xref: /dflybsd-src/sys/vfs/hammer2/hammer2_vfsops.c (revision 617f6bd6553bb7523bb0198e7aaa5d9c36fb6e05)
1 /*
2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
3  *
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)
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  *
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
17  *    distribution.
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.
21  *
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
33  * SUCH DAMAGE.
34  */
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>
42 #include <sys/buf.h>
43 #include <sys/uuid.h>
44 #include <sys/vfsops.h>
45 #include <sys/sysctl.h>
46 #include <sys/socket.h>
47 #include <sys/objcache.h>
48 
49 #include <sys/proc.h>
50 #include <sys/namei.h>
51 #include <sys/mountctl.h>
52 #include <sys/dirent.h>
53 #include <sys/uio.h>
54 
55 #include <sys/mutex.h>
56 #include <sys/mutex2.h>
57 
58 #include "hammer2.h"
59 #include "hammer2_disk.h"
60 #include "hammer2_mount.h"
61 #include "hammer2_lz4.h"
62 
63 #include "zlib/hammer2_zlib.h"
64 
65 #define REPORT_REFS_ERRORS 1	/* XXX remove me */
66 
67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
68 
69 struct hammer2_sync_info {
70 	int error;
71 	int waitfor;
72 };
73 
74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
75 static struct hammer2_mntlist hammer2_mntlist;
76 
77 struct hammer2_pfslist hammer2_pfslist;
78 struct lock hammer2_mntlk;
79 
80 int hammer2_debug;
81 int hammer2_cluster_meta_read = 1;	/* physical read-ahead */
82 int hammer2_cluster_data_read = 4;	/* physical read-ahead */
83 int hammer2_dedup_enable = 1;
84 int hammer2_always_compress = 0;	/* always try to compress */
85 int hammer2_inval_enable = 0;
86 int hammer2_flush_pipe = 100;
87 int hammer2_synchronous_flush = 1;
88 int hammer2_dio_count;
89 int hammer2_limit_dio = 256;
90 int hammer2_bulkfree_tps = 5000;
91 long hammer2_chain_allocs;
92 long hammer2_chain_frees;
93 long hammer2_limit_dirty_chains;
94 long hammer2_count_modified_chains;
95 long hammer2_iod_invals;
96 long hammer2_iod_file_read;
97 long hammer2_iod_meta_read;
98 long hammer2_iod_indr_read;
99 long hammer2_iod_fmap_read;
100 long hammer2_iod_volu_read;
101 long hammer2_iod_file_write;
102 long hammer2_iod_file_wembed;
103 long hammer2_iod_file_wzero;
104 long hammer2_iod_file_wdedup;
105 long hammer2_iod_meta_write;
106 long hammer2_iod_indr_write;
107 long hammer2_iod_fmap_write;
108 long hammer2_iod_volu_write;
109 
110 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
111 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
112 		"Buffer used for compression.");
113 
114 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
115 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
116 		"Buffer used for decompression.");
117 
118 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
119 
120 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
121 	   &hammer2_debug, 0, "");
122 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
123 	   &hammer2_cluster_meta_read, 0, "");
124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
125 	   &hammer2_cluster_data_read, 0, "");
126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
127 	   &hammer2_dedup_enable, 0, "");
128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
129 	   &hammer2_always_compress, 0, "");
130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
131 	   &hammer2_inval_enable, 0, "");
132 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
133 	   &hammer2_flush_pipe, 0, "");
134 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
135 	   &hammer2_synchronous_flush, 0, "");
136 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
137 	   &hammer2_bulkfree_tps, 0, "");
138 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
139 	   &hammer2_chain_allocs, 0, "");
140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
141 	   &hammer2_chain_frees, 0, "");
142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
143 	   &hammer2_limit_dirty_chains, 0, "");
144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
145 	   &hammer2_count_modified_chains, 0, "");
146 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
147 	   &hammer2_dio_count, 0, "");
148 SYSCTL_INT(_vfs_hammer2, OID_AUTO, limit_dio, CTLFLAG_RW,
149 	   &hammer2_limit_dio, 0, "");
150 
151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
152 	   &hammer2_iod_invals, 0, "");
153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
154 	   &hammer2_iod_file_read, 0, "");
155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
156 	   &hammer2_iod_meta_read, 0, "");
157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
158 	   &hammer2_iod_indr_read, 0, "");
159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
160 	   &hammer2_iod_fmap_read, 0, "");
161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
162 	   &hammer2_iod_volu_read, 0, "");
163 
164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
165 	   &hammer2_iod_file_write, 0, "");
166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
167 	   &hammer2_iod_file_wembed, 0, "");
168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
169 	   &hammer2_iod_file_wzero, 0, "");
170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
171 	   &hammer2_iod_file_wdedup, 0, "");
172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
173 	   &hammer2_iod_meta_write, 0, "");
174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
175 	   &hammer2_iod_indr_write, 0, "");
176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
177 	   &hammer2_iod_fmap_write, 0, "");
178 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
179 	   &hammer2_iod_volu_write, 0, "");
180 
181 long hammer2_check_icrc32;
182 long hammer2_check_xxhash64;
183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
184 	   &hammer2_check_icrc32, 0, "");
185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
186 	   &hammer2_check_xxhash64, 0, "");
187 
188 static int hammer2_vfs_init(struct vfsconf *conf);
189 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
190 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
191 				struct ucred *cred);
192 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
193 				struct vnode *, struct ucred *);
194 static int hammer2_recovery(hammer2_dev_t *hmp);
195 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
196 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
197 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
198 				struct ucred *cred);
199 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
200 				struct ucred *cred);
201 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
202 				struct fid *fhp, struct vnode **vpp);
203 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
204 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
205 				int *exflagsp, struct ucred **credanonp);
206 
207 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
208 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
209 
210 static void hammer2_update_pmps(hammer2_dev_t *hmp);
211 
212 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
213 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
214 				hammer2_dev_t *hmp);
215 
216 /*
217  * HAMMER2 vfs operations.
218  */
219 static struct vfsops hammer2_vfsops = {
220 	.vfs_init	= hammer2_vfs_init,
221 	.vfs_uninit	= hammer2_vfs_uninit,
222 	.vfs_sync	= hammer2_vfs_sync,
223 	.vfs_mount	= hammer2_vfs_mount,
224 	.vfs_unmount	= hammer2_vfs_unmount,
225 	.vfs_root 	= hammer2_vfs_root,
226 	.vfs_statfs	= hammer2_vfs_statfs,
227 	.vfs_statvfs	= hammer2_vfs_statvfs,
228 	.vfs_vget	= hammer2_vfs_vget,
229 	.vfs_vptofh	= hammer2_vfs_vptofh,
230 	.vfs_fhtovp	= hammer2_vfs_fhtovp,
231 	.vfs_checkexp	= hammer2_vfs_checkexp
232 };
233 
234 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
235 
236 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
237 MODULE_VERSION(hammer2, 1);
238 
239 static
240 int
241 hammer2_vfs_init(struct vfsconf *conf)
242 {
243 	static struct objcache_malloc_args margs_read;
244 	static struct objcache_malloc_args margs_write;
245 	static struct objcache_malloc_args margs_vop;
246 
247 	int error;
248 
249 	error = 0;
250 
251 	/*
252 	 * A large DIO cache is needed to retain dedup enablement masks.
253 	 * The bulkfree code clears related masks as part of the disk block
254 	 * recycling algorithm, preventing it from being used for a later
255 	 * dedup.
256 	 *
257 	 * NOTE: A large buffer cache can actually interfere with dedup
258 	 *	 operation because we dedup based on media physical buffers
259 	 *	 and not logical buffers.  Try to make the DIO chace large
260 	 *	 enough to avoid this problem, but also cap it.
261 	 */
262 	hammer2_limit_dio = nbuf * 2;
263 	if (hammer2_limit_dio > 100000)
264 		hammer2_limit_dio = 100000;
265 
266 	if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
267 		error = EINVAL;
268 	if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
269 		error = EINVAL;
270 	if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
271 		error = EINVAL;
272 
273 	if (error)
274 		kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
275 
276 	margs_read.objsize = 65536;
277 	margs_read.mtype = M_HAMMER2_DEBUFFER;
278 
279 	margs_write.objsize = 32768;
280 	margs_write.mtype = M_HAMMER2_CBUFFER;
281 
282 	margs_vop.objsize = sizeof(hammer2_xop_t);
283 	margs_vop.mtype = M_HAMMER2;
284 
285 	/*
286 	 * Note thaht for the XOPS cache we want backing store allocations
287 	 * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
288 	 * confusion), so use the backing store function that does it.  This
289 	 * means that initial XOPS objects are zerod but REUSED objects are
290 	 * not.  So we are responsible for cleaning the object up sufficiently
291 	 * for our needs before objcache_put()ing it back (typically just the
292 	 * FIFO indices).
293 	 */
294 	cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
295 				0, 1, NULL, NULL, NULL,
296 				objcache_malloc_alloc,
297 				objcache_malloc_free,
298 				&margs_read);
299 	cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
300 				0, 1, NULL, NULL, NULL,
301 				objcache_malloc_alloc,
302 				objcache_malloc_free,
303 				&margs_write);
304 	cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
305 				0, 1, NULL, NULL, NULL,
306 				objcache_malloc_alloc_zero,
307 				objcache_malloc_free,
308 				&margs_vop);
309 
310 
311 	lockinit(&hammer2_mntlk, "mntlk", 0, 0);
312 	TAILQ_INIT(&hammer2_mntlist);
313 	TAILQ_INIT(&hammer2_pfslist);
314 
315 	hammer2_limit_dirty_chains = maxvnodes / 10;
316 	if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
317 		hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
318 
319 	return (error);
320 }
321 
322 static
323 int
324 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
325 {
326 	objcache_destroy(cache_buffer_read);
327 	objcache_destroy(cache_buffer_write);
328 	objcache_destroy(cache_xops);
329 	return 0;
330 }
331 
332 /*
333  * Core PFS allocator.  Used to allocate or reference the pmp structure
334  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
335  * The pmp can be passed in or loaded by this function using the chain and
336  * inode data.
337  *
338  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
339  * transactions.  Note that synchronization does not use this field.
340  * (typically frontend operations and synchronization cannot run on the
341  * same PFS node at the same time).
342  *
343  * XXX check locking
344  */
345 hammer2_pfs_t *
346 hammer2_pfsalloc(hammer2_chain_t *chain,
347 		 const hammer2_inode_data_t *ripdata,
348 		 hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
349 {
350 	hammer2_pfs_t *pmp;
351 	hammer2_inode_t *iroot;
352 	int count;
353 	int i;
354 	int j;
355 
356 	pmp = NULL;
357 
358 	/*
359 	 * Locate or create the PFS based on the cluster id.  If ripdata
360 	 * is NULL this is a spmp which is unique and is always allocated.
361 	 *
362 	 * If the device is mounted in local mode all PFSs are considered
363 	 * independent and not part of any cluster (for debugging only).
