xref: /dflybsd-src/sys/vfs/hammer2/hammer2_vfsops.c (revision 14e5be3a8c11f2df26dba50c4a8aaa5e8d88133e)
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_count_modified_chains;
97 long hammer2_iod_invals;
98 long hammer2_iod_file_read;
99 long hammer2_iod_meta_read;
100 long hammer2_iod_indr_read;
101 long hammer2_iod_fmap_read;
102 long hammer2_iod_volu_read;
103 long hammer2_iod_file_write;
104 long hammer2_iod_file_wembed;
105 long hammer2_iod_file_wzero;
106 long hammer2_iod_file_wdedup;
107 long hammer2_iod_meta_write;
108 long hammer2_iod_indr_write;
109 long hammer2_iod_fmap_write;
110 long hammer2_iod_volu_write;
111 
112 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
113 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
114 		"Buffer used for compression.");
115 
116 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
117 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
118 		"Buffer used for decompression.");
119 
120 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
121 
122 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
123 	   &hammer2_supported_version, 0, "");
124 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
125 	   &hammer2_debug, 0, "");
126 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
127 	   &hammer2_cluster_meta_read, 0, "");
128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
129 	   &hammer2_cluster_data_read, 0, "");
130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
131 	   &hammer2_dedup_enable, 0, "");
132 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
133 	   &hammer2_always_compress, 0, "");
134 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
135 	   &hammer2_inval_enable, 0, "");
136 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
137 	   &hammer2_flush_pipe, 0, "");
138 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
139 	   &hammer2_bulkfree_tps, 0, "");
140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
141 	   &hammer2_chain_allocs, 0, "");
142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
143 	   &hammer2_chain_frees, 0, "");
144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
145 	   &hammer2_limit_dirty_chains, 0, "");
146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
147 	   &hammer2_count_modified_chains, 0, "");
148 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
149 	   &hammer2_dio_count, 0, "");
150 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
151 	   &hammer2_dio_limit, 0, "");
152 
153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
154 	   &hammer2_iod_invals, 0, "");
155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
156 	   &hammer2_iod_file_read, 0, "");
157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
158 	   &hammer2_iod_meta_read, 0, "");
159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
160 	   &hammer2_iod_indr_read, 0, "");
161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
162 	   &hammer2_iod_fmap_read, 0, "");
163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
164 	   &hammer2_iod_volu_read, 0, "");
165 
166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
167 	   &hammer2_iod_file_write, 0, "");
168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
169 	   &hammer2_iod_file_wembed, 0, "");
170 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
171 	   &hammer2_iod_file_wzero, 0, "");
172 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
173 	   &hammer2_iod_file_wdedup, 0, "");
174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
175 	   &hammer2_iod_meta_write, 0, "");
176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
177 	   &hammer2_iod_indr_write, 0, "");
178 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
179 	   &hammer2_iod_fmap_write, 0, "");
180 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
181 	   &hammer2_iod_volu_write, 0, "");
182 
183 long hammer2_process_icrc32;
184 long hammer2_process_xxhash64;
185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
186 	   &hammer2_process_icrc32, 0, "");
187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
188 	   &hammer2_process_xxhash64, 0, "");
189 
190 static int hammer2_vfs_init(struct vfsconf *conf);
191 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
192 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
193 				struct ucred *cred);
194 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
195 				struct vnode *, struct ucred *);
196 static int hammer2_recovery(hammer2_dev_t *hmp);
197 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
198 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
199 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
200 				struct ucred *cred);
201 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
202 				struct ucred *cred);
203 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
204 				struct fid *fhp, struct vnode **vpp);
205 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
206 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
207 				int *exflagsp, struct ucred **credanonp);
208 
209 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
210 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
211 
212 static void hammer2_update_pmps(hammer2_dev_t *hmp);
213 
214 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
215 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
216 				hammer2_dev_t *hmp);
217 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
218 
219 /*
220  * HAMMER2 vfs operations.
221  */
222 static struct vfsops hammer2_vfsops = {
223 	.vfs_init	= hammer2_vfs_init,
224 	.vfs_uninit	= hammer2_vfs_uninit,
225 	.vfs_sync	= hammer2_vfs_sync,
226 	.vfs_mount	= hammer2_vfs_mount,
227 	.vfs_unmount	= hammer2_vfs_unmount,
228 	.vfs_root 	= hammer2_vfs_root,
229 	.vfs_statfs	= hammer2_vfs_statfs,
230 	.vfs_statvfs	= hammer2_vfs_statvfs,
231 	.vfs_vget	= hammer2_vfs_vget,
232 	.vfs_vptofh	= hammer2_vfs_vptofh,
233 	.vfs_fhtovp	= hammer2_vfs_fhtovp,
234 	.vfs_checkexp	= hammer2_vfs_checkexp
235 };
236 
237 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
238 
239 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
240 MODULE_VERSION(hammer2, 1);
241 
242 static
243 int
244 hammer2_vfs_init(struct vfsconf *conf)
245 {
246 	static struct objcache_malloc_args margs_read;
247 	static struct objcache_malloc_args margs_write;
248 	static struct objcache_malloc_args margs_vop;
249 
250 	int error;
251 
252 	error = 0;
253 
254 	/*
255 	 * A large DIO cache is needed to retain dedup enablement masks.
256 	 * The bulkfree code clears related masks as part of the disk block
257 	 * recycling algorithm, preventing it from being used for a later
258 	 * dedup.
259 	 *
260 	 * NOTE: A large buffer cache can actually interfere with dedup
261 	 *	 operation because we dedup based on media physical buffers
262 	 *	 and not logical buffers.  Try to make the DIO case large
263 	 *	 enough to avoid this problem, but also cap it.
264 	 */
265 	hammer2_dio_limit = nbuf * 2;
266 	if (hammer2_dio_limit > 100000)
267 		hammer2_dio_limit = 100000;
268 
269 	if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
270 		error = EINVAL;
271 	if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
272 		error = EINVAL;
273 	if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
274 		error = EINVAL;
275 
276 	if (error)
277 		kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
278 
279 	margs_read.objsize = 65536;
280 	margs_read.mtype = M_HAMMER2_DEBUFFER;
281 
282 	margs_write.objsize = 32768;
283 	margs_write.mtype = M_HAMMER2_CBUFFER;
284 
285 	margs_vop.objsize = sizeof(hammer2_xop_t);
286 	margs_vop.mtype = M_HAMMER2;
287 
288 	/*
289 	 * Note thaht for the XOPS cache we want backing store allocations
290 	 * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
291 	 * confusion), so use the backing store function that does it.  This
292 	 * means that initial XOPS objects are zerod but REUSED objects are
293 	 * not.  So we are responsible for cleaning the object up sufficiently
294 	 * for our needs before objcache_put()ing it back (typically just the
295 	 * FIFO indices).
