup_enable = 1;
89 int hammer2_always_compress = 0; /* always try to compress */
90 int hammer2_inval_enable = 0;
91 int hammer2_flush_pipe = 100;
92 int hammer2_dio_count;
93 int hammer2_dio_limit = 256;
94 int hammer2_bulkfree_tps = 5000;
95 int hammer2_worker_rmask = 3;
96 long hammer2_chain_allocs;
97 long hammer2_chain_frees;
98 long hammer2_limit_dirty_chains;
99 long hammer2_limit_dirty_inodes;
100 long hammer2_count_modified_chains;
101 long hammer2_iod_invals;
102 long hammer2_iod_file_read;
103 long hammer2_iod_meta_read;
104 long hammer2_iod_indr_read;
105 long hammer2_iod_fmap_read;
106 long hammer2_iod_volu_read;
107 long hammer2_iod_file_write;
108 long hammer2_iod_file_wembed;
109 long hammer2_iod_file_wzero;
110 long hammer2_iod_file_wdedup;
111 long hammer2_iod_meta_write;
112 long hammer2_iod_indr_write;
113 long hammer2_iod_fmap_write;
114 long hammer2_iod_volu_write;
115 long hammer2_iod_inode_creates;
116 long hammer2_iod_inode_deletes;
118 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
119 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
120 "Buffer used for compression.");
122 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
123 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
124 "Buffer used for decompression.");
126 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
129 &hammer2_supported_version, 0, "");
130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
131 &hammer2_debug, 0, "");
132 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW,
133 &hammer2_debug_inode, 0, "");
134 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
135 &hammer2_cluster_meta_read, 0, "");
136 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
137 &hammer2_cluster_data_read, 0, "");
138 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
139 &hammer2_cluster_write, 0, "");
140 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
141 &hammer2_dedup_enable, 0, "");
142 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
143 &hammer2_always_compress, 0, "");
144 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
145 &hammer2_inval_enable, 0, "");
146 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
147 &hammer2_flush_pipe, 0, "");
148 SYSCTL_INT(_vfs_hammer2, OID_AUTO, worker_rmask, CTLFLAG_RW,
149 &hammer2_worker_rmask, 0, "");
150 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
151 &hammer2_bulkfree_tps, 0, "");
152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
153 &hammer2_chain_allocs, 0, "");
154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
155 &hammer2_chain_frees, 0, "");
156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
157 &hammer2_limit_dirty_chains, 0, "");
158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
159 &hammer2_limit_dirty_inodes, 0, "");
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
161 &hammer2_count_modified_chains, 0, "");
162 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
163 &hammer2_dio_count, 0, "");
164 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
165 &hammer2_dio_limit, 0, "");
167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
168 &hammer2_iod_invals, 0, "");
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
170 &hammer2_iod_file_read, 0, "");
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
172 &hammer2_iod_meta_read, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
174 &hammer2_iod_indr_read, 0, "");
175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
176 &hammer2_iod_fmap_read, 0, "");
177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
178 &hammer2_iod_volu_read, 0, "");
180 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
181 &hammer2_iod_file_write, 0, "");
182 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
183 &hammer2_iod_file_wembed, 0, "");
184 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
185 &hammer2_iod_file_wzero, 0, "");
186 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
187 &hammer2_iod_file_wdedup, 0, "");
188 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
189 &hammer2_iod_meta_write, 0, "");
190 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
191 &hammer2_iod_indr_write, 0, "");
192 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
193 &hammer2_iod_fmap_write, 0, "");
194 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
195 &hammer2_iod_volu_write, 0, "");
196 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW,
197 &hammer2_iod_inode_creates, 0, "");
198 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW,
199 &hammer2_iod_inode_deletes, 0, "");
201 long hammer2_process_icrc32;
202 long hammer2_process_xxhash64;
203 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
204 &hammer2_process_icrc32, 0, "");
205 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
206 &hammer2_process_xxhash64, 0, "");
208 static int hammer2_vfs_init(struct vfsconf *conf);
209 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
210 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
212 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
213 struct vnode *, struct ucred *);
214 static int hammer2_recovery(hammer2_dev_t *hmp);
215 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
216 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
217 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
219 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
221 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
222 struct fid *fhp, struct vnode **vpp);
223 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
224 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
225 int *exflagsp, struct ucred **credanonp);
226 static void hammer2_vfs_modifying(struct mount *mp);
228 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
230 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
233 static void hammer2_update_pmps(hammer2_dev_t *hmp);
235 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
236 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
238 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
241 * HAMMER2 vfs operations.
243 static struct vfsops hammer2_vfsops = {
244 .vfs_init = hammer2_vfs_init,
245 .vfs_uninit = hammer2_vfs_uninit,
246 .vfs_sync = hammer2_vfs_sync,
247 .vfs_mount = hammer2_vfs_mount,
248 .vfs_unmount = hammer2_vfs_unmount,
249 .vfs_root = hammer2_vfs_root,
250 .vfs_statfs = hammer2_vfs_statfs,
251 .vfs_statvfs = hammer2_vfs_statvfs,
252 .vfs_vget = hammer2_vfs_vget,
253 .vfs_vptofh = hammer2_vfs_vptofh,
254 .vfs_fhtovp = hammer2_vfs_fhtovp,
255 .vfs_checkexp = hammer2_vfs_checkexp,
256 .vfs_modifying = hammer2_vfs_modifying
259 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
261 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
262 MODULE_VERSION(hammer2, 1);
266 hammer2_vfs_init(struct vfsconf *conf)
268 static struct objcache_malloc_args margs_read;
269 static struct objcache_malloc_args margs_write;
270 static struct objcache_malloc_args margs_vop;
275 kmalloc_raise_limit(M_HAMMER2, 0); /* unlimited */
278 * A large DIO cache is needed to retain dedup enablement masks.
279 * The bulkfree code clears related masks as part of the disk block
280 * recycling algorithm, preventing it from being used for a later
283 * NOTE: A large buffer cache can actually interfere with dedup
284 * operation because we dedup based on media physical buffers
285 * and not logical buffers. Try to make the DIO case large
286 * enough to avoid this problem, but also cap it.
288 hammer2_dio_limit = nbuf * 2;
289 if (hammer2_dio_limit > 100000)
290 hammer2_dio_limit = 100000;
292 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
294 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
296 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
300 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
302 margs_read.objsize = 65536;
303 margs_read.mtype = M_HAMMER2_DEBUFFER;
305 margs_write.objsize = 32768;
306 margs_write.mtype = M_HAMMER2_CBUFFER;
308 margs_vop.objsize = sizeof(hammer2_xop_t);
309 margs_vop.mtype = M_HAMMER2;
312 * Note thaht for the XOPS cache we want backing store allocations
313 * to use M_ZERO. This is not allowed in objcache_get() (to avoid
314 * confusion), so use the backing store function that does it. This
315 * means that initial XOPS objects are zerod but REUSED objects are
316 * not. So we are responsible for cleaning the object up sufficiently
317 * for our needs before objcache_put()ing it back (typically just the
320 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
321 0, 1, NULL, NULL, NULL,
322 objcache_malloc_alloc,
323 objcache_malloc_free,
325 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
326 0, 1, NULL, NULL, NULL,
327 objcache_malloc_alloc,
328 objcache_malloc_free,
330 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
331 0, 1, NULL, NULL, NULL,
332 objcache_malloc_alloc_zero,
333 objcache_malloc_free,
337 lockinit(&hammer2_mntlk, "mntlk", 0, 0);
338 TAILQ_INIT(&hammer2_mntlist);
339 TAILQ_INIT(&hammer2_pfslist);
340 TAILQ_INIT(&hammer2_spmplist);
342 hammer2_limit_dirty_chains = maxvnodes / 10;
343 if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
344 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
345 if (hammer2_limit_dirty_chains < 1000)
346 hammer2_limit_dirty_chains = 1000;
348 hammer2_limit_dirty_inodes = maxvnodes / 25;
349 if (hammer2_limit_dirty_inodes < 100)
350 hammer2_limit_dirty_inodes = 100;
351 if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
352 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
359 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
361 objcache_destroy(cache_buffer_read);
362 objcache_destroy(cache_buffer_write);
363 objcache_destroy(cache_xops);
368 * Core PFS allocator. Used to allocate or reference the pmp structure
369 * for PFS cluster mounts and the spmp structure for media (hmp) structures.
370 * The pmp can be passed in or loaded by this function using the chain and
373 * pmp->modify_tid tracks new modify_tid transaction ids for front-end
374 * transactions. Note that synchronization does not use this field.
375 * (typically frontend operations and synchronization cannot run on the
376 * same PFS node at the same time).
381 hammer2_pfsalloc(hammer2_chain_t *chain,
382 const hammer2_inode_data_t *ripdata,
383 hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
386 hammer2_inode_t *iroot;
394 * Locate or create the PFS based on the cluster id. If ripdata
395 * is NULL this is a spmp which is unique and is always allocated.
397 * If the device is mounted in local mode all PFSs are considered
398 * independent and not part of any cluster (for debugging only).
