xref: /dragonfly/sys/vfs/hammer2/hammer2_vfsops.c (revision 8bf5b238)
1 /*
2  * Copyright (c) 2011-2018 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 	int pass;
73 };
74 
75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
76 static struct hammer2_mntlist hammer2_mntlist;
77 
78 struct hammer2_pfslist hammer2_pfslist;
79 struct hammer2_pfslist hammer2_spmplist;
80 struct lock hammer2_mntlk;
81 
82 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
83 int hammer2_debug;
84 int hammer2_cluster_meta_read = 1;	/* physical read-ahead */
85 int hammer2_cluster_data_read = 4;	/* physical read-ahead */
86 int hammer2_dedup_enable = 1;
87 int hammer2_always_compress = 0;	/* always try to compress */
88 int hammer2_inval_enable = 0;
89 int hammer2_flush_pipe = 100;
90 int hammer2_dio_count;
91 int hammer2_dio_limit = 256;
92 int hammer2_bulkfree_tps = 5000;
93 long hammer2_chain_allocs;
94 long hammer2_chain_frees;
95 long hammer2_limit_dirty_chains;
96 long hammer2_limit_dirty_inodes;
97 long hammer2_count_modified_chains;
98 long hammer2_iod_invals;
99 long hammer2_iod_file_read;
100 long hammer2_iod_meta_read;
101 long hammer2_iod_indr_read;
102 long hammer2_iod_fmap_read;
103 long hammer2_iod_volu_read;
104 long hammer2_iod_file_write;
105 long hammer2_iod_file_wembed;
106 long hammer2_iod_file_wzero;
107 long hammer2_iod_file_wdedup;
108 long hammer2_iod_meta_write;
109 long hammer2_iod_indr_write;
110 long hammer2_iod_fmap_write;
111 long hammer2_iod_volu_write;
112 
113 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
114 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
115 		"Buffer used for compression.");
116 
117 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
118 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
119 		"Buffer used for decompression.");
120 
121 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
122 
123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
124 	   &hammer2_supported_version, 0, "");
125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
126 	   &hammer2_debug, 0, "");
127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
128 	   &hammer2_cluster_meta_read, 0, "");
129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
130 	   &hammer2_cluster_data_read, 0, "");
131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
132 	   &hammer2_dedup_enable, 0, "");
133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
134 	   &hammer2_always_compress, 0, "");
135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
136 	   &hammer2_inval_enable, 0, "");
137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
138 	   &hammer2_flush_pipe, 0, "");
139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
140 	   &hammer2_bulkfree_tps, 0, "");
141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
142 	   &hammer2_chain_allocs, 0, "");
143 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
144 	   &hammer2_chain_frees, 0, "");
145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
146 	   &hammer2_limit_dirty_chains, 0, "");
147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
148 	   &hammer2_limit_dirty_inodes, 0, "");
149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
150 	   &hammer2_count_modified_chains, 0, "");
151 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
152 	   &hammer2_dio_count, 0, "");
153 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
154 	   &hammer2_dio_limit, 0, "");
155 
156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
157 	   &hammer2_iod_invals, 0, "");
158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
159 	   &hammer2_iod_file_read, 0, "");
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
161 	   &hammer2_iod_meta_read, 0, "");
162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
163 	   &hammer2_iod_indr_read, 0, "");
164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
165 	   &hammer2_iod_fmap_read, 0, "");
166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
167 	   &hammer2_iod_volu_read, 0, "");
168 
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
170 	   &hammer2_iod_file_write, 0, "");
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
172 	   &hammer2_iod_file_wembed, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
174 	   &hammer2_iod_file_wzero, 0, "");
175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
176 	   &hammer2_iod_file_wdedup, 0, "");
177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
178 	   &hammer2_iod_meta_write, 0, "");
179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
180 	   &hammer2_iod_indr_write, 0, "");
181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
182 	   &hammer2_iod_fmap_write, 0, "");
183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
184 	   &hammer2_iod_volu_write, 0, "");
185 
186 long hammer2_process_icrc32;
187 long hammer2_process_xxhash64;
188 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
189 	   &hammer2_process_icrc32, 0, "");
190 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
191 	   &hammer2_process_xxhash64, 0, "");
192 
193 static int hammer2_vfs_init(struct vfsconf *conf);
194 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
195 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
196 				struct ucred *cred);
197 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
198 				struct vnode *, struct ucred *);
199 static int hammer2_recovery(hammer2_dev_t *hmp);
200 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
201 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
202 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
203 				struct ucred *cred);
204 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
205 				struct ucred *cred);
206 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
207 				struct fid *fhp, struct vnode **vpp);
208 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
209 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
210 				int *exflagsp, struct ucred **credanonp);
211 
212 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
213 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
214 
215 static void hammer2_update_pmps(hammer2_dev_t *hmp);
216 
217 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
218 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
219 				hammer2_dev_t *hmp);
220 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
221 
222 /*
223  * HAMMER2 vfs operations.
224  */
225 static struct vfsops hammer2_vfsops = {
226 	.vfs_init	= hammer2_vfs_init,
227 	.vfs_uninit	= hammer2_vfs_uninit,
228 	.vfs_sync	= hammer2_vfs_sync,
229 	.vfs_mount	= hammer2_vfs_mount,
230 	.vfs_unmount	= hammer2_vfs_unmount,
231 	.vfs_root 	= hammer2_vfs_root,
232 	.vfs_statfs	= hammer2_vfs_statfs,
233 	.vfs_statvfs	= hammer2_vfs_statvfs,
234 	.vfs_vget	= hammer2_vfs_vget,
235 	.vfs_vptofh	= hammer2_vfs_vptofh,
236 	.vfs_fhtovp	= hammer2_vfs_fhtovp,
237 	.vfs_checkexp	= hammer2_vfs_checkexp
238 };
239 
240 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
241 
242 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
243 MODULE_VERSION(hammer2, 1);
244 
245 static
246 int
247 hammer2_vfs_init(struct vfsconf *conf)
248 {
249 	static struct objcache_malloc_args margs_read;
250 	static struct objcache_malloc_args margs_write;
251 	static struct objcache_malloc_args margs_vop;
252 
253 	int error;
254 
255 	error = 0;
256 
257 	/*
258 	 * A large DIO cache is needed to retain dedup enablement masks.
259 	 * The bulkfree code clears related masks as part of the disk block
260 	 * recycling algorithm, preventing it from being used for a later
261 	 * dedup.
262 	 *
263 	 * NOTE: A large buffer cache can actually interfere with dedup
264 	 *	 operation because we dedup based on media physical buffers
265 	 *	 and not logical buffers.  Try to make the DIO case large
266 	 *	 enough to avoid this problem, but also cap it.
267 	 */
268 	hammer2_dio_limit = nbuf * 2;
269 	if (hammer2_dio_limit > 100000)
270 		hammer2_dio_limit = 100000;
271 
272 	if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
273 		error = EINVAL;
274 	if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
275 		error = EINVAL;
276 	if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
277 		error = EINVAL;
278 
279 	if (error)
280 		kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
281 
282 	margs_read.objsize = 65536;
283 	margs_read.mtype = M_HAMMER2_DEBUFFER;
284 
285 	margs_write.objsize = 32768;
286 	margs_write.mtype = M_HAMMER2_CBUFFER;
287 
288 	margs_vop.objsize = sizeof(hammer2_xop_t);
289 	margs_vop.mtype = M_HAMMER2;
290 
291 	/*
292 	 * Note thaht for the XOPS cache we want backing store allocations
293 	 * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
294 	 * confusion), so use the backing store function that does it.  This
295 	 * means that initial XOPS objects are zerod but REUSED objects are
296 	 * not.  So we are responsible for cleaning the object up sufficiently
297 	 * for our needs before objcache_put()ing it back (typically just the
298 	 * FIFO indices).
299 	 */
300 	cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
301 				0, 1, NULL, NULL, NULL,
302 				objcache_malloc_alloc,
303 				objcache_malloc_free,
304 				&margs_read);
305 	cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
306 				0, 1, NULL, NULL, NULL,
307 				objcache_malloc_alloc,
308 				objcache_malloc_free,
309 				&margs_write);
310 	cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
311 				0, 1, NULL, NULL, NULL,
312 				objcache_malloc_alloc_zero,
313 				objcache_malloc_free,
314 				&margs_vop);
315 
316 
317 	lockinit(&hammer2_mntlk, "mntlk", 0, 0);
318 	TAILQ_INIT(&hammer2_mntlist);
319 	TAILQ_INIT(&hammer2_pfslist);
320 	TAILQ_INIT(&hammer2_spmplist);
321 
322 	hammer2_limit_dirty_chains = maxvnodes / 10;
323 	if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
324 		hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
325 
326 	hammer2_limit_dirty_inodes = maxvnodes / 100;
327 	if (hammer2_limit_dirty_inodes < 100)
328 		hammer2_limit_dirty_inodes = 100;
329 	if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
330 		hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
331 
332 	return (error);
333 }
334 
335 static
336 int
337 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
338 {
339 	objcache_destroy(cache_buffer_read);
340 	objcache_destroy(cache_buffer_write);
341 	objcache_destroy(cache_xops);
342 	return 0;
343 }
344 
345 /*
346  * Core PFS allocator.  Used to allocate or reference the pmp structure
347  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
348  * The pmp can be passed in or loaded by this function using the chain and
349  * inode data.
350  *
351  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
352  * transactions.  Note that synchronization does not use this field.
353  * (typically frontend operations and synchronization cannot run on the
354  * same PFS node at the same time).
355  *
356  * XXX check locking
357  */
358 hammer2_pfs_t *
359 hammer2_pfsalloc(hammer2_chain_t *chain,
360 		 const hammer2_inode_data_t *ripdata,
361 		 hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
362 {
363 	hammer2_pfs_t *pmp;
364 	hammer2_inode_t *iroot;
365 	int count;
366 	int i;
367 	int j;
368 
369 	pmp = NULL;
370 
371 	/*
372 	 * Locate or create the PFS based on the cluster id.  If ripdata
373 	 * is NULL this is a spmp which is unique and is always allocated.
374 	 *
375 	 * If the device is mounted in local mode all PFSs are considered
376 	 * independent and not part of any cluster (for debugging only).
377 	 */
378 	if (ripdata) {
379 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
380 			if (force_local != pmp->force_local)
381 				continue;
382 			if (force_local == NULL &&
383 			    bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
384 				 sizeof(pmp->pfs_clid)) == 0) {
385 					break;
386 			} else if (force_local && pmp->pfs_names[0] &&
387 			    strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
388 					break;
389 			}
390 		}
391 	}
392 
393 	if (pmp == NULL) {
394 		pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
395 		pmp->force_local = force_local;
396 		hammer2_trans_manage_init(pmp);
397 		kmalloc_create(&pmp->minode, "HAMMER2-inodes");
398 		kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
399 		lockinit(&pmp->lock, "pfslk", 0, 0);
400 		lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
401 		spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
402 		spin_init(&pmp->xop_spin, "h2xop");
403 		spin_init(&pmp->lru_spin, "h2lru");
404 		RB_INIT(&pmp->inum_tree);
405 		TAILQ_INIT(&pmp->sideq);
406 		TAILQ_INIT(&pmp->lru_list);
407 		spin_init(&pmp->list_spin, "hm2pfsalloc_list");
408 
409 		/*
410 		 * Distribute backend operations to threads
411 		 */
412 		for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
413 			hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
414 
415 		/*
416 		 * Save the last media transaction id for the flusher.  Set
417 		 * initial
418 		 */
419 		if (ripdata) {
420 			pmp->pfs_clid = ripdata->meta.pfs_clid;
421 			TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
422 		} else {
423 			pmp->flags |= HAMMER2_PMPF_SPMP;
424 			TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
425 		}
426 
427 		/*
428 		 * The synchronization thread may start too early, make
429 		 * sure it stays frozen until we are ready to let it go.
430 		 * XXX
431 		 */
432 		/*
433 		pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
434 					 HAMMER2_THREAD_REMASTER;
435 		*/
436 	}
437 
438 	/*
439 	 * Create the PFS's root inode and any missing XOP helper threads.