296 	 */
297 	cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
298 				0, 1, NULL, NULL, NULL,
299 				objcache_malloc_alloc,
300 				objcache_malloc_free,
301 				&margs_read);
302 	cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
303 				0, 1, NULL, NULL, NULL,
304 				objcache_malloc_alloc,
305 				objcache_malloc_free,
306 				&margs_write);
307 	cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
308 				0, 1, NULL, NULL, NULL,
309 				objcache_malloc_alloc_zero,
310 				objcache_malloc_free,
311 				&margs_vop);
312 
313 
314 	lockinit(&hammer2_mntlk, "mntlk", 0, 0);
315 	TAILQ_INIT(&hammer2_mntlist);
316 	TAILQ_INIT(&hammer2_pfslist);
317 	TAILQ_INIT(&hammer2_spmplist);
318 
319 	hammer2_limit_dirty_chains = maxvnodes / 10;
320 	if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
321 		hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
322 
323 	return (error);
324 }
325 
326 static
327 int
328 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
329 {
330 	objcache_destroy(cache_buffer_read);
331 	objcache_destroy(cache_buffer_write);
332 	objcache_destroy(cache_xops);
333 	return 0;
334 }
335 
336 /*
337  * Core PFS allocator.  Used to allocate or reference the pmp structure
338  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
339  * The pmp can be passed in or loaded by this function using the chain and
340  * inode data.
341  *
342  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
343  * transactions.  Note that synchronization does not use this field.
344  * (typically frontend operations and synchronization cannot run on the
345  * same PFS node at the same time).
346  *
347  * XXX check locking
348  */
349 hammer2_pfs_t *
350 hammer2_pfsalloc(hammer2_chain_t *chain,
351 		 const hammer2_inode_data_t *ripdata,
352 		 hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
353 {
354 	hammer2_pfs_t *pmp;
355 	hammer2_inode_t *iroot;
356 	int count;
357 	int i;
358 	int j;
359 
360 	pmp = NULL;
361 
362 	/*
363 	 * Locate or create the PFS based on the cluster id.  If ripdata
364 	 * is NULL this is a spmp which is unique and is always allocated.
365 	 *
366 	 * If the device is mounted in local mode all PFSs are considered
367 	 * independent and not part of any cluster (for debugging only).
368 	 */
369 	if (ripdata) {
370 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
371 			if (force_local != pmp->force_local)
372 				continue;
373 			if (force_local == NULL &&
374 			    bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
375 				 sizeof(pmp->pfs_clid)) == 0) {
376 					break;
377 			} else if (force_local && pmp->pfs_names[0] &&
378 			    strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
379 					break;
380 			}
381 		}
382 	}
383 
384 	if (pmp == NULL) {
385 		pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
386 		pmp->force_local = force_local;
387 		hammer2_trans_manage_init(pmp);
388 		kmalloc_create(&pmp->minode, "HAMMER2-inodes");
389 		kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
390 		lockinit(&pmp->lock, "pfslk", 0, 0);
391 		lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
392 		spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
393 		spin_init(&pmp->xop_spin, "h2xop");
394 		spin_init(&pmp->lru_spin, "h2lru");
395 		RB_INIT(&pmp->inum_tree);
396 		TAILQ_INIT(&pmp->sideq);
397 		TAILQ_INIT(&pmp->lru_list);
398 		spin_init(&pmp->list_spin, "hm2pfsalloc_list");
399 
400 		/*
401 		 * Distribute backend operations to threads
402 		 */
403 		for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
404 			hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
405 
406 		/*
407 		 * Save the last media transaction id for the flusher.  Set
408 		 * initial
409 		 */
410 		if (ripdata) {
411 			pmp->pfs_clid = ripdata->meta.pfs_clid;
412 			TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
413 		} else {
414 			pmp->flags |= HAMMER2_PMPF_SPMP;
415 			TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
416 		}
417 
418 		/*
419 		 * The synchronization thread may start too early, make
420 		 * sure it stays frozen until we are ready to let it go.
421 		 * XXX
422 		 */
423 		/*
424 		pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
425 					 HAMMER2_THREAD_REMASTER;
426 		*/
427 	}
428 
429 	/*
430 	 * Create the PFS's root inode and any missing XOP helper threads.
431 	 */
432 	if ((iroot = pmp->iroot) == NULL) {
433 		iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
434 		if (ripdata)
435 			iroot->meta = ripdata->meta;
436 		pmp->iroot = iroot;
437 		hammer2_inode_ref(iroot);
438 		hammer2_inode_unlock(iroot);
439 	}
440 
441 	/*
442 	 * Stop here if no chain is passed in.
443 	 */
444 	if (chain == NULL)
445 		goto done;
446 
447 	/*
448 	 * When a chain is passed in we must add it to the PFS's root
449 	 * inode, update pmp->pfs_types[], and update the syncronization
450 	 * threads.
451 	 *
452 	 * When forcing local mode, mark the PFS as a MASTER regardless.
453 	 *
454 	 * At the moment empty spots can develop due to removals or failures.
455 	 * Ultimately we want to re-fill these spots but doing so might
456 	 * confused running code. XXX
457 	 */
458 	hammer2_inode_ref(iroot);
459 	hammer2_mtx_ex(&iroot->lock);
460 	j = iroot->cluster.nchains;
461 
462 	if (j == HAMMER2_MAXCLUSTER) {
463 		kprintf("hammer2_mount: cluster full!\n");
464 		/* XXX fatal error? */
465 	} else {
466 		KKASSERT(chain->pmp == NULL);
467 		chain->pmp = pmp;
468 		hammer2_chain_ref(chain);
469 		iroot->cluster.array[j].chain = chain;
470 		if (force_local)
471 			pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
472 		else
473 			pmp->pfs_types[j] = ripdata->meta.pfs_type;
474 		pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
475 		pmp->pfs_hmps[j] = chain->hmp;
476 
477 		/*
478 		 * If the PFS is already mounted we must account
479 		 * for the mount_count here.
480 		 */
481 		if (pmp->mp)
482 			++chain->hmp->mount_count;
483 
484 		/*
485 		 * May have to fixup dirty chain tracking.  Previous
486 		 * pmp was NULL so nothing to undo.
487 		 */
488 		if (chain->flags & HAMMER2_CHAIN_MODIFIED)
489 			hammer2_pfs_memory_inc(pmp);
490 		++j;
491 	}
492 	iroot->cluster.nchains = j;
493 
494 	/*
495 	 * Update nmasters from any PFS inode which is part of the cluster.
496 	 * It is possible that this will result in a value which is too
497 	 * high.  MASTER PFSs are authoritative for pfs_nmasters and will
498 	 * override this value later on.
499 	 *
500 	 * (This informs us of masters that might not currently be
501 	 *  discoverable by this mount).
502 	 */
503 	if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
504 		pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
505 	}
506 
507 	/*
508 	 * Count visible masters.  Masters are usually added with
509 	 * ripdata->meta.pfs_nmasters set to 1.  This detects when there
510 	 * are more (XXX and must update the master inodes).
511 	 */
512 	count = 0;
513 	for (i = 0; i < iroot->cluster.nchains; ++i) {
514 		if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
515 			++count;
516 	}
517 	if (pmp->pfs_nmasters < count)
518 		pmp->pfs_nmasters = count;
519 
520 	/*
521 	 * Create missing synchronization and support threads.
522 	 *
523 	 * Single-node masters (including snapshots) have nothing to
524 	 * synchronize and do not require this thread.