401 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
402 if (force_local != pmp->force_local)
404 if (force_local == NULL &&
405 bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
406 sizeof(pmp->pfs_clid)) == 0) {
408 } else if (force_local && pmp->pfs_names[0] &&
409 strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
416 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
417 pmp->force_local = force_local;
418 hammer2_trans_manage_init(pmp);
419 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
420 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
421 lockinit(&pmp->lock, "pfslk", 0, 0);
422 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
423 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
424 spin_init(&pmp->xop_spin, "h2xop");
425 spin_init(&pmp->lru_spin, "h2lru");
426 RB_INIT(&pmp->inum_tree);
427 TAILQ_INIT(&pmp->syncq);
428 TAILQ_INIT(&pmp->depq);
429 TAILQ_INIT(&pmp->lru_list);
430 spin_init(&pmp->list_spin, "h2pfsalloc_list");
433 * Distribute backend operations to threads
435 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
436 hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
439 * Save the last media transaction id for the flusher. Set
443 pmp->pfs_clid = ripdata->meta.pfs_clid;
444 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
446 pmp->flags |= HAMMER2_PMPF_SPMP;
447 TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
451 * The synchronization thread may start too early, make
452 * sure it stays frozen until we are ready to let it go.
456 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
457 HAMMER2_THREAD_REMASTER;
462 * Create the PFS's root inode and any missing XOP helper threads.
464 if ((iroot = pmp->iroot) == NULL) {
465 iroot = hammer2_inode_get(pmp, NULL, 1, -1);
467 iroot->meta = ripdata->meta;
469 hammer2_inode_ref(iroot);
470 hammer2_inode_unlock(iroot);
474 * Stop here if no chain is passed in.
480 * When a chain is passed in we must add it to the PFS's root
481 * inode, update pmp->pfs_types[], and update the syncronization
484 * When forcing local mode, mark the PFS as a MASTER regardless.
486 * At the moment empty spots can develop due to removals or failures.
487 * Ultimately we want to re-fill these spots but doing so might
488 * confused running code. XXX
490 hammer2_inode_ref(iroot);
491 hammer2_mtx_ex(&iroot->lock);
492 j = iroot->cluster.nchains;
494 if (j == HAMMER2_MAXCLUSTER) {
495 kprintf("hammer2_mount: cluster full!\n");
496 /* XXX fatal error? */
498 KKASSERT(chain->pmp == NULL);
500 hammer2_chain_ref(chain);
501 iroot->cluster.array[j].chain = chain;
503 pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
505 pmp->pfs_types[j] = ripdata->meta.pfs_type;
506 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
507 pmp->pfs_hmps[j] = chain->hmp;
508 hammer2_spin_ex(&pmp->inum_spin);
509 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset;
510 hammer2_spin_unex(&pmp->inum_spin);
513 * If the PFS is already mounted we must account
514 * for the mount_count here.
517 ++chain->hmp->mount_count;
520 * May have to fixup dirty chain tracking. Previous
521 * pmp was NULL so nothing to undo.
523 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
524 hammer2_pfs_memory_inc(pmp);
527 iroot->cluster.nchains = j;
530 * Update nmasters from any PFS inode which is part of the cluster.
531 * It is possible that this will result in a value which is too
532 * high. MASTER PFSs are authoritative for pfs_nmasters and will
533 * override this value later on.
535 * (This informs us of masters that might not currently be
536 * discoverable by this mount).
538 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
539 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
543 * Count visible masters. Masters are usually added with
544 * ripdata->meta.pfs_nmasters set to 1. This detects when there
545 * are more (XXX and must update the master inodes).
548 for (i = 0; i < iroot->cluster.nchains; ++i) {
549 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
552 if (pmp->pfs_nmasters < count)
553 pmp->pfs_nmasters = count;
556 * Create missing synchronization and support threads.
558 * Single-node masters (including snapshots) have nothing to
559 * synchronize and do not require this thread.
561 * Multi-node masters or any number of soft masters, slaves, copy,
562 * or other PFS types need the thread.
564 * Each thread is responsible for its particular cluster index.
565 * We use independent threads so stalls or mismatches related to
566 * any given target do not affect other targets.
568 for (i = 0; i < iroot->cluster.nchains; ++i) {
570 * Single-node masters (including snapshots) have nothing
571 * to synchronize and will make direct xops support calls,
572 * thus they do not require this thread.
574 * Note that there can be thousands of snapshots. We do not
575 * want to create thousands of threads.
577 if (pmp->pfs_nmasters <= 1 &&
578 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
583 * Sync support thread
585 if (pmp->sync_thrs[i].td == NULL) {
586 hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
588 hammer2_primary_sync_thread);
593 * Create missing Xop threads
595 * NOTE: We create helper threads for all mounted PFSs or any
596 * PFSs with 2+ nodes (so the sync thread can update them,
597 * even if not mounted).
599 if (pmp->mp || iroot->cluster.nchains >= 2)
600 hammer2_xop_helper_create(pmp);
602 hammer2_mtx_unlock(&iroot->lock);
603 hammer2_inode_drop(iroot);
609 * Deallocate an element of a probed PFS. If destroying and this is a
610 * MASTER, adjust nmasters.
612 * This function does not physically destroy the PFS element in its device
613 * under the super-root (see hammer2_ioctl_pfs_delete()).
616 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
618 hammer2_inode_t *iroot;
619 hammer2_chain_t *chain;
623 * Cleanup our reference on iroot. iroot is (should) not be needed
631 * XXX flush after acquiring the iroot lock.
632 * XXX clean out the cluster index from all inode structures.
634 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
637 * Remove the cluster index from the group. If destroying
638 * the PFS and this is a master, adjust pfs_nmasters.
640 hammer2_mtx_ex(&iroot->lock);
641 chain = iroot->cluster.array[clindex].chain;
642 iroot->cluster.array[clindex].chain = NULL;
644 switch(pmp->pfs_types[clindex]) {
645 case HAMMER2_PFSTYPE_MASTER:
646 if (destroying && pmp->pfs_nmasters > 0)
648 /* XXX adjust ripdata->meta.pfs_nmasters */
653 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
655 hammer2_mtx_unlock(&iroot->lock);
661 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
662 hammer2_chain_drop(chain);
666 * Terminate all XOP threads for the cluster index.
668 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
669 hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
674 * Destroy a PFS, typically only occurs after the last mount on a device
678 hammer2_pfsfree(hammer2_pfs_t *pmp)
680 hammer2_inode_t *iroot;
681 hammer2_chain_t *chain;
682 int chains_still_present = 0;
687 * Cleanup our reference on iroot. iroot is (should) not be needed
690 if (pmp->flags & HAMMER2_PMPF_SPMP)
691 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
693 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
696 * Cleanup chains remaining on LRU list.
698 hammer2_spin_ex(&pmp->lru_spin);
699 while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
700 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
701 atomic_add_int(&pmp->lru_count, -1);
702 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
703 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
704 hammer2_chain_ref(chain);
705 hammer2_spin_unex(&pmp->lru_spin);
706 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
707 hammer2_chain_drop(chain);
708 hammer2_spin_ex(&pmp->lru_spin);
710 hammer2_spin_unex(&pmp->lru_spin);
717 for (i = 0; i < iroot->cluster.nchains; ++i) {
718 hammer2_thr_delete(&pmp->sync_thrs[i]);
719 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
720 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
721 chain = iroot->cluster.array[i].chain;
722 if (chain && !RB_EMPTY(&chain->core.rbtree)) {
723 kprintf("hammer2: Warning pmp %p still "
724 "has active chains\n", pmp);
725 chains_still_present = 1;
728 #if REPORT_REFS_ERRORS
729 if (iroot->refs != 1)
730 kprintf("PMP->IROOT %p REFS WRONG %d\n",
733 KKASSERT(iroot->refs == 1);
736 hammer2_inode_drop(iroot);
741 * Free remaining pmp resources
743 if (chains_still_present) {
744 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp);
746 kmalloc_destroy(&pmp->mmsg);
747 kmalloc_destroy(&pmp->minode);
748 kfree(pmp, M_HAMMER2);
753 * Remove all references to hmp from the pfs list. Any PFS which becomes
754 * empty is terminated and freed.
759 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
762 hammer2_inode_t *iroot;
763 hammer2_chain_t *rchain;
766 struct hammer2_pfslist *wlist;
769 wlist = &hammer2_pfslist;
771 wlist = &hammer2_spmplist;
773 TAILQ_FOREACH(pmp, wlist, mntentry) {
774 if ((iroot = pmp->iroot) == NULL)
778 * Determine if this PFS is affected. If it is we must
779 * freeze all management threads and lock its iroot.
781 * Freezing a management thread forces it idle, operations
782 * in-progress will be aborted and it will have to start
783 * over again when unfrozen, or exit if told to exit.
785 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
786 if (pmp->pfs_hmps[i] == hmp)
789 if (i == HAMMER2_MAXCLUSTER)
792 hammer2_vfs_sync_pmp(pmp, MNT_WAIT);
795 * Make sure all synchronization threads are locked
798 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
799 if (pmp->pfs_hmps[i] == NULL)
801 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
802 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
803 hammer2_thr_freeze_async(
804 &pmp->xop_groups[j].thrs[i]);
807 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
808 if (pmp->pfs_hmps[i] == NULL)
810 hammer2_thr_freeze(&pmp->sync_thrs[i]);
811 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
813 &pmp->xop_groups[j].thrs[i]);
818 * Lock the inode and clean out matching chains.
