2 * Copyright (c) 2011-2015 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 {
74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
75 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
76 static struct hammer2_mntlist hammer2_mntlist;
77 static struct hammer2_pfslist hammer2_pfslist;
78 static struct lock hammer2_mntlk;
81 int hammer2_cluster_enable = 1;
82 int hammer2_hardlink_enable = 1;
83 int hammer2_flush_pipe = 100;
84 int hammer2_synchronous_flush = 1;
85 int hammer2_dio_count;
86 long hammer2_limit_dirty_chains;
87 long hammer2_count_modified_chains;
88 long hammer2_iod_file_read;
89 long hammer2_iod_meta_read;
90 long hammer2_iod_indr_read;
91 long hammer2_iod_fmap_read;
92 long hammer2_iod_volu_read;
93 long hammer2_iod_file_write;
94 long hammer2_iod_file_wembed;
95 long hammer2_iod_file_wzero;
96 long hammer2_iod_file_wdedup;
97 long hammer2_iod_meta_write;
98 long hammer2_iod_indr_write;
99 long hammer2_iod_fmap_write;
100 long hammer2_iod_volu_write;
101 long hammer2_ioa_file_read;
102 long hammer2_ioa_meta_read;
103 long hammer2_ioa_indr_read;
104 long hammer2_ioa_fmap_read;
105 long hammer2_ioa_volu_read;
106 long hammer2_ioa_fmap_write;
107 long hammer2_ioa_file_write;
108 long hammer2_ioa_meta_write;
109 long hammer2_ioa_indr_write;
110 long hammer2_ioa_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, debug, CTLFLAG_RW,
123 &hammer2_debug, 0, "");
124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW,
125 &hammer2_cluster_enable, 0, "");
126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
127 &hammer2_hardlink_enable, 0, "");
128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
129 &hammer2_flush_pipe, 0, "");
130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
131 &hammer2_synchronous_flush, 0, "");
132 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
133 &hammer2_limit_dirty_chains, 0, "");
134 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
135 &hammer2_count_modified_chains, 0, "");
136 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
137 &hammer2_dio_count, 0, "");
139 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
140 &hammer2_iod_file_read, 0, "");
141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
142 &hammer2_iod_meta_read, 0, "");
143 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
144 &hammer2_iod_indr_read, 0, "");
145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
146 &hammer2_iod_fmap_read, 0, "");
147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
148 &hammer2_iod_volu_read, 0, "");
150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
151 &hammer2_iod_file_write, 0, "");
152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
153 &hammer2_iod_file_wembed, 0, "");
154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
155 &hammer2_iod_file_wzero, 0, "");
156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
157 &hammer2_iod_file_wdedup, 0, "");
158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
159 &hammer2_iod_meta_write, 0, "");
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
161 &hammer2_iod_indr_write, 0, "");
162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
163 &hammer2_iod_fmap_write, 0, "");
164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
165 &hammer2_iod_volu_write, 0, "");
167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
168 &hammer2_ioa_file_read, 0, "");
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
170 &hammer2_ioa_meta_read, 0, "");
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
172 &hammer2_ioa_indr_read, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
174 &hammer2_ioa_fmap_read, 0, "");
175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
176 &hammer2_ioa_volu_read, 0, "");
178 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
179 &hammer2_ioa_file_write, 0, "");
180 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
181 &hammer2_ioa_meta_write, 0, "");
182 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
183 &hammer2_ioa_indr_write, 0, "");
184 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
185 &hammer2_ioa_fmap_write, 0, "");
186 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
187 &hammer2_ioa_volu_write, 0, "");
189 static int hammer2_vfs_init(struct vfsconf *conf);
190 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
191 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
193 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
194 struct vnode *, struct ucred *);
195 static int hammer2_recovery(hammer2_dev_t *hmp);
196 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
197 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
198 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
200 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
202 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
203 ino_t ino, struct vnode **vpp);
204 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
205 struct fid *fhp, struct vnode **vpp);
206 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
207 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
208 int *exflagsp, struct ucred **credanonp);
210 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
211 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
213 static void hammer2_update_pmps(hammer2_dev_t *hmp);
215 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
216 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
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, 0);
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;
254 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
256 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
258 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
262 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
264 margs_read.objsize = 65536;
265 margs_read.mtype = M_HAMMER2_DEBUFFER;
267 margs_write.objsize = 32768;
268 margs_write.mtype = M_HAMMER2_CBUFFER;
270 margs_vop.objsize = sizeof(hammer2_xop_t);
271 margs_vop.mtype = M_HAMMER2;
274 * Note thaht for the XOPS cache we want backing store allocations
275 * to use M_ZERO. This is not allowed in objcache_get() (to avoid
276 * confusion), so use the backing store function that does it. This
277 * means that initial XOPS objects are zerod but REUSED objects are
278 * not. So we are responsible for cleaning the object up sufficiently
279 * for our needs before objcache_put()ing it back (typically just the
282 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
283 0, 1, NULL, NULL, NULL,
284 objcache_malloc_alloc,
285 objcache_malloc_free,
287 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
288 0, 1, NULL, NULL, NULL,
289 objcache_malloc_alloc,
290 objcache_malloc_free,
292 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
293 0, 1, NULL, NULL, NULL,
294 objcache_malloc_alloc_zero,
295 objcache_malloc_free,
299 lockinit(&hammer2_mntlk, "mntlk", 0, 0);
300 TAILQ_INIT(&hammer2_mntlist);
301 TAILQ_INIT(&hammer2_pfslist);
303 hammer2_limit_dirty_chains = desiredvnodes / 10;
310 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
312 objcache_destroy(cache_buffer_read);
313 objcache_destroy(cache_buffer_write);
314 objcache_destroy(cache_xops);
319 * Core PFS allocator. Used to allocate the pmp structure for PFS cluster
320 * mounts and the spmp structure for media (hmp) structures.
322 * pmp->modify_tid tracks new modify_tid transaction ids for front-end
323 * transactions. Note that synchronization does not use this field.
324 * (typically frontend operations and synchronization cannot run on the
325 * same PFS node at the same time).
330 hammer2_pfsalloc(hammer2_chain_t *chain, const hammer2_inode_data_t *ripdata,
331 hammer2_tid_t modify_tid)
333 hammer2_inode_t *iroot;
340 * Locate or create the PFS based on the cluster id. If ripdata
341 * is NULL this is a spmp which is unique and is always allocated.