364 	 */
365 	if (ripdata) {
366 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
367 			if (force_local != pmp->force_local)
368 				continue;
369 			if (force_local == NULL &&
370 			    bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
371 				 sizeof(pmp->pfs_clid)) == 0) {
372 					break;
373 			} else if (force_local && pmp->pfs_names[0] &&
374 			    strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
375 					break;
376 			}
377 		}
378 	}
379 
380 	if (pmp == NULL) {
381 		pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
382 		pmp->force_local = force_local;
383 		hammer2_trans_manage_init(pmp);
384 		kmalloc_create(&pmp->minode, "HAMMER2-inodes");
385 		kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
386 		lockinit(&pmp->lock, "pfslk", 0, 0);
387 		lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
388 		spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
389 		spin_init(&pmp->xop_spin, "h2xop");
390 		spin_init(&pmp->lru_spin, "h2lru");
391 		RB_INIT(&pmp->inum_tree);
392 		TAILQ_INIT(&pmp->sideq);
393 		TAILQ_INIT(&pmp->lru_list);
394 		spin_init(&pmp->list_spin, "hm2pfsalloc_list");
395 
396 		/*
397 		 * Distribute backend operations to threads
398 		 */
399 		for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
400 			hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
401 
402 		/*
403 		 * Save the last media transaction id for the flusher.  Set
404 		 * initial
405 		 */
406 		if (ripdata)
407 			pmp->pfs_clid = ripdata->meta.pfs_clid;
408 		TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
409 
410 		/*
411 		 * The synchronization thread may start too early, make
412 		 * sure it stays frozen until we are ready to let it go.
413 		 * XXX
414 		 */
415 		/*
416 		pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
417 					 HAMMER2_THREAD_REMASTER;
418 		*/
419 	}
420 
421 	/*
422 	 * Create the PFS's root inode and any missing XOP helper threads.
423 	 */
424 	if ((iroot = pmp->iroot) == NULL) {
425 		iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
426 		if (ripdata)
427 			iroot->meta = ripdata->meta;
428 		pmp->iroot = iroot;
429 		hammer2_inode_ref(iroot);
430 		hammer2_inode_unlock(iroot);
431 	}
432 
433 	/*
434 	 * Stop here if no chain is passed in.
435 	 */
436 	if (chain == NULL)
437 		goto done;
438 
439 	/*
440 	 * When a chain is passed in we must add it to the PFS's root
441 	 * inode, update pmp->pfs_types[], and update the syncronization
442 	 * threads.
443 	 *
444 	 * When forcing local mode, mark the PFS as a MASTER regardless.
445 	 *
446 	 * At the moment empty spots can develop due to removals or failures.
447 	 * Ultimately we want to re-fill these spots but doing so might
448 	 * confused running code. XXX
449 	 */
450 	hammer2_inode_ref(iroot);
451 	hammer2_mtx_ex(&iroot->lock);
452 	j = iroot->cluster.nchains;
453 
454 	if (j == HAMMER2_MAXCLUSTER) {
455 		kprintf("hammer2_mount: cluster full!\n");
456 		/* XXX fatal error? */
457 	} else {
458 		KKASSERT(chain->pmp == NULL);
459 		chain->pmp = pmp;
460 		hammer2_chain_ref(chain);
461 		iroot->cluster.array[j].chain = chain;
462 		if (force_local)
463 			pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
464 		else
465 			pmp->pfs_types[j] = ripdata->meta.pfs_type;
466 		pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
467 		pmp->pfs_hmps[j] = chain->hmp;
468 
469 		/*
470 		 * If the PFS is already mounted we must account
471 		 * for the mount_count here.
472 		 */
473 		if (pmp->mp)
474 			++chain->hmp->mount_count;
475 
476 		/*
477 		 * May have to fixup dirty chain tracking.  Previous
478 		 * pmp was NULL so nothing to undo.
479 		 */
480 		if (chain->flags & HAMMER2_CHAIN_MODIFIED)
481 			hammer2_pfs_memory_inc(pmp);
482 		++j;
483 	}
484 	iroot->cluster.nchains = j;
485 
486 	/*
487 	 * Update nmasters from any PFS inode which is part of the cluster.
488 	 * It is possible that this will result in a value which is too
489 	 * high.  MASTER PFSs are authoritative for pfs_nmasters and will
490 	 * override this value later on.
491 	 *
492 	 * (This informs us of masters that might not currently be
493 	 *  discoverable by this mount).
494 	 */
495 	if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
496 		pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
497 	}
498 
499 	/*
500 	 * Count visible masters.  Masters are usually added with
501 	 * ripdata->meta.pfs_nmasters set to 1.  This detects when there
502 	 * are more (XXX and must update the master inodes).
503 	 */
504 	count = 0;
505 	for (i = 0; i < iroot->cluster.nchains; ++i) {
506 		if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
507 			++count;
508 	}
509 	if (pmp->pfs_nmasters < count)
510 		pmp->pfs_nmasters = count;
511 
512 	/*
513 	 * Create missing synchronization and support threads.
514 	 *
515 	 * Single-node masters (including snapshots) have nothing to
516 	 * synchronize and do not require this thread.
517 	 *
518 	 * Multi-node masters or any number of soft masters, slaves, copy,
519 	 * or other PFS types need the thread.
520 	 *
521 	 * Each thread is responsible for its particular cluster index.
522 	 * We use independent threads so stalls or mismatches related to
523 	 * any given target do not affect other targets.
524 	 */
525 	for (i = 0; i < iroot->cluster.nchains; ++i) {
526 		/*
527 		 * Single-node masters (including snapshots) have nothing
528 		 * to synchronize and will make direct xops support calls,
529 		 * thus they do not require this thread.
530 		 *
531 		 * Note that there can be thousands of snapshots.  We do not
532 		 * want to create thousands of threads.
533 		 */
534 		if (pmp->pfs_nmasters <= 1 &&
535 		    pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
536 			continue;
537 		}
538 
539 		/*
540 		 * Sync support thread
541 		 */
542 		if (pmp->sync_thrs[i].td == NULL) {
543 			hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
544 					   "h2nod", i, -1,
545 					   hammer2_primary_sync_thread);
546 		}
547 	}
548 
549 	/*
550 	 * Create missing Xop threads
551 	 *
552 	 * NOTE: We create helper threads for all mounted PFSs or any
553 	 *	 PFSs with 2+ nodes (so the sync thread can update them,
554 	 *	 even if not mounted).
555 	 */
556 	if (pmp->mp || iroot->cluster.nchains >= 2)
557 		hammer2_xop_helper_create(pmp);
558 
559 	hammer2_mtx_unlock(&iroot->lock);
560 	hammer2_inode_drop(iroot);
561 done:
562 	return pmp;
563 }
564 
565 /*
566  * Deallocate an element of a probed PFS.  If destroying and this is a
567  * MASTER, adjust nmasters.
568  *
569  * This function does not physically destroy the PFS element in its device
570  * under the super-root  (see hammer2_ioctl_pfs_delete()).
571  */
572 void
573 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
574 {
575 	hammer2_inode_t *iroot;
576 	hammer2_chain_t *chain;
577 	int j;
578 
579 	/*
580 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
581 	 * by the flush code.
582 	 */
583 	iroot = pmp->iroot;
584 	if (iroot) {
585 		/*
586 		 * Stop synchronizing
587 		 *
588 		 * XXX flush after acquiring the iroot lock.
589 		 * XXX clean out the cluster index from all inode structures.
590 		 */
591 		hammer2_thr_delete(&pmp->sync_thrs[clindex]);
592 
593 		/*
594 		 * Remove the cluster index from the group.  If destroying
595 		 * the PFS and this is a master, adjust pfs_nmasters.
596 		 */
597 		hammer2_mtx_ex(&iroot->lock);
598 		chain = iroot->cluster.array[clindex].chain;
599 		iroot->cluster.array[clindex].chain = NULL;
600 
601 		switch(pmp->pfs_types[clindex]) {
602 		case HAMMER2_PFSTYPE_MASTER:
603 			if (destroying && pmp->pfs_nmasters > 0)
604 				--pmp->pfs_nmasters;
605 			/* XXX adjust ripdata->meta.pfs_nmasters */
606 			break;
607 		default:
608 			break;
609 		}
610 		pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
611 
612 		hammer2_mtx_unlock(&iroot->lock);
613 
614 		/*
615 		 * Release the chain.
616 		 */
617 		if (chain) {
618 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
619 			hammer2_chain_drop(chain);
620 		}
621 
622 		/*
623 		 * Terminate all XOP threads for the cluster index.
624 		 */
625 		for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
626 			hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
627 	}
628 }
629 
630 /*
631  * Destroy a PFS, typically only occurs after the last mount on a device
632  * has gone away.
633  */
634 static void
635 hammer2_pfsfree(hammer2_pfs_t *pmp)
636 {
637 	hammer2_inode_t *iroot;
638 	hammer2_chain_t *chain;
639 	int i;
640 	int j;
641 
642 	/*
643 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
644 	 * by the flush code.
645 	 */
646 	TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
647 
648 	iroot = pmp->iroot;
649 	if (iroot) {
650 		for (i = 0; i < iroot->cluster.nchains; ++i) {
651 			hammer2_thr_delete(&pmp->sync_thrs[i]);
652 			for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
653 				hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
654 		}
655 #if REPORT_REFS_ERRORS
656 		if (pmp->iroot->refs != 1)
657 			kprintf("PMP->IROOT %p REFS WRONG %d\n",
658 				pmp->iroot, pmp->iroot->refs);
659 #else
660 		KKASSERT(pmp->iroot->refs == 1);
661 #endif
662 		/* ref for pmp->iroot */
663 		hammer2_inode_drop(pmp->iroot);
664 		pmp->iroot = NULL;
665 	}
666 
667 	/*
668 	 * Cleanup chains remaining on LRU list.
669 	 */
670 	while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
671 		hammer2_chain_ref(chain);
672 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
673 		hammer2_chain_drop(chain);
674 	}
675 
676 	/*
677 	 * Free remaining pmp resources
678 	 */
679 	kmalloc_destroy(&pmp->mmsg);
680 	kmalloc_destroy(&pmp->minode);
681 
682 	kfree(pmp, M_HAMMER2);
683 }
684 
685 /*
686  * Remove all references to hmp from the pfs list.  Any PFS which becomes
687  * empty is terminated and freed.
688  *
689  * XXX inefficient.
690  */
691 static void
692 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
693 {
694 	hammer2_pfs_t *pmp;
695 	hammer2_inode_t *iroot;
696 	hammer2_chain_t *rchain;
697 	int didfreeze;
698 	int i;
699 	int j;
700 
701 again:
702 	TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
703 		if ((iroot = pmp->iroot) == NULL)
704 			continue;
705 		hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
706 		hammer2_inode_run_sideq(pmp, 1);
707 		hammer2_bioq_sync(pmp);
708 		hammer2_trans_done(pmp);
709 		if (hmp->spmp == pmp) {
710 			hmp->spmp = NULL;
711 			hmp->vchain.pmp = NULL;
712 			hmp->fchain.pmp = NULL;
713 		}
714 
715 		/*
716 		 * Determine if this PFS is affected.  If it is we must
717 		 * freeze all management threads and lock its iroot.
718 		 *
719 		 * Freezing a management thread forces it idle, operations
720 		 * in-progress will be aborted and it will have to start
721 		 * over again when unfrozen, or exit if told to exit.
722 		 */
723 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
724 			if (pmp->pfs_hmps[i] == hmp)
725 				break;
726 		}
727 		if (i != HAMMER2_MAXCLUSTER) {
728 			/*
729 			 * Make sure all synchronization threads are locked
730 			 * down.