525 	 *
526 	 * Multi-node masters or any number of soft masters, slaves, copy,
527 	 * or other PFS types need the thread.
528 	 *
529 	 * Each thread is responsible for its particular cluster index.
530 	 * We use independent threads so stalls or mismatches related to
531 	 * any given target do not affect other targets.
532 	 */
533 	for (i = 0; i < iroot->cluster.nchains; ++i) {
534 		/*
535 		 * Single-node masters (including snapshots) have nothing
536 		 * to synchronize and will make direct xops support calls,
537 		 * thus they do not require this thread.
538 		 *
539 		 * Note that there can be thousands of snapshots.  We do not
540 		 * want to create thousands of threads.
541 		 */
542 		if (pmp->pfs_nmasters <= 1 &&
543 		    pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
544 			continue;
545 		}
546 
547 		/*
548 		 * Sync support thread
549 		 */
550 		if (pmp->sync_thrs[i].td == NULL) {
551 			hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
552 					   "h2nod", i, -1,
553 					   hammer2_primary_sync_thread);
554 		}
555 	}
556 
557 	/*
558 	 * Create missing Xop threads
559 	 *
560 	 * NOTE: We create helper threads for all mounted PFSs or any
561 	 *	 PFSs with 2+ nodes (so the sync thread can update them,
562 	 *	 even if not mounted).
563 	 */
564 	if (pmp->mp || iroot->cluster.nchains >= 2)
565 		hammer2_xop_helper_create(pmp);
566 
567 	hammer2_mtx_unlock(&iroot->lock);
568 	hammer2_inode_drop(iroot);
569 done:
570 	return pmp;
571 }
572 
573 /*
574  * Deallocate an element of a probed PFS.  If destroying and this is a
575  * MASTER, adjust nmasters.
576  *
577  * This function does not physically destroy the PFS element in its device
578  * under the super-root  (see hammer2_ioctl_pfs_delete()).
579  */
580 void
581 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
582 {
583 	hammer2_inode_t *iroot;
584 	hammer2_chain_t *chain;
585 	int j;
586 
587 	/*
588 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
589 	 * by the flush code.
590 	 */
591 	iroot = pmp->iroot;
592 	if (iroot) {
593 		/*
594 		 * Stop synchronizing
595 		 *
596 		 * XXX flush after acquiring the iroot lock.
597 		 * XXX clean out the cluster index from all inode structures.
598 		 */
599 		hammer2_thr_delete(&pmp->sync_thrs[clindex]);
600 
601 		/*
602 		 * Remove the cluster index from the group.  If destroying
603 		 * the PFS and this is a master, adjust pfs_nmasters.
604 		 */
605 		hammer2_mtx_ex(&iroot->lock);
606 		chain = iroot->cluster.array[clindex].chain;
607 		iroot->cluster.array[clindex].chain = NULL;
608 
609 		switch(pmp->pfs_types[clindex]) {
610 		case HAMMER2_PFSTYPE_MASTER:
611 			if (destroying && pmp->pfs_nmasters > 0)
612 				--pmp->pfs_nmasters;
613 			/* XXX adjust ripdata->meta.pfs_nmasters */
614 			break;
615 		default:
616 			break;
617 		}
618 		pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
619 
620 		hammer2_mtx_unlock(&iroot->lock);
621 
622 		/*
623 		 * Release the chain.
624 		 */
625 		if (chain) {
626 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
627 			hammer2_chain_drop(chain);
628 		}
629 
630 		/*
631 		 * Terminate all XOP threads for the cluster index.
632 		 */
633 		for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
634 			hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
635 	}
636 }
637 
638 /*
639  * Destroy a PFS, typically only occurs after the last mount on a device
640  * has gone away.
641  */
642 static void
643 hammer2_pfsfree(hammer2_pfs_t *pmp)
644 {
645 	hammer2_inode_t *iroot;
646 	hammer2_chain_t *chain;
647 	int i;
648 	int j;
649 
650 	/*
651 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
652 	 * by the flush code.
653 	 */
654 	if (pmp->flags & HAMMER2_PMPF_SPMP)
655 		TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
656 	else
657 		TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
658 
659 	iroot = pmp->iroot;
660 	if (iroot) {
661 		for (i = 0; i < iroot->cluster.nchains; ++i) {
662 			hammer2_thr_delete(&pmp->sync_thrs[i]);
663 			for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
664 				hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
665 		}
666 #if REPORT_REFS_ERRORS
667 		if (pmp->iroot->refs != 1)
668 			kprintf("PMP->IROOT %p REFS WRONG %d\n",
669 				pmp->iroot, pmp->iroot->refs);
670 #else
671 		KKASSERT(pmp->iroot->refs == 1);
672 #endif
673 		/* ref for pmp->iroot */
674 		hammer2_inode_drop(pmp->iroot);
675 		pmp->iroot = NULL;
676 	}
677 
678 	/*
679 	 * Cleanup chains remaining on LRU list.
680 	 */
681 	hammer2_spin_ex(&pmp->lru_spin);
682 	while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
683 		KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
684 		atomic_add_int(&pmp->lru_count, -1);
685 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
686 		TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
687 		hammer2_chain_ref(chain);
688 		hammer2_spin_unex(&pmp->lru_spin);
689 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
690 		hammer2_chain_drop(chain);
691 		hammer2_spin_ex(&pmp->lru_spin);
692 	}
693 	hammer2_spin_unex(&pmp->lru_spin);
694 
695 	/*
696 	 * Free remaining pmp resources
697 	 */
698 	kmalloc_destroy(&pmp->mmsg);
699 	kmalloc_destroy(&pmp->minode);
700 
701 	kfree(pmp, M_HAMMER2);
702 }
703 
704 /*
705  * Remove all references to hmp from the pfs list.  Any PFS which becomes
706  * empty is terminated and freed.
707  *
708  * XXX inefficient.
709  */
710 static void
711 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
712 {
713 	hammer2_pfs_t *pmp;
714 	hammer2_inode_t *iroot;
715 	hammer2_chain_t *rchain;
716 	int didfreeze;
717 	int i;
718 	int j;
719 	struct hammer2_pfslist *wlist;
720 
721 	if (which == 0)
722 		wlist = &hammer2_pfslist;
723 	else
724 		wlist = &hammer2_spmplist;
725 again:
726 	TAILQ_FOREACH(pmp, wlist, mntentry) {
727 		if ((iroot = pmp->iroot) == NULL)
728 			continue;
729 		hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
730 		hammer2_inode_run_sideq(pmp, 1);
731 		hammer2_bioq_sync(pmp);
732 		hammer2_trans_done(pmp);
733 
734 		/*
735 		 * Determine if this PFS is affected.  If it is we must
736 		 * freeze all management threads and lock its iroot.
737 		 *
738 		 * Freezing a management thread forces it idle, operations
739 		 * in-progress will be aborted and it will have to start
740 		 * over again when unfrozen, or exit if told to exit.
741 		 */
742 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
743 			if (pmp->pfs_hmps[i] == hmp)
744 				break;
745 		}
746 		if (i != HAMMER2_MAXCLUSTER) {
747 			/*
748 			 * Make sure all synchronization threads are locked
749 			 * down.