344 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
345 if (bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
346 sizeof(pmp->pfs_clid)) == 0) {
355 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
356 hammer2_trans_manage_init(pmp);
357 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
358 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
359 lockinit(&pmp->lock, "pfslk", 0, 0);
360 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
361 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
362 spin_init(&pmp->xop_spin, "h2xop");
363 RB_INIT(&pmp->inum_tree);
364 TAILQ_INIT(&pmp->sideq);
365 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
368 * Distribute backend operations to threads
370 for (j = 0; j < HAMMER2_MAXCLUSTER; ++j)
371 TAILQ_INIT(&pmp->xopq[j]);
372 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
373 hammer2_xop_group_init(pmp, &pmp->xop_groups[j]);
376 * Save the last media transaction id for the flusher. Set
380 pmp->pfs_clid = ripdata->meta.pfs_clid;
381 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
384 * The synchronization thread may start too early, make
385 * sure it stays frozen until we are ready to let it go.
389 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
390 HAMMER2_THREAD_REMASTER;
395 * Create the PFS's root inode.
397 if ((iroot = pmp->iroot) == NULL) {
398 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
400 hammer2_inode_ref(iroot);
401 hammer2_inode_unlock(iroot);
405 * Stop here if no chain is passed in.
411 * When a chain is passed in we must add it to the PFS's root
412 * inode, update pmp->pfs_types[], and update the syncronization
415 * At the moment empty spots can develop due to removals or failures.
416 * Ultimately we want to re-fill these spots but doing so might
417 * confused running code. XXX
419 hammer2_inode_ref(iroot);
420 hammer2_mtx_ex(&iroot->lock);
421 j = iroot->cluster.nchains;
423 kprintf("add PFS to pmp %p[%d]\n", pmp, j);
425 if (j == HAMMER2_MAXCLUSTER) {
426 kprintf("hammer2_mount: cluster full!\n");
427 /* XXX fatal error? */
429 KKASSERT(chain->pmp == NULL);
431 hammer2_chain_ref(chain);
432 iroot->cluster.array[j].chain = chain;
433 pmp->pfs_types[j] = ripdata->meta.pfs_type;
434 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
435 pmp->pfs_hmps[j] = chain->hmp;
438 * If the PFS is already mounted we must account
439 * for the mount_count here.
442 ++chain->hmp->mount_count;
445 * May have to fixup dirty chain tracking. Previous
446 * pmp was NULL so nothing to undo.
448 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
449 hammer2_pfs_memory_inc(pmp);
452 iroot->cluster.nchains = j;
455 * Update nmasters from any PFS inode which is part of the cluster.
456 * It is possible that this will result in a value which is too
457 * high. MASTER PFSs are authoritative for pfs_nmasters and will
458 * override this value later on.
460 * (This informs us of masters that might not currently be
461 * discoverable by this mount).
463 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
464 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
468 * Count visible masters. Masters are usually added with
469 * ripdata->meta.pfs_nmasters set to 1. This detects when there
470 * are more (XXX and must update the master inodes).
473 for (i = 0; i < iroot->cluster.nchains; ++i) {
474 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
477 if (pmp->pfs_nmasters < count)
478 pmp->pfs_nmasters = count;
481 * Create missing synchronization and support threads.
483 * Single-node masters (including snapshots) have nothing to
484 * synchronize and do not require this thread.
486 * Multi-node masters or any number of soft masters, slaves, copy,
487 * or other PFS types need the thread.
489 * Each thread is responsible for its particular cluster index.
490 * We use independent threads so stalls or mismatches related to
491 * any given target do not affect other targets.
493 for (i = 0; i < iroot->cluster.nchains; ++i) {
495 * Single-node masters (including snapshots) have nothing
496 * to synchronize and will make direct xops support calls,
497 * thus they do not require this thread.
499 * Note that there can be thousands of snapshots. We do not
500 * want to create thousands of threads.
502 if (pmp->pfs_nmasters <= 1 &&
503 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
508 * Sync support thread
510 if (pmp->sync_thrs[i].td == NULL) {
511 hammer2_thr_create(&pmp->sync_thrs[i], pmp,
513 hammer2_primary_sync_thread);
518 * Create missing Xop threads
521 hammer2_xop_helper_create(pmp);
523 hammer2_mtx_unlock(&iroot->lock);
524 hammer2_inode_drop(iroot);
530 * Destroy a PFS, typically only occurs after the last mount on a device
534 hammer2_pfsfree(hammer2_pfs_t *pmp)
536 hammer2_inode_t *iroot;
541 * Cleanup our reference on iroot. iroot is (should) not be needed
544 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
548 for (i = 0; i < iroot->cluster.nchains; ++i) {
549 hammer2_thr_delete(&pmp->sync_thrs[i]);
550 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
551 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
553 #if REPORT_REFS_ERRORS
554 if (pmp->iroot->refs != 1)
555 kprintf("PMP->IROOT %p REFS WRONG %d\n",
556 pmp->iroot, pmp->iroot->refs);
558 KKASSERT(pmp->iroot->refs == 1);
560 /* ref for pmp->iroot */
561 hammer2_inode_drop(pmp->iroot);
565 kmalloc_destroy(&pmp->mmsg);
566 kmalloc_destroy(&pmp->minode);
568 kfree(pmp, M_HAMMER2);
572 * Remove all references to hmp from the pfs list. Any PFS which becomes
573 * empty is terminated and freed.
578 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
581 hammer2_inode_t *iroot;
582 hammer2_chain_t *rchain;
588 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
589 if ((iroot = pmp->iroot) == NULL)
591 if (hmp->spmp == pmp) {
592 kprintf("unmount hmp %p remove spmp %p\n",
598 * Determine if this PFS is affected. If it is we must
599 * freeze all management threads and lock its iroot.
601 * Freezing a management thread forces it idle, operations
602 * in-progress will be aborted and it will have to start
603 * over again when unfrozen, or exit if told to exit.
605 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
606 if (pmp->pfs_hmps[i] == hmp)
609 if (i != HAMMER2_MAXCLUSTER) {
611 * Make sure all synchronization threads are locked
614 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
615 if (pmp->pfs_hmps[i] == NULL)
617 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
618 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
619 hammer2_thr_freeze_async(
620 &pmp->xop_groups[j].thrs[i]);
623 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
624 if (pmp->pfs_hmps[i] == NULL)
626 hammer2_thr_freeze(&pmp->sync_thrs[i]);
627 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
629 &pmp->xop_groups[j].thrs[i]);
634 * Lock the inode and clean out matching chains.