731 			 */
732 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
733 				if (pmp->pfs_hmps[i] == NULL)
734 					continue;
735 				hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
736 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
737 					hammer2_thr_freeze_async(
738 						&pmp->xop_groups[j].thrs[i]);
739 				}
740 			}
741 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
742 				if (pmp->pfs_hmps[i] == NULL)
743 					continue;
744 				hammer2_thr_freeze(&pmp->sync_thrs[i]);
745 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
746 					hammer2_thr_freeze(
747 						&pmp->xop_groups[j].thrs[i]);
748 				}
749 			}
750 
751 			/*
752 			 * Lock the inode and clean out matching chains.
753 			 * Note that we cannot use hammer2_inode_lock_*()
754 			 * here because that would attempt to validate the
755 			 * cluster that we are in the middle of ripping
756 			 * apart.
757 			 *
758 			 * WARNING! We are working directly on the inodes
759 			 *	    embedded cluster.
760 			 */
761 			hammer2_mtx_ex(&iroot->lock);
762 
763 			/*
764 			 * Remove the chain from matching elements of the PFS.
765 			 */
766 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
767 				if (pmp->pfs_hmps[i] != hmp)
768 					continue;
769 				hammer2_thr_delete(&pmp->sync_thrs[i]);
770 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
771 					hammer2_thr_delete(
772 						&pmp->xop_groups[j].thrs[i]);
773 				}
774 				rchain = iroot->cluster.array[i].chain;
775 				iroot->cluster.array[i].chain = NULL;
776 				pmp->pfs_types[i] = 0;
777 				if (pmp->pfs_names[i]) {
778 					kfree(pmp->pfs_names[i], M_HAMMER2);
779 					pmp->pfs_names[i] = NULL;
780 				}
781 				if (rchain) {
782 					hammer2_chain_drop(rchain);
783 					/* focus hint */
784 					if (iroot->cluster.focus == rchain)
785 						iroot->cluster.focus = NULL;
786 				}
787 				pmp->pfs_hmps[i] = NULL;
788 			}
789 			hammer2_mtx_unlock(&iroot->lock);
790 			didfreeze = 1;	/* remaster, unfreeze down below */
791 		} else {
792 			didfreeze = 0;
793 		}
794 
795 		/*
796 		 * Cleanup trailing chains.  Gaps may remain.
797 		 */
798 		for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
799 			if (pmp->pfs_hmps[i])
800 				break;
801 		}
802 		iroot->cluster.nchains = i + 1;
803 
804 		/*
805 		 * If the PMP has no elements remaining we can destroy it.
806 		 * (this will transition management threads from frozen->exit).
807 		 */
808 		if (iroot->cluster.nchains == 0) {
809 			hammer2_pfsfree(pmp);
810 			goto again;
811 		}
812 
813 		/*
814 		 * If elements still remain we need to set the REMASTER
815 		 * flag and unfreeze it.
816 		 */
817 		if (didfreeze) {
818 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
819 				if (pmp->pfs_hmps[i] == NULL)
820 					continue;
821 				hammer2_thr_remaster(&pmp->sync_thrs[i]);
822 				hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
823 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
824 					hammer2_thr_remaster(
825 						&pmp->xop_groups[j].thrs[i]);
826 					hammer2_thr_unfreeze(
827 						&pmp->xop_groups[j].thrs[i]);
828 				}
829 			}
830 		}
831 	}
832 }
833 
834 /*
835  * Mount or remount HAMMER2 fileystem from physical media
836  *
837  *	mountroot
838  *		mp		mount point structure
839  *		path		NULL
840  *		data		<unused>
841  *		cred		<unused>
842  *
843  *	mount
844  *		mp		mount point structure
845  *		path		path to mount point
846  *		data		pointer to argument structure in user space
847  *			volume	volume path (device@LABEL form)
848  *			hflags	user mount flags
849  *		cred		user credentials
850  *
851  * RETURNS:	0	Success
852  *		!0	error number
853  */
854 static
855 int
856 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
857 		  struct ucred *cred)
858 {
859 	struct hammer2_mount_info info;
860 	hammer2_pfs_t *pmp;
861 	hammer2_pfs_t *spmp;
862 	hammer2_dev_t *hmp;
863 	hammer2_dev_t *force_local;
864 	hammer2_key_t key_next;
865 	hammer2_key_t key_dummy;
866 	hammer2_key_t lhc;
867 	struct vnode *devvp;
868 	struct nlookupdata nd;
869 	hammer2_chain_t *parent;
870 	hammer2_chain_t *chain;
871 	hammer2_cluster_t *cluster;
872 	const hammer2_inode_data_t *ripdata;
873 	hammer2_blockref_t bref;
874 	struct file *fp;
875 	char devstr[MNAMELEN];
876 	size_t size;
877 	size_t done;
878 	char *dev;
879 	char *label;
880 	int ronly = 1;
881 	int error;
882 	int i;
883 
884 	hmp = NULL;
885 	pmp = NULL;
886 	dev = NULL;
887 	label = NULL;
888 	devvp = NULL;
889 
890 	kprintf("hammer2_mount\n");
891 
892 	if (path == NULL) {
893 		/*
894 		 * Root mount
895 		 */
896 		bzero(&info, sizeof(info));
897 		info.cluster_fd = -1;
898 		ksnprintf(devstr, sizeof(devstr), "%s",
899 			  mp->mnt_stat.f_mntfromname);
900 		kprintf("hammer2_mount: root '%s'\n", devstr);
901 	} else {
902 		/*
903 		 * Non-root mount or updating a mount
904 		 */
905 		error = copyin(data, &info, sizeof(info));
906 		if (error)
907 			return (error);
908 
909 		error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
910 		if (error)
911 			return (error);
912 	}
913 
914 	/*
915 	 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
916 	 * if no label specified, based on the partition id.  Error out if no
917 	 * label or device (with partition id) is specified.  This is strictly
918 	 * a convenience to match the default label created by newfs_hammer2,
919 	 * our preference is that a label always be specified.
920 	 *
921 	 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
922 	 *	 that does not specify a device, as long as some H2 label
923 	 *	 has already been mounted from that device.  This makes
924 	 *	 mounting snapshots a lot easier.
925 	 */
926 	dev = devstr;
927 	label = strchr(devstr, '@');
928 	if (label && ((label + 1) - dev) > done)
929 		return (EINVAL);
930 	if (label == NULL || label[1] == 0) {
931 		char slice;
932 
933 		if (label == NULL)
934 			label = devstr + strlen(devstr);
935 		slice = label[-1];
936 		switch(slice) {
937 		case 'a':
938 			label = "BOOT";
939 			break;
940 		case 'd':
941 			label = "ROOT";
942 			break;
943 		default:
944 			label = "DATA";
945 			break;
946 		}
947 	} else {
948 		*label = '\0';
949 		label++;
950 	}
951 
952 	kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
953 		dev, label, (mp->mnt_flag & MNT_RDONLY));
954 
955 	if (mp->mnt_flag & MNT_UPDATE) {
956 		/*
957 		 * Update mount.  Note that pmp->iroot->cluster is
958 		 * an inode-embedded cluster and thus cannot be
959 		 * directly locked.
960 		 *
961 		 * XXX HAMMER2 needs to implement NFS export via
962 		 *     mountctl.
963 		 */
964 		pmp = MPTOPMP(mp);
965 		pmp->hflags = info.hflags;
966 		cluster = &pmp->iroot->cluster;
967 		for (i = 0; i < cluster->nchains; ++i) {
968 			if (cluster->array[i].chain == NULL)
969 				continue;
970 			hmp = cluster->array[i].chain->hmp;
971 			devvp = hmp->devvp;
972 			error = hammer2_remount(hmp, mp, path,
973 						devvp, cred);
974 			if (error)
975 				break;
976 		}
977 
978 		return error;
979 	}
980 
981 	/*
982 	 * HMP device mount
983 	 *
984 	 * If a path is specified and dev is not an empty string, lookup the
985 	 * name and verify that it referes to a block device.
986 	 *
987 	 * If a path is specified and dev is an empty string we fall through
988 	 * and locate the label in the hmp search.
989 	 */
990 	if (path && *dev != 0) {
991 		error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
992 		if (error == 0)
993 			error = nlookup(&nd);
994 		if (error == 0)
995 			error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
996 		nlookup_done(&nd);
997 	} else if (path == NULL) {
998 		/* root mount */
999 		cdev_t cdev = kgetdiskbyname(dev);
1000 		error = bdevvp(cdev, &devvp);
1001 		if (error)
1002 			kprintf("hammer2: cannot find '%s'\n", dev);
1003 	} else {
1004 		/*
1005 		 * We will locate the hmp using the label in the hmp loop.
1006 		 */
1007 		error = 0;
1008 	}
1009 
1010 	/*
1011 	 * Make sure its a block device.  Do not check to see if it is
1012 	 * already mounted until we determine that its a fresh H2 device.
1013 	 */
1014 	if (error == 0 && devvp) {
1015 		vn_isdisk(devvp, &error);
1016 	}
1017 
1018 	/*
1019 	 * Determine if the device has already been mounted.  After this
1020 	 * check hmp will be non-NULL if we are doing the second or more
1021 	 * hammer2 mounts from the same device.
1022 	 */
1023 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1024 	if (devvp) {
1025 		/*
1026 		 * Match the device.  Due to the way devfs works,
1027 		 * we may not be able to directly match the vnode pointer,
1028 		 * so also check to see if the underlying device matches.
1029 		 */
1030 		TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1031 			if (hmp->devvp == devvp)
1032 				break;
1033 			if (devvp->v_rdev &&
1034 			    hmp->devvp->v_rdev == devvp->v_rdev) {
1035 				break;
1036 			}
1037 		}
1038 
1039 		/*
1040 		 * If no match this may be a fresh H2 mount, make sure
1041 		 * the device is not mounted on anything else.
1042 		 */
1043 		if (hmp == NULL)
1044 			error = vfs_mountedon(devvp);
1045 	} else if (error == 0) {
1046 		/*
1047 		 * Match the label to a pmp already probed.
1048 		 */
1049 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1050 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1051 				if (pmp->pfs_names[i] &&
1052 				    strcmp(pmp->pfs_names[i], label) == 0) {
1053 					hmp = pmp->pfs_hmps[i];
1054 					break;
1055 				}
1056 			}
1057 			if (hmp)
1058 				break;
1059 		}
1060 		if (hmp == NULL)
1061 			error = ENOENT;
1062 	}
1063 
1064 	/*
1065 	 * Open the device if this isn't a secondary mount and construct
1066 	 * the H2 device mount (hmp).
1067 	 */
1068 	if (hmp == NULL) {
1069 		hammer2_chain_t *schain;
1070 		hammer2_xid_t xid;
1071 
1072 		if (error == 0 && vcount(devvp) > 0) {
1073 			kprintf("Primary device already has references\n");
1074 			error = EBUSY;
1075 		}
1076 
1077 		/*
1078 		 * Now open the device
1079 		 */
1080 		if (error == 0) {
1081 			ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1082 			vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1083 			error = vinvalbuf(devvp, V_SAVE, 0, 0);
1084 			if (error == 0) {
1085 				error = VOP_OPEN(devvp,
1086 					     (ronly ? FREAD : FREAD | FWRITE),
1087 					     FSCRED, NULL);
1088 			}
1089 			vn_unlock(devvp);
1090 		}
1091 		if (error && devvp) {
1092 			vrele(devvp);
1093 			devvp = NULL;
1094 		}
1095 		if (error) {
1096 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1097 			return error;
1098 		}
1099 		hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1100 		ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1101 		hmp->ronly = ronly;
1102 		hmp->devvp = devvp;
1103 		hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1104 		kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1105 		TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1106 		RB_INIT(&hmp->iotree);
1107 		spin_init(&hmp->io_spin, "hm2mount_io");
1108 		spin_init(&hmp->list_spin, "hm2mount_list");
1109 		TAILQ_INIT(&hmp->flushq);
1110 
1111 		lockinit(&hmp->vollk, "h2vol", 0, 0);
1112 		lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1113 		lockinit(&hmp->bflock, "h2bflk", 0, 0);
1114 
1115 		/*
1116 		 * vchain setup. vchain.data is embedded.