819 * Note that we cannot use hammer2_inode_lock_*()
820 * here because that would attempt to validate the
821 * cluster that we are in the middle of ripping
824 * WARNING! We are working directly on the inodes
827 hammer2_mtx_ex(&iroot->lock);
830 * Remove the chain from matching elements of the PFS.
832 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
833 if (pmp->pfs_hmps[i] != hmp)
835 hammer2_thr_delete(&pmp->sync_thrs[i]);
836 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
838 &pmp->xop_groups[j].thrs[i]);
840 rchain = iroot->cluster.array[i].chain;
841 iroot->cluster.array[i].chain = NULL;
842 pmp->pfs_types[i] = 0;
843 if (pmp->pfs_names[i]) {
844 kfree(pmp->pfs_names[i], M_HAMMER2);
845 pmp->pfs_names[i] = NULL;
848 hammer2_chain_drop(rchain);
850 if (iroot->cluster.focus == rchain)
851 iroot->cluster.focus = NULL;
853 pmp->pfs_hmps[i] = NULL;
855 hammer2_mtx_unlock(&iroot->lock);
858 * Cleanup trailing chains. Gaps may remain.
860 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
861 if (pmp->pfs_hmps[i])
864 iroot->cluster.nchains = i + 1;
867 * If the PMP has no elements remaining we can destroy it.
868 * (this will transition management threads from frozen->exit).
870 if (iroot->cluster.nchains == 0) {
872 * If this was the hmp's spmp, we need to clean
873 * a little more stuff out.
875 if (hmp->spmp == pmp) {
877 hmp->vchain.pmp = NULL;
878 hmp->fchain.pmp = NULL;
882 * Free the pmp and restart the loop
884 KKASSERT(TAILQ_EMPTY(&pmp->syncq));
885 KKASSERT(TAILQ_EMPTY(&pmp->depq));
886 hammer2_pfsfree(pmp);
891 * If elements still remain we need to set the REMASTER
892 * flag and unfreeze it.
894 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
895 if (pmp->pfs_hmps[i] == NULL)
897 hammer2_thr_remaster(&pmp->sync_thrs[i]);
898 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
899 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
900 hammer2_thr_remaster(
901 &pmp->xop_groups[j].thrs[i]);
902 hammer2_thr_unfreeze(
903 &pmp->xop_groups[j].thrs[i]);
910 * Mount or remount HAMMER2 fileystem from physical media
913 * mp mount point structure
919 * mp mount point structure
920 * path path to mount point
921 * data pointer to argument structure in user space
922 * volume volume path (device@LABEL form)
923 * hflags user mount flags
924 * cred user credentials
931 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
934 struct hammer2_mount_info info;
938 hammer2_dev_t *force_local;
939 hammer2_key_t key_next;
940 hammer2_key_t key_dummy;
943 struct nlookupdata nd;
944 hammer2_chain_t *parent;
945 hammer2_chain_t *chain;
946 const hammer2_inode_data_t *ripdata;
947 hammer2_blockref_t bref;
949 char devstr[MNAMELEN];
968 bzero(&info, sizeof(info));
969 info.cluster_fd = -1;
970 ksnprintf(devstr, sizeof(devstr), "%s",
971 mp->mnt_stat.f_mntfromname);
972 kprintf("hammer2_mount: root '%s'\n", devstr);
973 done = strlen(devstr) + 1;
976 * Non-root mount or updating a mount
978 error = copyin(data, &info, sizeof(info));
982 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
985 kprintf("hammer2_mount: '%s'\n", devstr);
989 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
990 * if no label specified, based on the partition id. Error out if no
991 * label or device (with partition id) is specified. This is strictly
992 * a convenience to match the default label created by newfs_hammer2,
993 * our preference is that a label always be specified.
995 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
996 * that does not specify a device, as long as some H2 label
997 * has already been mounted from that device. This makes
998 * mounting snapshots a lot easier.
1001 label = strchr(devstr, '@');
1002 if (label && ((label + 1) - dev) > done) {
1003 kprintf("hammer2: mount: bad label %s/%zd\n",
1007 if (label == NULL || label[1] == 0) {
1011 label = devstr + strlen(devstr);
1013 *label = '\0'; /* clean up trailing @ */
1032 kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
1033 dev, label, (mp->mnt_flag & MNT_RDONLY));
1035 if (mp->mnt_flag & MNT_UPDATE) {
1037 * Update mount. Note that pmp->iroot->cluster is
1038 * an inode-embedded cluster and thus cannot be
1041 * XXX HAMMER2 needs to implement NFS export via
1044 hammer2_cluster_t *cluster;
1047 pmp->hflags = info.hflags;
1048 cluster = &pmp->iroot->cluster;
1049 for (i = 0; i < cluster->nchains; ++i) {
1050 if (cluster->array[i].chain == NULL)
1052 hmp = cluster->array[i].chain->hmp;
1054 error = hammer2_remount(hmp, mp, path,
1066 * If a path is specified and dev is not an empty string, lookup the
1067 * name and verify that it referes to a block device.
1069 * If a path is specified and dev is an empty string we fall through
1070 * and locate the label in the hmp search.
1072 if (path && *dev != 0) {
1073 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1075 error = nlookup(&nd);
1077 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1079 } else if (path == NULL) {
1081 cdev_t cdev = kgetdiskbyname(dev);
1082 error = bdevvp(cdev, &devvp);
1084 kprintf("hammer2: cannot find '%s'\n", dev);
1087 * We will locate the hmp using the label in the hmp loop.
1093 * Make sure its a block device. Do not check to see if it is
1094 * already mounted until we determine that its a fresh H2 device.
1096 if (error == 0 && devvp) {
1097 vn_isdisk(devvp, &error);
1101 * Determine if the device has already been mounted. After this
1102 * check hmp will be non-NULL if we are doing the second or more
1103 * hammer2 mounts from the same device.
1105 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1108 * Match the device. Due to the way devfs works,
1109 * we may not be able to directly match the vnode pointer,
1110 * so also check to see if the underlying device matches.
1112 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1113 if (hmp->devvp == devvp)
1115 if (devvp->v_rdev &&
1116 hmp->devvp->v_rdev == devvp->v_rdev) {
1122 * If no match this may be a fresh H2 mount, make sure
1123 * the device is not mounted on anything else.
1126 error = vfs_mountedon(devvp);
1127 } else if (error == 0) {
1129 * Match the label to a pmp already probed.
1131 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1132 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1133 if (pmp->pfs_names[i] &&
1134 strcmp(pmp->pfs_names[i], label) == 0) {
1135 hmp = pmp->pfs_hmps[i];
1147 * Open the device if this isn't a secondary mount and construct
1148 * the H2 device mount (hmp).
1151 hammer2_chain_t *schain;
1153 hammer2_xop_head_t xop;
1155 if (error == 0 && vcount(devvp) > 0) {
1156 kprintf("Primary device already has references\n");
1161 * Now open the device
1164 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1165 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1166 error = vinvalbuf(devvp, V_SAVE, 0, 0);
1168 error = VOP_OPEN(devvp,
1169 (ronly ? FREAD : FREAD | FWRITE),
1174 if (error && devvp) {
1179 lockmgr(&hammer2_mntlk, LK_RELEASE);
1182 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1183 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1186 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1187 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1188 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1189 RB_INIT(&hmp->iotree);
1190 spin_init(&hmp->io_spin, "h2mount_io");
1191 spin_init(&hmp->list_spin, "h2mount_list");
1193 lockinit(&hmp->vollk, "h2vol", 0, 0);
1194 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1195 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1198 * vchain setup. vchain.data is embedded.
1199 * vchain.refs is initialized and will never drop to 0.
1201 * NOTE! voldata is not yet loaded.
1203 hmp->vchain.hmp = hmp;
1204 hmp->vchain.refs = 1;
1205 hmp->vchain.data = (void *)&hmp->voldata;
1206 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1207 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1208 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1210 hammer2_chain_core_init(&hmp->vchain);
1211 /* hmp->vchain.u.xxx is left NULL */
1214 * fchain setup. fchain.data is embedded.
1215 * fchain.refs is initialized and will never drop to 0.
1217 * The data is not used but needs to be initialized to
1218 * pass assertion muster. We use this chain primarily
1219 * as a placeholder for the freemap's top-level RBTREE
1220 * so it does not interfere with the volume's topology
1223 hmp->fchain.hmp = hmp;
1224 hmp->fchain.refs = 1;
1225 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1226 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1227 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1228 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1229 hmp->fchain.bref.methods =
1230 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1231 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1233 hammer2_chain_core_init(&hmp->fchain);
1234 /* hmp->fchain.u.xxx is left NULL */
1237 * Install the volume header and initialize fields from
1240 error = hammer2_install_volume_header(hmp);
1242 hammer2_unmount_helper(mp, NULL, hmp);
1243 lockmgr(&hammer2_mntlk, LK_RELEASE);
1244 hammer2_vfs_unmount(mp, MNT_FORCE);
1249 * Really important to get these right or the flush and
1250 * teardown code will get confused.