635 * Note that we cannot use hammer2_inode_lock_*()
636 * here because that would attempt to validate the
637 * cluster that we are in the middle of ripping
640 * WARNING! We are working directly on the inodes
643 hammer2_mtx_ex(&iroot->lock);
646 * Remove the chain from matching elements of the PFS.
648 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
649 if (pmp->pfs_hmps[i] != hmp)
651 hammer2_thr_delete(&pmp->sync_thrs[i]);
652 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
654 &pmp->xop_groups[j].thrs[i]);
656 rchain = iroot->cluster.array[i].chain;
657 iroot->cluster.array[i].chain = NULL;
658 pmp->pfs_types[i] = 0;
659 if (pmp->pfs_names[i]) {
660 kfree(pmp->pfs_names[i], M_HAMMER2);
661 pmp->pfs_names[i] = NULL;
664 hammer2_chain_drop(rchain);
666 if (iroot->cluster.focus == rchain)
667 iroot->cluster.focus = NULL;
669 pmp->pfs_hmps[i] = NULL;
671 hammer2_mtx_unlock(&iroot->lock);
672 didfreeze = 1; /* remaster, unfreeze down below */
678 * Cleanup trailing chains. Gaps may remain.
680 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
681 if (pmp->pfs_hmps[i])
684 iroot->cluster.nchains = i + 1;
687 * If the PMP has no elements remaining we can destroy it.
688 * (this will transition management threads from frozen->exit).
690 if (iroot->cluster.nchains == 0) {
691 kprintf("unmount hmp %p last ref to PMP=%p\n",
693 hammer2_pfsfree(pmp);
698 * If elements still remain we need to set the REMASTER
699 * flag and unfreeze it.
702 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
703 if (pmp->pfs_hmps[i] == NULL)
705 hammer2_thr_remaster(&pmp->sync_thrs[i]);
706 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
707 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
708 hammer2_thr_remaster(
709 &pmp->xop_groups[j].thrs[i]);
710 hammer2_thr_unfreeze(
711 &pmp->xop_groups[j].thrs[i]);
719 * Mount or remount HAMMER2 fileystem from physical media
722 * mp mount point structure
728 * mp mount point structure
729 * path path to mount point
730 * data pointer to argument structure in user space
731 * volume volume path (device@LABEL form)
732 * hflags user mount flags
733 * cred user credentials
740 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
743 struct hammer2_mount_info info;
747 hammer2_key_t key_next;
748 hammer2_key_t key_dummy;
751 struct nlookupdata nd;
752 hammer2_chain_t *parent;
753 hammer2_chain_t *chain;
754 hammer2_cluster_t *cluster;
755 const hammer2_inode_data_t *ripdata;
756 hammer2_blockref_t bref;
758 char devstr[MNAMELEN];
775 kprintf("hammer2_mount\n");
781 bzero(&info, sizeof(info));
782 info.cluster_fd = -1;
783 ksnprintf(devstr, sizeof(devstr), "%s",
784 mp->mnt_stat.f_mntfromname);
785 kprintf("hammer2_mount: root '%s'\n", devstr);
788 * Non-root mount or updating a mount
790 error = copyin(data, &info, sizeof(info));
794 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
799 /* Extract device and label */
801 label = strchr(devstr, '@');
803 ((label + 1) - dev) > done) {
811 if (mp->mnt_flag & MNT_UPDATE) {
813 * Update mount. Note that pmp->iroot->cluster is
814 * an inode-embedded cluster and thus cannot be
817 * XXX HAMMER2 needs to implement NFS export via
821 cluster = &pmp->iroot->cluster;
822 for (i = 0; i < cluster->nchains; ++i) {
823 if (cluster->array[i].chain == NULL)
825 hmp = cluster->array[i].chain->hmp;
827 error = hammer2_remount(hmp, mp, path,
839 * Lookup name and verify it refers to a block device.
842 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
844 error = nlookup(&nd);
846 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
850 cdev_t cdev = kgetdiskbyname(dev);
851 error = bdevvp(cdev, &devvp);
853 kprintf("hammer2: cannot find '%s'\n", dev);
857 if (vn_isdisk(devvp, &error))
858 error = vfs_mountedon(devvp);
862 * Determine if the device has already been mounted. After this
863 * check hmp will be non-NULL if we are doing the second or more
864 * hammer2 mounts from the same device.
866 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
867 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
868 if (hmp->devvp == devvp)
873 * Open the device if this isn't a secondary mount and construct
874 * the H2 device mount (hmp).
877 hammer2_chain_t *schain;
880 if (error == 0 && vcount(devvp) > 0)
884 * Now open the device
887 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
888 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
889 error = vinvalbuf(devvp, V_SAVE, 0, 0);
891 error = VOP_OPEN(devvp,
892 ronly ? FREAD : FREAD | FWRITE,
897 if (error && devvp) {
902 lockmgr(&hammer2_mntlk, LK_RELEASE);
905 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
906 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
909 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
910 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
911 RB_INIT(&hmp->iotree);
912 spin_init(&hmp->io_spin, "hm2mount_io");
913 spin_init(&hmp->list_spin, "hm2mount_list");
914 TAILQ_INIT(&hmp->flushq);
916 lockinit(&hmp->vollk, "h2vol", 0, 0);
917 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
920 * vchain setup. vchain.data is embedded.
921 * vchain.refs is initialized and will never drop to 0.
923 * NOTE! voldata is not yet loaded.
925 hmp->vchain.hmp = hmp;
926 hmp->vchain.refs = 1;
927 hmp->vchain.data = (void *)&hmp->voldata;
928 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
929 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
930 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
932 hammer2_chain_core_init(&hmp->vchain);
933 /* hmp->vchain.u.xxx is left NULL */
936 * fchain setup. fchain.data is embedded.
937 * fchain.refs is initialized and will never drop to 0.