1117 		 * vchain.refs is initialized and will never drop to 0.
1118 		 *
1119 		 * NOTE! voldata is not yet loaded.
1120 		 */
1121 		hmp->vchain.hmp = hmp;
1122 		hmp->vchain.refs = 1;
1123 		hmp->vchain.data = (void *)&hmp->voldata;
1124 		hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1125 		hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1126 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1127 
1128 		hammer2_chain_core_init(&hmp->vchain);
1129 		/* hmp->vchain.u.xxx is left NULL */
1130 
1131 		/*
1132 		 * fchain setup.  fchain.data is embedded.
1133 		 * fchain.refs is initialized and will never drop to 0.
1134 		 *
1135 		 * The data is not used but needs to be initialized to
1136 		 * pass assertion muster.  We use this chain primarily
1137 		 * as a placeholder for the freemap's top-level RBTREE
1138 		 * so it does not interfere with the volume's topology
1139 		 * RBTREE.
1140 		 */
1141 		hmp->fchain.hmp = hmp;
1142 		hmp->fchain.refs = 1;
1143 		hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1144 		hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1145 		hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1146 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1147 		hmp->fchain.bref.methods =
1148 			HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1149 			HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1150 
1151 		hammer2_chain_core_init(&hmp->fchain);
1152 		/* hmp->fchain.u.xxx is left NULL */
1153 
1154 		/*
1155 		 * Install the volume header and initialize fields from
1156 		 * voldata.
1157 		 */
1158 		error = hammer2_install_volume_header(hmp);
1159 		if (error) {
1160 			hammer2_unmount_helper(mp, NULL, hmp);
1161 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1162 			hammer2_vfs_unmount(mp, MNT_FORCE);
1163 			return error;
1164 		}
1165 
1166 		/*
1167 		 * Really important to get these right or flush will get
1168 		 * confused.
1169 		 */
1170 		hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1171 		kprintf("alloc spmp %p tid %016jx\n",
1172 			hmp->spmp, hmp->voldata.mirror_tid);
1173 		spmp = hmp->spmp;
1174 
1175 		/*
1176 		 * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1177 		 * is inherited from the volume header.
1178 		 */
1179 		xid = 0;
1180 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1181 		hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1182 		hmp->vchain.pmp = spmp;
1183 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1184 		hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1185 		hmp->fchain.pmp = spmp;
1186 
1187 		/*
1188 		 * First locate the super-root inode, which is key 0
1189 		 * relative to the volume header's blockset.
1190 		 *
1191 		 * Then locate the root inode by scanning the directory keyspace
1192 		 * represented by the label.
1193 		 */
1194 		parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1195 		schain = hammer2_chain_lookup(&parent, &key_dummy,
1196 				      HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1197 				      &error, 0);
1198 		hammer2_chain_lookup_done(parent);
1199 		if (schain == NULL) {
1200 			kprintf("hammer2_mount: invalid super-root\n");
1201 			hammer2_unmount_helper(mp, NULL, hmp);
1202 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1203 			hammer2_vfs_unmount(mp, MNT_FORCE);
1204 			return EINVAL;
1205 		}
1206 		if (schain->error) {
1207 			kprintf("hammer2_mount: error %s reading super-root\n",
1208 				hammer2_error_str(schain->error));
1209 			hammer2_chain_unlock(schain);
1210 			hammer2_chain_drop(schain);
1211 			schain = NULL;
1212 			hammer2_unmount_helper(mp, NULL, hmp);
1213 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1214 			hammer2_vfs_unmount(mp, MNT_FORCE);
1215 			return EINVAL;
1216 		}
1217 
1218 		/*
1219 		 * The super-root always uses an inode_tid of 1 when
1220 		 * creating PFSs.
1221 		 */
1222 		spmp->inode_tid = 1;
1223 		spmp->modify_tid = schain->bref.modify_tid + 1;
1224 
1225 		/*
1226 		 * Sanity-check schain's pmp and finish initialization.
1227 		 * Any chain belonging to the super-root topology should
1228 		 * have a NULL pmp (not even set to spmp).
1229 		 */
1230 		ripdata = &hammer2_chain_rdata(schain)->ipdata;
1231 		KKASSERT(schain->pmp == NULL);
1232 		spmp->pfs_clid = ripdata->meta.pfs_clid;
1233 
1234 		/*
1235 		 * Replace the dummy spmp->iroot with a real one.  It's
1236 		 * easier to just do a wholesale replacement than to try
1237 		 * to update the chain and fixup the iroot fields.
1238 		 *
1239 		 * The returned inode is locked with the supplied cluster.
1240 		 */
1241 		cluster = hammer2_cluster_from_chain(schain);
1242 		hammer2_inode_drop(spmp->iroot);
1243 		spmp->iroot = NULL;
1244 		spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1245 		spmp->spmp_hmp = hmp;
1246 		spmp->pfs_types[0] = ripdata->meta.pfs_type;
1247 		spmp->pfs_hmps[0] = hmp;
1248 		hammer2_inode_ref(spmp->iroot);
1249 		hammer2_inode_unlock(spmp->iroot);
1250 		hammer2_cluster_unlock(cluster);
1251 		hammer2_cluster_drop(cluster);
1252 		schain = NULL;
1253 		/* leave spmp->iroot with one ref */
1254 
1255 		if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1256 			error = hammer2_recovery(hmp);
1257 			/* XXX do something with error */
1258 		}
1259 		hammer2_update_pmps(hmp);
1260 		hammer2_iocom_init(hmp);
1261 		hammer2_bulkfree_init(hmp);
1262 
1263 		/*
1264 		 * Ref the cluster management messaging descriptor.  The mount
1265 		 * program deals with the other end of the communications pipe.
1266 		 *
1267 		 * Root mounts typically do not supply one.
1268 		 */
1269 		if (info.cluster_fd >= 0) {
1270 			fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1271 			if (fp) {
1272 				hammer2_cluster_reconnect(hmp, fp);
1273 			} else {
1274 				kprintf("hammer2_mount: bad cluster_fd!\n");
1275 			}
1276 		}
1277 	} else {
1278 		spmp = hmp->spmp;
1279 		if (info.hflags & HMNT2_DEVFLAGS) {
1280 			kprintf("hammer2: Warning: mount flags pertaining "
1281 				"to the whole device may only be specified "
1282 				"on the first mount of the device: %08x\n",
1283 				info.hflags & HMNT2_DEVFLAGS);
1284 		}
1285 	}
1286 
1287 	/*
1288 	 * Force local mount (disassociate all PFSs from their clusters).
1289 	 * Used primarily for debugging.
1290 	 */
1291 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1292 
1293 	/*
1294 	 * Lookup the mount point under the media-localized super-root.
1295 	 * Scanning hammer2_pfslist doesn't help us because it represents
1296 	 * PFS cluster ids which can aggregate several named PFSs together.
1297 	 *
1298 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1299 	 * up later on.
1300 	 */
1301 	hammer2_inode_lock(spmp->iroot, 0);
1302 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1303 	lhc = hammer2_dirhash(label, strlen(label));
1304 	chain = hammer2_chain_lookup(&parent, &key_next,
1305 				     lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1306 				     &error, 0);
1307 	while (chain) {
1308 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1309 		    strcmp(label, chain->data->ipdata.filename) == 0) {
1310 			break;
1311 		}
1312 		chain = hammer2_chain_next(&parent, chain, &key_next,
1313 					    key_next,
1314 					    lhc + HAMMER2_DIRHASH_LOMASK,
1315 					    &error, 0);
1316 	}
1317 	if (parent) {
1318 		hammer2_chain_unlock(parent);
1319 		hammer2_chain_drop(parent);
1320 	}
1321 	hammer2_inode_unlock(spmp->iroot);
1322 
1323 	/*
1324 	 * PFS could not be found?
1325 	 */
1326 	if (chain == NULL) {
1327 		if (error)
1328 			kprintf("hammer2_mount: PFS label I/O error\n");
1329 		else
1330 			kprintf("hammer2_mount: PFS label not found\n");
1331 		hammer2_unmount_helper(mp, NULL, hmp);
1332 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1333 		hammer2_vfs_unmount(mp, MNT_FORCE);
1334 
1335 		return EINVAL;
1336 	}
1337 
1338 	/*
1339 	 * Acquire the pmp structure (it should have already been allocated
1340 	 * via hammer2_update_pmps() so do not pass cluster in to add to
1341 	 * available chains).
1342 	 *
1343 	 * Check if the cluster has already been mounted.  A cluster can
1344 	 * only be mounted once, use null mounts to mount additional copies.
1345 	 */
1346 	if (chain->error) {
1347 		kprintf("hammer2_mount: PFS label I/O error\n");
1348 	} else {
1349 		ripdata = &chain->data->ipdata;
1350 		bref = chain->bref;
1351 		pmp = hammer2_pfsalloc(NULL, ripdata,
1352 				       bref.modify_tid, force_local);
1353 	}
1354 	hammer2_chain_unlock(chain);
1355 	hammer2_chain_drop(chain);
1356 
1357 	/*
1358 	 * Finish the mount
1359 	 */
1360         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1361 
1362 	if (pmp->mp) {
1363 		kprintf("hammer2_mount: PFS already mounted!\n");
1364 		hammer2_unmount_helper(mp, NULL, hmp);
1365 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1366 		hammer2_vfs_unmount(mp, MNT_FORCE);
1367 
1368 		return EBUSY;
1369 	}
1370 
1371 	pmp->hflags = info.hflags;
1372         mp->mnt_flag |= MNT_LOCAL;
1373         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1374         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1375 
1376         /*
1377          * required mount structure initializations
1378          */
1379         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1380         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1381 
1382         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1383         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1384 
1385         /*
1386          * Optional fields
1387          */
1388         mp->mnt_iosize_max = MAXPHYS;
1389 
1390 	/*
1391 	 * Connect up mount pointers.
1392 	 */
1393 	hammer2_mount_helper(mp, pmp);
1394 
1395         lockmgr(&hammer2_mntlk, LK_RELEASE);
1396 
1397 	/*
1398 	 * Finish setup
1399 	 */
1400 	vfs_getnewfsid(mp);
1401 	vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1402 	vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1403 	vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1404 
1405 	if (path) {
1406 		copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1407 			  MNAMELEN - 1, &size);
1408 		bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1409 	} /* else root mount, already in there */
1410 
1411 	bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1412 	if (path) {
1413 		copyinstr(path, mp->mnt_stat.f_mntonname,
1414 			  sizeof(mp->mnt_stat.f_mntonname) - 1,
1415 			  &size);
1416 	} else {
1417 		/* root mount */
1418 		mp->mnt_stat.f_mntonname[0] = '/';
1419 	}
1420 
1421 	/*
1422 	 * Initial statfs to prime mnt_stat.