1252 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1254 spmp->pfs_hmps[0] = hmp;
1257 * Dummy-up vchain and fchain's modify_tid. mirror_tid
1258 * is inherited from the volume header.
1261 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1262 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1263 hmp->vchain.pmp = spmp;
1264 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1265 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1266 hmp->fchain.pmp = spmp;
1269 * First locate the super-root inode, which is key 0
1270 * relative to the volume header's blockset.
1272 * Then locate the root inode by scanning the directory keyspace
1273 * represented by the label.
1275 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1276 schain = hammer2_chain_lookup(&parent, &key_dummy,
1277 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1279 hammer2_chain_lookup_done(parent);
1280 if (schain == NULL) {
1281 kprintf("hammer2_mount: invalid super-root\n");
1282 hammer2_unmount_helper(mp, NULL, hmp);
1283 lockmgr(&hammer2_mntlk, LK_RELEASE);
1284 hammer2_vfs_unmount(mp, MNT_FORCE);
1287 if (schain->error) {
1288 kprintf("hammer2_mount: error %s reading super-root\n",
1289 hammer2_error_str(schain->error));
1290 hammer2_chain_unlock(schain);
1291 hammer2_chain_drop(schain);
1293 hammer2_unmount_helper(mp, NULL, hmp);
1294 lockmgr(&hammer2_mntlk, LK_RELEASE);
1295 hammer2_vfs_unmount(mp, MNT_FORCE);
1300 * The super-root always uses an inode_tid of 1 when
1303 spmp->inode_tid = 1;
1304 spmp->modify_tid = schain->bref.modify_tid + 1;
1307 * Sanity-check schain's pmp and finish initialization.
1308 * Any chain belonging to the super-root topology should
1309 * have a NULL pmp (not even set to spmp).
1311 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1312 KKASSERT(schain->pmp == NULL);
1313 spmp->pfs_clid = ripdata->meta.pfs_clid;
1316 * Replace the dummy spmp->iroot with a real one. It's
1317 * easier to just do a wholesale replacement than to try
1318 * to update the chain and fixup the iroot fields.
1320 * The returned inode is locked with the supplied cluster.
1322 hammer2_dummy_xop_from_chain(&xop, schain);
1323 hammer2_inode_drop(spmp->iroot);
1325 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1);
1326 spmp->spmp_hmp = hmp;
1327 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1328 spmp->pfs_hmps[0] = hmp;
1329 hammer2_inode_ref(spmp->iroot);
1330 hammer2_inode_unlock(spmp->iroot);
1331 hammer2_cluster_unlock(&xop.cluster);
1332 hammer2_chain_drop(schain);
1333 /* do not call hammer2_cluster_drop() on an embedded cluster */
1334 schain = NULL; /* now invalid */
1335 /* leave spmp->iroot with one ref */
1337 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1338 error = hammer2_recovery(hmp);
1340 error |= hammer2_fixup_pfses(hmp);
1341 /* XXX do something with error */
1343 hammer2_update_pmps(hmp);
1344 hammer2_iocom_init(hmp);
1345 hammer2_bulkfree_init(hmp);
1348 * Ref the cluster management messaging descriptor. The mount
1349 * program deals with the other end of the communications pipe.
1351 * Root mounts typically do not supply one.
1353 if (info.cluster_fd >= 0) {
1354 fp = holdfp(curthread, info.cluster_fd, -1);
1356 hammer2_cluster_reconnect(hmp, fp);
1358 kprintf("hammer2_mount: bad cluster_fd!\n");
1363 if (info.hflags & HMNT2_DEVFLAGS) {
1364 kprintf("hammer2: Warning: mount flags pertaining "
1365 "to the whole device may only be specified "
1366 "on the first mount of the device: %08x\n",
1367 info.hflags & HMNT2_DEVFLAGS);
1372 * Force local mount (disassociate all PFSs from their clusters).
1373 * Used primarily for debugging.
1375 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1378 * Lookup the mount point under the media-localized super-root.
1379 * Scanning hammer2_pfslist doesn't help us because it represents
1380 * PFS cluster ids which can aggregate several named PFSs together.
1382 * cluster->pmp will incorrectly point to spmp and must be fixed
1385 hammer2_inode_lock(spmp->iroot, 0);
1386 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1387 lhc = hammer2_dirhash(label, strlen(label));
1388 chain = hammer2_chain_lookup(&parent, &key_next,
1389 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1392 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1393 strcmp(label, chain->data->ipdata.filename) == 0) {
1396 chain = hammer2_chain_next(&parent, chain, &key_next,
1398 lhc + HAMMER2_DIRHASH_LOMASK,
1402 hammer2_chain_unlock(parent);
1403 hammer2_chain_drop(parent);
1405 hammer2_inode_unlock(spmp->iroot);
1408 * PFS could not be found?
1410 if (chain == NULL) {
1412 kprintf("hammer2_mount: PFS label I/O error\n");
1414 kprintf("hammer2_mount: PFS label not found\n");
1415 hammer2_unmount_helper(mp, NULL, hmp);
1416 lockmgr(&hammer2_mntlk, LK_RELEASE);
1417 hammer2_vfs_unmount(mp, MNT_FORCE);
1423 * Acquire the pmp structure (it should have already been allocated
1424 * via hammer2_update_pmps() so do not pass cluster in to add to
1425 * available chains).
1427 * Check if the cluster has already been mounted. A cluster can
1428 * only be mounted once, use null mounts to mount additional copies.
1431 kprintf("hammer2_mount: PFS label I/O error\n");
1433 ripdata = &chain->data->ipdata;
1435 pmp = hammer2_pfsalloc(NULL, ripdata,
1436 bref.modify_tid, force_local);
1438 hammer2_chain_unlock(chain);
1439 hammer2_chain_drop(chain);
1444 kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1447 kprintf("hammer2_mount: PFS already mounted!\n");
1448 hammer2_unmount_helper(mp, NULL, hmp);
1449 lockmgr(&hammer2_mntlk, LK_RELEASE);
1450 hammer2_vfs_unmount(mp, MNT_FORCE);
1455 pmp->hflags = info.hflags;
1456 mp->mnt_flag |= MNT_LOCAL;
1457 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */
1458 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */
1461 * required mount structure initializations
1463 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1464 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1466 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1467 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1472 mp->mnt_iosize_max = MAXPHYS;
1475 * Connect up mount pointers.
1477 hammer2_mount_helper(mp, pmp);
1479 lockmgr(&hammer2_mntlk, LK_RELEASE);
1485 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1486 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1487 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1490 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1491 MNAMELEN - 1, &size);
1492 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1493 } /* else root mount, already in there */
1495 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1497 copyinstr(path, mp->mnt_stat.f_mntonname,
1498 sizeof(mp->mnt_stat.f_mntonname) - 1,
1502 mp->mnt_stat.f_mntonname[0] = '/';
1506 * Initial statfs to prime mnt_stat.
1508 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1514 * Scan PFSs under the super-root and create hammer2_pfs structures.
1518 hammer2_update_pmps(hammer2_dev_t *hmp)
1520 const hammer2_inode_data_t *ripdata;
1521 hammer2_chain_t *parent;
1522 hammer2_chain_t *chain;
1523 hammer2_blockref_t bref;
1524 hammer2_dev_t *force_local;
1525 hammer2_pfs_t *spmp;
1527 hammer2_key_t key_next;
1531 * Force local mount (disassociate all PFSs from their clusters).
1532 * Used primarily for debugging.
1534 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1537 * Lookup mount point under the media-localized super-root.
1539 * cluster->pmp will incorrectly point to spmp and must be fixed
1543 hammer2_inode_lock(spmp->iroot, 0);
1544 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1545 chain = hammer2_chain_lookup(&parent, &key_next,
1546 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1549 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1552 kprintf("I/O error scanning PFS labels\n");
1554 ripdata = &chain->data->ipdata;
1557 pmp = hammer2_pfsalloc(chain, ripdata,
1558 bref.modify_tid, force_local);
1560 chain = hammer2_chain_next(&parent, chain, &key_next,
1561 key_next, HAMMER2_KEY_MAX,
1565 hammer2_chain_unlock(parent);
1566 hammer2_chain_drop(parent);
1568 hammer2_inode_unlock(spmp->iroot);
1573 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1574 struct vnode *devvp, struct ucred *cred)
1578 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1579 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1580 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1582 error = hammer2_recovery(hmp);
1584 error |= hammer2_fixup_pfses(hmp);
1585 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1587 VOP_CLOSE(devvp, FREAD, NULL);
1590 VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1601 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1612 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1615 * If mount initialization proceeded far enough we must flush
1616 * its vnodes and sync the underlying mount points. Three syncs
1617 * are required to fully flush the filesystem (freemap updates lag
1618 * by one flush, and one extra for safety).
1620 if (mntflags & MNT_FORCE)
1625 error = vflush(mp, 0, flags);
1628 hammer2_vfs_sync(mp, MNT_WAIT);
1629 hammer2_vfs_sync(mp, MNT_WAIT);
1630 hammer2_vfs_sync(mp, MNT_WAIT);
1634 * Cleanup the frontend support XOPS threads
1636 hammer2_xop_helper_cleanup(pmp);
1639 hammer2_unmount_helper(mp, pmp, NULL);
1643 lockmgr(&hammer2_mntlk, LK_RELEASE);
1649 * Mount helper, hook the system mount into our PFS.