939 * The data is not used but needs to be initialized to
940 * pass assertion muster. We use this chain primarily
941 * as a placeholder for the freemap's top-level RBTREE
942 * so it does not interfere with the volume's topology
945 hmp->fchain.hmp = hmp;
946 hmp->fchain.refs = 1;
947 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
948 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
949 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
950 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
951 hmp->fchain.bref.methods =
952 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
953 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
955 hammer2_chain_core_init(&hmp->fchain);
956 /* hmp->fchain.u.xxx is left NULL */
959 * Install the volume header and initialize fields from
962 error = hammer2_install_volume_header(hmp);
964 hammer2_unmount_helper(mp, NULL, hmp);
965 lockmgr(&hammer2_mntlk, LK_RELEASE);
966 hammer2_vfs_unmount(mp, MNT_FORCE);
971 * Really important to get these right or flush will get
974 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0);
975 kprintf("alloc spmp %p tid %016jx\n",
976 hmp->spmp, hmp->voldata.mirror_tid);
980 * Dummy-up vchain and fchain's modify_tid. mirror_tid
981 * is inherited from the volume header.
984 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
985 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
986 hmp->vchain.pmp = spmp;
987 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
988 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
989 hmp->fchain.pmp = spmp;
992 * First locate the super-root inode, which is key 0
993 * relative to the volume header's blockset.
995 * Then locate the root inode by scanning the directory keyspace
996 * represented by the label.
998 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
999 schain = hammer2_chain_lookup(&parent, &key_dummy,
1000 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1002 hammer2_chain_lookup_done(parent);
1003 if (schain == NULL) {
1004 kprintf("hammer2_mount: invalid super-root\n");
1005 hammer2_unmount_helper(mp, NULL, hmp);
1006 lockmgr(&hammer2_mntlk, LK_RELEASE);
1007 hammer2_vfs_unmount(mp, MNT_FORCE);
1010 if (schain->error) {
1011 kprintf("hammer2_mount: error %s reading super-root\n",
1012 hammer2_error_str(schain->error));
1013 hammer2_chain_unlock(schain);
1014 hammer2_chain_drop(schain);
1016 hammer2_unmount_helper(mp, NULL, hmp);
1017 lockmgr(&hammer2_mntlk, LK_RELEASE);
1018 hammer2_vfs_unmount(mp, MNT_FORCE);
1023 * The super-root always uses an inode_tid of 1 when
1026 spmp->inode_tid = 1;
1027 spmp->modify_tid = schain->bref.modify_tid + 1;
1030 * Sanity-check schain's pmp and finish initialization.
1031 * Any chain belonging to the super-root topology should
1032 * have a NULL pmp (not even set to spmp).
1034 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1035 KKASSERT(schain->pmp == NULL);
1036 spmp->pfs_clid = ripdata->meta.pfs_clid;
1039 * Replace the dummy spmp->iroot with a real one. It's
1040 * easier to just do a wholesale replacement than to try
1041 * to update the chain and fixup the iroot fields.
1043 * The returned inode is locked with the supplied cluster.
1045 cluster = hammer2_cluster_from_chain(schain);
1046 hammer2_inode_drop(spmp->iroot);
1048 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1049 spmp->spmp_hmp = hmp;
1050 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1051 spmp->pfs_hmps[0] = hmp;
1052 hammer2_inode_ref(spmp->iroot);
1053 hammer2_inode_unlock(spmp->iroot);
1054 hammer2_cluster_unlock(cluster);
1055 hammer2_cluster_drop(cluster);
1057 /* leave spmp->iroot with one ref */
1059 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1060 error = hammer2_recovery(hmp);
1061 /* XXX do something with error */
1063 hammer2_update_pmps(hmp);
1064 hammer2_iocom_init(hmp);
1067 * Ref the cluster management messaging descriptor. The mount
1068 * program deals with the other end of the communications pipe.
1070 * Root mounts typically do not supply one.
1072 if (info.cluster_fd >= 0) {
1073 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1075 hammer2_cluster_reconnect(hmp, fp);
1077 kprintf("hammer2_mount: bad cluster_fd!\n");
1085 * Lookup the mount point under the media-localized super-root.
1086 * Scanning hammer2_pfslist doesn't help us because it represents
1087 * PFS cluster ids which can aggregate several named PFSs together.
1089 * cluster->pmp will incorrectly point to spmp and must be fixed
1092 hammer2_inode_lock(spmp->iroot, 0);
1093 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1094 lhc = hammer2_dirhash(label, strlen(label));
1095 chain = hammer2_chain_lookup(&parent, &key_next,
1096 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1099 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1100 strcmp(label, chain->data->ipdata.filename) == 0) {
1103 chain = hammer2_chain_next(&parent, chain, &key_next,
1105 lhc + HAMMER2_DIRHASH_LOMASK,
1109 hammer2_chain_unlock(parent);
1110 hammer2_chain_drop(parent);
1112 hammer2_inode_unlock(spmp->iroot);
1115 * PFS could not be found?
1117 if (chain == NULL) {
1118 kprintf("hammer2_mount: PFS label not found\n");
1119 hammer2_unmount_helper(mp, NULL, hmp);
1120 lockmgr(&hammer2_mntlk, LK_RELEASE);
1121 hammer2_vfs_unmount(mp, MNT_FORCE);
1127 * Acquire the pmp structure (it should have already been allocated
1128 * via hammer2_update_pmps() so do not pass cluster in to add to
1129 * available chains).
1131 * Check if the cluster has already been mounted. A cluster can
1132 * only be mounted once, use null mounts to mount additional copies.
1134 ripdata = &chain->data->ipdata;
1136 pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid);
1137 hammer2_chain_unlock(chain);
1138 hammer2_chain_drop(chain);
1141 kprintf("hammer2_mount: PFS already mounted!\n");
1142 hammer2_unmount_helper(mp, NULL, hmp);
1143 lockmgr(&hammer2_mntlk, LK_RELEASE);
1144 hammer2_vfs_unmount(mp, MNT_FORCE);
1152 kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1154 mp->mnt_flag = MNT_LOCAL;
1155 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */
1156 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */
1159 * required mount structure initializations
1161 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1162 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1164 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1165 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1170 mp->mnt_iosize_max = MAXPHYS;
1173 * Connect up mount pointers.
1175 hammer2_mount_helper(mp, pmp);
1177 lockmgr(&hammer2_mntlk, LK_RELEASE);
1183 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1184 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1185 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1188 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1189 MNAMELEN - 1, &size);
1190 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1191 } /* else root mount, already in there */
1193 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1195 copyinstr(path, mp->mnt_stat.f_mntonname,
1196 sizeof(mp->mnt_stat.f_mntonname) - 1,
1200 mp->mnt_stat.f_mntonname[0] = '/';
1204 * Initial statfs to prime mnt_stat.