440 	 */
441 	if ((iroot = pmp->iroot) == NULL) {
442 		iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
443 		if (ripdata)
444 			iroot->meta = ripdata->meta;
445 		pmp->iroot = iroot;
446 		hammer2_inode_ref(iroot);
447 		hammer2_inode_unlock(iroot);
448 	}
449 
450 	/*
451 	 * Stop here if no chain is passed in.
452 	 */
453 	if (chain == NULL)
454 		goto done;
455 
456 	/*
457 	 * When a chain is passed in we must add it to the PFS's root
458 	 * inode, update pmp->pfs_types[], and update the syncronization
459 	 * threads.
460 	 *
461 	 * When forcing local mode, mark the PFS as a MASTER regardless.
462 	 *
463 	 * At the moment empty spots can develop due to removals or failures.
464 	 * Ultimately we want to re-fill these spots but doing so might
465 	 * confused running code. XXX
466 	 */
467 	hammer2_inode_ref(iroot);
468 	hammer2_mtx_ex(&iroot->lock);
469 	j = iroot->cluster.nchains;
470 
471 	if (j == HAMMER2_MAXCLUSTER) {
472 		kprintf("hammer2_mount: cluster full!\n");
473 		/* XXX fatal error? */
474 	} else {
475 		KKASSERT(chain->pmp == NULL);
476 		chain->pmp = pmp;
477 		hammer2_chain_ref(chain);
478 		iroot->cluster.array[j].chain = chain;
479 		if (force_local)
480 			pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
481 		else
482 			pmp->pfs_types[j] = ripdata->meta.pfs_type;
483 		pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
484 		pmp->pfs_hmps[j] = chain->hmp;
485 
486 		/*
487 		 * If the PFS is already mounted we must account
488 		 * for the mount_count here.
489 		 */
490 		if (pmp->mp)
491 			++chain->hmp->mount_count;
492 
493 		/*
494 		 * May have to fixup dirty chain tracking.  Previous
495 		 * pmp was NULL so nothing to undo.
496 		 */
497 		if (chain->flags & HAMMER2_CHAIN_MODIFIED)
498 			hammer2_pfs_memory_inc(pmp);
499 		++j;
500 	}
501 	iroot->cluster.nchains = j;
502 
503 	/*
504 	 * Update nmasters from any PFS inode which is part of the cluster.
505 	 * It is possible that this will result in a value which is too
506 	 * high.  MASTER PFSs are authoritative for pfs_nmasters and will
507 	 * override this value later on.
508 	 *
509 	 * (This informs us of masters that might not currently be
510 	 *  discoverable by this mount).
511 	 */
512 	if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
513 		pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
514 	}
515 
516 	/*
517 	 * Count visible masters.  Masters are usually added with
518 	 * ripdata->meta.pfs_nmasters set to 1.  This detects when there
519 	 * are more (XXX and must update the master inodes).
520 	 */
521 	count = 0;
522 	for (i = 0; i < iroot->cluster.nchains; ++i) {
523 		if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
524 			++count;
525 	}
526 	if (pmp->pfs_nmasters < count)
527 		pmp->pfs_nmasters = count;
528 
529 	/*
530 	 * Create missing synchronization and support threads.
531 	 *
532 	 * Single-node masters (including snapshots) have nothing to
533 	 * synchronize and do not require this thread.
534 	 *
535 	 * Multi-node masters or any number of soft masters, slaves, copy,
536 	 * or other PFS types need the thread.
537 	 *
538 	 * Each thread is responsible for its particular cluster index.
539 	 * We use independent threads so stalls or mismatches related to
540 	 * any given target do not affect other targets.
541 	 */
542 	for (i = 0; i < iroot->cluster.nchains; ++i) {
543 		/*
544 		 * Single-node masters (including snapshots) have nothing
545 		 * to synchronize and will make direct xops support calls,
546 		 * thus they do not require this thread.
547 		 *
548 		 * Note that there can be thousands of snapshots.  We do not
549 		 * want to create thousands of threads.
550 		 */
551 		if (pmp->pfs_nmasters <= 1 &&
552 		    pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
553 			continue;
554 		}
555 
556 		/*
557 		 * Sync support thread
558 		 */
559 		if (pmp->sync_thrs[i].td == NULL) {
560 			hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
561 					   "h2nod", i, -1,
562 					   hammer2_primary_sync_thread);
563 		}
564 	}
565 
566 	/*
567 	 * Create missing Xop threads
568 	 *
569 	 * NOTE: We create helper threads for all mounted PFSs or any
570 	 *	 PFSs with 2+ nodes (so the sync thread can update them,
571 	 *	 even if not mounted).
572 	 */
573 	if (pmp->mp || iroot->cluster.nchains >= 2)
574 		hammer2_xop_helper_create(pmp);
575 
576 	hammer2_mtx_unlock(&iroot->lock);
577 	hammer2_inode_drop(iroot);
578 done:
579 	return pmp;
580 }
581 
582 /*
583  * Deallocate an element of a probed PFS.  If destroying and this is a
584  * MASTER, adjust nmasters.
585  *
586  * This function does not physically destroy the PFS element in its device
587  * under the super-root  (see hammer2_ioctl_pfs_delete()).
588  */
589 void
590 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
591 {
592 	hammer2_inode_t *iroot;
593 	hammer2_chain_t *chain;
594 	int j;
595 
596 	/*
597 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
598 	 * by the flush code.
599 	 */
600 	iroot = pmp->iroot;
601 	if (iroot) {
602 		/*
603 		 * Stop synchronizing
604 		 *
605 		 * XXX flush after acquiring the iroot lock.
606 		 * XXX clean out the cluster index from all inode structures.
607 		 */
608 		hammer2_thr_delete(&pmp->sync_thrs[clindex]);
609 
610 		/*
611 		 * Remove the cluster index from the group.  If destroying
612 		 * the PFS and this is a master, adjust pfs_nmasters.
613 		 */
614 		hammer2_mtx_ex(&iroot->lock);
615 		chain = iroot->cluster.array[clindex].chain;
616 		iroot->cluster.array[clindex].chain = NULL;
617 
618 		switch(pmp->pfs_types[clindex]) {
619 		case HAMMER2_PFSTYPE_MASTER:
620 			if (destroying && pmp->pfs_nmasters > 0)
621 				--pmp->pfs_nmasters;
622 			/* XXX adjust ripdata->meta.pfs_nmasters */
623 			break;
624 		default:
625 			break;
626 		}
627 		pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
628 
629 		hammer2_mtx_unlock(&iroot->lock);
630 
631 		/*
632 		 * Release the chain.
633 		 */
634 		if (chain) {
635 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
636 			hammer2_chain_drop(chain);
637 		}
638 
639 		/*
640 		 * Terminate all XOP threads for the cluster index.
641 		 */
642 		for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
643 			hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
644 	}
645 }
646 
647 /*
648  * Destroy a PFS, typically only occurs after the last mount on a device
649  * has gone away.
650  */
651 static void
652 hammer2_pfsfree(hammer2_pfs_t *pmp)
653 {
654 	hammer2_inode_t *iroot;
655 	hammer2_chain_t *chain;
656 	int i;
657 	int j;
658 
659 	/*
660 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
661 	 * by the flush code.
662 	 */
663 	if (pmp->flags & HAMMER2_PMPF_SPMP)
664 		TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
665 	else
666 		TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
667 
668 	iroot = pmp->iroot;
669 	if (iroot) {
670 		for (i = 0; i < iroot->cluster.nchains; ++i) {
671 			hammer2_thr_delete(&pmp->sync_thrs[i]);
672 			for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
673 				hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
674 		}
675 #if REPORT_REFS_ERRORS
676 		if (pmp->iroot->refs != 1)
677 			kprintf("PMP->IROOT %p REFS WRONG %d\n",
678 				pmp->iroot, pmp->iroot->refs);
679 #else
680 		KKASSERT(pmp->iroot->refs == 1);
681 #endif
682 		/* ref for pmp->iroot */
683 		hammer2_inode_drop(pmp->iroot);
684 		pmp->iroot = NULL;
685 	}
686 
687 	/*
688 	 * Cleanup chains remaining on LRU list.
689 	 */
690 	hammer2_spin_ex(&pmp->lru_spin);
691 	while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
692 		KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
693 		atomic_add_int(&pmp->lru_count, -1);
694 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
695 		TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
696 		hammer2_chain_ref(chain);
697 		hammer2_spin_unex(&pmp->lru_spin);
698 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
699 		hammer2_chain_drop(chain);
700 		hammer2_spin_ex(&pmp->lru_spin);
701 	}
702 	hammer2_spin_unex(&pmp->lru_spin);
703 
704 	/*
705 	 * Free remaining pmp resources
706 	 */
707 	kmalloc_destroy(&pmp->mmsg);
708 	kmalloc_destroy(&pmp->minode);
709 
710 	kfree(pmp, M_HAMMER2);
711 }
712 
713 /*
714  * Remove all references to hmp from the pfs list.  Any PFS which becomes
715  * empty is terminated and freed.
716  *
717  * XXX inefficient.
718  */
719 static void
720 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
721 {
722 	hammer2_pfs_t *pmp;
723 	hammer2_inode_t *iroot;
724 	hammer2_chain_t *rchain;
725 	int didfreeze;
726 	int i;
727 	int j;
728 	struct hammer2_pfslist *wlist;
729 
730 	if (which == 0)
731 		wlist = &hammer2_pfslist;
732 	else
733 		wlist = &hammer2_spmplist;
734 again:
735 	TAILQ_FOREACH(pmp, wlist, mntentry) {
736 		if ((iroot = pmp->iroot) == NULL)
737 			continue;
738 		hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
739 		hammer2_inode_run_sideq(pmp, 1);
740 		hammer2_bioq_sync(pmp);
741 		hammer2_trans_done(pmp, 0);
742 
743 		/*
744 		 * Determine if this PFS is affected.  If it is we must
745 		 * freeze all management threads and lock its iroot.
746 		 *
747 		 * Freezing a management thread forces it idle, operations
748 		 * in-progress will be aborted and it will have to start
749 		 * over again when unfrozen, or exit if told to exit.
750 		 */
751 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
752 			if (pmp->pfs_hmps[i] == hmp)
753 				break;
754 		}
755 		if (i != HAMMER2_MAXCLUSTER) {
756 			/*
757 			 * Make sure all synchronization threads are locked
758 			 * down.
759 			 */
760 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
761 				if (pmp->pfs_hmps[i] == NULL)
762 					continue;
763 				hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
764 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
765 					hammer2_thr_freeze_async(
766 						&pmp->xop_groups[j].thrs[i]);
767 				}
768 			}
769 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
770 				if (pmp->pfs_hmps[i] == NULL)
771 					continue;
772 				hammer2_thr_freeze(&pmp->sync_thrs[i]);
773 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
774 					hammer2_thr_freeze(
775 						&pmp->xop_groups[j].thrs[i]);
776 				}
777 			}
778 
779 			/*
780 			 * Lock the inode and clean out matching chains.
781 			 * Note that we cannot use hammer2_inode_lock_*()
782 			 * here because that would attempt to validate the
783 			 * cluster that we are in the middle of ripping
784 			 * apart.
785 			 *
786 			 * WARNING! We are working directly on the inodes
787 			 *	    embedded cluster.
788 			 */
789 			hammer2_mtx_ex(&iroot->lock);
790 
791 			/*
792 			 * Remove the chain from matching elements of the PFS.
793 			 */
794 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
795 				if (pmp->pfs_hmps[i] != hmp)
796 					continue;
797 				hammer2_thr_delete(&pmp->sync_thrs[i]);
798 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
799 					hammer2_thr_delete(
800 						&pmp->xop_groups[j].thrs[i]);
801 				}
802 				rchain = iroot->cluster.array[i].chain;
803 				iroot->cluster.array[i].chain = NULL;
804 				pmp->pfs_types[i] = 0;
805 				if (pmp->pfs_names[i]) {
806 					kfree(pmp->pfs_names[i], M_HAMMER2);
807 					pmp->pfs_names[i] = NULL;
808 				}
809 				if (rchain) {
810 					hammer2_chain_drop(rchain);
811 					/* focus hint */
812 					if (iroot->cluster.focus == rchain)
813 						iroot->cluster.focus = NULL;
814 				}
815 				pmp->pfs_hmps[i] = NULL;
816 			}
817 			hammer2_mtx_unlock(&iroot->lock);
818 			didfreeze = 1;	/* remaster, unfreeze down below */
819 		} else {
820 			didfreeze = 0;
821 		}
822 
823 		/*
824 		 * Cleanup trailing chains.  Gaps may remain.