750 			 */
751 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
752 				if (pmp->pfs_hmps[i] == NULL)
753 					continue;
754 				hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
755 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
756 					hammer2_thr_freeze_async(
757 						&pmp->xop_groups[j].thrs[i]);
758 				}
759 			}
760 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
761 				if (pmp->pfs_hmps[i] == NULL)
762 					continue;
763 				hammer2_thr_freeze(&pmp->sync_thrs[i]);
764 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
765 					hammer2_thr_freeze(
766 						&pmp->xop_groups[j].thrs[i]);
767 				}
768 			}
769 
770 			/*
771 			 * Lock the inode and clean out matching chains.
772 			 * Note that we cannot use hammer2_inode_lock_*()
773 			 * here because that would attempt to validate the
774 			 * cluster that we are in the middle of ripping
775 			 * apart.
776 			 *
777 			 * WARNING! We are working directly on the inodes
778 			 *	    embedded cluster.
779 			 */
780 			hammer2_mtx_ex(&iroot->lock);
781 
782 			/*
783 			 * Remove the chain from matching elements of the PFS.
784 			 */
785 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
786 				if (pmp->pfs_hmps[i] != hmp)
787 					continue;
788 				hammer2_thr_delete(&pmp->sync_thrs[i]);
789 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
790 					hammer2_thr_delete(
791 						&pmp->xop_groups[j].thrs[i]);
792 				}
793 				rchain = iroot->cluster.array[i].chain;
794 				iroot->cluster.array[i].chain = NULL;
795 				pmp->pfs_types[i] = 0;
796 				if (pmp->pfs_names[i]) {
797 					kfree(pmp->pfs_names[i], M_HAMMER2);
798 					pmp->pfs_names[i] = NULL;
799 				}
800 				if (rchain) {
801 					hammer2_chain_drop(rchain);
802 					/* focus hint */
803 					if (iroot->cluster.focus == rchain)
804 						iroot->cluster.focus = NULL;
805 				}
806 				pmp->pfs_hmps[i] = NULL;
807 			}
808 			hammer2_mtx_unlock(&iroot->lock);
809 			didfreeze = 1;	/* remaster, unfreeze down below */
810 		} else {
811 			didfreeze = 0;
812 		}
813 
814 		/*
815 		 * Cleanup trailing chains.  Gaps may remain.
816 		 */
817 		for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
818 			if (pmp->pfs_hmps[i])
819 				break;
820 		}
821 		iroot->cluster.nchains = i + 1;
822 
823 		/*
824 		 * If the PMP has no elements remaining we can destroy it.
825 		 * (this will transition management threads from frozen->exit).
826 		 */
827 		if (iroot->cluster.nchains == 0) {
828 			/*
829 			 * If this was the hmp's spmp, we need to clean
830 			 * a little more stuff out.
831 			 */
832 			if (hmp->spmp == pmp) {
833 				hmp->spmp = NULL;
834 				hmp->vchain.pmp = NULL;
835 				hmp->fchain.pmp = NULL;
836 			}
837 
838 			/*
839 			 * Free the pmp and restart the loop
840 			 */
841 			hammer2_pfsfree(pmp);
842 			goto again;
843 		}
844 
845 		/*
846 		 * If elements still remain we need to set the REMASTER
847 		 * flag and unfreeze it.
848 		 */
849 		if (didfreeze) {
850 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
851 				if (pmp->pfs_hmps[i] == NULL)
852 					continue;
853 				hammer2_thr_remaster(&pmp->sync_thrs[i]);
854 				hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
855 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
856 					hammer2_thr_remaster(
857 						&pmp->xop_groups[j].thrs[i]);
858 					hammer2_thr_unfreeze(
859 						&pmp->xop_groups[j].thrs[i]);
860 				}
861 			}
862 		}
863 	}
864 }
865 
866 /*
867  * Mount or remount HAMMER2 fileystem from physical media
868  *
869  *	mountroot
870  *		mp		mount point structure
871  *		path		NULL
872  *		data		<unused>
873  *		cred		<unused>
874  *
875  *	mount
876  *		mp		mount point structure
877  *		path		path to mount point
878  *		data		pointer to argument structure in user space
879  *			volume	volume path (device@LABEL form)
880  *			hflags	user mount flags
881  *		cred		user credentials
882  *
883  * RETURNS:	0	Success
884  *		!0	error number
885  */
886 static
887 int
888 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
889 		  struct ucred *cred)
890 {
891 	struct hammer2_mount_info info;
892 	hammer2_pfs_t *pmp;
893 	hammer2_pfs_t *spmp;
894 	hammer2_dev_t *hmp;
895 	hammer2_dev_t *force_local;
896 	hammer2_key_t key_next;
897 	hammer2_key_t key_dummy;
898 	hammer2_key_t lhc;
899 	struct vnode *devvp;
900 	struct nlookupdata nd;
901 	hammer2_chain_t *parent;
902 	hammer2_chain_t *chain;
903 	hammer2_cluster_t *cluster;
904 	const hammer2_inode_data_t *ripdata;
905 	hammer2_blockref_t bref;
906 	struct file *fp;
907 	char devstr[MNAMELEN];
908 	size_t size;
909 	size_t done;
910 	char *dev;
911 	char *label;
912 	int ronly = 1;
913 	int error;
914 	int i;
915 
916 	hmp = NULL;
917 	pmp = NULL;
918 	dev = NULL;
919 	label = NULL;
920 	devvp = NULL;
921 
922 	kprintf("hammer2_mount\n");
923 
924 	if (path == NULL) {
925 		/*
926 		 * Root mount
927 		 */
928 		bzero(&info, sizeof(info));
929 		info.cluster_fd = -1;
930 		ksnprintf(devstr, sizeof(devstr), "%s",
931 			  mp->mnt_stat.f_mntfromname);
932 		kprintf("hammer2_mount: root '%s'\n", devstr);
933 	} else {
934 		/*
935 		 * Non-root mount or updating a mount
936 		 */
937 		error = copyin(data, &info, sizeof(info));
938 		if (error)
939 			return (error);
940 
941 		error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
942 		if (error)
943 			return (error);
944 	}
945 
946 	/*
947 	 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
948 	 * if no label specified, based on the partition id.  Error out if no
949 	 * label or device (with partition id) is specified.  This is strictly
950 	 * a convenience to match the default label created by newfs_hammer2,
951 	 * our preference is that a label always be specified.
952 	 *
953 	 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
954 	 *	 that does not specify a device, as long as some H2 label
955 	 *	 has already been mounted from that device.  This makes
956 	 *	 mounting snapshots a lot easier.
957 	 */
958 	dev = devstr;
959 	label = strchr(devstr, '@');
960 	if (label && ((label + 1) - dev) > done)
961 		return (EINVAL);
962 	if (label == NULL || label[1] == 0) {
963 		char slice;
964 
965 		if (label == NULL)
966 			label = devstr + strlen(devstr);
967 		slice = label[-1];
968 		switch(slice) {
969 		case 'a':
970 			label = "BOOT";
971 			break;
972 		case 'd':
973 			label = "ROOT";
974 			break;
975 		default:
976 			label = "DATA";
977 			break;
978 		}
979 	} else {
980 		*label = '\0';
981 		label++;
982 	}
983 
984 	kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
985 		dev, label, (mp->mnt_flag & MNT_RDONLY));
986 
987 	if (mp->mnt_flag & MNT_UPDATE) {
988 		/*
989 		 * Update mount.  Note that pmp->iroot->cluster is
990 		 * an inode-embedded cluster and thus cannot be
991 		 * directly locked.