1206 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1212 * Scan PFSs under the super-root and create hammer2_pfs structures.
1216 hammer2_update_pmps(hammer2_dev_t *hmp)
1218 const hammer2_inode_data_t *ripdata;
1219 hammer2_chain_t *parent;
1220 hammer2_chain_t *chain;
1221 hammer2_blockref_t bref;
1222 hammer2_pfs_t *spmp;
1224 hammer2_key_t key_next;
1225 int cache_index = -1;
1228 * Lookup mount point under the media-localized super-root.
1230 * cluster->pmp will incorrectly point to spmp and must be fixed
1234 hammer2_inode_lock(spmp->iroot, 0);
1235 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1236 chain = hammer2_chain_lookup(&parent, &key_next,
1237 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1240 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1242 ripdata = &chain->data->ipdata;
1244 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1246 pmp = hammer2_pfsalloc(chain, ripdata, bref.modify_tid);
1247 chain = hammer2_chain_next(&parent, chain, &key_next,
1248 key_next, HAMMER2_KEY_MAX,
1252 hammer2_chain_unlock(parent);
1253 hammer2_chain_drop(parent);
1255 hammer2_inode_unlock(spmp->iroot);
1260 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1261 struct vnode *devvp, struct ucred *cred)
1265 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1266 error = hammer2_recovery(hmp);
1275 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1286 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1289 * If mount initialization proceeded far enough we must flush
1290 * its vnodes and sync the underlying mount points. Three syncs
1291 * are required to fully flush the filesystem (freemap updates lag
1292 * by one flush, and one extra for safety).
1294 if (mntflags & MNT_FORCE)
1299 error = vflush(mp, 0, flags);
1302 hammer2_vfs_sync(mp, MNT_WAIT);
1303 hammer2_vfs_sync(mp, MNT_WAIT);
1304 hammer2_vfs_sync(mp, MNT_WAIT);
1308 * Cleanup the frontend support XOPS threads
1310 hammer2_xop_helper_cleanup(pmp);
1313 * Cleanup our reference on ihidden.
1316 hammer2_inode_drop(pmp->ihidden);
1317 pmp->ihidden = NULL;
1320 hammer2_unmount_helper(mp, pmp, NULL);
1324 lockmgr(&hammer2_mntlk, LK_RELEASE);
1330 * Mount helper, hook the system mount into our PFS.
1331 * The mount lock is held.
1333 * We must bump the mount_count on related devices for any
1338 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1340 hammer2_cluster_t *cluster;
1341 hammer2_chain_t *rchain;
1344 mp->mnt_data = (qaddr_t)pmp;
1348 * After pmp->mp is set we have to adjust hmp->mount_count.
1350 cluster = &pmp->iroot->cluster;
1351 for (i = 0; i < cluster->nchains; ++i) {
1352 rchain = cluster->array[i].chain;
1355 ++rchain->hmp->mount_count;
1356 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1357 rchain->hmp, rchain->hmp->mount_count);
1361 * Create missing Xop threads
1363 hammer2_xop_helper_create(pmp);
1367 * Mount helper, unhook the system mount from our PFS.
1368 * The mount lock is held.
1370 * If hmp is supplied a mount responsible for being the first to open
1371 * the block device failed and the block device and all PFSs using the
1372 * block device must be cleaned up.
1374 * If pmp is supplied multiple devices might be backing the PFS and each
1375 * must be disconnect. This might not be the last PFS using some of the
1376 * underlying devices. Also, we have to adjust our hmp->mount_count
1377 * accounting for the devices backing the pmp which is now undergoing an
1382 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1384 hammer2_cluster_t *cluster;
1385 hammer2_chain_t *rchain;
1386 struct vnode *devvp;
1392 * If no device supplied this is a high-level unmount and we have to
1393 * to disconnect the mount, adjust mount_count, and locate devices
1394 * that might now have no mounts.
1397 KKASSERT(hmp == NULL);
1398 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1400 mp->mnt_data = NULL;
1403 * After pmp->mp is cleared we have to account for
1406 cluster = &pmp->iroot->cluster;
1407 for (i = 0; i < cluster->nchains; ++i) {
1408 rchain = cluster->array[i].chain;
1411 --rchain->hmp->mount_count;
1412 kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1413 rchain->hmp, rchain->hmp->mount_count);
1414 /* scrapping hmp now may invalidate the pmp */
1417 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1418 if (hmp->mount_count == 0) {
1419 hammer2_unmount_helper(NULL, NULL, hmp);
1427 * Try to terminate the block device. We can't terminate it if
1428 * there are still PFSs referencing it.
1430 kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1431 hmp, hmp->mount_count);
1432 if (hmp->mount_count)
1435 hammer2_pfsfree_scan(hmp);
1436 hammer2_dev_exlock(hmp); /* XXX order */
1439 * Cycle the volume data lock as a safety (probably not needed any
1440 * more). To ensure everything is out we need to flush at least
1441 * three times. (1) The running of the sideq can dirty the
1442 * filesystem, (2) A normal flush can dirty the freemap, and
1443 * (3) ensure that the freemap is fully synchronized.
1445 * The next mount's recovery scan can clean everything up but we want
1446 * to leave the filesystem in a 100% clean state on a normal unmount.