825 		 */
826 		for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
827 			if (pmp->pfs_hmps[i])
828 				break;
829 		}
830 		iroot->cluster.nchains = i + 1;
831 
832 		/*
833 		 * If the PMP has no elements remaining we can destroy it.
834 		 * (this will transition management threads from frozen->exit).
835 		 */
836 		if (iroot->cluster.nchains == 0) {
837 			/*
838 			 * If this was the hmp's spmp, we need to clean
839 			 * a little more stuff out.
840 			 */
841 			if (hmp->spmp == pmp) {
842 				hmp->spmp = NULL;
843 				hmp->vchain.pmp = NULL;
844 				hmp->fchain.pmp = NULL;
845 			}
846 
847 			/*
848 			 * Free the pmp and restart the loop
849 			 */
850 			hammer2_pfsfree(pmp);
851 			goto again;
852 		}
853 
854 		/*
855 		 * If elements still remain we need to set the REMASTER
856 		 * flag and unfreeze it.
857 		 */
858 		if (didfreeze) {
859 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
860 				if (pmp->pfs_hmps[i] == NULL)
861 					continue;
862 				hammer2_thr_remaster(&pmp->sync_thrs[i]);
863 				hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
864 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
865 					hammer2_thr_remaster(
866 						&pmp->xop_groups[j].thrs[i]);
867 					hammer2_thr_unfreeze(
868 						&pmp->xop_groups[j].thrs[i]);
869 				}
870 			}
871 		}
872 	}
873 }
874 
875 /*
876  * Mount or remount HAMMER2 fileystem from physical media
877  *
878  *	mountroot
879  *		mp		mount point structure
880  *		path		NULL
881  *		data		<unused>
882  *		cred		<unused>
883  *
884  *	mount
885  *		mp		mount point structure
886  *		path		path to mount point
887  *		data		pointer to argument structure in user space
888  *			volume	volume path (device@LABEL form)
889  *			hflags	user mount flags
890  *		cred		user credentials
891  *
892  * RETURNS:	0	Success
893  *		!0	error number
894  */
895 static
896 int
897 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
898 		  struct ucred *cred)
899 {
900 	struct hammer2_mount_info info;
901 	hammer2_pfs_t *pmp;
902 	hammer2_pfs_t *spmp;
903 	hammer2_dev_t *hmp;
904 	hammer2_dev_t *force_local;
905 	hammer2_key_t key_next;
906 	hammer2_key_t key_dummy;
907 	hammer2_key_t lhc;
908 	struct vnode *devvp;
909 	struct nlookupdata nd;
910 	hammer2_chain_t *parent;
911 	hammer2_chain_t *chain;
912 	const hammer2_inode_data_t *ripdata;
913 	hammer2_blockref_t bref;
914 	struct file *fp;
915 	char devstr[MNAMELEN];
916 	size_t size;
917 	size_t done;
918 	char *dev;
919 	char *label;
920 	int ronly = 1;
921 	int error;
922 	int i;
923 
924 	hmp = NULL;
925 	pmp = NULL;
926 	dev = NULL;
927 	label = NULL;
928 	devvp = NULL;
929 
930 	if (path == NULL) {
931 		/*
932 		 * Root mount
933 		 */
934 		bzero(&info, sizeof(info));
935 		info.cluster_fd = -1;
936 		ksnprintf(devstr, sizeof(devstr), "%s",
937 			  mp->mnt_stat.f_mntfromname);
938 		kprintf("hammer2_mount: root '%s'\n", devstr);
939 		done = strlen(devstr) + 1;
940 	} else {
941 		/*
942 		 * Non-root mount or updating a mount
943 		 */
944 		error = copyin(data, &info, sizeof(info));
945 		if (error)
946 			return (error);
947 
948 		error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
949 		if (error)
950 			return (error);
951 		kprintf("hammer2_mount: '%s'\n", devstr);
952 	}
953 
954 	/*
955 	 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
956 	 * if no label specified, based on the partition id.  Error out if no
957 	 * label or device (with partition id) is specified.  This is strictly
958 	 * a convenience to match the default label created by newfs_hammer2,
959 	 * our preference is that a label always be specified.
960 	 *
961 	 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
962 	 *	 that does not specify a device, as long as some H2 label
963 	 *	 has already been mounted from that device.  This makes
964 	 *	 mounting snapshots a lot easier.
965 	 */
966 	dev = devstr;
967 	label = strchr(devstr, '@');
968 	if (label && ((label + 1) - dev) > done) {
969 		kprintf("hammer2: mount: bad label %s/%zd\n",
970 			devstr, done);
971 		return (EINVAL);
972 	}
973 	if (label == NULL || label[1] == 0) {
974 		char slice;
975 
976 		if (label == NULL)
977 			label = devstr + strlen(devstr);
978 		else
979 			*label = '\0';		/* clean up trailing @ */
980 
981 		slice = label[-1];
982 		switch(slice) {
983 		case 'a':
984 			label = "BOOT";
985 			break;
986 		case 'd':
987 			label = "ROOT";
988 			break;
989 		default:
990 			label = "DATA";
991 			break;
992 		}
993 	} else {
994 		*label = '\0';
995 		label++;
996 	}
997 
998 	kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
999 		dev, label, (mp->mnt_flag & MNT_RDONLY));
1000 
1001 	if (mp->mnt_flag & MNT_UPDATE) {
1002 		/*
1003 		 * Update mount.  Note that pmp->iroot->cluster is
1004 		 * an inode-embedded cluster and thus cannot be
1005 		 * directly locked.
1006 		 *
1007 		 * XXX HAMMER2 needs to implement NFS export via
1008 		 *     mountctl.
1009 		 */
1010 		hammer2_cluster_t *cluster;
1011 
1012 		pmp = MPTOPMP(mp);
1013 		pmp->hflags = info.hflags;
1014 		cluster = &pmp->iroot->cluster;
1015 		for (i = 0; i < cluster->nchains; ++i) {
1016 			if (cluster->array[i].chain == NULL)
1017 				continue;
1018 			hmp = cluster->array[i].chain->hmp;
1019 			devvp = hmp->devvp;
1020 			error = hammer2_remount(hmp, mp, path,
1021 						devvp, cred);
1022 			if (error)
1023 				break;
1024 		}
1025 
1026 		return error;
1027 	}
1028 
1029 	/*
1030 	 * HMP device mount
1031 	 *
1032 	 * If a path is specified and dev is not an empty string, lookup the
1033 	 * name and verify that it referes to a block device.
1034 	 *
1035 	 * If a path is specified and dev is an empty string we fall through
1036 	 * and locate the label in the hmp search.
1037 	 */
1038 	if (path && *dev != 0) {
1039 		error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1040 		if (error == 0)
1041 			error = nlookup(&nd);
1042 		if (error == 0)
1043 			error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1044 		nlookup_done(&nd);
1045 	} else if (path == NULL) {
1046 		/* root mount */
1047 		cdev_t cdev = kgetdiskbyname(dev);
1048 		error = bdevvp(cdev, &devvp);
1049 		if (error)
1050 			kprintf("hammer2: cannot find '%s'\n", dev);
1051 	} else {
1052 		/*
1053 		 * We will locate the hmp using the label in the hmp loop.
1054 		 */
1055 		error = 0;
1056 	}
1057 
1058 	/*
1059 	 * Make sure its a block device.  Do not check to see if it is
1060 	 * already mounted until we determine that its a fresh H2 device.
1061 	 */
1062 	if (error == 0 && devvp) {
1063 		vn_isdisk(devvp, &error);
1064 	}
1065 
1066 	/*
1067 	 * Determine if the device has already been mounted.  After this
1068 	 * check hmp will be non-NULL if we are doing the second or more
1069 	 * hammer2 mounts from the same device.
1070 	 */
1071 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1072 	if (devvp) {
1073 		/*
1074 		 * Match the device.  Due to the way devfs works,
1075 		 * we may not be able to directly match the vnode pointer,
1076 		 * so also check to see if the underlying device matches.
1077 		 */
1078 		TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1079 			if (hmp->devvp == devvp)
1080 				break;
1081 			if (devvp->v_rdev &&
1082 			    hmp->devvp->v_rdev == devvp->v_rdev) {
1083 				break;
1084 			}
1085 		}
1086 
1087 		/*
1088 		 * If no match this may be a fresh H2 mount, make sure
1089 		 * the device is not mounted on anything else.
1090 		 */
1091 		if (hmp == NULL)
1092 			error = vfs_mountedon(devvp);
1093 	} else if (error == 0) {
1094 		/*
1095 		 * Match the label to a pmp already probed.
1096 		 */
1097 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1098 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1099 				if (pmp->pfs_names[i] &&
1100 				    strcmp(pmp->pfs_names[i], label) == 0) {
1101 					hmp = pmp->pfs_hmps[i];
1102 					break;
1103 				}
1104 			}
1105 			if (hmp)
1106 				break;
1107 		}
1108 		if (hmp == NULL)
1109 			error = ENOENT;
1110 	}
1111 
1112 	/*
1113 	 * Open the device if this isn't a secondary mount and construct
1114 	 * the H2 device mount (hmp).
1115 	 */
1116 	if (hmp == NULL) {
1117 		hammer2_chain_t *schain;
1118 		hammer2_xid_t xid;
1119 		hammer2_xop_head_t xop;
1120 
1121 		if (error == 0 && vcount(devvp) > 0) {
1122 			kprintf("Primary device already has references\n");
1123 			error = EBUSY;
1124 		}
1125 
1126 		/*
1127 		 * Now open the device
1128 		 */
1129 		if (error == 0) {
1130 			ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1131 			vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1132 			error = vinvalbuf(devvp, V_SAVE, 0, 0);
1133 			if (error == 0) {
1134 				error = VOP_OPEN(devvp,
1135 					     (ronly ? FREAD : FREAD | FWRITE),
1136 					     FSCRED, NULL);
1137 			}
1138 			vn_unlock(devvp);
1139 		}
1140 		if (error && devvp) {
1141 			vrele(devvp);
1142 			devvp = NULL;
1143 		}
1144 		if (error) {
1145 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1146 			return error;
1147 		}
1148 		hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1149 		ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1150 		hmp->ronly = ronly;
1151 		hmp->devvp = devvp;
1152 		hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1153 		kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1154 		TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1155 		RB_INIT(&hmp->iotree);
1156 		spin_init(&hmp->io_spin, "hm2mount_io");
1157 		spin_init(&hmp->list_spin, "hm2mount_list");
1158 		TAILQ_INIT(&hmp->flushq);
1159 
1160 		lockinit(&hmp->vollk, "h2vol", 0, 0);
1161 		lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1162 		lockinit(&hmp->bflock, "h2bflk", 0, 0);
1163 
1164 		/*
1165 		 * vchain setup. vchain.data is embedded.
1166 		 * vchain.refs is initialized and will never drop to 0.
1167 		 *
1168 		 * NOTE! voldata is not yet loaded.
1169 		 */
1170 		hmp->vchain.hmp = hmp;
1171 		hmp->vchain.refs = 1;
1172 		hmp->vchain.data = (void *)&hmp->voldata;
1173 		hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1174 		hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1175 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1176 
1177 		hammer2_chain_core_init(&hmp->vchain);
1178 		/* hmp->vchain.u.xxx is left NULL */
1179 
1180 		/*
1181 		 * fchain setup.  fchain.data is embedded.
1182 		 * fchain.refs is initialized and will never drop to 0.
1183 		 *
1184 		 * The data is not used but needs to be initialized to
1185 		 * pass assertion muster.  We use this chain primarily
1186 		 * as a placeholder for the freemap's top-level RBTREE
1187 		 * so it does not interfere with the volume's topology
1188 		 * RBTREE.
1189 		 */
1190 		hmp->fchain.hmp = hmp;
1191 		hmp->fchain.refs = 1;
1192 		hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1193 		hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1194 		hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1195 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1196 		hmp->fchain.bref.methods =
1197 			HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1198 			HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1199 
1200 		hammer2_chain_core_init(&hmp->fchain);
1201 		/* hmp->fchain.u.xxx is left NULL */
1202 
1203 		/*
1204 		 * Install the volume header and initialize fields from
1205 		 * voldata.