1650 * The mount lock is held.
1652 * We must bump the mount_count on related devices for any
1657 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1659 hammer2_cluster_t *cluster;
1660 hammer2_chain_t *rchain;
1663 mp->mnt_data = (qaddr_t)pmp;
1667 * After pmp->mp is set we have to adjust hmp->mount_count.
1669 cluster = &pmp->iroot->cluster;
1670 for (i = 0; i < cluster->nchains; ++i) {
1671 rchain = cluster->array[i].chain;
1674 ++rchain->hmp->mount_count;
1678 * Create missing Xop threads
1680 hammer2_xop_helper_create(pmp);
1684 * Mount helper, unhook the system mount from our PFS.
1685 * The mount lock is held.
1687 * If hmp is supplied a mount responsible for being the first to open
1688 * the block device failed and the block device and all PFSs using the
1689 * block device must be cleaned up.
1691 * If pmp is supplied multiple devices might be backing the PFS and each
1692 * must be disconnected. This might not be the last PFS using some of the
1693 * underlying devices. Also, we have to adjust our hmp->mount_count
1694 * accounting for the devices backing the pmp which is now undergoing an
1699 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1701 hammer2_cluster_t *cluster;
1702 hammer2_chain_t *rchain;
1703 struct vnode *devvp;
1709 * If no device supplied this is a high-level unmount and we have to
1710 * to disconnect the mount, adjust mount_count, and locate devices
1711 * that might now have no mounts.
1714 KKASSERT(hmp == NULL);
1715 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1717 mp->mnt_data = NULL;
1720 * After pmp->mp is cleared we have to account for
1723 cluster = &pmp->iroot->cluster;
1724 for (i = 0; i < cluster->nchains; ++i) {
1725 rchain = cluster->array[i].chain;
1728 --rchain->hmp->mount_count;
1729 /* scrapping hmp now may invalidate the pmp */
1732 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1733 if (hmp->mount_count == 0) {
1734 hammer2_unmount_helper(NULL, NULL, hmp);
1742 * Try to terminate the block device. We can't terminate it if
1743 * there are still PFSs referencing it.
1745 if (hmp->mount_count)
1749 * Decomission the network before we start messing with the
1752 hammer2_iocom_uninit(hmp);
1754 hammer2_bulkfree_uninit(hmp);
1755 hammer2_pfsfree_scan(hmp, 0);
1757 hammer2_dev_exlock(hmp); /* XXX order */
1761 * Cycle the volume data lock as a safety (probably not needed any
1762 * more). To ensure everything is out we need to flush at least
1763 * three times. (1) The running of the sideq can dirty the
1764 * filesystem, (2) A normal flush can dirty the freemap, and
1765 * (3) ensure that the freemap is fully synchronized.
1767 * The next mount's recovery scan can clean everything up but we want
1768 * to leave the filesystem in a 100% clean state on a normal unmount.
1771 hammer2_voldata_lock(hmp);
1772 hammer2_voldata_unlock(hmp);
1776 * Flush whatever is left. Unmounted but modified PFS's might still
1777 * have some dirty chains on them.
1779 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1780 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1782 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1783 hammer2_voldata_modify(hmp);
1784 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1787 hammer2_chain_unlock(&hmp->fchain);
1789 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1790 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1793 hammer2_chain_unlock(&hmp->vchain);
1795 if ((hmp->vchain.flags | hmp->fchain.flags) &
1796 HAMMER2_CHAIN_FLUSH_MASK) {
1797 kprintf("hammer2_unmount: chains left over "
1798 "after final sync\n");
1799 kprintf(" vchain %08x\n", hmp->vchain.flags);
1800 kprintf(" fchain %08x\n", hmp->fchain.flags);
1802 if (hammer2_debug & 0x0010)
1803 Debugger("entered debugger");
1806 hammer2_pfsfree_scan(hmp, 1);
1808 KKASSERT(hmp->spmp == NULL);
1811 * Finish up with the device vnode
1813 if ((devvp = hmp->devvp) != NULL) {
1815 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1816 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1817 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1819 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1820 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1821 hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1823 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1830 * Clear vchain/fchain flags that might prevent final cleanup
1833 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1834 atomic_add_long(&hammer2_count_modified_chains, -1);
1835 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1836 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1838 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1839 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1842 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1843 atomic_add_long(&hammer2_count_modified_chains, -1);
1844 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1845 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1847 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1848 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1852 * Final drop of embedded freemap root chain to
1853 * clean up fchain.core (fchain structure is not
1854 * flagged ALLOCATED so it is cleaned out and then
1857 hammer2_chain_drop(&hmp->fchain);
1860 * Final drop of embedded volume root chain to clean
1861 * up vchain.core (vchain structure is not flagged
1862 * ALLOCATED so it is cleaned out and then left to
1866 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1868 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1870 hammer2_dev_unlock(hmp);
1872 hammer2_chain_drop(&hmp->vchain);
1874 hammer2_io_cleanup(hmp, &hmp->iotree);
1875 if (hmp->iofree_count) {
1876 kprintf("io_cleanup: %d I/O's left hanging\n",
1880 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1881 kmalloc_destroy(&hmp->mchain);
1882 kfree(hmp, M_HAMMER2);
1886 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1887 ino_t ino, struct vnode **vpp)
1889 hammer2_xop_lookup_t *xop;
1891 hammer2_inode_t *ip;
1895 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1901 * Easy if we already have it cached
1903 ip = hammer2_inode_lookup(pmp, inum);
1905 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1906 *vpp = hammer2_igetv(ip, &error);
1907 hammer2_inode_unlock(ip);
1908 hammer2_inode_drop(ip); /* from lookup */
1914 * Otherwise we have to find the inode
1916 xop = hammer2_xop_alloc(pmp->iroot, 0);
1918 hammer2_xop_start(&xop->head, &hammer2_lookup_desc);
1919 error = hammer2_xop_collect(&xop->head, 0);
1922 ip = hammer2_inode_get(pmp, &xop->head, -1, -1);
1923 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1926 *vpp = hammer2_igetv(ip, &error);
1927 hammer2_inode_unlock(ip);
1937 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1944 if (pmp->iroot == NULL) {
1945 kprintf("hammer2 (%s): no root inode\n",
1946 mp->mnt_stat.f_mntfromname);
1952 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1954 while (pmp->inode_tid == 0) {
1955 hammer2_xop_ipcluster_t *xop;
1956 const hammer2_inode_meta_t *meta;
1958 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1959 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc);
1960 error = hammer2_xop_collect(&xop->head, 0);
1963 meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1964 pmp->iroot->meta = *meta;
1965 pmp->inode_tid = meta->pfs_inum + 1;
1966 hammer2_xop_pdata(&xop->head);
1969 if (pmp->inode_tid < HAMMER2_INODE_START)
1970 pmp->inode_tid = HAMMER2_INODE_START;
1972 xop->head.cluster.focus->bref.modify_tid + 1;
1974 kprintf("PFS: Starting inode %jd\n",
1975 (intmax_t)pmp->inode_tid);
1976 kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1977 pmp->inode_tid, pmp->modify_tid);
1979 wakeup(&pmp->iroot);
1981 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1984 * Prime the mount info.
1986 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1993 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1994 hammer2_inode_unlock(pmp->iroot);
1995 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1996 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
2002 hammer2_inode_unlock(pmp->iroot);
2005 vp = hammer2_igetv(pmp->iroot, &error);
2006 hammer2_inode_unlock(pmp->iroot);
2016 * XXX incorporate ipdata->meta.inode_quota and data_quota
2020 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
2024 hammer2_blockref_t bref;
2029 * NOTE: iroot might not have validated the cluster yet.
2033 bzero(&tmp, sizeof(tmp));
2035 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2036 hmp = pmp->pfs_hmps[i];
2039 if (pmp->iroot->cluster.array[i].chain)
2040 bref = pmp->iroot->cluster.array[i].chain->bref;
2042 bzero(&bref, sizeof(bref));
2044 tmp.f_files = bref.embed.stats.inode_count;
2046 tmp.f_blocks = hmp->voldata.allocator_size /
2047 mp->mnt_vstat.f_bsize;
2048 tmp.f_bfree = hmp->voldata.allocator_free /
2049 mp->mnt_vstat.f_bsize;
2050 tmp.f_bavail = tmp.f_bfree;
2052 if (cred && cred->cr_uid != 0) {
2056 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2057 tmp.f_blocks -= adj;
2059 tmp.f_bavail -= adj;
2062 mp->mnt_stat.f_blocks = tmp.f_blocks;
2063 mp->mnt_stat.f_bfree = tmp.f_bfree;
2064 mp->mnt_stat.f_bavail = tmp.f_bavail;
2065 mp->mnt_stat.f_files = tmp.f_files;
2066 mp->mnt_stat.f_ffree = tmp.f_ffree;
2068 *sbp = mp->mnt_stat;
2075 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2079 hammer2_blockref_t bref;
2084 * NOTE: iroot might not have validated the cluster yet.