1449 hammer2_voldata_lock(hmp);
1450 hammer2_voldata_unlock(hmp);
1452 hammer2_iocom_uninit(hmp);
1454 if ((hmp->vchain.flags | hmp->fchain.flags) &
1455 HAMMER2_CHAIN_FLUSH_MASK) {
1456 kprintf("hammer2_unmount: chains left over "
1457 "after final sync\n");
1458 kprintf(" vchain %08x\n", hmp->vchain.flags);
1459 kprintf(" fchain %08x\n", hmp->fchain.flags);
1461 if (hammer2_debug & 0x0010)
1462 Debugger("entered debugger");
1465 KKASSERT(hmp->spmp == NULL);
1468 * Finish up with the device vnode
1470 if ((devvp = hmp->devvp) != NULL) {
1471 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1472 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1474 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1481 * Clear vchain/fchain flags that might prevent final cleanup
1484 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1485 atomic_add_long(&hammer2_count_modified_chains, -1);
1486 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1487 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1489 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1490 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1493 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1494 atomic_add_long(&hammer2_count_modified_chains, -1);
1495 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1496 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1498 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1499 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1503 * Final drop of embedded freemap root chain to
1504 * clean up fchain.core (fchain structure is not
1505 * flagged ALLOCATED so it is cleaned out and then
1508 hammer2_chain_drop(&hmp->fchain);
1511 * Final drop of embedded volume root chain to clean
1512 * up vchain.core (vchain structure is not flagged
1513 * ALLOCATED so it is cleaned out and then left to
1517 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1519 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1520 hammer2_dev_unlock(hmp);
1521 hammer2_chain_drop(&hmp->vchain);
1523 hammer2_io_cleanup(hmp, &hmp->iotree);
1524 if (hmp->iofree_count) {
1525 kprintf("io_cleanup: %d I/O's left hanging\n",
1529 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1530 kmalloc_destroy(&hmp->mchain);
1531 kfree(hmp, M_HAMMER2);
1536 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1537 ino_t ino, struct vnode **vpp)
1539 kprintf("hammer2_vget\n");
1540 return (EOPNOTSUPP);
1545 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1552 if (pmp->iroot == NULL) {
1558 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1560 while (pmp->inode_tid == 0) {
1561 hammer2_xop_ipcluster_t *xop;
1562 hammer2_inode_meta_t *meta;
1564 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1565 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1566 error = hammer2_xop_collect(&xop->head, 0);
1569 meta = &xop->head.cluster.focus->data->ipdata.meta;
1570 pmp->iroot->meta = *meta;
1571 pmp->inode_tid = meta->pfs_inum + 1;
1572 if (pmp->inode_tid < HAMMER2_INODE_START)
1573 pmp->inode_tid = HAMMER2_INODE_START;
1575 xop->head.cluster.focus->bref.modify_tid + 1;
1576 kprintf("PFS: Starting inode %jd\n",
1577 (intmax_t)pmp->inode_tid);
1578 kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1579 pmp->inode_tid, pmp->modify_tid);
1580 wakeup(&pmp->iroot);
1582 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1585 * Prime the mount info.
1587 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1590 * With the cluster operational, check for and
1591 * install ihidden if needed. The install_hidden
1592 * code needs to get a transaction so we must unlock
1595 * This is only applicable PFS mounts, there is no
1596 * hidden directory in the spmp.
1598 hammer2_inode_unlock(pmp->iroot);
1599 hammer2_inode_install_hidden(pmp);
1600 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1608 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1609 hammer2_inode_unlock(pmp->iroot);
1610 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1611 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1617 hammer2_inode_unlock(pmp->iroot);
1620 vp = hammer2_igetv(pmp->iroot, &error);
1621 hammer2_inode_unlock(pmp->iroot);
1631 * XXX incorporate ipdata->meta.inode_quota and data_quota
1635 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1639 hammer2_blockref_t bref;
1643 * NOTE: iroot might not have validated the cluster yet.
1647 mp->mnt_stat.f_files = 0;
1648 mp->mnt_stat.f_ffree = 0;
1649 mp->mnt_stat.f_blocks = 0;
1650 mp->mnt_stat.f_bfree = 0;
1651 mp->mnt_stat.f_bavail = 0;
1653 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1654 hmp = pmp->pfs_hmps[i];
1657 if (pmp->iroot->cluster.array[i].chain)
1658 bref = pmp->iroot->cluster.array[i].chain->bref;
1660 bzero(&bref, sizeof(bref));
1662 mp->mnt_stat.f_files = bref.inode_count;
1663 mp->mnt_stat.f_ffree = 0;
1664 mp->mnt_stat.f_blocks = (bref.data_count +
1665 hmp->voldata.allocator_free) /
1666 mp->mnt_vstat.f_bsize;
1667 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1668 mp->mnt_vstat.f_bsize;
1669 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1671 *sbp = mp->mnt_stat;
1678 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1682 hammer2_blockref_t bref;
1686 * NOTE: iroot might not have validated the cluster yet.
1690 mp->mnt_vstat.f_bsize = 0;
1691 mp->mnt_vstat.f_files = 0;
1692 mp->mnt_vstat.f_ffree = 0;
1693 mp->mnt_vstat.f_blocks = 0;
1694 mp->mnt_vstat.f_bfree = 0;
1695 mp->mnt_vstat.f_bavail = 0;
1697 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1698 hmp = pmp->pfs_hmps[i];
1701 if (pmp->iroot->cluster.array[i].chain)
1702 bref = pmp->iroot->cluster.array[i].chain->bref;
1704 bzero(&bref, sizeof(bref));
1706 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1707 mp->mnt_vstat.f_files = bref.inode_count;
1708 mp->mnt_vstat.f_ffree = 0;
1709 mp->mnt_vstat.f_blocks = (bref.data_count +
1710 hmp->voldata.allocator_free) /
1711 mp->mnt_vstat.f_bsize;
1712 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1713 mp->mnt_vstat.f_bsize;
1714 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1716 *sbp = mp->mnt_vstat;
1722 * Mount-time recovery (RW mounts)
1724 * Updates to the free block table are allowed to lag flushes by one
1725 * transaction. In case of a crash, then on a fresh mount we must do an
1726 * incremental scan of the last committed transaction id and make sure that
1727 * all related blocks have been marked allocated.
1729 * The super-root topology and each PFS has its own transaction id domain,
1730 * so we must track PFS boundary transitions.