1206 		 */
1207 		error = hammer2_install_volume_header(hmp);
1208 		if (error) {
1209 			hammer2_unmount_helper(mp, NULL, hmp);
1210 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1211 			hammer2_vfs_unmount(mp, MNT_FORCE);
1212 			return error;
1213 		}
1214 
1215 		/*
1216 		 * Really important to get these right or flush will get
1217 		 * confused.
1218 		 */
1219 		hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1220 		spmp = hmp->spmp;
1221 
1222 		/*
1223 		 * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1224 		 * is inherited from the volume header.
1225 		 */
1226 		xid = 0;
1227 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1228 		hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1229 		hmp->vchain.pmp = spmp;
1230 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1231 		hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1232 		hmp->fchain.pmp = spmp;
1233 
1234 		/*
1235 		 * First locate the super-root inode, which is key 0
1236 		 * relative to the volume header's blockset.
1237 		 *
1238 		 * Then locate the root inode by scanning the directory keyspace
1239 		 * represented by the label.
1240 		 */
1241 		parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1242 		schain = hammer2_chain_lookup(&parent, &key_dummy,
1243 				      HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1244 				      &error, 0);
1245 		hammer2_chain_lookup_done(parent);
1246 		if (schain == NULL) {
1247 			kprintf("hammer2_mount: invalid super-root\n");
1248 			hammer2_unmount_helper(mp, NULL, hmp);
1249 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1250 			hammer2_vfs_unmount(mp, MNT_FORCE);
1251 			return EINVAL;
1252 		}
1253 		if (schain->error) {
1254 			kprintf("hammer2_mount: error %s reading super-root\n",
1255 				hammer2_error_str(schain->error));
1256 			hammer2_chain_unlock(schain);
1257 			hammer2_chain_drop(schain);
1258 			schain = NULL;
1259 			hammer2_unmount_helper(mp, NULL, hmp);
1260 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1261 			hammer2_vfs_unmount(mp, MNT_FORCE);
1262 			return EINVAL;
1263 		}
1264 
1265 		/*
1266 		 * The super-root always uses an inode_tid of 1 when
1267 		 * creating PFSs.
1268 		 */
1269 		spmp->inode_tid = 1;
1270 		spmp->modify_tid = schain->bref.modify_tid + 1;
1271 
1272 		/*
1273 		 * Sanity-check schain's pmp and finish initialization.
1274 		 * Any chain belonging to the super-root topology should
1275 		 * have a NULL pmp (not even set to spmp).
1276 		 */
1277 		ripdata = &hammer2_chain_rdata(schain)->ipdata;
1278 		KKASSERT(schain->pmp == NULL);
1279 		spmp->pfs_clid = ripdata->meta.pfs_clid;
1280 
1281 		/*
1282 		 * Replace the dummy spmp->iroot with a real one.  It's
1283 		 * easier to just do a wholesale replacement than to try
1284 		 * to update the chain and fixup the iroot fields.
1285 		 *
1286 		 * The returned inode is locked with the supplied cluster.
1287 		 */
1288 		hammer2_dummy_xop_from_chain(&xop, schain);
1289 		hammer2_inode_drop(spmp->iroot);
1290 		spmp->iroot = NULL;
1291 		spmp->iroot = hammer2_inode_get(spmp, NULL, &xop, -1);
1292 		spmp->spmp_hmp = hmp;
1293 		spmp->pfs_types[0] = ripdata->meta.pfs_type;
1294 		spmp->pfs_hmps[0] = hmp;
1295 		hammer2_inode_ref(spmp->iroot);
1296 		hammer2_inode_unlock(spmp->iroot);
1297 		hammer2_cluster_unlock(&xop.cluster);
1298 		hammer2_chain_drop(schain);
1299 		/* do not call hammer2_cluster_drop() on an embedded cluster */
1300 		schain = NULL;	/* now invalid */
1301 		/* leave spmp->iroot with one ref */
1302 
1303 		if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1304 			error = hammer2_recovery(hmp);
1305 			if (error == 0)
1306 				error |= hammer2_fixup_pfses(hmp);
1307 			/* XXX do something with error */
1308 		}
1309 		hammer2_update_pmps(hmp);
1310 		hammer2_iocom_init(hmp);
1311 		hammer2_bulkfree_init(hmp);
1312 
1313 		/*
1314 		 * Ref the cluster management messaging descriptor.  The mount
1315 		 * program deals with the other end of the communications pipe.
1316 		 *
1317 		 * Root mounts typically do not supply one.
1318 		 */
1319 		if (info.cluster_fd >= 0) {
1320 			fp = holdfp(curthread, info.cluster_fd, -1);
1321 			if (fp) {
1322 				hammer2_cluster_reconnect(hmp, fp);
1323 			} else {
1324 				kprintf("hammer2_mount: bad cluster_fd!\n");
1325 			}
1326 		}
1327 	} else {
1328 		spmp = hmp->spmp;
1329 		if (info.hflags & HMNT2_DEVFLAGS) {
1330 			kprintf("hammer2: Warning: mount flags pertaining "
1331 				"to the whole device may only be specified "
1332 				"on the first mount of the device: %08x\n",
1333 				info.hflags & HMNT2_DEVFLAGS);
1334 		}
1335 	}
1336 
1337 	/*
1338 	 * Force local mount (disassociate all PFSs from their clusters).
1339 	 * Used primarily for debugging.
1340 	 */
1341 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1342 
1343 	/*
1344 	 * Lookup the mount point under the media-localized super-root.
1345 	 * Scanning hammer2_pfslist doesn't help us because it represents
1346 	 * PFS cluster ids which can aggregate several named PFSs together.
1347 	 *
1348 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1349 	 * up later on.
1350 	 */
1351 	hammer2_inode_lock(spmp->iroot, 0);
1352 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1353 	lhc = hammer2_dirhash(label, strlen(label));
1354 	chain = hammer2_chain_lookup(&parent, &key_next,
1355 				     lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1356 				     &error, 0);
1357 	while (chain) {
1358 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1359 		    strcmp(label, chain->data->ipdata.filename) == 0) {
1360 			break;
1361 		}
1362 		chain = hammer2_chain_next(&parent, chain, &key_next,
1363 					    key_next,
1364 					    lhc + HAMMER2_DIRHASH_LOMASK,
1365 					    &error, 0);
1366 	}
1367 	if (parent) {
1368 		hammer2_chain_unlock(parent);
1369 		hammer2_chain_drop(parent);
1370 	}
1371 	hammer2_inode_unlock(spmp->iroot);
1372 
1373 	/*
1374 	 * PFS could not be found?
1375 	 */
1376 	if (chain == NULL) {
1377 		if (error)
1378 			kprintf("hammer2_mount: PFS label I/O error\n");
1379 		else
1380 			kprintf("hammer2_mount: PFS label not found\n");
1381 		hammer2_unmount_helper(mp, NULL, hmp);
1382 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1383 		hammer2_vfs_unmount(mp, MNT_FORCE);
1384 
1385 		return EINVAL;
1386 	}
1387 
1388 	/*
1389 	 * Acquire the pmp structure (it should have already been allocated
1390 	 * via hammer2_update_pmps() so do not pass cluster in to add to
1391 	 * available chains).
1392 	 *
1393 	 * Check if the cluster has already been mounted.  A cluster can
1394 	 * only be mounted once, use null mounts to mount additional copies.
1395 	 */
1396 	if (chain->error) {
1397 		kprintf("hammer2_mount: PFS label I/O error\n");
1398 	} else {
1399 		ripdata = &chain->data->ipdata;
1400 		bref = chain->bref;
1401 		pmp = hammer2_pfsalloc(NULL, ripdata,
1402 				       bref.modify_tid, force_local);
1403 	}
1404 	hammer2_chain_unlock(chain);
1405 	hammer2_chain_drop(chain);
1406 
1407 	/*
1408 	 * Finish the mount
1409 	 */
1410         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1411 
1412 	if (pmp->mp) {
1413 		kprintf("hammer2_mount: PFS already mounted!\n");
1414 		hammer2_unmount_helper(mp, NULL, hmp);
1415 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1416 		hammer2_vfs_unmount(mp, MNT_FORCE);
1417 
1418 		return EBUSY;
1419 	}
1420 
1421 	pmp->hflags = info.hflags;
1422         mp->mnt_flag |= MNT_LOCAL;
1423         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1424         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1425 
1426         /*
1427          * required mount structure initializations
1428          */
1429         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1430         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1431 
1432         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1433         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1434 
1435         /*
1436          * Optional fields
1437          */
1438         mp->mnt_iosize_max = MAXPHYS;
1439 
1440 	/*
1441 	 * Connect up mount pointers.
1442 	 */
1443 	hammer2_mount_helper(mp, pmp);
1444 
1445         lockmgr(&hammer2_mntlk, LK_RELEASE);
1446 
1447 	/*
1448 	 * Finish setup
1449 	 */
1450 	vfs_getnewfsid(mp);
1451 	vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1452 	vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1453 	vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1454 
1455 	if (path) {
1456 		copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1457 			  MNAMELEN - 1, &size);
1458 		bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1459 	} /* else root mount, already in there */
1460 
1461 	bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1462 	if (path) {
1463 		copyinstr(path, mp->mnt_stat.f_mntonname,
1464 			  sizeof(mp->mnt_stat.f_mntonname) - 1,
1465 			  &size);
1466 	} else {
1467 		/* root mount */
1468 		mp->mnt_stat.f_mntonname[0] = '/';
1469 	}
1470 
1471 	/*
1472 	 * Initial statfs to prime mnt_stat.
1473 	 */
1474 	hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1475 
1476 	return 0;
1477 }
1478 
1479 /*
1480  * Scan PFSs under the super-root and create hammer2_pfs structures.
1481  */
1482 static
1483 void
1484 hammer2_update_pmps(hammer2_dev_t *hmp)
1485 {
1486 	const hammer2_inode_data_t *ripdata;
1487 	hammer2_chain_t *parent;
1488 	hammer2_chain_t *chain;
1489 	hammer2_blockref_t bref;
1490 	hammer2_dev_t *force_local;
1491 	hammer2_pfs_t *spmp;
1492 	hammer2_pfs_t *pmp;
1493 	hammer2_key_t key_next;
1494 	int error;
1495 
1496 	/*
1497 	 * Force local mount (disassociate all PFSs from their clusters).
1498 	 * Used primarily for debugging.
1499 	 */
1500 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1501 
1502 	/*
1503 	 * Lookup mount point under the media-localized super-root.
1504 	 *
1505 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1506 	 * up later on.
1507 	 */
1508 	spmp = hmp->spmp;
1509 	hammer2_inode_lock(spmp->iroot, 0);
1510 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1511 	chain = hammer2_chain_lookup(&parent, &key_next,
1512 					 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1513 					 &error, 0);
1514 	while (chain) {
1515 		if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1516 			continue;
1517 		if (chain->error) {
1518 			kprintf("I/O error scanning PFS labels\n");
1519 		} else {
1520 			ripdata = &chain->data->ipdata;
1521 			bref = chain->bref;
1522 
1523 			pmp = hammer2_pfsalloc(chain, ripdata,
1524 					       bref.modify_tid, force_local);
1525 		}
1526 		chain = hammer2_chain_next(&parent, chain, &key_next,
1527 					   key_next, HAMMER2_KEY_MAX,
1528 					   &error, 0);
1529 	}
1530 	if (parent) {
1531 		hammer2_chain_unlock(parent);
1532 		hammer2_chain_drop(parent);
1533 	}
1534 	hammer2_inode_unlock(spmp->iroot);
1535 }
1536 
1537 static
1538 int
1539 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1540 		struct vnode *devvp, struct ucred *cred)
1541 {
1542 	int error;
1543 
1544 	if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1545 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1546 		VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1547 		vn_unlock(devvp);
1548 		error = hammer2_recovery(hmp);
1549 		if (error == 0)
1550 			error |= hammer2_fixup_pfses(hmp);
1551 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1552 		if (error == 0) {
1553 			VOP_CLOSE(devvp, FREAD, NULL);
1554 			hmp->ronly = 0;
1555 		} else {
1556 			VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1557 		}
1558 		vn_unlock(devvp);
1559 	} else {
1560 		error = 0;
1561 	}
1562 	return error;
1563 }
1564 
1565 static
1566 int
1567 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1568 {
1569 	hammer2_pfs_t *pmp;
1570 	int flags;
1571 	int error = 0;
1572 
1573 	pmp = MPTOPMP(mp);
1574 
1575 	if (pmp == NULL)
1576 		return(0);
1577 
1578 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1579 
1580 	/*
1581 	 * If mount initialization proceeded far enough we must flush
1582 	 * its vnodes and sync the underlying mount points.  Three syncs
1583 	 * are required to fully flush the filesystem (freemap updates lag
1584 	 * by one flush, and one extra for safety).