992 		 *
993 		 * XXX HAMMER2 needs to implement NFS export via
994 		 *     mountctl.
995 		 */
996 		pmp = MPTOPMP(mp);
997 		pmp->hflags = info.hflags;
998 		cluster = &pmp->iroot->cluster;
999 		for (i = 0; i < cluster->nchains; ++i) {
1000 			if (cluster->array[i].chain == NULL)
1001 				continue;
1002 			hmp = cluster->array[i].chain->hmp;
1003 			devvp = hmp->devvp;
1004 			error = hammer2_remount(hmp, mp, path,
1005 						devvp, cred);
1006 			if (error)
1007 				break;
1008 		}
1009 
1010 		return error;
1011 	}
1012 
1013 	/*
1014 	 * HMP device mount
1015 	 *
1016 	 * If a path is specified and dev is not an empty string, lookup the
1017 	 * name and verify that it referes to a block device.
1018 	 *
1019 	 * If a path is specified and dev is an empty string we fall through
1020 	 * and locate the label in the hmp search.
1021 	 */
1022 	if (path && *dev != 0) {
1023 		error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1024 		if (error == 0)
1025 			error = nlookup(&nd);
1026 		if (error == 0)
1027 			error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1028 		nlookup_done(&nd);
1029 	} else if (path == NULL) {
1030 		/* root mount */
1031 		cdev_t cdev = kgetdiskbyname(dev);
1032 		error = bdevvp(cdev, &devvp);
1033 		if (error)
1034 			kprintf("hammer2: cannot find '%s'\n", dev);
1035 	} else {
1036 		/*
1037 		 * We will locate the hmp using the label in the hmp loop.
1038 		 */
1039 		error = 0;
1040 	}
1041 
1042 	/*
1043 	 * Make sure its a block device.  Do not check to see if it is
1044 	 * already mounted until we determine that its a fresh H2 device.
1045 	 */
1046 	if (error == 0 && devvp) {
1047 		vn_isdisk(devvp, &error);
1048 	}
1049 
1050 	/*
1051 	 * Determine if the device has already been mounted.  After this
1052 	 * check hmp will be non-NULL if we are doing the second or more
1053 	 * hammer2 mounts from the same device.
1054 	 */
1055 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1056 	if (devvp) {
1057 		/*
1058 		 * Match the device.  Due to the way devfs works,
1059 		 * we may not be able to directly match the vnode pointer,
1060 		 * so also check to see if the underlying device matches.
1061 		 */
1062 		TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1063 			if (hmp->devvp == devvp)
1064 				break;
1065 			if (devvp->v_rdev &&
1066 			    hmp->devvp->v_rdev == devvp->v_rdev) {
1067 				break;
1068 			}
1069 		}
1070 
1071 		/*
1072 		 * If no match this may be a fresh H2 mount, make sure
1073 		 * the device is not mounted on anything else.
1074 		 */
1075 		if (hmp == NULL)
1076 			error = vfs_mountedon(devvp);
1077 	} else if (error == 0) {
1078 		/*
1079 		 * Match the label to a pmp already probed.
1080 		 */
1081 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1082 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1083 				if (pmp->pfs_names[i] &&
1084 				    strcmp(pmp->pfs_names[i], label) == 0) {
1085 					hmp = pmp->pfs_hmps[i];
1086 					break;
1087 				}
1088 			}
1089 			if (hmp)
1090 				break;
1091 		}
1092 		if (hmp == NULL)
1093 			error = ENOENT;
1094 	}
1095 
1096 	/*
1097 	 * Open the device if this isn't a secondary mount and construct
1098 	 * the H2 device mount (hmp).
1099 	 */
1100 	if (hmp == NULL) {
1101 		hammer2_chain_t *schain;
1102 		hammer2_xid_t xid;
1103 
1104 		if (error == 0 && vcount(devvp) > 0) {
1105 			kprintf("Primary device already has references\n");
1106 			error = EBUSY;
1107 		}
1108 
1109 		/*
1110 		 * Now open the device
1111 		 */
1112 		if (error == 0) {
1113 			ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1114 			vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1115 			error = vinvalbuf(devvp, V_SAVE, 0, 0);
1116 			if (error == 0) {
1117 				error = VOP_OPEN(devvp,
1118 					     (ronly ? FREAD : FREAD | FWRITE),
1119 					     FSCRED, NULL);
1120 			}
1121 			vn_unlock(devvp);
1122 		}
1123 		if (error && devvp) {
1124 			vrele(devvp);
1125 			devvp = NULL;
1126 		}
1127 		if (error) {
1128 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1129 			return error;
1130 		}
1131 		hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1132 		ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1133 		hmp->ronly = ronly;
1134 		hmp->devvp = devvp;
1135 		hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1136 		kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1137 		TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1138 		RB_INIT(&hmp->iotree);
1139 		spin_init(&hmp->io_spin, "hm2mount_io");
1140 		spin_init(&hmp->list_spin, "hm2mount_list");
1141 		TAILQ_INIT(&hmp->flushq);
1142 
1143 		lockinit(&hmp->vollk, "h2vol", 0, 0);
1144 		lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1145 		lockinit(&hmp->bflock, "h2bflk", 0, 0);
1146 
1147 		/*
1148 		 * vchain setup. vchain.data is embedded.
1149 		 * vchain.refs is initialized and will never drop to 0.
1150 		 *
1151 		 * NOTE! voldata is not yet loaded.
1152 		 */
1153 		hmp->vchain.hmp = hmp;
1154 		hmp->vchain.refs = 1;
1155 		hmp->vchain.data = (void *)&hmp->voldata;
1156 		hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1157 		hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1158 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1159 
1160 		hammer2_chain_core_init(&hmp->vchain);
1161 		/* hmp->vchain.u.xxx is left NULL */
1162 
1163 		/*
1164 		 * fchain setup.  fchain.data is embedded.
1165 		 * fchain.refs is initialized and will never drop to 0.
1166 		 *
1167 		 * The data is not used but needs to be initialized to
1168 		 * pass assertion muster.  We use this chain primarily
1169 		 * as a placeholder for the freemap's top-level RBTREE
1170 		 * so it does not interfere with the volume's topology
1171 		 * RBTREE.
1172 		 */
1173 		hmp->fchain.hmp = hmp;
1174 		hmp->fchain.refs = 1;
1175 		hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1176 		hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1177 		hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1178 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1179 		hmp->fchain.bref.methods =
1180 			HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1181 			HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1182 
1183 		hammer2_chain_core_init(&hmp->fchain);
1184 		/* hmp->fchain.u.xxx is left NULL */
1185 
1186 		/*
1187 		 * Install the volume header and initialize fields from
1188 		 * voldata.