1423 	 */
1424 	hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1425 
1426 	return 0;
1427 }
1428 
1429 /*
1430  * Scan PFSs under the super-root and create hammer2_pfs structures.
1431  */
1432 static
1433 void
1434 hammer2_update_pmps(hammer2_dev_t *hmp)
1435 {
1436 	const hammer2_inode_data_t *ripdata;
1437 	hammer2_chain_t *parent;
1438 	hammer2_chain_t *chain;
1439 	hammer2_blockref_t bref;
1440 	hammer2_dev_t *force_local;
1441 	hammer2_pfs_t *spmp;
1442 	hammer2_pfs_t *pmp;
1443 	hammer2_key_t key_next;
1444 	int error;
1445 
1446 	/*
1447 	 * Force local mount (disassociate all PFSs from their clusters).
1448 	 * Used primarily for debugging.
1449 	 */
1450 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1451 
1452 	/*
1453 	 * Lookup mount point under the media-localized super-root.
1454 	 *
1455 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1456 	 * up later on.
1457 	 */
1458 	spmp = hmp->spmp;
1459 	hammer2_inode_lock(spmp->iroot, 0);
1460 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1461 	chain = hammer2_chain_lookup(&parent, &key_next,
1462 					 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1463 					 &error, 0);
1464 	while (chain) {
1465 		if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1466 			continue;
1467 		if (chain->error) {
1468 			kprintf("I/O error scanning PFS labels\n");
1469 		} else {
1470 			ripdata = &chain->data->ipdata;
1471 			bref = chain->bref;
1472 
1473 			pmp = hammer2_pfsalloc(chain, ripdata,
1474 					       bref.modify_tid, force_local);
1475 		}
1476 		chain = hammer2_chain_next(&parent, chain, &key_next,
1477 					   key_next, HAMMER2_KEY_MAX,
1478 					   &error, 0);
1479 	}
1480 	if (parent) {
1481 		hammer2_chain_unlock(parent);
1482 		hammer2_chain_drop(parent);
1483 	}
1484 	hammer2_inode_unlock(spmp->iroot);
1485 }
1486 
1487 static
1488 int
1489 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1490 		struct vnode *devvp, struct ucred *cred)
1491 {
1492 	int error;
1493 
1494 	if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1495 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1496 		VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1497 		vn_unlock(devvp);
1498 		error = hammer2_recovery(hmp);
1499 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1500 		if (error == 0) {
1501 			VOP_CLOSE(devvp, FREAD, NULL);
1502 			hmp->ronly = 0;
1503 		} else {
1504 			VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1505 		}
1506 		vn_unlock(devvp);
1507 	} else {
1508 		error = 0;
1509 	}
1510 	return error;
1511 }
1512 
1513 static
1514 int
1515 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1516 {
1517 	hammer2_pfs_t *pmp;
1518 	int flags;
1519 	int error = 0;
1520 
1521 	pmp = MPTOPMP(mp);
1522 
1523 	if (pmp == NULL)
1524 		return(0);
1525 
1526 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1527 
1528 	/*
1529 	 * If mount initialization proceeded far enough we must flush
1530 	 * its vnodes and sync the underlying mount points.  Three syncs
1531 	 * are required to fully flush the filesystem (freemap updates lag
1532 	 * by one flush, and one extra for safety).
1533 	 */
1534 	if (mntflags & MNT_FORCE)
1535 		flags = FORCECLOSE;
1536 	else
1537 		flags = 0;
1538 	if (pmp->iroot) {
1539 		error = vflush(mp, 0, flags);
1540 		if (error)
1541 			goto failed;
1542 		hammer2_vfs_sync(mp, MNT_WAIT);
1543 		hammer2_vfs_sync(mp, MNT_WAIT);
1544 		hammer2_vfs_sync(mp, MNT_WAIT);
1545 	}
1546 
1547 	/*
1548 	 * Cleanup the frontend support XOPS threads
1549 	 */
1550 	hammer2_xop_helper_cleanup(pmp);
1551 
1552 	if (pmp->mp)
1553 		hammer2_unmount_helper(mp, pmp, NULL);
1554 
1555 	error = 0;
1556 failed:
1557 	lockmgr(&hammer2_mntlk, LK_RELEASE);
1558 
1559 	return (error);
1560 }
1561 
1562 /*
1563  * Mount helper, hook the system mount into our PFS.
1564  * The mount lock is held.
1565  *
1566  * We must bump the mount_count on related devices for any
1567  * mounted PFSs.
1568  */
1569 static
1570 void
1571 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1572 {
1573 	hammer2_cluster_t *cluster;
1574 	hammer2_chain_t *rchain;
1575 	int i;
1576 
1577         mp->mnt_data = (qaddr_t)pmp;
1578 	pmp->mp = mp;
1579 
1580 	/*
1581 	 * After pmp->mp is set we have to adjust hmp->mount_count.
1582 	 */
1583 	cluster = &pmp->iroot->cluster;
1584 	for (i = 0; i < cluster->nchains; ++i) {
1585 		rchain = cluster->array[i].chain;
1586 		if (rchain == NULL)
1587 			continue;
1588 		++rchain->hmp->mount_count;
1589 		kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1590 			rchain->hmp, rchain->hmp->mount_count);
1591 	}
1592 
1593 	/*
1594 	 * Create missing Xop threads
1595 	 */
1596 	hammer2_xop_helper_create(pmp);
1597 }
1598 
1599 /*
1600  * Mount helper, unhook the system mount from our PFS.
1601  * The mount lock is held.
1602  *
1603  * If hmp is supplied a mount responsible for being the first to open
1604  * the block device failed and the block device and all PFSs using the
1605  * block device must be cleaned up.
1606  *
1607  * If pmp is supplied multiple devices might be backing the PFS and each
1608  * must be disconnected.  This might not be the last PFS using some of the
1609  * underlying devices.  Also, we have to adjust our hmp->mount_count
1610  * accounting for the devices backing the pmp which is now undergoing an
1611  * unmount.
1612  */
1613 static
1614 void
1615 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1616 {
1617 	hammer2_cluster_t *cluster;
1618 	hammer2_chain_t *rchain;
1619 	struct vnode *devvp;
1620 	int dumpcnt;
1621 	int ronly;
1622 	int i;
1623 
1624 	/*
1625 	 * If no device supplied this is a high-level unmount and we have to
1626 	 * to disconnect the mount, adjust mount_count, and locate devices
1627 	 * that might now have no mounts.
1628 	 */
1629 	if (pmp) {
1630 		KKASSERT(hmp == NULL);
1631 		KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1632 		pmp->mp = NULL;
1633 		mp->mnt_data = NULL;
1634 
1635 		/*
1636 		 * After pmp->mp is cleared we have to account for
1637 		 * mount_count.
1638 		 */
1639 		cluster = &pmp->iroot->cluster;
1640 		for (i = 0; i < cluster->nchains; ++i) {
1641 			rchain = cluster->array[i].chain;
1642 			if (rchain == NULL)
1643 				continue;
1644 			--rchain->hmp->mount_count;
1645 			/* scrapping hmp now may invalidate the pmp */
1646 		}
1647 again:
1648 		TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1649 			if (hmp->mount_count == 0) {
1650 				hammer2_unmount_helper(NULL, NULL, hmp);
1651 				goto again;
1652 			}
1653 		}
1654 		return;
1655 	}
1656 
1657 	/*
1658 	 * Try to terminate the block device.  We can't terminate it if
1659 	 * there are still PFSs referencing it.
1660 	 */
1661 	if (hmp->mount_count)
1662 		return;
1663 
1664 	/*
1665 	 * Decomission the network before we start messing with the
1666 	 * device and PFS.
1667 	 */
1668 	hammer2_iocom_uninit(hmp);
1669 
1670 	hammer2_bulkfree_uninit(hmp);
1671 	hammer2_pfsfree_scan(hmp);
1672 	hammer2_dev_exlock(hmp);	/* XXX order */
1673 
1674 	/*
1675 	 * Cycle the volume data lock as a safety (probably not needed any
1676 	 * more).  To ensure everything is out we need to flush at least
1677 	 * three times.  (1) The running of the sideq can dirty the
1678 	 * filesystem, (2) A normal flush can dirty the freemap, and
1679 	 * (3) ensure that the freemap is fully synchronized.
1680 	 *
1681 	 * The next mount's recovery scan can clean everything up but we want
1682 	 * to leave the filesystem in a 100% clean state on a normal unmount.
1683 	 */
1684 #if 0
1685 	hammer2_voldata_lock(hmp);
1686 	hammer2_voldata_unlock(hmp);
1687 #endif
1688 
1689 	/*
1690 	 * Flush whatever is left.  Unmounted but modified PFS's might still
1691 	 * have some dirty chains on them.
1692 	 */
1693 	hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1694 	hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1695 	hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP | HAMMER2_FLUSH_ALL);
1696 	hammer2_chain_unlock(&hmp->fchain);
1697 	hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP | HAMMER2_FLUSH_ALL);
1698 	hammer2_chain_unlock(&hmp->vchain);
1699 
1700 	if ((hmp->vchain.flags | hmp->fchain.flags) &
1701 	    HAMMER2_CHAIN_FLUSH_MASK) {
1702 		kprintf("hammer2_unmount: chains left over "
1703 			"after final sync\n");
1704 		kprintf("    vchain %08x\n", hmp->vchain.flags);
1705 		kprintf("    fchain %08x\n", hmp->fchain.flags);
1706 
1707 		if (hammer2_debug & 0x0010)
1708 			Debugger("entered debugger");
1709 	}
1710 
1711 	KKASSERT(hmp->spmp == NULL);
1712 
1713 	/*
1714 	 * Finish up with the device vnode
1715 	 */
1716 	if ((devvp = hmp->devvp) != NULL) {
1717 		ronly = hmp->ronly;
1718 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1719 		kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1720 			hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1721 			ronly);
1722 		vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1723 		kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1724 			hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1725 		hmp->devvp = NULL;
1726 		VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1727 		vn_unlock(devvp);
1728 		vrele(devvp);
1729 		devvp = NULL;
1730 	}
1731 
1732 	/*
1733 	 * Clear vchain/fchain flags that might prevent final cleanup
1734 	 * of these chains.
1735 	 */
1736 	if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1737 		atomic_add_long(&hammer2_count_modified_chains, -1);
1738 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1739 		hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1740 	}
1741 	if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1742 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1743 	}
1744 
1745 	if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1746 		atomic_add_long(&hammer2_count_modified_chains, -1);
1747 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1748 		hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1749 	}
1750 	if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1751 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1752 	}
1753 
1754 	/*
1755 	 * Final drop of embedded freemap root chain to
1756 	 * clean up fchain.core (fchain structure is not
1757 	 * flagged ALLOCATED so it is cleaned out and then
1758 	 * left to rot).
1759 	 */
1760 	hammer2_chain_drop(&hmp->fchain);
1761 
1762 	/*
1763 	 * Final drop of embedded volume root chain to clean
1764 	 * up vchain.core (vchain structure is not flagged
1765 	 * ALLOCATED so it is cleaned out and then left to
1766 	 * rot).