1585 	 */
1586 	if (mntflags & MNT_FORCE)
1587 		flags = FORCECLOSE;
1588 	else
1589 		flags = 0;
1590 	if (pmp->iroot) {
1591 		error = vflush(mp, 0, flags);
1592 		if (error)
1593 			goto failed;
1594 		hammer2_vfs_sync(mp, MNT_WAIT);
1595 		hammer2_vfs_sync(mp, MNT_WAIT);
1596 		hammer2_vfs_sync(mp, MNT_WAIT);
1597 	}
1598 
1599 	/*
1600 	 * Cleanup the frontend support XOPS threads
1601 	 */
1602 	hammer2_xop_helper_cleanup(pmp);
1603 
1604 	if (pmp->mp)
1605 		hammer2_unmount_helper(mp, pmp, NULL);
1606 
1607 	error = 0;
1608 failed:
1609 	lockmgr(&hammer2_mntlk, LK_RELEASE);
1610 
1611 	return (error);
1612 }
1613 
1614 /*
1615  * Mount helper, hook the system mount into our PFS.
1616  * The mount lock is held.
1617  *
1618  * We must bump the mount_count on related devices for any
1619  * mounted PFSs.
1620  */
1621 static
1622 void
1623 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1624 {
1625 	hammer2_cluster_t *cluster;
1626 	hammer2_chain_t *rchain;
1627 	int i;
1628 
1629         mp->mnt_data = (qaddr_t)pmp;
1630 	pmp->mp = mp;
1631 
1632 	/*
1633 	 * After pmp->mp is set we have to adjust hmp->mount_count.
1634 	 */
1635 	cluster = &pmp->iroot->cluster;
1636 	for (i = 0; i < cluster->nchains; ++i) {
1637 		rchain = cluster->array[i].chain;
1638 		if (rchain == NULL)
1639 			continue;
1640 		++rchain->hmp->mount_count;
1641 	}
1642 
1643 	/*
1644 	 * Create missing Xop threads
1645 	 */
1646 	hammer2_xop_helper_create(pmp);
1647 }
1648 
1649 /*
1650  * Mount helper, unhook the system mount from our PFS.
1651  * The mount lock is held.
1652  *
1653  * If hmp is supplied a mount responsible for being the first to open
1654  * the block device failed and the block device and all PFSs using the
1655  * block device must be cleaned up.
1656  *
1657  * If pmp is supplied multiple devices might be backing the PFS and each
1658  * must be disconnected.  This might not be the last PFS using some of the
1659  * underlying devices.  Also, we have to adjust our hmp->mount_count
1660  * accounting for the devices backing the pmp which is now undergoing an
1661  * unmount.
1662  */
1663 static
1664 void
1665 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1666 {
1667 	hammer2_cluster_t *cluster;
1668 	hammer2_chain_t *rchain;
1669 	struct vnode *devvp;
1670 	int dumpcnt;
1671 	int ronly;
1672 	int i;
1673 
1674 	/*
1675 	 * If no device supplied this is a high-level unmount and we have to
1676 	 * to disconnect the mount, adjust mount_count, and locate devices
1677 	 * that might now have no mounts.
1678 	 */
1679 	if (pmp) {
1680 		KKASSERT(hmp == NULL);
1681 		KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1682 		pmp->mp = NULL;
1683 		mp->mnt_data = NULL;
1684 
1685 		/*
1686 		 * After pmp->mp is cleared we have to account for
1687 		 * mount_count.
1688 		 */
1689 		cluster = &pmp->iroot->cluster;
1690 		for (i = 0; i < cluster->nchains; ++i) {
1691 			rchain = cluster->array[i].chain;
1692 			if (rchain == NULL)
1693 				continue;
1694 			--rchain->hmp->mount_count;
1695 			/* scrapping hmp now may invalidate the pmp */
1696 		}
1697 again:
1698 		TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1699 			if (hmp->mount_count == 0) {
1700 				hammer2_unmount_helper(NULL, NULL, hmp);
1701 				goto again;
1702 			}
1703 		}
1704 		return;
1705 	}
1706 
1707 	/*
1708 	 * Try to terminate the block device.  We can't terminate it if
1709 	 * there are still PFSs referencing it.
1710 	 */
1711 	if (hmp->mount_count)
1712 		return;
1713 
1714 	/*
1715 	 * Decomission the network before we start messing with the
1716 	 * device and PFS.
1717 	 */
1718 	hammer2_iocom_uninit(hmp);
1719 
1720 	hammer2_bulkfree_uninit(hmp);
1721 	hammer2_pfsfree_scan(hmp, 0);
1722 	hammer2_dev_exlock(hmp);	/* XXX order */
1723 
1724 	/*
1725 	 * Cycle the volume data lock as a safety (probably not needed any
1726 	 * more).  To ensure everything is out we need to flush at least
1727 	 * three times.  (1) The running of the sideq can dirty the
1728 	 * filesystem, (2) A normal flush can dirty the freemap, and
1729 	 * (3) ensure that the freemap is fully synchronized.
1730 	 *
1731 	 * The next mount's recovery scan can clean everything up but we want
1732 	 * to leave the filesystem in a 100% clean state on a normal unmount.
1733 	 */
1734 #if 0
1735 	hammer2_voldata_lock(hmp);
1736 	hammer2_voldata_unlock(hmp);
1737 #endif
1738 
1739 	/*
1740 	 * Flush whatever is left.  Unmounted but modified PFS's might still
1741 	 * have some dirty chains on them.
1742 	 */
1743 	hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1744 	hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1745 
1746 	if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1747 		hammer2_voldata_modify(hmp);
1748 		hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1749 					    HAMMER2_FLUSH_ALL);
1750 	}
1751 	hammer2_chain_unlock(&hmp->fchain);
1752 
1753 	if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1754 		hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1755 					    HAMMER2_FLUSH_ALL);
1756 	}
1757 	hammer2_chain_unlock(&hmp->vchain);
1758 
1759 	if ((hmp->vchain.flags | hmp->fchain.flags) &
1760 	    HAMMER2_CHAIN_FLUSH_MASK) {
1761 		kprintf("hammer2_unmount: chains left over "
1762 			"after final sync\n");
1763 		kprintf("    vchain %08x\n", hmp->vchain.flags);
1764 		kprintf("    fchain %08x\n", hmp->fchain.flags);
1765 
1766 		if (hammer2_debug & 0x0010)
1767 			Debugger("entered debugger");
1768 	}
1769 
1770 	hammer2_pfsfree_scan(hmp, 1);
1771 
1772 	KKASSERT(hmp->spmp == NULL);
1773 
1774 	/*
1775 	 * Finish up with the device vnode
1776 	 */
1777 	if ((devvp = hmp->devvp) != NULL) {
1778 		ronly = hmp->ronly;
1779 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1780 		kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1781 			hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1782 			ronly);
1783 		vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1784 		kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1785 			hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1786 		hmp->devvp = NULL;
1787 		VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1788 		vn_unlock(devvp);
1789 		vrele(devvp);
1790 		devvp = NULL;
1791 	}
1792 
1793 	/*
1794 	 * Clear vchain/fchain flags that might prevent final cleanup
1795 	 * of these chains.
1796 	 */
1797 	if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1798 		atomic_add_long(&hammer2_count_modified_chains, -1);
1799 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1800 		hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1801 	}
1802 	if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1803 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1804 	}
1805 
1806 	if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1807 		atomic_add_long(&hammer2_count_modified_chains, -1);
1808 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1809 		hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1810 	}
1811 	if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1812 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1813 	}
1814 
1815 	/*
1816 	 * Final drop of embedded freemap root chain to
1817 	 * clean up fchain.core (fchain structure is not
1818 	 * flagged ALLOCATED so it is cleaned out and then
1819 	 * left to rot).
1820 	 */
1821 	hammer2_chain_drop(&hmp->fchain);
1822 
1823 	/*
1824 	 * Final drop of embedded volume root chain to clean
1825 	 * up vchain.core (vchain structure is not flagged
1826 	 * ALLOCATED so it is cleaned out and then left to
1827 	 * rot).
1828 	 */
1829 	dumpcnt = 50;
1830 	hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1831 	dumpcnt = 50;
1832 	hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1833 	hammer2_dev_unlock(hmp);
1834 	hammer2_chain_drop(&hmp->vchain);
1835 
1836 	hammer2_io_cleanup(hmp, &hmp->iotree);
1837 	if (hmp->iofree_count) {
1838 		kprintf("io_cleanup: %d I/O's left hanging\n",
1839 			hmp->iofree_count);
1840 	}
1841 
1842 	TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1843 	kmalloc_destroy(&hmp->mchain);
1844 	kfree(hmp, M_HAMMER2);
1845 }
1846 
1847 int
1848 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1849 		 ino_t ino, struct vnode **vpp)
1850 {
1851 	hammer2_xop_lookup_t *xop;
1852 	hammer2_pfs_t *pmp;
1853 	hammer2_inode_t *ip;
1854 	hammer2_tid_t inum;
1855 	int error;
1856 
1857 	inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1858 
1859 	error = 0;
1860 	pmp = MPTOPMP(mp);
1861 
1862 	/*
1863 	 * Easy if we already have it cached
1864 	 */
1865 	ip = hammer2_inode_lookup(pmp, inum);
1866 	if (ip) {
1867 		hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1868 		*vpp = hammer2_igetv(ip, &error);
1869 		hammer2_inode_unlock(ip);
1870 		hammer2_inode_drop(ip);		/* from lookup */
1871 
1872 		return error;
1873 	}
1874 
1875 	/*
1876 	 * Otherwise we have to find the inode
1877 	 */
1878 	xop = hammer2_xop_alloc(pmp->iroot, 0);
1879 	xop->lhc = inum;
1880 	hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1881 	error = hammer2_xop_collect(&xop->head, 0);
1882 
1883 	if (error == 0)
1884 		ip = hammer2_inode_get(pmp, NULL, &xop->head, -1);
1885 	hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1886 
1887 	if (ip) {
1888 		*vpp = hammer2_igetv(ip, &error);
1889 		hammer2_inode_unlock(ip);
1890 	} else {
1891 		*vpp = NULL;
1892 		error = ENOENT;
1893 	}
1894 	return (error);
1895 }
1896 
1897 static
1898 int
1899 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1900 {
1901 	hammer2_pfs_t *pmp;
1902 	struct vnode *vp;
1903 	int error;
1904 
1905 	pmp = MPTOPMP(mp);
1906 	if (pmp->iroot == NULL) {
1907 		kprintf("hammer2 (%s): no root inode\n",
1908 			mp->mnt_stat.f_mntfromname);
1909 		*vpp = NULL;
1910 		return EINVAL;
1911 	}
1912 
1913 	error = 0;
1914 	hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1915 
1916 	while (pmp->inode_tid == 0) {
1917 		hammer2_xop_ipcluster_t *xop;
1918 		const hammer2_inode_meta_t *meta;
1919 
1920 		xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1921 		hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1922 		error = hammer2_xop_collect(&xop->head, 0);
1923 
1924 		if (error == 0) {
1925 			meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1926 			pmp->iroot->meta = *meta;
1927 			pmp->inode_tid = meta->pfs_inum + 1;
1928 			hammer2_xop_pdata(&xop->head);
1929 			/* meta invalid */
1930 
1931 			if (pmp->inode_tid < HAMMER2_INODE_START)
1932 				pmp->inode_tid = HAMMER2_INODE_START;
1933 			pmp->modify_tid =
1934 				xop->head.cluster.focus->bref.modify_tid + 1;
1935 #if 0
1936 			kprintf("PFS: Starting inode %jd\n",
1937 				(intmax_t)pmp->inode_tid);
1938 			kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1939 				pmp->inode_tid, pmp->modify_tid);
1940 #endif
1941 			wakeup(&pmp->iroot);
1942 
1943 			hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1944 
1945 			/*
1946 			 * Prime the mount info.