1189 		 */
1190 		error = hammer2_install_volume_header(hmp);
1191 		if (error) {
1192 			hammer2_unmount_helper(mp, NULL, hmp);
1193 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1194 			hammer2_vfs_unmount(mp, MNT_FORCE);
1195 			return error;
1196 		}
1197 
1198 		/*
1199 		 * Really important to get these right or flush will get
1200 		 * confused.
1201 		 */
1202 		hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1203 		spmp = hmp->spmp;
1204 
1205 		/*
1206 		 * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1207 		 * is inherited from the volume header.
1208 		 */
1209 		xid = 0;
1210 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1211 		hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1212 		hmp->vchain.pmp = spmp;
1213 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1214 		hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1215 		hmp->fchain.pmp = spmp;
1216 
1217 		/*
1218 		 * First locate the super-root inode, which is key 0
1219 		 * relative to the volume header's blockset.
1220 		 *
1221 		 * Then locate the root inode by scanning the directory keyspace
1222 		 * represented by the label.
1223 		 */
1224 		parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1225 		schain = hammer2_chain_lookup(&parent, &key_dummy,
1226 				      HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1227 				      &error, 0);
1228 		hammer2_chain_lookup_done(parent);
1229 		if (schain == NULL) {
1230 			kprintf("hammer2_mount: invalid super-root\n");
1231 			hammer2_unmount_helper(mp, NULL, hmp);
1232 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1233 			hammer2_vfs_unmount(mp, MNT_FORCE);
1234 			return EINVAL;
1235 		}
1236 		if (schain->error) {
1237 			kprintf("hammer2_mount: error %s reading super-root\n",
1238 				hammer2_error_str(schain->error));
1239 			hammer2_chain_unlock(schain);
1240 			hammer2_chain_drop(schain);
1241 			schain = NULL;
1242 			hammer2_unmount_helper(mp, NULL, hmp);
1243 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1244 			hammer2_vfs_unmount(mp, MNT_FORCE);
1245 			return EINVAL;
1246 		}
1247 
1248 		/*
1249 		 * The super-root always uses an inode_tid of 1 when
1250 		 * creating PFSs.
1251 		 */
1252 		spmp->inode_tid = 1;
1253 		spmp->modify_tid = schain->bref.modify_tid + 1;
1254 
1255 		/*
1256 		 * Sanity-check schain's pmp and finish initialization.
1257 		 * Any chain belonging to the super-root topology should
1258 		 * have a NULL pmp (not even set to spmp).
1259 		 */
1260 		ripdata = &hammer2_chain_rdata(schain)->ipdata;
1261 		KKASSERT(schain->pmp == NULL);
1262 		spmp->pfs_clid = ripdata->meta.pfs_clid;
1263 
1264 		/*
1265 		 * Replace the dummy spmp->iroot with a real one.  It's
1266 		 * easier to just do a wholesale replacement than to try
1267 		 * to update the chain and fixup the iroot fields.
1268 		 *
1269 		 * The returned inode is locked with the supplied cluster.
1270 		 */
1271 		cluster = hammer2_cluster_from_chain(schain);
1272 		hammer2_inode_drop(spmp->iroot);
1273 		spmp->iroot = NULL;
1274 		spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1275 		spmp->spmp_hmp = hmp;
1276 		spmp->pfs_types[0] = ripdata->meta.pfs_type;
1277 		spmp->pfs_hmps[0] = hmp;
1278 		hammer2_inode_ref(spmp->iroot);
1279 		hammer2_inode_unlock(spmp->iroot);
1280 		hammer2_cluster_unlock(cluster);
1281 		hammer2_cluster_drop(cluster);
1282 		schain = NULL;
1283 		/* leave spmp->iroot with one ref */
1284 
1285 		if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1286 			error = hammer2_recovery(hmp);
1287 			if (error == 0)
1288 				error |= hammer2_fixup_pfses(hmp);
1289 			/* XXX do something with error */
1290 		}
1291 		hammer2_update_pmps(hmp);
1292 		hammer2_iocom_init(hmp);
1293 		hammer2_bulkfree_init(hmp);
1294 
1295 		/*
1296 		 * Ref the cluster management messaging descriptor.  The mount
1297 		 * program deals with the other end of the communications pipe.
1298 		 *
1299 		 * Root mounts typically do not supply one.
1300 		 */
1301 		if (info.cluster_fd >= 0) {
1302 			fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1303 			if (fp) {
1304 				hammer2_cluster_reconnect(hmp, fp);
1305 			} else {
1306 				kprintf("hammer2_mount: bad cluster_fd!\n");
1307 			}
1308 		}
1309 	} else {
1310 		spmp = hmp->spmp;
1311 		if (info.hflags & HMNT2_DEVFLAGS) {
1312 			kprintf("hammer2: Warning: mount flags pertaining "
1313 				"to the whole device may only be specified "
1314 				"on the first mount of the device: %08x\n",
1315 				info.hflags & HMNT2_DEVFLAGS);
1316 		}
1317 	}
1318 
1319 	/*
1320 	 * Force local mount (disassociate all PFSs from their clusters).
1321 	 * Used primarily for debugging.
1322 	 */
1323 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1324 
1325 	/*
1326 	 * Lookup the mount point under the media-localized super-root.
1327 	 * Scanning hammer2_pfslist doesn't help us because it represents
1328 	 * PFS cluster ids which can aggregate several named PFSs together.
1329 	 *
1330 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1331 	 * up later on.
1332 	 */
1333 	hammer2_inode_lock(spmp->iroot, 0);
1334 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1335 	lhc = hammer2_dirhash(label, strlen(label));
1336 	chain = hammer2_chain_lookup(&parent, &key_next,
1337 				     lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1338 				     &error, 0);
1339 	while (chain) {
1340 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1341 		    strcmp(label, chain->data->ipdata.filename) == 0) {
1342 			break;
1343 		}
1344 		chain = hammer2_chain_next(&parent, chain, &key_next,
1345 					    key_next,
1346 					    lhc + HAMMER2_DIRHASH_LOMASK,
1347 					    &error, 0);
1348 	}
1349 	if (parent) {
1350 		hammer2_chain_unlock(parent);
1351 		hammer2_chain_drop(parent);
1352 	}
1353 	hammer2_inode_unlock(spmp->iroot);
1354 
1355 	/*
1356 	 * PFS could not be found?
1357 	 */
1358 	if (chain == NULL) {
1359 		if (error)
1360 			kprintf("hammer2_mount: PFS label I/O error\n");
1361 		else
1362 			kprintf("hammer2_mount: PFS label not found\n");
1363 		hammer2_unmount_helper(mp, NULL, hmp);
1364 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1365 		hammer2_vfs_unmount(mp, MNT_FORCE);
1366 
1367 		return EINVAL;
1368 	}
1369 
1370 	/*
1371 	 * Acquire the pmp structure (it should have already been allocated
1372 	 * via hammer2_update_pmps() so do not pass cluster in to add to
1373 	 * available chains).
1374 	 *
1375 	 * Check if the cluster has already been mounted.  A cluster can
1376 	 * only be mounted once, use null mounts to mount additional copies.