1767 	 */
1768 	dumpcnt = 50;
1769 	hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1770 	dumpcnt = 50;
1771 	hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1772 	hammer2_dev_unlock(hmp);
1773 	hammer2_chain_drop(&hmp->vchain);
1774 
1775 	hammer2_io_cleanup(hmp, &hmp->iotree);
1776 	if (hmp->iofree_count) {
1777 		kprintf("io_cleanup: %d I/O's left hanging\n",
1778 			hmp->iofree_count);
1779 	}
1780 
1781 	TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1782 	kmalloc_destroy(&hmp->mchain);
1783 	kfree(hmp, M_HAMMER2);
1784 }
1785 
1786 int
1787 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1788 		 ino_t ino, struct vnode **vpp)
1789 {
1790 	hammer2_xop_lookup_t *xop;
1791 	hammer2_pfs_t *pmp;
1792 	hammer2_inode_t *ip;
1793 	hammer2_tid_t inum;
1794 	int error;
1795 
1796 	inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1797 
1798 	error = 0;
1799 	pmp = MPTOPMP(mp);
1800 
1801 	/*
1802 	 * Easy if we already have it cached
1803 	 */
1804 	ip = hammer2_inode_lookup(pmp, inum);
1805 	if (ip) {
1806 		hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1807 		*vpp = hammer2_igetv(ip, &error);
1808 		hammer2_inode_unlock(ip);
1809 		hammer2_inode_drop(ip);		/* from lookup */
1810 
1811 		return error;
1812 	}
1813 
1814 	/*
1815 	 * Otherwise we have to find the inode
1816 	 */
1817 	xop = hammer2_xop_alloc(pmp->iroot, 0);
1818 	xop->lhc = inum;
1819 	hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1820 	error = hammer2_xop_collect(&xop->head, 0);
1821 
1822 	if (error == 0) {
1823 		if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1824 			kprintf("vget: no collect error but also no rdata\n");
1825 			kprintf("xop %p\n", xop);
1826 			while ((hammer2_debug & 0x80000) == 0) {
1827 				tsleep(xop, PCATCH, "wait", hz * 10);
1828 			}
1829 			ip = NULL;
1830 		} else {
1831 			ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1832 		}
1833 	}
1834 	hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1835 
1836 	if (ip) {
1837 		*vpp = hammer2_igetv(ip, &error);
1838 		hammer2_inode_unlock(ip);
1839 	} else {
1840 		*vpp = NULL;
1841 		error = ENOENT;
1842 	}
1843 	return (error);
1844 }
1845 
1846 static
1847 int
1848 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1849 {
1850 	hammer2_pfs_t *pmp;
1851 	struct vnode *vp;
1852 	int error;
1853 
1854 	pmp = MPTOPMP(mp);
1855 	if (pmp->iroot == NULL) {
1856 		*vpp = NULL;
1857 		return EINVAL;
1858 	}
1859 
1860 	error = 0;
1861 	hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1862 
1863 	while (pmp->inode_tid == 0) {
1864 		hammer2_xop_ipcluster_t *xop;
1865 		hammer2_inode_meta_t *meta;
1866 
1867 		xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1868 		hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1869 		error = hammer2_xop_collect(&xop->head, 0);
1870 
1871 		if (error == 0) {
1872 			meta = &xop->head.cluster.focus->data->ipdata.meta;
1873 			pmp->iroot->meta = *meta;
1874 			pmp->inode_tid = meta->pfs_inum + 1;
1875 			if (pmp->inode_tid < HAMMER2_INODE_START)
1876 				pmp->inode_tid = HAMMER2_INODE_START;
1877 			pmp->modify_tid =
1878 				xop->head.cluster.focus->bref.modify_tid + 1;
1879 			kprintf("PFS: Starting inode %jd\n",
1880 				(intmax_t)pmp->inode_tid);
1881 			kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1882 				pmp->inode_tid, pmp->modify_tid);
1883 			wakeup(&pmp->iroot);
1884 
1885 			hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1886 
1887 			/*
1888 			 * Prime the mount info.
1889 			 */
1890 			hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1891 			break;
1892 		}
1893 
1894 		/*
1895 		 * Loop, try again
1896 		 */
1897 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1898 		hammer2_inode_unlock(pmp->iroot);
1899 		error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1900 		hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1901 		if (error == EINTR)
1902 			break;
1903 	}
1904 
1905 	if (error) {
1906 		hammer2_inode_unlock(pmp->iroot);
1907 		*vpp = NULL;
1908 	} else {
1909 		vp = hammer2_igetv(pmp->iroot, &error);
1910 		hammer2_inode_unlock(pmp->iroot);
1911 		*vpp = vp;
1912 	}
1913 
1914 	return (error);
1915 }
1916 
1917 /*
1918  * Filesystem status
1919  *
1920  * XXX incorporate ipdata->meta.inode_quota and data_quota
1921  */
1922 static
1923 int
1924 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1925 {
1926 	hammer2_pfs_t *pmp;
1927 	hammer2_dev_t *hmp;
1928 	hammer2_blockref_t bref;
1929 	int i;
1930 
1931 	/*
1932 	 * NOTE: iroot might not have validated the cluster yet.
1933 	 */
1934 	pmp = MPTOPMP(mp);
1935 
1936 	mp->mnt_stat.f_files = 0;
1937 	mp->mnt_stat.f_ffree = 0;
1938 	mp->mnt_stat.f_blocks = 0;
1939 	mp->mnt_stat.f_bfree = 0;
1940 	mp->mnt_stat.f_bavail = 0;
1941 
1942 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1943 		hmp = pmp->pfs_hmps[i];
1944 		if (hmp == NULL)
1945 			continue;
1946 		if (pmp->iroot->cluster.array[i].chain)
1947 			bref = pmp->iroot->cluster.array[i].chain->bref;
1948 		else
1949 			bzero(&bref, sizeof(bref));
1950 
1951 		mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1952 		mp->mnt_stat.f_ffree = 0;
1953 		mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1954 					mp->mnt_vstat.f_bsize;
1955 		mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1956 					mp->mnt_vstat.f_bsize;
1957 		mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1958 
1959 		if (cred && cred->cr_uid != 0) {
1960 			uint64_t adj;
1961 
1962 			/* 5% */
1963 			adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
1964 			mp->mnt_stat.f_blocks -= adj;
1965 			mp->mnt_stat.f_bfree -= adj;
1966 			mp->mnt_stat.f_bavail -= adj;
1967 		}
1968 
1969 		*sbp = mp->mnt_stat;
1970 	}
1971 	return (0);
1972 }
1973 
1974 static
1975 int
1976 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1977 {
1978 	hammer2_pfs_t *pmp;
1979 	hammer2_dev_t *hmp;
1980 	hammer2_blockref_t bref;
1981 	int i;
1982 
1983 	/*
1984 	 * NOTE: iroot might not have validated the cluster yet.
1985 	 */
1986 	pmp = MPTOPMP(mp);
1987 
1988 	mp->mnt_vstat.f_bsize = 0;
1989 	mp->mnt_vstat.f_files = 0;
1990 	mp->mnt_vstat.f_ffree = 0;
1991 	mp->mnt_vstat.f_blocks = 0;
1992 	mp->mnt_vstat.f_bfree = 0;
1993 	mp->mnt_vstat.f_bavail = 0;
1994 
1995 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1996 		hmp = pmp->pfs_hmps[i];
1997 		if (hmp == NULL)
1998 			continue;
1999 		if (pmp->iroot->cluster.array[i].chain)
2000 			bref = pmp->iroot->cluster.array[i].chain->bref;
2001 		else
2002 			bzero(&bref, sizeof(bref));
2003 
2004 		mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
2005 		mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
2006 		mp->mnt_vstat.f_ffree = 0;
2007 		mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
2008 					mp->mnt_vstat.f_bsize;
2009 		mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
2010 					mp->mnt_vstat.f_bsize;
2011 		mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
2012 
2013 		if (cred && cred->cr_uid != 0) {
2014 			uint64_t adj;
2015 
2016 			/* 5% */
2017 			adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2018 			mp->mnt_vstat.f_blocks -= adj;
2019 			mp->mnt_vstat.f_bfree -= adj;
2020 			mp->mnt_vstat.f_bavail -= adj;
2021 		}
2022 
2023 		*sbp = mp->mnt_vstat;
2024 	}
2025 	return (0);
2026 }
2027 
2028 /*
2029  * Mount-time recovery (RW mounts)
2030  *
2031  * Updates to the free block table are allowed to lag flushes by one
2032  * transaction.  In case of a crash, then on a fresh mount we must do an
2033  * incremental scan of the last committed transaction id and make sure that
2034  * all related blocks have been marked allocated.
2035  *
2036  * The super-root topology and each PFS has its own transaction id domain,
2037  * so we must track PFS boundary transitions.
2038  */
2039 struct hammer2_recovery_elm {
2040 	TAILQ_ENTRY(hammer2_recovery_elm) entry;
2041 	hammer2_chain_t *chain;
2042 	hammer2_tid_t sync_tid;
2043 };
2044 
2045 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2046 
2047 struct hammer2_recovery_info {
2048 	struct hammer2_recovery_list list;
2049 	hammer2_tid_t	mtid;
2050 	int	depth;
2051 };
2052 
2053 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2054 			hammer2_chain_t *parent,
2055 			struct hammer2_recovery_info *info,
2056 			hammer2_tid_t sync_tid);
2057 
2058 #define HAMMER2_RECOVERY_MAXDEPTH	10
2059 
2060 static
2061 int
2062 hammer2_recovery(hammer2_dev_t *hmp)
2063 {
2064 	struct hammer2_recovery_info info;
2065 	struct hammer2_recovery_elm *elm;
2066 	hammer2_chain_t *parent;
2067 	hammer2_tid_t sync_tid;
2068 	hammer2_tid_t mirror_tid;
2069 	int error;
2070 
2071 	hammer2_trans_init(hmp->spmp, 0);
2072 
2073 	sync_tid = hmp->voldata.freemap_tid;
2074 	mirror_tid = hmp->voldata.mirror_tid;
2075 
2076 	kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2077 	if (sync_tid >= mirror_tid) {
2078 		kprintf(" no recovery needed\n");
2079 	} else {
2080 		kprintf(" freemap recovery %016jx-%016jx\n",
2081 			sync_tid + 1, mirror_tid);
2082 	}
2083 
2084 	TAILQ_INIT(&info.list);
2085 	info.depth = 0;
2086 	parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2087 	error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2088 	hammer2_chain_lookup_done(parent);
2089 
2090 	while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2091 		TAILQ_REMOVE(&info.list, elm, entry);
2092 		parent = elm->chain;
2093 		sync_tid = elm->sync_tid;
2094 		kfree(elm, M_HAMMER2);
2095 
2096 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2097 		error |= hammer2_recovery_scan(hmp, parent, &info,
2098 					      hmp->voldata.freemap_tid);
2099 		hammer2_chain_unlock(parent);
2100 		hammer2_chain_drop(parent);	/* drop elm->chain ref */
2101 	}
2102 	hammer2_trans_done(hmp->spmp);
2103 
2104 	return error;
2105 }
2106 
2107 static
2108 int
2109 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2110 		      struct hammer2_recovery_info *info,
2111 		      hammer2_tid_t sync_tid)
2112 {
2113 	const hammer2_inode_data_t *ripdata;
2114 	hammer2_chain_t *chain;
2115 	hammer2_blockref_t bref;
2116 	int tmp_error;
2117 	int rup_error;
2118 	int error;
2119 	int first;
2120 
2121 	/*
2122 	 * Adjust freemap to ensure that the block(s) are marked allocated.