1732 struct hammer2_recovery_elm {
1733 TAILQ_ENTRY(hammer2_recovery_elm) entry;
1734 hammer2_chain_t *chain;
1735 hammer2_tid_t sync_tid;
1738 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1740 struct hammer2_recovery_info {
1741 struct hammer2_recovery_list list;
1746 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1747 hammer2_chain_t *parent,
1748 struct hammer2_recovery_info *info,
1749 hammer2_tid_t sync_tid);
1751 #define HAMMER2_RECOVERY_MAXDEPTH 10
1755 hammer2_recovery(hammer2_dev_t *hmp)
1757 struct hammer2_recovery_info info;
1758 struct hammer2_recovery_elm *elm;
1759 hammer2_chain_t *parent;
1760 hammer2_tid_t sync_tid;
1761 hammer2_tid_t mirror_tid;
1763 int cumulative_error = 0;
1765 hammer2_trans_init(hmp->spmp, 0);
1767 sync_tid = hmp->voldata.freemap_tid;
1768 mirror_tid = hmp->voldata.mirror_tid;
1770 kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1771 if (sync_tid >= mirror_tid) {
1772 kprintf(" no recovery needed\n");
1774 kprintf(" freemap recovery %016jx-%016jx\n",
1775 sync_tid + 1, mirror_tid);
1778 TAILQ_INIT(&info.list);
1780 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1781 cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
1782 hammer2_chain_lookup_done(parent);
1784 while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1785 TAILQ_REMOVE(&info.list, elm, entry);
1786 parent = elm->chain;
1787 sync_tid = elm->sync_tid;
1788 kfree(elm, M_HAMMER2);
1790 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1791 error = hammer2_recovery_scan(hmp, parent, &info,
1792 hmp->voldata.freemap_tid);
1793 hammer2_chain_unlock(parent);
1794 hammer2_chain_drop(parent); /* drop elm->chain ref */
1796 cumulative_error = error;
1798 hammer2_trans_done(hmp->spmp);
1800 return cumulative_error;
1805 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1806 struct hammer2_recovery_info *info,
1807 hammer2_tid_t sync_tid)
1809 const hammer2_inode_data_t *ripdata;
1810 hammer2_chain_t *chain;
1811 hammer2_blockref_t bref;
1813 int cumulative_error = 0;
1818 * Adjust freemap to ensure that the block(s) are marked allocated.
1820 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1821 hammer2_freemap_adjust(hmp, &parent->bref,
1822 HAMMER2_FREEMAP_DORECOVER);
1826 * Check type for recursive scan
1828 switch(parent->bref.type) {
1829 case HAMMER2_BREF_TYPE_VOLUME:
1830 /* data already instantiated */
1832 case HAMMER2_BREF_TYPE_INODE:
1834 * Must instantiate data for DIRECTDATA test and also
1837 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1838 ripdata = &hammer2_chain_rdata(parent)->ipdata;
1839 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1840 /* not applicable to recovery scan */
1841 hammer2_chain_unlock(parent);
1844 hammer2_chain_unlock(parent);
1846 case HAMMER2_BREF_TYPE_INDIRECT:
1848 * Must instantiate data for recursion
1850 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1851 hammer2_chain_unlock(parent);
1853 case HAMMER2_BREF_TYPE_DATA:
1854 case HAMMER2_BREF_TYPE_FREEMAP:
1855 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1856 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1857 /* not applicable to recovery scan */
1865 * Defer operation if depth limit reached or if we are crossing a
1868 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1869 struct hammer2_recovery_elm *elm;
1871 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1872 elm->chain = parent;
1873 elm->sync_tid = sync_tid;
1874 hammer2_chain_ref(parent);
1875 TAILQ_INSERT_TAIL(&info->list, elm, entry);
1876 /* unlocked by caller */
1883 * Recursive scan of the last flushed transaction only. We are
1884 * doing this without pmp assignments so don't leave the chains
1885 * hanging around after we are done with them.
1891 while (hammer2_chain_scan(parent, &chain, &bref,
1892 &first, &cache_index,
1893 HAMMER2_LOOKUP_NODATA) != NULL) {
1897 if (chain == NULL) {
1898 if (bref.mirror_tid > sync_tid) {
1899 hammer2_freemap_adjust(hmp, &bref,
1900 HAMMER2_FREEMAP_DORECOVER);
1906 * This may or may not be a recursive node.
1908 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
1909 if (bref.mirror_tid > sync_tid) {
1911 error = hammer2_recovery_scan(hmp, chain,
1915 cumulative_error = error;
1919 * Flush the recovery at the PFS boundary to stage it for
1920 * the final flush of the super-root topology.
1922 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
1923 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
1924 hammer2_flush(chain, HAMMER2_FLUSH_TOP);
1928 return cumulative_error;
1932 * Sync a mount point; this is called on a per-mount basis from the
1933 * filesystem syncer process periodically and whenever a user issues
1937 hammer2_vfs_sync(struct mount *mp, int waitfor)
1939 hammer2_xop_flush_t *xop;
1940 struct hammer2_sync_info info;
1941 hammer2_inode_t *iroot;
1949 KKASSERT(iroot->pmp == pmp);
1952 * We can't acquire locks on existing vnodes while in a transaction
1953 * without risking a deadlock. This assumes that vfsync() can be
1954 * called without the vnode locked (which it can in DragonFly).
1955 * Otherwise we'd have to implement a multi-pass or flag the lock
1956 * failures and retry.
1958 * The reclamation code interlocks with the sync list's token
1959 * (by removing the vnode from the scan list) before unlocking
1960 * the inode, giving us time to ref the inode.
1962 /*flags = VMSC_GETVP;*/
1964 if (waitfor & MNT_LAZY)
1965 flags |= VMSC_ONEPASS;
1969 * Preflush the vnodes using a normal transaction before interlocking
1970 * with a flush transaction.
1972 hammer2_trans_init(pmp, 0);
1974 info.waitfor = MNT_NOWAIT;
1975 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1976 hammer2_trans_done(pmp);
1980 * Start our flush transaction. This does not return until all
1981 * concurrent transactions have completed and will prevent any
1982 * new transactions from running concurrently, except for the
1983 * buffer cache transactions.
1985 * For efficiency do an async pass before making sure with a
1986 * synchronous pass on all related buffer cache buffers. It
1987 * should theoretically not be possible for any new file buffers
1988 * to be instantiated during this sequence.
1990 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH |
1991 HAMMER2_TRANS_PREFLUSH);
1992 hammer2_inode_run_sideq(pmp);
1995 info.waitfor = MNT_NOWAIT;
1996 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1997 info.waitfor = MNT_WAIT;
1998 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2001 * Clear PREFLUSH. This prevents (or asserts on) any new logical
2002 * buffer cache flushes which occur during the flush. Device buffers
2005 hammer2_bioq_sync(pmp);
2006 hammer2_trans_clear_preflush(pmp);
2009 * Use the XOP interface to concurrently flush all nodes to
2010 * synchronize the PFSROOT subtopology to the media. A standard
2011 * end-of-scan ENOENT error indicates cluster sufficiency.
2013 * Note that this flush will not be visible on crash recovery until
2014 * we flush the super-root topology in the next loop.
2016 * XXX For now wait for all flushes to complete.