2087 bzero(&tmp, sizeof(tmp));
2089 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2090 hmp = pmp->pfs_hmps[i];
2093 if (pmp->iroot->cluster.array[i].chain)
2094 bref = pmp->iroot->cluster.array[i].chain->bref;
2096 bzero(&bref, sizeof(bref));
2098 tmp.f_files = bref.embed.stats.inode_count;
2100 tmp.f_blocks = hmp->voldata.allocator_size /
2101 mp->mnt_vstat.f_bsize;
2102 tmp.f_bfree = hmp->voldata.allocator_free /
2103 mp->mnt_vstat.f_bsize;
2104 tmp.f_bavail = tmp.f_bfree;
2106 if (cred && cred->cr_uid != 0) {
2110 adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2111 tmp.f_blocks -= adj;
2113 tmp.f_bavail -= adj;
2116 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2117 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2118 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2119 mp->mnt_vstat.f_files = tmp.f_files;
2120 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2122 *sbp = mp->mnt_vstat;
2128 * Mount-time recovery (RW mounts)
2130 * Updates to the free block table are allowed to lag flushes by one
2131 * transaction. In case of a crash, then on a fresh mount we must do an
2132 * incremental scan of the last committed transaction id and make sure that
2133 * all related blocks have been marked allocated.
2135 * The super-root topology and each PFS has its own transaction id domain,
2136 * so we must track PFS boundary transitions.
2138 struct hammer2_recovery_elm {
2139 TAILQ_ENTRY(hammer2_recovery_elm) entry;
2140 hammer2_chain_t *chain;
2141 hammer2_tid_t sync_tid;
2144 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2146 struct hammer2_recovery_info {
2147 struct hammer2_recovery_list list;
2152 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2153 hammer2_chain_t *parent,
2154 struct hammer2_recovery_info *info,
2155 hammer2_tid_t sync_tid);
2157 #define HAMMER2_RECOVERY_MAXDEPTH 10
2161 hammer2_recovery(hammer2_dev_t *hmp)
2163 struct hammer2_recovery_info info;
2164 struct hammer2_recovery_elm *elm;
2165 hammer2_chain_t *parent;
2166 hammer2_tid_t sync_tid;
2167 hammer2_tid_t mirror_tid;
2170 hammer2_trans_init(hmp->spmp, 0);
2172 sync_tid = hmp->voldata.freemap_tid;
2173 mirror_tid = hmp->voldata.mirror_tid;
2175 kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2176 if (sync_tid >= mirror_tid) {
2177 kprintf(" no recovery needed\n");
2179 kprintf(" freemap recovery %016jx-%016jx\n",
2180 sync_tid + 1, mirror_tid);
2183 TAILQ_INIT(&info.list);
2185 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2186 error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2187 hammer2_chain_lookup_done(parent);
2189 while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2190 TAILQ_REMOVE(&info.list, elm, entry);
2191 parent = elm->chain;
2192 sync_tid = elm->sync_tid;
2193 kfree(elm, M_HAMMER2);
2195 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2196 error |= hammer2_recovery_scan(hmp, parent, &info,
2197 hmp->voldata.freemap_tid);
2198 hammer2_chain_unlock(parent);
2199 hammer2_chain_drop(parent); /* drop elm->chain ref */
2202 hammer2_trans_done(hmp->spmp, 0);
2209 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2210 struct hammer2_recovery_info *info,
2211 hammer2_tid_t sync_tid)
2213 const hammer2_inode_data_t *ripdata;
2214 hammer2_chain_t *chain;
2215 hammer2_blockref_t bref;
2222 * Adjust freemap to ensure that the block(s) are marked allocated.
2224 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2225 hammer2_freemap_adjust(hmp, &parent->bref,
2226 HAMMER2_FREEMAP_DORECOVER);
2230 * Check type for recursive scan
2232 switch(parent->bref.type) {
2233 case HAMMER2_BREF_TYPE_VOLUME:
2234 /* data already instantiated */
2236 case HAMMER2_BREF_TYPE_INODE:
2238 * Must instantiate data for DIRECTDATA test and also
2241 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2242 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2243 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2244 /* not applicable to recovery scan */
2245 hammer2_chain_unlock(parent);
2248 hammer2_chain_unlock(parent);
2250 case HAMMER2_BREF_TYPE_INDIRECT:
2252 * Must instantiate data for recursion
2254 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2255 hammer2_chain_unlock(parent);
2257 case HAMMER2_BREF_TYPE_DIRENT:
2258 case HAMMER2_BREF_TYPE_DATA:
2259 case HAMMER2_BREF_TYPE_FREEMAP:
2260 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2261 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2262 /* not applicable to recovery scan */
2266 return HAMMER2_ERROR_BADBREF;
2270 * Defer operation if depth limit reached or if we are crossing a
2273 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2274 struct hammer2_recovery_elm *elm;
2276 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2277 elm->chain = parent;
2278 elm->sync_tid = sync_tid;
2279 hammer2_chain_ref(parent);
2280 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2281 /* unlocked by caller */
2288 * Recursive scan of the last flushed transaction only. We are
2289 * doing this without pmp assignments so don't leave the chains
2290 * hanging around after we are done with them.
2292 * error Cumulative error this level only
2293 * rup_error Cumulative error for recursion
2294 * tmp_error Specific non-cumulative recursion error
2302 error |= hammer2_chain_scan(parent, &chain, &bref,
2304 HAMMER2_LOOKUP_NODATA);
2307 * Problem during scan or EOF
2315 if (chain == NULL) {
2316 if (bref.mirror_tid > sync_tid) {
2317 hammer2_freemap_adjust(hmp, &bref,
2318 HAMMER2_FREEMAP_DORECOVER);
2324 * This may or may not be a recursive node.
2326 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2327 if (bref.mirror_tid > sync_tid) {
2329 tmp_error = hammer2_recovery_scan(hmp, chain,
2337 * Flush the recovery at the PFS boundary to stage it for
2338 * the final flush of the super-root topology.
2340 if (tmp_error == 0 &&
2341 (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2342 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2343 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2346 rup_error |= tmp_error;
2348 return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2352 * This fixes up an error introduced in earlier H2 implementations where
2353 * moving a PFS inode into an indirect block wound up causing the
2354 * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2358 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2360 const hammer2_inode_data_t *ripdata;
2361 hammer2_chain_t *parent;
2362 hammer2_chain_t *chain;
2363 hammer2_key_t key_next;
2364 hammer2_pfs_t *spmp;
2370 * Lookup mount point under the media-localized super-root.
2372 * cluster->pmp will incorrectly point to spmp and must be fixed
2376 hammer2_inode_lock(spmp->iroot, 0);
2377 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2378 chain = hammer2_chain_lookup(&parent, &key_next,
2379 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2382 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2385 kprintf("I/O error scanning PFS labels\n");
2386 error |= chain->error;
2387 } else if ((chain->bref.flags &
2388 HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2391 ripdata = &chain->data->ipdata;
2392 hammer2_trans_init(hmp->spmp, 0);
2393 error2 = hammer2_chain_modify(chain,
2394 chain->bref.modify_tid,
2397 kprintf("hammer2: Correct mis-flagged PFS %s\n",
2399 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2403 hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2405 hammer2_trans_done(hmp->spmp, 0);
2407 chain = hammer2_chain_next(&parent, chain, &key_next,
2408 key_next, HAMMER2_KEY_MAX,
2412 hammer2_chain_unlock(parent);
2413 hammer2_chain_drop(parent);
2415 hammer2_inode_unlock(spmp->iroot);
2421 * Sync a mount point; this is called periodically on a per-mount basis from
2422 * the filesystem syncer, and whenever a user issues a sync.
2425 hammer2_vfs_sync(struct mount *mp, int waitfor)
2429 error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor);
2435 * Because frontend operations lock vnodes before we get a chance to
2436 * lock the related inode, we can't just acquire a vnode lock without
2437 * risking a deadlock. The frontend may be holding a vnode lock while
2438 * also blocked on our SYNCQ flag while trying to get the inode lock.
2440 * To deal with this situation we can check the vnode lock situation
2441 * after locking the inode and perform a work-around.
2444 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor)
2447 /*hammer2_xop_flush_t *xop;*/
2448 /*struct hammer2_sync_info info;*/
2449 hammer2_inode_t *ip;
2450 hammer2_depend_t *depend;
2451 hammer2_depend_t *depend_next;
2460 * Move all inodes on sideq to syncq. This will clear sideq.
2461 * This should represent all flushable inodes. These inodes
2462 * will already have refs due to being on syncq or sideq. We
2463 * must do this all at once with the spinlock held to ensure that
2464 * all inode dependencies are part of the same flush.
2466 * We should be able to do this asynchronously from frontend
2467 * operations because we will be locking the inodes later on
2468 * to actually flush them, and that will partition any frontend
2469 * op using the same inode. Either it has already locked the
2470 * inode and we will block, or it has not yet locked the inode
2471 * and it will block until we are finished flushing that inode.
2473 * When restarting, only move the inodes flagged as PASS2 from
2474 * SIDEQ to SYNCQ. PASS2 propagation by inode_lock4() and
2475 * inode_depend() are atomic with the spin-lock.
2477 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2478 #ifdef HAMMER2_DEBUG_SYNC
2479 kprintf("FILESYSTEM SYNC BOUNDARY\n");
2484 * Move inodes from depq to syncq, releasing the related
2485 * depend structures.