1947 			 */
1948 			hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1949 			break;
1950 		}
1951 
1952 		/*
1953 		 * Loop, try again
1954 		 */
1955 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1956 		hammer2_inode_unlock(pmp->iroot);
1957 		error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1958 		hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1959 		if (error == EINTR)
1960 			break;
1961 	}
1962 
1963 	if (error) {
1964 		hammer2_inode_unlock(pmp->iroot);
1965 		*vpp = NULL;
1966 	} else {
1967 		vp = hammer2_igetv(pmp->iroot, &error);
1968 		hammer2_inode_unlock(pmp->iroot);
1969 		*vpp = vp;
1970 	}
1971 
1972 	return (error);
1973 }
1974 
1975 /*
1976  * Filesystem status
1977  *
1978  * XXX incorporate ipdata->meta.inode_quota and data_quota
1979  */
1980 static
1981 int
1982 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1983 {
1984 	hammer2_pfs_t *pmp;
1985 	hammer2_dev_t *hmp;
1986 	hammer2_blockref_t bref;
1987 	struct statfs tmp;
1988 	int i;
1989 
1990 	/*
1991 	 * NOTE: iroot might not have validated the cluster yet.
1992 	 */
1993 	pmp = MPTOPMP(mp);
1994 
1995 	bzero(&tmp, sizeof(tmp));
1996 
1997 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1998 		hmp = pmp->pfs_hmps[i];
1999 		if (hmp == NULL)
2000 			continue;
2001 		if (pmp->iroot->cluster.array[i].chain)
2002 			bref = pmp->iroot->cluster.array[i].chain->bref;
2003 		else
2004 			bzero(&bref, sizeof(bref));
2005 
2006 		tmp.f_files = bref.embed.stats.inode_count;
2007 		tmp.f_ffree = 0;
2008 		tmp.f_blocks = hmp->voldata.allocator_size /
2009 			       mp->mnt_vstat.f_bsize;
2010 		tmp.f_bfree = hmp->voldata.allocator_free /
2011 			      mp->mnt_vstat.f_bsize;
2012 		tmp.f_bavail = tmp.f_bfree;
2013 
2014 		if (cred && cred->cr_uid != 0) {
2015 			uint64_t adj;
2016 
2017 			/* 5% */
2018 			adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2019 			tmp.f_blocks -= adj;
2020 			tmp.f_bfree -= adj;
2021 			tmp.f_bavail -= adj;
2022 		}
2023 
2024 		mp->mnt_stat.f_blocks = tmp.f_blocks;
2025 		mp->mnt_stat.f_bfree = tmp.f_bfree;
2026 		mp->mnt_stat.f_bavail = tmp.f_bavail;
2027 		mp->mnt_stat.f_files = tmp.f_files;
2028 		mp->mnt_stat.f_ffree = tmp.f_ffree;
2029 
2030 		*sbp = mp->mnt_stat;
2031 	}
2032 	return (0);
2033 }
2034 
2035 static
2036 int
2037 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2038 {
2039 	hammer2_pfs_t *pmp;
2040 	hammer2_dev_t *hmp;
2041 	hammer2_blockref_t bref;
2042 	struct statvfs tmp;
2043 	int i;
2044 
2045 	/*
2046 	 * NOTE: iroot might not have validated the cluster yet.
2047 	 */
2048 	pmp = MPTOPMP(mp);
2049 	bzero(&tmp, sizeof(tmp));
2050 
2051 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2052 		hmp = pmp->pfs_hmps[i];
2053 		if (hmp == NULL)
2054 			continue;
2055 		if (pmp->iroot->cluster.array[i].chain)
2056 			bref = pmp->iroot->cluster.array[i].chain->bref;
2057 		else
2058 			bzero(&bref, sizeof(bref));
2059 
2060 		tmp.f_files = bref.embed.stats.inode_count;
2061 		tmp.f_ffree = 0;
2062 		tmp.f_blocks = hmp->voldata.allocator_size /
2063 			       mp->mnt_vstat.f_bsize;
2064 		tmp.f_bfree = hmp->voldata.allocator_free /
2065 			      mp->mnt_vstat.f_bsize;
2066 		tmp.f_bavail = tmp.f_bfree;
2067 
2068 		if (cred && cred->cr_uid != 0) {
2069 			uint64_t adj;
2070 
2071 			/* 5% */
2072 			adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2073 			tmp.f_blocks -= adj;
2074 			tmp.f_bfree -= adj;
2075 			tmp.f_bavail -= adj;
2076 		}
2077 
2078 		mp->mnt_vstat.f_blocks = tmp.f_blocks;
2079 		mp->mnt_vstat.f_bfree = tmp.f_bfree;
2080 		mp->mnt_vstat.f_bavail = tmp.f_bavail;
2081 		mp->mnt_vstat.f_files = tmp.f_files;
2082 		mp->mnt_vstat.f_ffree = tmp.f_ffree;
2083 
2084 		*sbp = mp->mnt_vstat;
2085 	}
2086 	return (0);
2087 }
2088 
2089 /*
2090  * Mount-time recovery (RW mounts)
2091  *
2092  * Updates to the free block table are allowed to lag flushes by one
2093  * transaction.  In case of a crash, then on a fresh mount we must do an
2094  * incremental scan of the last committed transaction id and make sure that
2095  * all related blocks have been marked allocated.
2096  *
2097  * The super-root topology and each PFS has its own transaction id domain,
2098  * so we must track PFS boundary transitions.
2099  */
2100 struct hammer2_recovery_elm {
2101 	TAILQ_ENTRY(hammer2_recovery_elm) entry;
2102 	hammer2_chain_t *chain;
2103 	hammer2_tid_t sync_tid;
2104 };
2105 
2106 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2107 
2108 struct hammer2_recovery_info {
2109 	struct hammer2_recovery_list list;
2110 	hammer2_tid_t	mtid;
2111 	int	depth;
2112 };
2113 
2114 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2115 			hammer2_chain_t *parent,
2116 			struct hammer2_recovery_info *info,
2117 			hammer2_tid_t sync_tid);
2118 
2119 #define HAMMER2_RECOVERY_MAXDEPTH	10
2120 
2121 static
2122 int
2123 hammer2_recovery(hammer2_dev_t *hmp)
2124 {
2125 	struct hammer2_recovery_info info;
2126 	struct hammer2_recovery_elm *elm;
2127 	hammer2_chain_t *parent;
2128 	hammer2_tid_t sync_tid;
2129 	hammer2_tid_t mirror_tid;
2130 	int error;
2131 
2132 	hammer2_trans_init(hmp->spmp, 0);
2133 
2134 	sync_tid = hmp->voldata.freemap_tid;
2135 	mirror_tid = hmp->voldata.mirror_tid;
2136 
2137 	kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2138 	if (sync_tid >= mirror_tid) {
2139 		kprintf(" no recovery needed\n");
2140 	} else {
2141 		kprintf(" freemap recovery %016jx-%016jx\n",
2142 			sync_tid + 1, mirror_tid);
2143 	}
2144 
2145 	TAILQ_INIT(&info.list);
2146 	info.depth = 0;
2147 	parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2148 	error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2149 	hammer2_chain_lookup_done(parent);
2150 
2151 	while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2152 		TAILQ_REMOVE(&info.list, elm, entry);
2153 		parent = elm->chain;
2154 		sync_tid = elm->sync_tid;
2155 		kfree(elm, M_HAMMER2);
2156 
2157 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2158 		error |= hammer2_recovery_scan(hmp, parent, &info,
2159 					      hmp->voldata.freemap_tid);
2160 		hammer2_chain_unlock(parent);
2161 		hammer2_chain_drop(parent);	/* drop elm->chain ref */
2162 	}
2163 
2164 	hammer2_trans_done(hmp->spmp, 0);
2165 
2166 	return error;
2167 }
2168 
2169 static
2170 int
2171 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2172 		      struct hammer2_recovery_info *info,
2173 		      hammer2_tid_t sync_tid)
2174 {
2175 	const hammer2_inode_data_t *ripdata;
2176 	hammer2_chain_t *chain;
2177 	hammer2_blockref_t bref;
2178 	int tmp_error;
2179 	int rup_error;
2180 	int error;
2181 	int first;
2182 
2183 	/*
2184 	 * Adjust freemap to ensure that the block(s) are marked allocated.
2185 	 */
2186 	if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2187 		hammer2_freemap_adjust(hmp, &parent->bref,
2188 				       HAMMER2_FREEMAP_DORECOVER);
2189 	}
2190 
2191 	/*
2192 	 * Check type for recursive scan
2193 	 */
2194 	switch(parent->bref.type) {
2195 	case HAMMER2_BREF_TYPE_VOLUME:
2196 		/* data already instantiated */
2197 		break;
2198 	case HAMMER2_BREF_TYPE_INODE:
2199 		/*
2200 		 * Must instantiate data for DIRECTDATA test and also
2201 		 * for recursion.
2202 		 */
2203 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2204 		ripdata = &hammer2_chain_rdata(parent)->ipdata;
2205 		if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2206 			/* not applicable to recovery scan */
2207 			hammer2_chain_unlock(parent);
2208 			return 0;
2209 		}
2210 		hammer2_chain_unlock(parent);
2211 		break;
2212 	case HAMMER2_BREF_TYPE_INDIRECT:
2213 		/*
2214 		 * Must instantiate data for recursion
2215 		 */
2216 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2217 		hammer2_chain_unlock(parent);
2218 		break;
2219 	case HAMMER2_BREF_TYPE_DIRENT:
2220 	case HAMMER2_BREF_TYPE_DATA:
2221 	case HAMMER2_BREF_TYPE_FREEMAP:
2222 	case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2223 	case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2224 		/* not applicable to recovery scan */
2225 		return 0;
2226 		break;
2227 	default:
2228 		return HAMMER2_ERROR_BADBREF;
2229 	}
2230 
2231 	/*
2232 	 * Defer operation if depth limit reached or if we are crossing a
2233 	 * PFS boundary.
2234 	 */
2235 	if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2236 		struct hammer2_recovery_elm *elm;
2237 
2238 		elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2239 		elm->chain = parent;
2240 		elm->sync_tid = sync_tid;
2241 		hammer2_chain_ref(parent);
2242 		TAILQ_INSERT_TAIL(&info->list, elm, entry);
2243 		/* unlocked by caller */
2244 
2245 		return(0);
2246 	}
2247 
2248 
2249 	/*
2250 	 * Recursive scan of the last flushed transaction only.  We are
2251 	 * doing this without pmp assignments so don't leave the chains
2252 	 * hanging around after we are done with them.
2253 	 *
2254 	 * error	Cumulative error this level only
2255 	 * rup_error	Cumulative error for recursion
2256 	 * tmp_error	Specific non-cumulative recursion error
2257 	 */
2258 	chain = NULL;
2259 	first = 1;
2260 	rup_error = 0;
2261 	error = 0;
2262 
2263 	for (;;) {
2264 		error |= hammer2_chain_scan(parent, &chain, &bref,
2265 					    &first,
2266 					    HAMMER2_LOOKUP_NODATA);
2267 
2268 		/*
2269 		 * Problem during scan or EOF
2270 		 */
2271 		if (error)
2272 			break;
2273 
2274 		/*
2275 		 * If this is a leaf
2276 		 */
2277 		if (chain == NULL) {
2278 			if (bref.mirror_tid > sync_tid) {
2279 				hammer2_freemap_adjust(hmp, &bref,
2280 						     HAMMER2_FREEMAP_DORECOVER);
2281 			}
2282 			continue;
2283 		}
2284 
2285 		/*
2286 		 * This may or may not be a recursive node.
2287 		 */
2288 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2289 		if (bref.mirror_tid > sync_tid) {
2290 			++info->depth;
2291 			tmp_error = hammer2_recovery_scan(hmp, chain,
2292 							   info, sync_tid);
2293 			--info->depth;
2294 		} else {
2295 			tmp_error = 0;
2296 		}
2297 
2298 		/*
2299 		 * Flush the recovery at the PFS boundary to stage it for
2300 		 * the final flush of the super-root topology.