2019 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2020 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2021 error = hammer2_xop_collect(&xop->head,
2022 HAMMER2_XOP_COLLECT_WAITALL);
2023 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2024 if (error == ENOENT)
2029 hammer2_trans_done(pmp);
2037 * Note that we ignore the tranasction mtid we got above. Instead,
2038 * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2042 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2044 struct hammer2_sync_info *info = data;
2045 hammer2_inode_t *ip;
2049 * Degenerate cases. Note that ip == NULL typically means the
2050 * syncer vnode itself and we don't want to vclrisdirty() in that
2057 if (vp->v_type == VNON || vp->v_type == VBAD) {
2063 * VOP_FSYNC will start a new transaction so replicate some code
2064 * here to do it inline (see hammer2_vop_fsync()).
2066 * WARNING: The vfsync interacts with the buffer cache and might
2067 * block, we can't hold the inode lock at that time.
2068 * However, we MUST ref ip before blocking to ensure that
2069 * it isn't ripped out from under us (since we do not
2070 * hold a lock on the vnode).
2072 hammer2_inode_ref(ip);
2073 if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2074 !RB_EMPTY(&vp->v_rbdirty_tree)) {
2075 vfsync(vp, info->waitfor, 1, NULL, NULL);
2076 if (ip->flags & (HAMMER2_INODE_RESIZED |
2077 HAMMER2_INODE_MODIFIED)) {
2078 hammer2_inode_lock(ip, 0);
2079 hammer2_inode_chain_sync(ip);
2080 hammer2_inode_unlock(ip);
2083 if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2084 RB_EMPTY(&vp->v_rbdirty_tree)) {
2088 hammer2_inode_drop(ip);
2092 info->error = error;
2099 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2106 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2107 struct fid *fhp, struct vnode **vpp)
2114 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2115 int *exflagsp, struct ucred **credanonp)
2121 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume
2122 * header into the HMP
2124 * XXX read four volhdrs and use the one with the highest TID whos CRC
2129 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2130 * nonexistant locations.
2132 * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2136 hammer2_install_volume_header(hammer2_dev_t *hmp)
2138 hammer2_volume_data_t *vd;
2140 hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2152 * There are up to 4 copies of the volume header (syncs iterate
2153 * between them so there is no single master). We don't trust the
2154 * volu_size field so we don't know precisely how large the filesystem
2155 * is, so depend on the OS to return an error if we go beyond the
2156 * block device's EOF.
2158 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2159 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2160 HAMMER2_VOLUME_BYTES, &bp);
2167 vd = (struct hammer2_volume_data *) bp->b_data;
2168 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2169 (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2175 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2176 /* XXX: Reversed-endianness filesystem */
2177 kprintf("hammer2: reverse-endian filesystem detected");
2183 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2184 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2185 HAMMER2_VOLUME_ICRC0_SIZE);
2186 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2187 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2188 HAMMER2_VOLUME_ICRC1_SIZE);
2189 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2190 kprintf("hammer2 volume header crc "
2191 "mismatch copy #%d %08x/%08x\n",
2198 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2207 hmp->volsync = hmp->voldata;
2209 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2210 kprintf("hammer2: using volume header #%d\n",
2215 kprintf("hammer2: no valid volume headers found!\n");
2221 * This handles hysteresis on regular file flushes. Because the BIOs are
2222 * routed to a thread it is possible for an excessive number to build up
2223 * and cause long front-end stalls long before the runningbuffspace limit
2224 * is hit, so we implement hammer2_flush_pipe to control the
2227 * This is a particular problem when compression is used.
2230 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2232 atomic_add_int(&pmp->count_lwinprog, 1);
2236 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2240 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2241 if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2242 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2243 atomic_clear_int(&pmp->count_lwinprog,
2244 HAMMER2_LWINPROG_WAITING);
2245 wakeup(&pmp->count_lwinprog);
2247 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2248 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2249 atomic_clear_int(&pmp->count_lwinprog,
2250 HAMMER2_LWINPROG_WAITING0);
2251 wakeup(&pmp->count_lwinprog);
2256 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2259 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2260 HAMMER2_LWINPROG_WAITING0;
2263 lwinprog = pmp->count_lwinprog;
2265 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2267 tsleep_interlock(&pmp->count_lwinprog, 0);
2268 atomic_set_int(&pmp->count_lwinprog, lwflag);
2269 lwinprog = pmp->count_lwinprog;
2270 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2272 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2277 * Manage excessive memory resource use for chain and related
2281 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2291 * Atomic check condition and wait. Also do an early speedup of
2292 * the syncer to try to avoid hitting the wait.
2295 waiting = pmp->inmem_dirty_chains;
2297 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2299 limit = pmp->mp->mnt_nvnodelistsize / 10;
2300 if (limit < hammer2_limit_dirty_chains)
2301 limit = hammer2_limit_dirty_chains;
2306 if ((int)(ticks - zzticks) > hz) {
2308 kprintf("count %ld %ld\n", count, limit);
2313 * Block if there are too many dirty chains present, wait
2314 * for the flush to clean some out.
2316 if (count > limit) {
2317 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2318 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2320 waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2321 speedup_syncer(pmp->mp);
2322 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2325 continue; /* loop on success or fail */
2329 * Try to start an early flush before we are forced to block.
2331 if (count > limit * 7 / 10)
2332 speedup_syncer(pmp->mp);
2338 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2341 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2346 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2354 waiting = pmp->inmem_dirty_chains;
2356 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2359 ~HAMMER2_DIRTYCHAIN_WAITING)) {
2364 if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2365 wakeup(&pmp->inmem_dirty_chains);
2372 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2374 hammer2_chain_t *scan;
2375 hammer2_chain_t *parent;
2379 kprintf("%*.*s...\n", tab, tab, "");
2384 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2386 chain, chain->bref.type,
2387 chain->bref.key, chain->bref.keybits,
2388 chain->bref.mirror_tid);
2390 kprintf("%*.*s [%08x] (%s) refs=%d",
2393 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2394 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
2397 parent = chain->parent;
2399 kprintf("\n%*.*s p=%p [pflags %08x prefs %d",
2401 parent, parent->flags, parent->refs);
2402 if (RB_EMPTY(&chain->core.rbtree)) {
2406 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2407 hammer2_dump_chain(scan, tab + 4, countp, 'a');
2408 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2409 kprintf("%*.*s}(%s)\n", tab, tab, "",
2410 chain->data->ipdata.filename);
2412 kprintf("%*.*s}\n", tab, tab, "");