2488 #ifdef HAMMER2_DEBUG_SYNC
2489 kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart);
2491 hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/);
2492 hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN);
2495 * Move inodes from depq to syncq. When restarting, only depq's
2496 * marked pass2 are moved.
2498 hammer2_spin_ex(&pmp->list_spin);
2499 depend_next = TAILQ_FIRST(&pmp->depq);
2501 while ((depend = depend_next) != NULL) {
2502 depend_next = TAILQ_NEXT(depend, entry);
2503 if (dorestart && depend->pass2 == 0)
2505 TAILQ_FOREACH(ip, &depend->sideq, entry) {
2506 KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ);
2507 atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ);
2508 atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ);
2511 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry);
2512 pmp->sideq_count -= depend->count;
2515 TAILQ_REMOVE(&pmp->depq, depend, entry);
2518 hammer2_spin_unex(&pmp->list_spin);
2519 hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/
2520 HAMMER2_TRANS_WAITING);
2524 * sideq_count may have dropped enough to allow us to unstall
2527 hammer2_pfs_memory_inc(pmp);
2528 hammer2_pfs_memory_wakeup(pmp);
2531 * Now run through all inodes on syncq.
2533 * Flush transactions only interlock with other flush transactions.
2534 * Any conflicting frontend operations will block on the inode, but
2535 * may hold a vnode lock while doing so.
2537 hammer2_spin_ex(&pmp->list_spin);
2538 while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) {
2540 * Remove the inode from the SYNCQ, transfer the syncq ref
2541 * to us. We must clear SYNCQ to allow any potential
2542 * front-end deadlock to proceed. We must set PASS2 so
2543 * the dependency code knows what to do.
2547 if (atomic_cmpset_int(&ip->flags,
2549 (pass2 & ~(HAMMER2_INODE_SYNCQ |
2550 HAMMER2_INODE_SYNCQ_WAKEUP)) |
2551 HAMMER2_INODE_SYNCQ_PASS2) == 0) {
2554 TAILQ_REMOVE(&pmp->syncq, ip, entry);
2555 hammer2_spin_unex(&pmp->list_spin);
2556 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP)
2560 * Relock the inode, and we inherit a ref from the above.
2561 * We will check for a race after we acquire the vnode.
2563 hammer2_mtx_ex(&ip->lock);
2566 * We need the vp in order to vfsync() dirty buffers, so if
2567 * one isn't attached we can skip it.
2569 * Ordering the inode lock and then the vnode lock has the
2570 * potential to deadlock. If we had left SYNCQ set that could
2571 * also deadlock us against the frontend even if we don't hold
2572 * any locks, but the latter is not a problem now since we
2573 * cleared it. igetv will temporarily release the inode lock
2574 * in a safe manner to work-around the deadlock.
2576 * Unfortunately it is still possible to deadlock when the
2577 * frontend obtains multiple inode locks, because all the
2578 * related vnodes are already locked (nor can the vnode locks
2579 * be released and reacquired without messing up RECLAIM and
2580 * INACTIVE sequencing).
2582 * The solution for now is to move the vp back onto SIDEQ
2583 * and set dorestart, which will restart the flush after we
2584 * exhaust the current SYNCQ. Note that additional
2585 * dependencies may build up, so we definitely need to move
2586 * the whole SIDEQ back to SYNCQ when we restart.
2590 if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) {
2592 * Failed to get the vnode, requeue the inode
2593 * (PASS2 is already set so it will be found
2594 * again on the restart).
2596 * Then unlock, possibly sleep, and retry
2597 * later. We sleep if PASS2 was *previously*
2598 * set, before we set it again above.
2602 #ifdef HAMMER2_DEBUG_SYNC
2603 kprintf("inum %ld (sync delayed by vnode)\n",
2604 (long)ip->meta.inum);
2606 hammer2_inode_delayed_sideq(ip);
2608 hammer2_mtx_unlock(&ip->lock);
2609 hammer2_inode_drop(ip);
2611 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) {
2612 tsleep(&dorestart, 0, "h2syndel", 2);
2614 hammer2_spin_ex(&pmp->list_spin);
2622 * If the inode wound up on a SIDEQ again it will already be
2623 * prepped for another PASS2. In this situation if we flush
2624 * it now we will just wind up flushing it again in the same
2625 * syncer run, so we might as well not flush it now.
2627 if (ip->flags & HAMMER2_INODE_SIDEQ) {
2628 hammer2_mtx_unlock(&ip->lock);
2629 hammer2_inode_drop(ip);
2633 hammer2_spin_ex(&pmp->list_spin);
2638 * Ok we have the inode exclusively locked and if vp is
2639 * not NULL that will also be exclusively locked. Do the
2640 * meat of the flush.
2642 * vp token needed for v_rbdirty_tree check / vclrisdirty
2643 * sequencing. Though we hold the vnode exclusively so
2644 * we shouldn't need to hold the token also in this case.
2647 vfsync(vp, MNT_WAIT, 1, NULL, NULL);
2648 bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */
2652 * If the inode has not yet been inserted into the tree
2653 * we must do so. Then sync and flush it. The flush should
2654 * update the parent.
2656 if (ip->flags & HAMMER2_INODE_DELETING) {
2657 #ifdef HAMMER2_DEBUG_SYNC
2658 kprintf("inum %ld destroy\n", (long)ip->meta.inum);
2660 hammer2_inode_chain_des(ip);
2661 atomic_add_long(&hammer2_iod_inode_deletes, 1);
2662 } else if (ip->flags & HAMMER2_INODE_CREATING) {
2663 #ifdef HAMMER2_DEBUG_SYNC
2664 kprintf("inum %ld insert\n", (long)ip->meta.inum);
2666 hammer2_inode_chain_ins(ip);
2667 atomic_add_long(&hammer2_iod_inode_creates, 1);
2669 #ifdef HAMMER2_DEBUG_SYNC
2670 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum);
2674 * Because I kinda messed up the design and index the inodes
2675 * under the root inode, along side the directory entries,
2676 * we can't flush the inode index under the iroot until the
2677 * end. If we do it now we might miss effects created by
2678 * other inodes on the SYNCQ.
2680 * Do a normal (non-FSSYNC) flush instead, which allows the
2681 * vnode code to work the same. We don't want to force iroot
2682 * back onto the SIDEQ, and we also don't want the flush code
2683 * to update pfs_iroot_blocksets until the final flush later.
2685 * XXX at the moment this will likely result in a double-flush
2686 * of the iroot chain.
2688 hammer2_inode_chain_sync(ip);
2689 if (ip == pmp->iroot) {
2690 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP);
2692 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2693 HAMMER2_XOP_FSSYNC);
2696 lwkt_gettoken(&vp->v_token);
2697 if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2698 HAMMER2_INODE_RESIZED |
2699 HAMMER2_INODE_DIRTYDATA)) == 0 &&
2700 RB_EMPTY(&vp->v_rbdirty_tree) &&
2701 !bio_track_active(&vp->v_track_write)) {
2704 hammer2_inode_delayed_sideq(ip);
2706 lwkt_reltoken(&vp->v_token);
2708 vp = NULL; /* safety */
2710 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2);
2711 hammer2_inode_unlock(ip); /* unlock+drop */
2712 /* ip pointer invalid */
2715 * If the inode got dirted after we dropped our locks,
2716 * it will have already been moved back to the SIDEQ.
2718 hammer2_spin_ex(&pmp->list_spin);
2720 hammer2_spin_unex(&pmp->list_spin);
2721 if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) {
2722 #ifdef HAMMER2_DEBUG_SYNC
2723 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n");
2724 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/
2729 #ifdef HAMMER2_DEBUG_SYNC
2730 kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n");
2731 /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/
2735 * We have to flush the PFS root last, even if it does not appear to
2736 * be dirty, because all the inodes in the PFS are indexed under it.
2737 * The normal flushing of iroot above would only occur if directory
2738 * entries under the root were changed.
2740 * Specifying VOLHDR will cause an additionl flush of hmp->spmp
2741 * for the media making up the cluster.
2743 if ((ip = pmp->iroot) != NULL) {
2744 hammer2_inode_ref(ip);
2745 hammer2_mtx_ex(&ip->lock);
2746 hammer2_inode_chain_sync(ip);
2747 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2748 HAMMER2_XOP_FSSYNC |
2749 HAMMER2_XOP_VOLHDR);
2750 hammer2_inode_unlock(ip); /* unlock+drop */
2752 #ifdef HAMMER2_DEBUG_SYNC
2753 kprintf("FILESYSTEM SYNC STAGE 2 DONE\n");
2759 hammer2_bioq_sync(pmp);
2763 info.waitfor = MNT_WAIT;
2764 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2767 info.waitfor = MNT_WAIT;
2768 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2772 * Generally speaking we now want to flush the media topology from
2773 * the iroot through to the inodes. The flush stops at any inode
2774 * boundary, which allows the frontend to continue running concurrent
2775 * modifying operations on inodes (including kernel flushes of
2776 * buffers) without interfering with the main sync.
2778 * Use the XOP interface to concurrently flush all nodes to
2779 * synchronize the PFSROOT subtopology to the media. A standard
2780 * end-of-scan ENOENT error indicates cluster sufficiency.