2123 	 */
2124 	if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2125 		hammer2_freemap_adjust(hmp, &parent->bref,
2126 				       HAMMER2_FREEMAP_DORECOVER);
2127 	}
2128 
2129 	/*
2130 	 * Check type for recursive scan
2131 	 */
2132 	switch(parent->bref.type) {
2133 	case HAMMER2_BREF_TYPE_VOLUME:
2134 		/* data already instantiated */
2135 		break;
2136 	case HAMMER2_BREF_TYPE_INODE:
2137 		/*
2138 		 * Must instantiate data for DIRECTDATA test and also
2139 		 * for recursion.
2140 		 */
2141 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2142 		ripdata = &hammer2_chain_rdata(parent)->ipdata;
2143 		if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2144 			/* not applicable to recovery scan */
2145 			hammer2_chain_unlock(parent);
2146 			return 0;
2147 		}
2148 		hammer2_chain_unlock(parent);
2149 		break;
2150 	case HAMMER2_BREF_TYPE_INDIRECT:
2151 		/*
2152 		 * Must instantiate data for recursion
2153 		 */
2154 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2155 		hammer2_chain_unlock(parent);
2156 		break;
2157 	case HAMMER2_BREF_TYPE_DIRENT:
2158 	case HAMMER2_BREF_TYPE_DATA:
2159 	case HAMMER2_BREF_TYPE_FREEMAP:
2160 	case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2161 	case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2162 		/* not applicable to recovery scan */
2163 		return 0;
2164 		break;
2165 	default:
2166 		return HAMMER2_ERROR_BADBREF;
2167 	}
2168 
2169 	/*
2170 	 * Defer operation if depth limit reached or if we are crossing a
2171 	 * PFS boundary.
2172 	 */
2173 	if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2174 		struct hammer2_recovery_elm *elm;
2175 
2176 		elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2177 		elm->chain = parent;
2178 		elm->sync_tid = sync_tid;
2179 		hammer2_chain_ref(parent);
2180 		TAILQ_INSERT_TAIL(&info->list, elm, entry);
2181 		/* unlocked by caller */
2182 
2183 		return(0);
2184 	}
2185 
2186 
2187 	/*
2188 	 * Recursive scan of the last flushed transaction only.  We are
2189 	 * doing this without pmp assignments so don't leave the chains
2190 	 * hanging around after we are done with them.
2191 	 *
2192 	 * error	Cumulative error this level only
2193 	 * rup_error	Cumulative error for recursion
2194 	 * tmp_error	Specific non-cumulative recursion error
2195 	 */
2196 	chain = NULL;
2197 	first = 1;
2198 	rup_error = 0;
2199 	error = 0;
2200 
2201 	for (;;) {
2202 		error |= hammer2_chain_scan(parent, &chain, &bref,
2203 					    &first,
2204 					    HAMMER2_LOOKUP_NODATA);
2205 
2206 		/*
2207 		 * Problem during scan or EOF
2208 		 */
2209 		if (error)
2210 			break;
2211 
2212 		/*
2213 		 * If this is a leaf
2214 		 */
2215 		if (chain == NULL) {
2216 			if (bref.mirror_tid > sync_tid) {
2217 				hammer2_freemap_adjust(hmp, &bref,
2218 						     HAMMER2_FREEMAP_DORECOVER);
2219 			}
2220 			continue;
2221 		}
2222 
2223 		/*
2224 		 * This may or may not be a recursive node.
2225 		 */
2226 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2227 		if (bref.mirror_tid > sync_tid) {
2228 			++info->depth;
2229 			tmp_error = hammer2_recovery_scan(hmp, chain,
2230 							   info, sync_tid);
2231 			--info->depth;
2232 		} else {
2233 			tmp_error = 0;
2234 		}
2235 
2236 		/*
2237 		 * Flush the recovery at the PFS boundary to stage it for
2238 		 * the final flush of the super-root topology.
2239 		 */
2240 		if (tmp_error == 0 &&
2241 		    (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2242 		    (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2243 			hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2244 		}
2245 		rup_error |= tmp_error;
2246 	}
2247 	return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2248 }
2249 
2250 /*
2251  * Sync a mount point; this is called on a per-mount basis from the
2252  * filesystem syncer process periodically and whenever a user issues
2253  * a sync.
2254  */
2255 int
2256 hammer2_vfs_sync(struct mount *mp, int waitfor)
2257 {
2258 	hammer2_xop_flush_t *xop;
2259 	struct hammer2_sync_info info;
2260 	hammer2_inode_t *iroot;
2261 	hammer2_pfs_t *pmp;
2262 	int flags;
2263 	int error;
2264 
2265 	pmp = MPTOPMP(mp);
2266 	iroot = pmp->iroot;
2267 	KKASSERT(iroot);
2268 	KKASSERT(iroot->pmp == pmp);
2269 
2270 	/*
2271 	 * We can't acquire locks on existing vnodes while in a transaction
2272 	 * without risking a deadlock.  This assumes that vfsync() can be
2273 	 * called without the vnode locked (which it can in DragonFly).
2274 	 * Otherwise we'd have to implement a multi-pass or flag the lock
2275 	 * failures and retry.
2276 	 *
2277 	 * The reclamation code interlocks with the sync list's token
2278 	 * (by removing the vnode from the scan list) before unlocking
2279 	 * the inode, giving us time to ref the inode.
2280 	 */
2281 	/*flags = VMSC_GETVP;*/
2282 	flags = 0;
2283 	if (waitfor & MNT_LAZY)
2284 		flags |= VMSC_ONEPASS;
2285 
2286 	/*
2287 	 * Preflush the vnodes using a normal transaction before interlocking
2288 	 * with a flush transaction.  We do this to try to run as much of
2289 	 * the compression as possible outside the flush transaction.
2290 	 *
2291 	 * For efficiency do an async pass before making sure with a
2292 	 * synchronous pass on all related buffer cache buffers.
2293 	 */
2294 	hammer2_trans_init(pmp, 0);
2295 	info.error = 0;
2296 	info.waitfor = MNT_NOWAIT;
2297 	vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2298 	info.waitfor = MNT_WAIT;
2299 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2300 	hammer2_trans_done(pmp);
2301 
2302 	/*
2303 	 * Start our flush transaction.  This does not return until all
2304 	 * concurrent transactions have completed and will prevent any
2305 	 * new transactions from running concurrently, except for the
2306 	 * buffer cache transactions.
2307 	 *
2308 	 * (1) vfsync() all dirty vnodes via vfsyncscan().
2309 	 *
2310 	 * (2) Flush any remaining dirty inodes (the sideq), including any
2311 	 *     which may have been created during or raced against the
2312 	 *     vfsync().  To catch all cases this must be done after the
2313 	 *     vfsync().
2314 	 *
2315 	 * (3) Wait for any pending BIO I/O to complete (hammer2_bioq_sync()).
2316 	 *
2317 	 * NOTE! It is still possible for the paging code to push pages
2318 	 *	 out via a UIO_NOCOPY hammer2_vop_write() during the main
2319 	 *	 flush.
2320 	 */
2321 	hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2322 
2323 	info.error = 0;
2324 	info.waitfor = MNT_NOWAIT;
2325 	vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2326 	info.waitfor = MNT_WAIT;
2327 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2328 	hammer2_inode_run_sideq(pmp, 1);
2329 	hammer2_bioq_sync(pmp);
2330 
2331 	/*
2332 	 * Use the XOP interface to concurrently flush all nodes to
2333 	 * synchronize the PFSROOT subtopology to the media.  A standard
2334 	 * end-of-scan ENOENT error indicates cluster sufficiency.
2335 	 *
2336 	 * Note that this flush will not be visible on crash recovery until
2337 	 * we flush the super-root topology in the next loop.
2338 	 *
2339 	 * XXX For now wait for all flushes to complete.
2340 	 */
2341 	if (iroot) {
2342 		xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2343 		hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2344 		error = hammer2_xop_collect(&xop->head,
2345 					    HAMMER2_XOP_COLLECT_WAITALL);
2346 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2347 		if (error == HAMMER2_ERROR_ENOENT)
2348 			error = 0;
2349 		else
2350 			error = hammer2_error_to_errno(error);
2351 	} else {
2352 		error = 0;
2353 	}
2354 	hammer2_trans_done(pmp);
2355 
2356 	return (error);
2357 }
2358 
2359 /*
2360  * Sync passes.
2361  *
2362  * Note that we ignore the tranasction mtid we got above.  Instead,
2363  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2364  * transactions.
2365  */
2366 static int
2367 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2368 {
2369 	struct hammer2_sync_info *info = data;
2370 	hammer2_inode_t *ip;
2371 	int error;
2372 
2373 	/*
2374 	 * Degenerate cases.  Note that ip == NULL typically means the
2375 	 * syncer vnode itself and we don't want to vclrisdirty() in that
2376 	 * situation.
2377 	 */
2378 	ip = VTOI(vp);
2379 	if (ip == NULL) {
2380 		return(0);
2381 	}
2382 	if (vp->v_type == VNON || vp->v_type == VBAD) {
2383 		vclrisdirty(vp);
2384 		return(0);
2385 	}
2386 
2387 	/*
2388 	 * VOP_FSYNC will start a new transaction so replicate some code
2389 	 * here to do it inline (see hammer2_vop_fsync()).
2390 	 *
2391 	 * WARNING: The vfsync interacts with the buffer cache and might
2392 	 *          block, we can't hold the inode lock at that time.
2393 	 *	    However, we MUST ref ip before blocking to ensure that
2394 	 *	    it isn't ripped out from under us (since we do not
2395 	 *	    hold a lock on the vnode).
2396 	 */
2397 	hammer2_inode_ref(ip);
2398 	if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2399 	    !RB_EMPTY(&vp->v_rbdirty_tree)) {
2400 		vfsync(vp, info->waitfor, 1, NULL, NULL);
2401 		if (ip->flags & (HAMMER2_INODE_RESIZED |
2402 				 HAMMER2_INODE_MODIFIED)) {
2403 			hammer2_inode_lock(ip, 0);
2404 			if (ip->flags & (HAMMER2_INODE_RESIZED |
2405 					 HAMMER2_INODE_MODIFIED)) {
2406 				hammer2_inode_chain_sync(ip);
2407 			}
2408 			hammer2_inode_unlock(ip);
2409 		}
2410 	}
2411 	if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2412 	    RB_EMPTY(&vp->v_rbdirty_tree)) {
2413 		vclrisdirty(vp);
2414 	}
2415 
2416 	hammer2_inode_drop(ip);
2417 #if 1
2418 	error = 0;
2419 	if (error)
2420 		info->error = error;
2421 #endif
2422 	return(0);
2423 }
2424 
2425 static
2426 int
2427 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2428 {
2429 	hammer2_inode_t *ip;
2430 
2431 	KKASSERT(MAXFIDSZ >= 16);
2432 	ip = VTOI(vp);
2433 	fhp->fid_len = offsetof(struct fid, fid_data[16]);
2434 	fhp->fid_ext = 0;
2435 	((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2436 	((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2437 
2438 	return 0;
2439 }
2440 
2441 static
2442 int
2443 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2444 	       struct fid *fhp, struct vnode **vpp)
2445 {
2446 	hammer2_pfs_t *pmp;
2447 	hammer2_tid_t inum;
2448 	int error;
2449 
2450 	pmp = MPTOPMP(mp);
2451 	inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2452 	if (vpp) {
2453 		if (inum == 1)
2454 			error = hammer2_vfs_root(mp, vpp);
2455 		else
2456 			error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2457 	} else {
2458 		error = 0;
2459 	}
2460 	if (error)
2461 		kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2462 	return error;
2463 }
2464 
2465 static
2466 int
2467 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2468 		 int *exflagsp, struct ucred **credanonp)
2469 {
2470 	hammer2_pfs_t *pmp;
2471 	struct netcred *np;
2472 	int error;
2473 
2474 	pmp = MPTOPMP(mp);
2475 	np = vfs_export_lookup(mp, &pmp->export, nam);
2476 	if (np) {
2477 		*exflagsp = np->netc_exflags;
2478 		*credanonp = &np->netc_anon;
2479 		error = 0;
2480 	} else {
2481 		error = EACCES;
2482 	}
2483 	return error;
2484 }
2485 
2486 /*
2487  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2488  * header into the HMP
2489  *
2490  * XXX read four volhdrs and use the one with the highest TID whos CRC
2491  *     matches.