1377 	 */
1378 	if (chain->error) {
1379 		kprintf("hammer2_mount: PFS label I/O error\n");
1380 	} else {
1381 		ripdata = &chain->data->ipdata;
1382 		bref = chain->bref;
1383 		pmp = hammer2_pfsalloc(NULL, ripdata,
1384 				       bref.modify_tid, force_local);
1385 	}
1386 	hammer2_chain_unlock(chain);
1387 	hammer2_chain_drop(chain);
1388 
1389 	/*
1390 	 * Finish the mount
1391 	 */
1392         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1393 
1394 	if (pmp->mp) {
1395 		kprintf("hammer2_mount: PFS already mounted!\n");
1396 		hammer2_unmount_helper(mp, NULL, hmp);
1397 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1398 		hammer2_vfs_unmount(mp, MNT_FORCE);
1399 
1400 		return EBUSY;
1401 	}
1402 
1403 	pmp->hflags = info.hflags;
1404         mp->mnt_flag |= MNT_LOCAL;
1405         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1406         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1407 
1408         /*
1409          * required mount structure initializations
1410          */
1411         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1412         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1413 
1414         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1415         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1416 
1417         /*
1418          * Optional fields
1419          */
1420         mp->mnt_iosize_max = MAXPHYS;
1421 
1422 	/*
1423 	 * Connect up mount pointers.
1424 	 */
1425 	hammer2_mount_helper(mp, pmp);
1426 
1427         lockmgr(&hammer2_mntlk, LK_RELEASE);
1428 
1429 	/*
1430 	 * Finish setup
1431 	 */
1432 	vfs_getnewfsid(mp);
1433 	vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1434 	vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1435 	vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1436 
1437 	if (path) {
1438 		copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1439 			  MNAMELEN - 1, &size);
1440 		bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1441 	} /* else root mount, already in there */
1442 
1443 	bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1444 	if (path) {
1445 		copyinstr(path, mp->mnt_stat.f_mntonname,
1446 			  sizeof(mp->mnt_stat.f_mntonname) - 1,
1447 			  &size);
1448 	} else {
1449 		/* root mount */
1450 		mp->mnt_stat.f_mntonname[0] = '/';
1451 	}
1452 
1453 	/*
1454 	 * Initial statfs to prime mnt_stat.
1455 	 */
1456 	hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1457 
1458 	return 0;
1459 }
1460 
1461 /*
1462  * Scan PFSs under the super-root and create hammer2_pfs structures.
1463  */
1464 static
1465 void
1466 hammer2_update_pmps(hammer2_dev_t *hmp)
1467 {
1468 	const hammer2_inode_data_t *ripdata;
1469 	hammer2_chain_t *parent;
1470 	hammer2_chain_t *chain;
1471 	hammer2_blockref_t bref;
1472 	hammer2_dev_t *force_local;
1473 	hammer2_pfs_t *spmp;
1474 	hammer2_pfs_t *pmp;
1475 	hammer2_key_t key_next;
1476 	int error;
1477 
1478 	/*
1479 	 * Force local mount (disassociate all PFSs from their clusters).
1480 	 * Used primarily for debugging.
1481 	 */
1482 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1483 
1484 	/*
1485 	 * Lookup mount point under the media-localized super-root.
1486 	 *
1487 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1488 	 * up later on.
1489 	 */
1490 	spmp = hmp->spmp;
1491 	hammer2_inode_lock(spmp->iroot, 0);
1492 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1493 	chain = hammer2_chain_lookup(&parent, &key_next,
1494 					 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1495 					 &error, 0);
1496 	while (chain) {
1497 		if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1498 			continue;
1499 		if (chain->error) {
1500 			kprintf("I/O error scanning PFS labels\n");
1501 		} else {
1502 			ripdata = &chain->data->ipdata;
1503 			bref = chain->bref;
1504 
1505 			pmp = hammer2_pfsalloc(chain, ripdata,
1506 					       bref.modify_tid, force_local);
1507 		}
1508 		chain = hammer2_chain_next(&parent, chain, &key_next,
1509 					   key_next, HAMMER2_KEY_MAX,
1510 					   &error, 0);
1511 	}
1512 	if (parent) {
1513 		hammer2_chain_unlock(parent);
1514 		hammer2_chain_drop(parent);
1515 	}
1516 	hammer2_inode_unlock(spmp->iroot);
1517 }
1518 
1519 static
1520 int
1521 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1522 		struct vnode *devvp, struct ucred *cred)
1523 {
1524 	int error;
1525 
1526 	if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1527 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1528 		VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1529 		vn_unlock(devvp);
1530 		error = hammer2_recovery(hmp);
1531 		if (error == 0)
1532 			error |= hammer2_fixup_pfses(hmp);
1533 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1534 		if (error == 0) {
1535 			VOP_CLOSE(devvp, FREAD, NULL);
1536 			hmp->ronly = 0;
1537 		} else {
1538 			VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1539 		}
1540 		vn_unlock(devvp);
1541 	} else {
1542 		error = 0;
1543 	}
1544 	return error;
1545 }
1546 
1547 static
1548 int
1549 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1550 {
1551 	hammer2_pfs_t *pmp;
1552 	int flags;
1553 	int error = 0;
1554 
1555 	pmp = MPTOPMP(mp);
1556 
1557 	if (pmp == NULL)
1558 		return(0);
1559 
1560 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1561 
1562 	/*
1563 	 * If mount initialization proceeded far enough we must flush
1564 	 * its vnodes and sync the underlying mount points.  Three syncs
1565 	 * are required to fully flush the filesystem (freemap updates lag
1566 	 * by one flush, and one extra for safety).
1567 	 */
1568 	if (mntflags & MNT_FORCE)
1569 		flags = FORCECLOSE;
1570 	else
1571 		flags = 0;
1572 	if (pmp->iroot) {
1573 		error = vflush(mp, 0, flags);
1574 		if (error)
1575 			goto failed;
1576 		hammer2_vfs_sync(mp, MNT_WAIT);
1577 		hammer2_vfs_sync(mp, MNT_WAIT);
1578 		hammer2_vfs_sync(mp, MNT_WAIT);
1579 	}
1580 
1581 	/*
1582 	 * Cleanup the frontend support XOPS threads
1583 	 */
1584 	hammer2_xop_helper_cleanup(pmp);
1585 
1586 	if (pmp->mp)
1587 		hammer2_unmount_helper(mp, pmp, NULL);
1588 
1589 	error = 0;
1590 failed:
1591 	lockmgr(&hammer2_mntlk, LK_RELEASE);
1592 
1593 	return (error);
1594 }
1595 
1596 /*
1597  * Mount helper, hook the system mount into our PFS.
1598  * The mount lock is held.
1599  *
1600  * We must bump the mount_count on related devices for any
1601  * mounted PFSs.
1602  */
1603 static
1604 void
1605 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1606 {
1607 	hammer2_cluster_t *cluster;
1608 	hammer2_chain_t *rchain;
1609 	int i;
1610 
1611         mp->mnt_data = (qaddr_t)pmp;
1612 	pmp->mp = mp;
1613 
1614 	/*
1615 	 * After pmp->mp is set we have to adjust hmp->mount_count.