2782 * Note that this flush will not be visible on crash recovery until
2783 * we flush the super-root topology in the next loop.
2785 * XXX For now wait for all flushes to complete.
2787 if (mp && (ip = pmp->iroot) != NULL) {
2789 * If unmounting try to flush everything including any
2790 * sub-trees under inodes, just in case there is dangling
2791 * modified data, as a safety. Otherwise just flush up to
2792 * the inodes in this stage.
2794 kprintf("MP & IROOT\n");
2795 #ifdef HAMMER2_DEBUG_SYNC
2796 kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n");
2798 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2799 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2800 HAMMER2_XOP_VOLHDR |
2801 HAMMER2_XOP_FSSYNC |
2802 HAMMER2_XOP_INODE_STOP);
2804 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2805 HAMMER2_XOP_INODE_STOP |
2806 HAMMER2_XOP_VOLHDR |
2807 HAMMER2_XOP_FSSYNC |
2808 HAMMER2_XOP_INODE_STOP);
2810 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc);
2811 error = hammer2_xop_collect(&xop->head,
2812 HAMMER2_XOP_COLLECT_WAITALL);
2813 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2814 #ifdef HAMMER2_DEBUG_SYNC
2815 kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n");
2817 if (error == HAMMER2_ERROR_ENOENT)
2820 error = hammer2_error_to_errno(error);
2825 error = 0; /* XXX */
2826 hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH);
2833 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2835 hammer2_inode_t *ip;
2837 KKASSERT(MAXFIDSZ >= 16);
2839 fhp->fid_len = offsetof(struct fid, fid_data[16]);
2841 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2842 ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2849 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2850 struct fid *fhp, struct vnode **vpp)
2857 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2860 error = hammer2_vfs_root(mp, vpp);
2862 error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2867 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2873 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2874 int *exflagsp, struct ucred **credanonp)
2881 np = vfs_export_lookup(mp, &pmp->export, nam);
2883 *exflagsp = np->netc_exflags;
2884 *credanonp = &np->netc_anon;
2893 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume
2894 * header into the HMP
2896 * XXX read four volhdrs and use the one with the highest TID whos CRC
2901 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2902 * nonexistant locations.
2904 * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2908 hammer2_install_volume_header(hammer2_dev_t *hmp)
2910 hammer2_volume_data_t *vd;
2912 hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2924 * There are up to 4 copies of the volume header (syncs iterate
2925 * between them so there is no single master). We don't trust the
2926 * volu_size field so we don't know precisely how large the filesystem
2927 * is, so depend on the OS to return an error if we go beyond the
2928 * block device's EOF.
2930 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2931 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2932 HAMMER2_VOLUME_BYTES, &bp);
2939 vd = (struct hammer2_volume_data *) bp->b_data;
2940 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2941 (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2947 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2948 /* XXX: Reversed-endianness filesystem */
2949 kprintf("hammer2: reverse-endian filesystem detected");
2955 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2956 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2957 HAMMER2_VOLUME_ICRC0_SIZE);
2958 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2959 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2960 HAMMER2_VOLUME_ICRC1_SIZE);
2961 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2962 kprintf("hammer2 volume header crc "
2963 "mismatch copy #%d %08x/%08x\n",
2970 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2979 hmp->volsync = hmp->voldata;
2980 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2982 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2983 kprintf("hammer2: using volume header #%d\n",
2988 kprintf("hammer2: no valid volume headers found!\n");
2994 * This handles hysteresis on regular file flushes. Because the BIOs are
2995 * routed to a thread it is possible for an excessive number to build up
2996 * and cause long front-end stalls long before the runningbuffspace limit
2997 * is hit, so we implement hammer2_flush_pipe to control the
3000 * This is a particular problem when compression is used.
3003 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
3005 atomic_add_int(&pmp->count_lwinprog, 1);
3009 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
3013 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
3014 if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
3015 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
3016 atomic_clear_int(&pmp->count_lwinprog,
3017 HAMMER2_LWINPROG_WAITING);
3018 wakeup(&pmp->count_lwinprog);
3020 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
3021 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
3022 atomic_clear_int(&pmp->count_lwinprog,
3023 HAMMER2_LWINPROG_WAITING0);
3024 wakeup(&pmp->count_lwinprog);
3029 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
3032 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
3033 HAMMER2_LWINPROG_WAITING0;
3036 lwinprog = pmp->count_lwinprog;
3038 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3040 tsleep_interlock(&pmp->count_lwinprog, 0);
3041 atomic_set_int(&pmp->count_lwinprog, lwflag);
3042 lwinprog = pmp->count_lwinprog;
3043 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3045 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
3050 * It is possible for an excessive number of dirty chains or dirty inodes
3051 * to build up. When this occurs we start an asynchronous filesystem sync.
3052 * If the level continues to build up, we stall, waiting for it to drop,
3053 * with some hysteresis.
3055 * We limit the stall to two seconds per call.
3057 * This relies on the kernel calling hammer2_vfs_modifying() prior to
3058 * obtaining any vnode locks before making a modifying VOP call.
3061 hammer2_vfs_modifying(struct mount *mp)
3063 hammer2_pfs_memory_wait(MPTOPMP(mp));
3067 * Initiate an asynchronous filesystem sync and, with hysteresis,
3068 * stall if the internal data structure count becomes too bloated.
3071 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
3076 if (pmp == NULL || pmp->mp == NULL)
3079 for (loops = 0; loops < 2; ++loops) {
3080 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK;
3084 * Start the syncer running at 1/2 the limit
3086 if (waiting > hammer2_limit_dirty_chains / 2 ||
3087 pmp->sideq_count > hammer2_limit_dirty_inodes / 2) {
3088 trigger_syncer(pmp->mp);
3092 * Stall at the limit waiting for the counts to drop.
3093 * This code will typically be woken up once the count
3094 * drops below 3/4 the limit, or in one second.
3096 if (waiting < hammer2_limit_dirty_chains &&
3097 pmp->sideq_count < hammer2_limit_dirty_inodes) {
3100 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
3101 atomic_set_int(&pmp->inmem_dirty_chains,
3102 HAMMER2_DIRTYCHAIN_WAITING);
3103 if (waiting < hammer2_limit_dirty_chains &&
3104 pmp->sideq_count < hammer2_limit_dirty_inodes) {
3107 trigger_syncer(pmp->mp);
3108 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED, "h2memw", hz);
3110 limit = pmp->mp->mnt_nvnodelistsize / 10;
3111 if (limit < hammer2_limit_dirty_chains)
3112 limit = hammer2_limit_dirty_chains;
3120 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
3123 atomic_add_int(&pmp->inmem_dirty_chains, 1);
3128 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
3133 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
3134 /* don't need --waiting to test flag */
3136 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) &&
3137 (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <=
3138 hammer2_limit_dirty_chains * 2 / 3 &&
3139 pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) {
3140 atomic_clear_int(&pmp->inmem_dirty_chains,
3141 HAMMER2_DIRTYCHAIN_WAITING);
3142 wakeup(&pmp->inmem_dirty_chains);
3148 * Returns 0 if the filesystem has tons of free space
3149 * Returns 1 if the filesystem has less than 10% remaining
3150 * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
3153 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
3157 hammer2_off_t free_reserved;
3158 hammer2_off_t free_nominal;
3163 if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
3164 free_reserved = HAMMER2_SEGSIZE;
3165 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
3166 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
3167 hmp = pmp->pfs_hmps[i];
3170 if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
3171 pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
3174 if (free_nominal > hmp->voldata.allocator_free)
3175 free_nominal = hmp->voldata.allocator_free;
3176 if (free_reserved < hmp->free_reserved)
3177 free_reserved = hmp->free_reserved;
3183 pmp->free_reserved = free_reserved;
3184 pmp->free_nominal = free_nominal;
3185 pmp->free_ticks = ticks;
3187 free_reserved = pmp->free_reserved;
3188 free_nominal = pmp->free_nominal;
3190 if (cred && cred->cr_uid != 0) {
3191 if ((int64_t)(free_nominal - bytes) <
3192 (int64_t)free_reserved) {
3196 if ((int64_t)(free_nominal - bytes) <
3197 (int64_t)free_reserved / 2) {
3201 if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
3210 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
3213 hammer2_chain_t *scan;
3214 hammer2_chain_t *parent;
3218 kprintf("%*.*s...\n", tab, tab, "");
3223 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
3225 chain, chain->bref.type,
3226 chain->bref.key, chain->bref.keybits,
3227 chain->bref.mirror_tid);
3229 kprintf("%*.*s [%08x] (%s) refs=%d",
3232 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3233 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
3236 parent = chain->parent;
3238 kprintf("\n%*.*s p=%p [pflags %08x prefs %d",
3240 parent, parent->flags, parent->refs);
3241 if (RB_EMPTY(&chain->core.rbtree)) {
3245 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3246 if ((scan->flags & flags) || flags == (u_int)-1) {
3247 hammer2_dump_chain(scan, tab + 4, countp, 'a',
3251 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3252 kprintf("%*.*s}(%s)\n", tab, tab, "",
3253 chain->data->ipdata.filename);
3255 kprintf("%*.*s}\n", tab, tab, "");