2301 		 */
2302 		if (tmp_error == 0 &&
2303 		    (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2304 		    (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2305 			hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2306 					     HAMMER2_FLUSH_ALL);
2307 		}
2308 		rup_error |= tmp_error;
2309 	}
2310 	return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2311 }
2312 
2313 /*
2314  * This fixes up an error introduced in earlier H2 implementations where
2315  * moving a PFS inode into an indirect block wound up causing the
2316  * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2317  */
2318 static
2319 int
2320 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2321 {
2322 	const hammer2_inode_data_t *ripdata;
2323 	hammer2_chain_t *parent;
2324 	hammer2_chain_t *chain;
2325 	hammer2_key_t key_next;
2326 	hammer2_pfs_t *spmp;
2327 	int error;
2328 
2329 	error = 0;
2330 
2331 	/*
2332 	 * Lookup mount point under the media-localized super-root.
2333 	 *
2334 	 * cluster->pmp will incorrectly point to spmp and must be fixed
2335 	 * up later on.
2336 	 */
2337 	spmp = hmp->spmp;
2338 	hammer2_inode_lock(spmp->iroot, 0);
2339 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2340 	chain = hammer2_chain_lookup(&parent, &key_next,
2341 					 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2342 					 &error, 0);
2343 	while (chain) {
2344 		if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2345 			continue;
2346 		if (chain->error) {
2347 			kprintf("I/O error scanning PFS labels\n");
2348 			error |= chain->error;
2349 		} else if ((chain->bref.flags &
2350 			    HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2351 			int error2;
2352 
2353 			ripdata = &chain->data->ipdata;
2354 			hammer2_trans_init(hmp->spmp, 0);
2355 			error2 = hammer2_chain_modify(chain,
2356 						      chain->bref.modify_tid,
2357 						      0, 0);
2358 			if (error2 == 0) {
2359 				kprintf("hammer2: Correct mis-flagged PFS %s\n",
2360 					ripdata->filename);
2361 				chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2362 			} else {
2363 				error |= error2;
2364 			}
2365 			hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2366 					     HAMMER2_FLUSH_ALL);
2367 			hammer2_trans_done(hmp->spmp, 0);
2368 		}
2369 		chain = hammer2_chain_next(&parent, chain, &key_next,
2370 					   key_next, HAMMER2_KEY_MAX,
2371 					   &error, 0);
2372 	}
2373 	if (parent) {
2374 		hammer2_chain_unlock(parent);
2375 		hammer2_chain_drop(parent);
2376 	}
2377 	hammer2_inode_unlock(spmp->iroot);
2378 
2379 	return error;
2380 }
2381 
2382 /*
2383  * Sync a mount point; this is called periodically on a per-mount basis from
2384  * the filesystem syncer, and whenever a user issues a sync.
2385  */
2386 int
2387 hammer2_vfs_sync(struct mount *mp, int waitfor)
2388 {
2389 	hammer2_xop_flush_t *xop;
2390 	struct hammer2_sync_info info;
2391 	hammer2_inode_t *iroot;
2392 	hammer2_pfs_t *pmp;
2393 	int flags;
2394 	int error;
2395 
2396 	pmp = MPTOPMP(mp);
2397 	iroot = pmp->iroot;
2398 	KKASSERT(iroot);
2399 	KKASSERT(iroot->pmp == pmp);
2400 
2401 	/*
2402 	 * We can't acquire locks on existing vnodes while in a transaction
2403 	 * without risking a deadlock.  This assumes that vfsync() can be
2404 	 * called without the vnode locked (which it can in DragonFly).
2405 	 * Otherwise we'd have to implement a multi-pass or flag the lock
2406 	 * failures and retry.
2407 	 *
2408 	 * The reclamation code interlocks with the sync list's token
2409 	 * (by removing the vnode from the scan list) before unlocking
2410 	 * the inode, giving us time to ref the inode.
2411 	 */
2412 	/*flags = VMSC_GETVP;*/
2413 	flags = 0;
2414 	if (waitfor & MNT_LAZY)
2415 		flags |= VMSC_ONEPASS;
2416 
2417 	/*
2418 	 * Flush vnodes individually using a normal transaction to avoid
2419 	 * stalling any concurrent operations.  This will flush the related
2420 	 * buffer cache buffers and inodes to the media.
2421 	 *
2422 	 * For efficiency do an async pass before making sure with a
2423 	 * synchronous pass on all related buffer cache buffers.
2424 	 */
2425 	hammer2_trans_init(pmp, 0);
2426 
2427 	info.error = 0;
2428 
2429 	info.waitfor = MNT_NOWAIT;
2430 	info.pass = 1;
2431 	vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2432 
2433 	/*
2434 	 * Now do two passes making sure we get everything.  The first pass
2435 	 * vfsync()s dirty vnodes.  The second pass waits for their I/O's
2436 	 * to finish and cleans up the dirty flag on the vnode.
2437 	 */
2438 	info.pass = 1;
2439 	info.waitfor = MNT_WAIT;
2440 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2441 
2442 	info.pass = 2;
2443 	info.waitfor = MNT_WAIT;
2444 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2445 
2446 	/*
2447 	 * We must also run the sideq to handle any disconnected inodes
2448 	 * as the vnode scan will not see these.
2449 	 */
2450 	hammer2_inode_run_sideq(pmp, 1);
2451 	hammer2_trans_done(pmp, 0);
2452 
2453 	/*
2454 	 * Start our flush transaction and flush the root topology down to
2455 	 * the inodes, but not the inodes themselves (which we already flushed
2456 	 * above).  Any concurrent activity effecting inode contents will not
2457 	 *
2458 	 * The flush sequence will
2459 	 *
2460 	 * NOTE! It is still possible for the paging code to push pages
2461 	 *	 out via a UIO_NOCOPY hammer2_vop_write() during the main
2462 	 *	 flush.
2463 	 */
2464 	hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2465 
2466 	/*
2467 	 * sync dirty vnodes again while in the flush transaction.  This is
2468 	 * currently an expensive shim to makre sure the logical topology is
2469 	 * completely consistent before we flush the volume header.
2470 	 */
2471 	info.pass = 1;
2472 	info.waitfor = MNT_WAIT;
2473 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2474 
2475 	info.pass = 2;
2476 	info.waitfor = MNT_WAIT;
2477 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2478 
2479 	/*
2480 	 * Use the XOP interface to concurrently flush all nodes to
2481 	 * synchronize the PFSROOT subtopology to the media.  A standard
2482 	 * end-of-scan ENOENT error indicates cluster sufficiency.
2483 	 *
2484 	 * Note that this flush will not be visible on crash recovery until
2485 	 * we flush the super-root topology in the next loop.
2486 	 *
2487 	 * XXX For now wait for all flushes to complete.
2488 	 */
2489 	if (iroot) {
2490 		/*
2491 		 * If unmounting try to flush everything including any
2492 		 * sub-trees under inodes, just in case there is dangling
2493 		 * modified data, as a safety.  Otherwise just flush up to
2494 		 * the inodes in this stage.
2495 		 */
2496 		if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2497 			xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2498 						       HAMMER2_XOP_VOLHDR);
2499 		} else {
2500 			xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2501 						       HAMMER2_XOP_INODE_STOP |
2502 						       HAMMER2_XOP_VOLHDR);
2503 		}
2504 		hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2505 		error = hammer2_xop_collect(&xop->head,
2506 					    HAMMER2_XOP_COLLECT_WAITALL);
2507 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2508 		if (error == HAMMER2_ERROR_ENOENT)
2509 			error = 0;
2510 		else
2511 			error = hammer2_error_to_errno(error);
2512 	} else {
2513 		error = 0;
2514 	}
2515 	hammer2_trans_done(pmp, 0);
2516 
2517 	return (error);
2518 }
2519 
2520 /*
2521  * Sync passes.
2522  *
2523  * Note that we ignore the tranasction mtid we got above.  Instead,
2524  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2525  * transactions.
2526  *
2527  * WARNING! The frontend might be waiting on chnmem (limit_dirty_chains)
2528  * while holding a vnode locked.  When this situation occurs we cannot
2529  * safely test whether it is ok to clear the dirty bit on the vnode.
2530  * However, we can still flush the inode's topology.
2531  */
2532 static int
2533 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2534 {
2535 	struct hammer2_sync_info *info = data;
2536 	hammer2_inode_t *ip;
2537 	int error;
2538 
2539 	/*
2540 	 * Degenerate cases.  Note that ip == NULL typically means the
2541 	 * syncer vnode itself and we don't want to vclrisdirty() in that
2542 	 * situation.
2543 	 */
2544 	ip = VTOI(vp);
2545 	if (ip == NULL) {
2546 		return(0);
2547 	}
2548 	if (vp->v_type == VNON || vp->v_type == VBAD) {
2549 		vclrisdirty(vp);
2550 		return(0);
2551 	}
2552 
2553 	/*
2554 	 * Synchronize the buffer cche and inode meta-data to the backing
2555 	 * chain topology.
2556 	 *
2557 	 * vfsync is not necessarily synchronous, so it is best NOT to try
2558 	 * to flush the backing topology to media at this point.
2559 	 */
2560 	hammer2_inode_ref(ip);
2561 	if ((ip->flags & (HAMMER2_INODE_RESIZED|HAMMER2_INODE_MODIFIED)) ||
2562 	    !RB_EMPTY(&vp->v_rbdirty_tree)) {
2563 		if (info->pass == 1)
2564 			vfsync(vp, info->waitfor, 1, NULL, NULL);
2565 		else
2566 			bio_track_wait(&vp->v_track_write, 0, 0);
2567 	}
2568 	if (info->pass == 2 && (vp->v_flag & VISDIRTY)) {
2569 		/*
2570 		 * v_token is needed to interlock v_rbdirty_tree.
2571 		 */
2572 		lwkt_gettoken(&vp->v_token);
2573 		hammer2_inode_lock(ip, 0);
2574 		hammer2_inode_chain_sync(ip);
2575 		hammer2_inode_chain_flush(ip);
2576 		if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2577 				  HAMMER2_INODE_RESIZED |
2578 				  HAMMER2_INODE_DIRTYDATA)) == 0 &&
2579 		    RB_EMPTY(&vp->v_rbdirty_tree) &&
2580 		    !bio_track_active(&vp->v_track_write)) {
2581 			vclrisdirty(vp);
2582 		}
2583 		hammer2_inode_unlock(ip);
2584 		lwkt_reltoken(&vp->v_token);
2585 	}
2586 	hammer2_inode_drop(ip);
2587 #if 1
2588 	error = 0;
2589 	if (error)
2590 		info->error = error;
2591 #endif
2592 	return(0);
2593 }
2594 
2595 static
2596 int
2597 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2598 {
2599 	hammer2_inode_t *ip;
2600 
2601 	KKASSERT(MAXFIDSZ >= 16);
2602 	ip = VTOI(vp);
2603 	fhp->fid_len = offsetof(struct fid, fid_data[16]);
2604 	fhp->fid_ext = 0;
2605 	((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2606 	((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2607 
2608 	return 0;
2609 }
2610 
2611 static
2612 int
2613 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2614 	       struct fid *fhp, struct vnode **vpp)
2615 {
2616 	hammer2_pfs_t *pmp;
2617 	hammer2_tid_t inum;
2618 	int error;
2619 
2620 	pmp = MPTOPMP(mp);
2621 	inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2622 	if (vpp) {
2623 		if (inum == 1)
2624 			error = hammer2_vfs_root(mp, vpp);
2625 		else
2626 			error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2627 	} else {
2628 		error = 0;
2629 	}
2630 	if (error)
2631 		kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2632 	return error;
2633 }
2634 
2635 static
2636 int
2637 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2638 		 int *exflagsp, struct ucred **credanonp)
2639 {
2640 	hammer2_pfs_t *pmp;
2641 	struct netcred *np;
2642 	int error;
2643 
2644 	pmp = MPTOPMP(mp);
2645 	np = vfs_export_lookup(mp, &pmp->export, nam);
2646 	if (np) {
2647 		*exflagsp = np->netc_exflags;
2648 		*credanonp = &np->netc_anon;
2649 		error = 0;
2650 	} else {
2651 		error = EACCES;
2652 	}
2653 	return error;
2654 }
2655 
2656 /*
2657  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2658  * header into the HMP
2659  *
2660  * XXX read four volhdrs and use the one with the highest TID whos CRC
2661  *     matches.