2492  *
2493  * XXX check iCRCs.
2494  *
2495  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2496  *     nonexistant locations.
2497  *
2498  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2499  */
2500 static
2501 int
2502 hammer2_install_volume_header(hammer2_dev_t *hmp)
2503 {
2504 	hammer2_volume_data_t *vd;
2505 	struct buf *bp;
2506 	hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2507 	int error_reported;
2508 	int error;
2509 	int valid;
2510 	int i;
2511 
2512 	error_reported = 0;
2513 	error = 0;
2514 	valid = 0;
2515 	bp = NULL;
2516 
2517 	/*
2518 	 * There are up to 4 copies of the volume header (syncs iterate
2519 	 * between them so there is no single master).  We don't trust the
2520 	 * volu_size field so we don't know precisely how large the filesystem
2521 	 * is, so depend on the OS to return an error if we go beyond the
2522 	 * block device's EOF.
2523 	 */
2524 	for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2525 		error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2526 			      HAMMER2_VOLUME_BYTES, &bp);
2527 		if (error) {
2528 			brelse(bp);
2529 			bp = NULL;
2530 			continue;
2531 		}
2532 
2533 		vd = (struct hammer2_volume_data *) bp->b_data;
2534 		if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2535 		    (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2536 			brelse(bp);
2537 			bp = NULL;
2538 			continue;
2539 		}
2540 
2541 		if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2542 			/* XXX: Reversed-endianness filesystem */
2543 			kprintf("hammer2: reverse-endian filesystem detected");
2544 			brelse(bp);
2545 			bp = NULL;
2546 			continue;
2547 		}
2548 
2549 		crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2550 		crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2551 				      HAMMER2_VOLUME_ICRC0_SIZE);
2552 		bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2553 		bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2554 				       HAMMER2_VOLUME_ICRC1_SIZE);
2555 		if ((crc0 != crc) || (bcrc0 != bcrc)) {
2556 			kprintf("hammer2 volume header crc "
2557 				"mismatch copy #%d %08x/%08x\n",
2558 				i, crc0, crc);
2559 			error_reported = 1;
2560 			brelse(bp);
2561 			bp = NULL;
2562 			continue;
2563 		}
2564 		if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2565 			valid = 1;
2566 			hmp->voldata = *vd;
2567 			hmp->volhdrno = i;
2568 		}
2569 		brelse(bp);
2570 		bp = NULL;
2571 	}
2572 	if (valid) {
2573 		hmp->volsync = hmp->voldata;
2574 		hmp->free_reserved = hmp->voldata.allocator_size / 20;
2575 		error = 0;
2576 		if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2577 			kprintf("hammer2: using volume header #%d\n",
2578 				hmp->volhdrno);
2579 		}
2580 	} else {
2581 		error = EINVAL;
2582 		kprintf("hammer2: no valid volume headers found!\n");
2583 	}
2584 	return (error);
2585 }
2586 
2587 /*
2588  * This handles hysteresis on regular file flushes.  Because the BIOs are
2589  * routed to a thread it is possible for an excessive number to build up
2590  * and cause long front-end stalls long before the runningbuffspace limit
2591  * is hit, so we implement hammer2_flush_pipe to control the
2592  * hysteresis.
2593  *
2594  * This is a particular problem when compression is used.
2595  */
2596 void
2597 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2598 {
2599 	atomic_add_int(&pmp->count_lwinprog, 1);
2600 }
2601 
2602 void
2603 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2604 {
2605 	int lwinprog;
2606 
2607 	lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2608 	if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2609 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2610 		atomic_clear_int(&pmp->count_lwinprog,
2611 				 HAMMER2_LWINPROG_WAITING);
2612 		wakeup(&pmp->count_lwinprog);
2613 	}
2614 	if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2615 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2616 		atomic_clear_int(&pmp->count_lwinprog,
2617 				 HAMMER2_LWINPROG_WAITING0);
2618 		wakeup(&pmp->count_lwinprog);
2619 	}
2620 }
2621 
2622 void
2623 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2624 {
2625 	int lwinprog;
2626 	int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2627 				    HAMMER2_LWINPROG_WAITING0;
2628 
2629 	for (;;) {
2630 		lwinprog = pmp->count_lwinprog;
2631 		cpu_ccfence();
2632 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2633 			break;
2634 		tsleep_interlock(&pmp->count_lwinprog, 0);
2635 		atomic_set_int(&pmp->count_lwinprog, lwflag);
2636 		lwinprog = pmp->count_lwinprog;
2637 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2638 			break;
2639 		tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2640 	}
2641 }
2642 
2643 /*
2644  * Manage excessive memory resource use for chain and related
2645  * structures.
2646  */
2647 void
2648 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2649 {
2650 	uint32_t waiting;
2651 	uint32_t count;
2652 	uint32_t limit;
2653 #if 0
2654 	static int zzticks;
2655 #endif
2656 
2657 	/*
2658 	 * Atomic check condition and wait.  Also do an early speedup of
2659 	 * the syncer to try to avoid hitting the wait.
2660 	 */
2661 	for (;;) {
2662 		waiting = pmp->inmem_dirty_chains;
2663 		cpu_ccfence();
2664 		count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2665 
2666 		limit = pmp->mp->mnt_nvnodelistsize / 10;
2667 		if (limit < hammer2_limit_dirty_chains)
2668 			limit = hammer2_limit_dirty_chains;
2669 		if (limit < 1000)
2670 			limit = 1000;
2671 
2672 #if 0
2673 		if ((int)(ticks - zzticks) > hz) {
2674 			zzticks = ticks;
2675 			kprintf("count %ld %ld\n", count, limit);
2676 		}
2677 #endif
2678 
2679 		/*
2680 		 * Block if there are too many dirty chains present, wait
2681 		 * for the flush to clean some out.
2682 		 */
2683 		if (count > limit) {
2684 			tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2685 			if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2686 					       waiting,
2687 				       waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2688 				speedup_syncer(pmp->mp);
2689 				tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2690 				       "chnmem", hz);
2691 			}
2692 			continue;	/* loop on success or fail */
2693 		}
2694 
2695 		/*
2696 		 * Try to start an early flush before we are forced to block.
2697 		 */
2698 		if (count > limit * 7 / 10)
2699 			speedup_syncer(pmp->mp);
2700 		break;
2701 	}
2702 }
2703 
2704 void
2705 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2706 {
2707 	if (pmp) {
2708 		atomic_add_int(&pmp->inmem_dirty_chains, 1);
2709 	}
2710 }
2711 
2712 void
2713 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2714 {
2715 	uint32_t waiting;
2716 
2717 	if (pmp == NULL)
2718 		return;
2719 
2720 	for (;;) {
2721 		waiting = pmp->inmem_dirty_chains;
2722 		cpu_ccfence();
2723 		if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2724 				       waiting,
2725 				       (waiting - 1) &
2726 					~HAMMER2_DIRTYCHAIN_WAITING)) {
2727 			break;
2728 		}
2729 	}
2730 
2731 	if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2732 		wakeup(&pmp->inmem_dirty_chains);
2733 }
2734 
2735 /*
2736  * Returns 0 if the filesystem has tons of free space
2737  * Returns 1 if the filesystem has less than 10% remaining
2738  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2739  */
2740 int
2741 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2742 {
2743 	hammer2_pfs_t *pmp;
2744 	hammer2_dev_t *hmp;
2745 	hammer2_off_t free_reserved;
2746 	hammer2_off_t free_nominal;
2747 	int i;
2748 
2749 	pmp = ip->pmp;
2750 
2751 	if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2752 		free_reserved = HAMMER2_SEGSIZE;
2753 		free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2754 		for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2755 			hmp = pmp->pfs_hmps[i];
2756 			if (hmp == NULL)
2757 				continue;
2758 			if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2759 			    pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2760 				continue;
2761 
2762 			if (free_nominal > hmp->voldata.allocator_free)
2763 				free_nominal = hmp->voldata.allocator_free;
2764 			if (free_reserved < hmp->free_reserved)
2765 				free_reserved = hmp->free_reserved;
2766 		}
2767 
2768 		/*
2769 		 * SMP races ok
2770 		 */
2771 		pmp->free_reserved = free_reserved;
2772 		pmp->free_nominal = free_nominal;
2773 		pmp->free_ticks = ticks;
2774 	} else {
2775 		free_reserved = pmp->free_reserved;
2776 		free_nominal = pmp->free_nominal;
2777 	}
2778 	if (cred && cred->cr_uid != 0) {
2779 		if ((int64_t)(free_nominal - bytes) <
2780 		    (int64_t)free_reserved) {
2781 			return 2;
2782 		}
2783 	} else {
2784 		if ((int64_t)(free_nominal - bytes) <
2785 		    (int64_t)free_reserved / 2) {
2786 			return 2;
2787 		}
2788 	}
2789 	if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2790 		return 1;
2791 	return 0;
2792 }
2793 
2794 /*
2795  * Debugging
2796  */
2797 void
2798 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2799 {
2800 	hammer2_chain_t *scan;
2801 	hammer2_chain_t *parent;
2802 
2803 	--*countp;
2804 	if (*countp == 0) {
2805 		kprintf("%*.*s...\n", tab, tab, "");
2806 		return;
2807 	}
2808 	if (*countp < 0)
2809 		return;
2810 	kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2811 		tab, tab, "", pfx,
2812 		chain, chain->bref.type,
2813 		chain->bref.key, chain->bref.keybits,
2814 		chain->bref.mirror_tid);
2815 
2816 	kprintf("%*.*s      [%08x] (%s) refs=%d",
2817 		tab, tab, "",
2818 		chain->flags,
2819 		((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2820 		chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2821 		chain->refs);
2822 
2823 	parent = chain->parent;
2824 	if (parent)
2825 		kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2826 			tab, tab, "",
2827 			parent, parent->flags, parent->refs);
2828 	if (RB_EMPTY(&chain->core.rbtree)) {
2829 		kprintf("\n");
2830 	} else {
2831 		kprintf(" {\n");
2832 		RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2833 			hammer2_dump_chain(scan, tab + 4, countp, 'a');
2834 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2835 			kprintf("%*.*s}(%s)\n", tab, tab, "",
2836 				chain->data->ipdata.filename);
2837 		else
2838 			kprintf("%*.*s}\n", tab, tab, "");
2839 	}
2840 }
2841