2662  *
2663  * XXX check iCRCs.
2664  *
2665  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2666  *     nonexistant locations.
2667  *
2668  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2669  */
2670 static
2671 int
2672 hammer2_install_volume_header(hammer2_dev_t *hmp)
2673 {
2674 	hammer2_volume_data_t *vd;
2675 	struct buf *bp;
2676 	hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2677 	int error_reported;
2678 	int error;
2679 	int valid;
2680 	int i;
2681 
2682 	error_reported = 0;
2683 	error = 0;
2684 	valid = 0;
2685 	bp = NULL;
2686 
2687 	/*
2688 	 * There are up to 4 copies of the volume header (syncs iterate
2689 	 * between them so there is no single master).  We don't trust the
2690 	 * volu_size field so we don't know precisely how large the filesystem
2691 	 * is, so depend on the OS to return an error if we go beyond the
2692 	 * block device's EOF.
2693 	 */
2694 	for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2695 		error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2696 			      HAMMER2_VOLUME_BYTES, &bp);
2697 		if (error) {
2698 			brelse(bp);
2699 			bp = NULL;
2700 			continue;
2701 		}
2702 
2703 		vd = (struct hammer2_volume_data *) bp->b_data;
2704 		if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2705 		    (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2706 			brelse(bp);
2707 			bp = NULL;
2708 			continue;
2709 		}
2710 
2711 		if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2712 			/* XXX: Reversed-endianness filesystem */
2713 			kprintf("hammer2: reverse-endian filesystem detected");
2714 			brelse(bp);
2715 			bp = NULL;
2716 			continue;
2717 		}
2718 
2719 		crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2720 		crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2721 				      HAMMER2_VOLUME_ICRC0_SIZE);
2722 		bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2723 		bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2724 				       HAMMER2_VOLUME_ICRC1_SIZE);
2725 		if ((crc0 != crc) || (bcrc0 != bcrc)) {
2726 			kprintf("hammer2 volume header crc "
2727 				"mismatch copy #%d %08x/%08x\n",
2728 				i, crc0, crc);
2729 			error_reported = 1;
2730 			brelse(bp);
2731 			bp = NULL;
2732 			continue;
2733 		}
2734 		if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2735 			valid = 1;
2736 			hmp->voldata = *vd;
2737 			hmp->volhdrno = i;
2738 		}
2739 		brelse(bp);
2740 		bp = NULL;
2741 	}
2742 	if (valid) {
2743 		hmp->volsync = hmp->voldata;
2744 		hmp->free_reserved = hmp->voldata.allocator_size / 20;
2745 		error = 0;
2746 		if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2747 			kprintf("hammer2: using volume header #%d\n",
2748 				hmp->volhdrno);
2749 		}
2750 	} else {
2751 		error = EINVAL;
2752 		kprintf("hammer2: no valid volume headers found!\n");
2753 	}
2754 	return (error);
2755 }
2756 
2757 /*
2758  * This handles hysteresis on regular file flushes.  Because the BIOs are
2759  * routed to a thread it is possible for an excessive number to build up
2760  * and cause long front-end stalls long before the runningbuffspace limit
2761  * is hit, so we implement hammer2_flush_pipe to control the
2762  * hysteresis.
2763  *
2764  * This is a particular problem when compression is used.
2765  */
2766 void
2767 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2768 {
2769 	atomic_add_int(&pmp->count_lwinprog, 1);
2770 }
2771 
2772 void
2773 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2774 {
2775 	int lwinprog;
2776 
2777 	lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2778 	if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2779 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2780 		atomic_clear_int(&pmp->count_lwinprog,
2781 				 HAMMER2_LWINPROG_WAITING);
2782 		wakeup(&pmp->count_lwinprog);
2783 	}
2784 	if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2785 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2786 		atomic_clear_int(&pmp->count_lwinprog,
2787 				 HAMMER2_LWINPROG_WAITING0);
2788 		wakeup(&pmp->count_lwinprog);
2789 	}
2790 }
2791 
2792 void
2793 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2794 {
2795 	int lwinprog;
2796 	int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2797 				    HAMMER2_LWINPROG_WAITING0;
2798 
2799 	for (;;) {
2800 		lwinprog = pmp->count_lwinprog;
2801 		cpu_ccfence();
2802 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2803 			break;
2804 		tsleep_interlock(&pmp->count_lwinprog, 0);
2805 		atomic_set_int(&pmp->count_lwinprog, lwflag);
2806 		lwinprog = pmp->count_lwinprog;
2807 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2808 			break;
2809 		tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2810 	}
2811 }
2812 
2813 /*
2814  * Attempt to proactively fsync dirty vnodes if we have too many.  This
2815  * solves an issue where the kernel syncer thread can get seriously behind
2816  * when multiple user processes/threads are furiously modifying inodes.
2817  * This situation can occur on slow storage and is only limited by
2818  * kern.maxvnodes without the moderation code below.  It is made worse
2819  * when the device buffers underlying the modified inodes (which are clean)
2820  * get evicted before the flush can occur, forcing a re-read.
2821  *
2822  * We do not want sysads to feel that they have to torpedo kern.maxvnodes
2823  * to solve this problem, so we implement vfs.hammer2.limit_dirty_inodes
2824  * (per-mount-basis) and default it to something reasonable.
2825  */
2826 static void
2827 hammer2_pfs_moderate(hammer2_inode_t *ip, int always_moderate)
2828 {
2829 	hammer2_pfs_t *pmp = ip->pmp;
2830 	struct mount *mp = pmp->mp;
2831 
2832 	if (mp && vn_syncer_count(mp) > hammer2_limit_dirty_inodes) {
2833 		vn_syncer_one(mp);
2834 	}
2835 }
2836 
2837 /*
2838  * Manage excessive memory resource use for chain and related
2839  * structures.
2840  *
2841  * Called without any inode locks or transaction locks.  VNodes
2842  * might be locked by the kernel in the call stack.
2843  */
2844 void
2845 hammer2_pfs_memory_wait(hammer2_inode_t *ip, int always_moderate)
2846 {
2847 	hammer2_pfs_t *pmp = ip->pmp;
2848 	uint32_t waiting;
2849 	uint32_t count;
2850 	uint32_t limit;
2851 #if 0
2852 	static int zzticks;
2853 #endif
2854 
2855 	/*
2856 	 * Moderate the number of dirty inodes
2857 	 */
2858 	hammer2_pfs_moderate(ip, always_moderate);
2859 
2860 	/*
2861 	 * Atomic check condition and wait.  Also do an early speedup of
2862 	 * the syncer to try to avoid hitting the wait.
2863 	 */
2864 	for (;;) {
2865 		waiting = pmp->inmem_dirty_chains;
2866 		cpu_ccfence();
2867 		count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2868 
2869 		limit = pmp->mp->mnt_nvnodelistsize / 10;
2870 		if (limit < hammer2_limit_dirty_chains)
2871 			limit = hammer2_limit_dirty_chains;
2872 		if (limit < 1000)
2873 			limit = 1000;
2874 
2875 #if 0
2876 		if ((int)(ticks - zzticks) > hz) {
2877 			zzticks = ticks;
2878 			kprintf("count %ld %ld\n", count, limit);
2879 		}
2880 #endif
2881 
2882 		/*
2883 		 * Block if there are too many dirty chains present, wait
2884 		 * for the flush to clean some out.
2885 		 */
2886 		if (count > limit) {
2887 			hammer2_pfs_moderate(ip, always_moderate);
2888 			tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2889 			if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2890 					       waiting,
2891 				       waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2892 				if (ticks != pmp->speedup_ticks) {
2893 					pmp->speedup_ticks = ticks;
2894 					speedup_syncer(pmp->mp);
2895 				}
2896 				tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2897 				       "chnmem", hz);
2898 			}
2899 			continue;	/* loop on success or fail */
2900 		}
2901 
2902 		/*
2903 		 * Try to start an early flush before we are forced to block.
2904 		 */
2905 		if (count > limit * 5 / 10 &&
2906 		    ticks != pmp->speedup_ticks) {
2907 			pmp->speedup_ticks = ticks;
2908 			speedup_syncer(pmp->mp);
2909 		}
2910 		break;
2911 	}
2912 }
2913 
2914 void
2915 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2916 {
2917 	if (pmp) {
2918 		atomic_add_int(&pmp->inmem_dirty_chains, 1);
2919 	}
2920 }
2921 
2922 void
2923 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2924 {
2925 	uint32_t waiting;
2926 
2927 	if (pmp) {
2928 		waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
2929 		/* don't need --waiting to test flag */
2930 		if (waiting & HAMMER2_DIRTYCHAIN_WAITING) {
2931 			atomic_clear_int(&pmp->inmem_dirty_chains,
2932 					 HAMMER2_DIRTYCHAIN_WAITING);
2933 			wakeup(&pmp->inmem_dirty_chains);
2934 		}
2935 	}
2936 }
2937 
2938 /*
2939  * Returns 0 if the filesystem has tons of free space
2940  * Returns 1 if the filesystem has less than 10% remaining
2941  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2942  */
2943 int
2944 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2945 {
2946 	hammer2_pfs_t *pmp;
2947 	hammer2_dev_t *hmp;
2948 	hammer2_off_t free_reserved;
2949 	hammer2_off_t free_nominal;
2950 	int i;
2951 
2952 	pmp = ip->pmp;
2953 
2954 	if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2955 		free_reserved = HAMMER2_SEGSIZE;
2956 		free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2957 		for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2958 			hmp = pmp->pfs_hmps[i];
2959 			if (hmp == NULL)
2960 				continue;
2961 			if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2962 			    pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2963 				continue;
2964 
2965 			if (free_nominal > hmp->voldata.allocator_free)
2966 				free_nominal = hmp->voldata.allocator_free;
2967 			if (free_reserved < hmp->free_reserved)
2968 				free_reserved = hmp->free_reserved;
2969 		}
2970 
2971 		/*
2972 		 * SMP races ok
2973 		 */
2974 		pmp->free_reserved = free_reserved;
2975 		pmp->free_nominal = free_nominal;
2976 		pmp->free_ticks = ticks;
2977 	} else {
2978 		free_reserved = pmp->free_reserved;
2979 		free_nominal = pmp->free_nominal;
2980 	}
2981 	if (cred && cred->cr_uid != 0) {
2982 		if ((int64_t)(free_nominal - bytes) <
2983 		    (int64_t)free_reserved) {
2984 			return 2;
2985 		}
2986 	} else {
2987 		if ((int64_t)(free_nominal - bytes) <
2988 		    (int64_t)free_reserved / 2) {
2989 			return 2;
2990 		}
2991 	}
2992 	if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2993 		return 1;
2994 	return 0;
2995 }
2996 
2997 /*
2998  * Debugging
2999  */
3000 void
3001 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
3002 		   u_int flags)
3003 {
3004 	hammer2_chain_t *scan;
3005 	hammer2_chain_t *parent;
3006 
3007 	--*countp;
3008 	if (*countp == 0) {
3009 		kprintf("%*.*s...\n", tab, tab, "");
3010 		return;
3011 	}
3012 	if (*countp < 0)
3013 		return;
3014 	kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
3015 		tab, tab, "", pfx,
3016 		chain, chain->bref.type,
3017 		chain->bref.key, chain->bref.keybits,
3018 		chain->bref.mirror_tid);
3019 
3020 	kprintf("%*.*s      [%08x] (%s) refs=%d",
3021 		tab, tab, "",
3022 		chain->flags,
3023 		((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3024 		chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
3025 		chain->refs);
3026 
3027 	parent = chain->parent;
3028 	if (parent)
3029 		kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
3030 			tab, tab, "",
3031 			parent, parent->flags, parent->refs);
3032 	if (RB_EMPTY(&chain->core.rbtree)) {
3033 		kprintf("\n");
3034 	} else {
3035 		kprintf(" {\n");
3036 		RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3037 			if ((scan->flags & flags) || flags == (u_int)-1) {
3038 				hammer2_dump_chain(scan, tab + 4, countp, 'a',
3039 						   flags);
3040 			}
3041 		}
3042 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3043 			kprintf("%*.*s}(%s)\n", tab, tab, "",
3044 				chain->data->ipdata.filename);
3045 		else
3046 			kprintf("%*.*s}\n", tab, tab, "");
3047 	}
3048 }
3049