1616 	 */
1617 	cluster = &pmp->iroot->cluster;
1618 	for (i = 0; i < cluster->nchains; ++i) {
1619 		rchain = cluster->array[i].chain;
1620 		if (rchain == NULL)
1621 			continue;
1622 		++rchain->hmp->mount_count;
1623 	}
1624 
1625 	/*
1626 	 * Create missing Xop threads
1627 	 */
1628 	hammer2_xop_helper_create(pmp);
1629 }
1630 
1631 /*
1632  * Mount helper, unhook the system mount from our PFS.
1633  * The mount lock is held.
1634  *
1635  * If hmp is supplied a mount responsible for being the first to open
1636  * the block device failed and the block device and all PFSs using the
1637  * block device must be cleaned up.
1638  *
1639  * If pmp is supplied multiple devices might be backing the PFS and each
1640  * must be disconnected.  This might not be the last PFS using some of the
1641  * underlying devices.  Also, we have to adjust our hmp->mount_count
1642  * accounting for the devices backing the pmp which is now undergoing an
1643  * unmount.
1644  */
1645 static
1646 void
1647 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1648 {
1649 	hammer2_cluster_t *cluster;
1650 	hammer2_chain_t *rchain;
1651 	struct vnode *devvp;
1652 	int dumpcnt;
1653 	int ronly;
1654 	int i;
1655 
1656 	/*
1657 	 * If no device supplied this is a high-level unmount and we have to
1658 	 * to disconnect the mount, adjust mount_count, and locate devices
1659 	 * that might now have no mounts.
1660 	 */
1661 	if (pmp) {
1662 		KKASSERT(hmp == NULL);
1663 		KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1664 		pmp->mp = NULL;
1665 		mp->mnt_data = NULL;
1666 
1667 		/*
1668 		 * After pmp->mp is cleared we have to account for
1669 		 * mount_count.
1670 		 */
1671 		cluster = &pmp->iroot->cluster;
1672 		for (i = 0; i < cluster->nchains; ++i) {
1673 			rchain = cluster->array[i].chain;
1674 			if (rchain == NULL)
1675 				continue;
1676 			--rchain->hmp->mount_count;
1677 			/* scrapping hmp now may invalidate the pmp */
1678 		}
1679 again:
1680 		TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1681 			if (hmp->mount_count == 0) {
1682 				hammer2_unmount_helper(NULL, NULL, hmp);
1683 				goto again;
1684 			}
1685 		}
1686 		return;
1687 	}
1688 
1689 	/*
1690 	 * Try to terminate the block device.  We can't terminate it if
1691 	 * there are still PFSs referencing it.
1692 	 */
1693 	if (hmp->mount_count)
1694 		return;
1695 
1696 	/*
1697 	 * Decomission the network before we start messing with the
1698 	 * device and PFS.
1699 	 */
1700 	hammer2_iocom_uninit(hmp);
1701 
1702 	hammer2_bulkfree_uninit(hmp);
1703 	hammer2_pfsfree_scan(hmp, 0);
1704 	hammer2_dev_exlock(hmp);	/* XXX order */
1705 
1706 	/*
1707 	 * Cycle the volume data lock as a safety (probably not needed any
1708 	 * more).  To ensure everything is out we need to flush at least
1709 	 * three times.  (1) The running of the sideq can dirty the
1710 	 * filesystem, (2) A normal flush can dirty the freemap, and
1711 	 * (3) ensure that the freemap is fully synchronized.
1712 	 *
1713 	 * The next mount's recovery scan can clean everything up but we want
1714 	 * to leave the filesystem in a 100% clean state on a normal unmount.
1715 	 */
1716 #if 0
1717 	hammer2_voldata_lock(hmp);
1718 	hammer2_voldata_unlock(hmp);
1719 #endif
1720 
1721 	/*
1722 	 * Flush whatever is left.  Unmounted but modified PFS's might still
1723 	 * have some dirty chains on them.
1724 	 */
1725 	hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1726 	hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1727 
1728 	if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1729 		hammer2_voldata_modify(hmp);
1730 		hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1731 					    HAMMER2_FLUSH_ALL);
1732 	}
1733 	hammer2_chain_unlock(&hmp->fchain);
1734 
1735 	if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1736 		hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1737 					    HAMMER2_FLUSH_ALL);
1738 	}
1739 	hammer2_chain_unlock(&hmp->vchain);
1740 
1741 	if ((hmp->vchain.flags | hmp->fchain.flags) &
1742 	    HAMMER2_CHAIN_FLUSH_MASK) {
1743 		kprintf("hammer2_unmount: chains left over "
1744 			"after final sync\n");
1745 		kprintf("    vchain %08x\n", hmp->vchain.flags);
1746 		kprintf("    fchain %08x\n", hmp->fchain.flags);
1747 
1748 		if (hammer2_debug & 0x0010)
1749 			Debugger("entered debugger");
1750 	}
1751 
1752 	hammer2_pfsfree_scan(hmp, 1);
1753 
1754 	KKASSERT(hmp->spmp == NULL);
1755 
1756 	/*
1757 	 * Finish up with the device vnode
1758 	 */
1759 	if ((devvp = hmp->devvp) != NULL) {
1760 		ronly = hmp->ronly;
1761 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1762 		kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1763 			hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1764 			ronly);
1765 		vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1766 		kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1767 			hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1768 		hmp->devvp = NULL;
1769 		VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1770 		vn_unlock(devvp);
1771 		vrele(devvp);
1772 		devvp = NULL;
1773 	}
1774 
1775 	/*
1776 	 * Clear vchain/fchain flags that might prevent final cleanup
1777 	 * of these chains.
1778 	 */
1779 	if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1780 		atomic_add_long(&hammer2_count_modified_chains, -1);
1781 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1782 		hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1783 	}
1784 	if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1785 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1786 	}
1787 
1788 	if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1789 		atomic_add_long(&hammer2_count_modified_chains, -1);
1790 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1791 		hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1792 	}
1793 	if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1794 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1795 	}
1796 
1797 	/*
1798 	 * Final drop of embedded freemap root chain to
1799 	 * clean up fchain.core (fchain structure is not
1800 	 * flagged ALLOCATED so it is cleaned out and then
1801 	 * left to rot).
1802 	 */
1803 	hammer2_chain_drop(&hmp->fchain);
1804 
1805 	/*
1806 	 * Final drop of embedded volume root chain to clean
1807 	 * up vchain.core (vchain structure is not flagged
1808 	 * ALLOCATED so it is cleaned out and then left to
1809 	 * rot).
1810 	 */
1811 	dumpcnt = 50;
1812 	hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1813 	dumpcnt = 50;
1814 	hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1815 	hammer2_dev_unlock(hmp);
1816 	hammer2_chain_drop(&hmp->vchain);
1817 
1818 	hammer2_io_cleanup(hmp, &hmp->iotree);
1819 	if (hmp->iofree_count) {
1820 		kprintf("io_cleanup: %d I/O's left hanging\n",
1821 			hmp->iofree_count);
1822 	}
1823 
1824 	TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1825 	kmalloc_destroy(&hmp->mchain);
1826 	kfree(hmp, M_HAMMER2);
1827 }
1828 
1829 int
1830 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1831 		 ino_t ino, struct vnode **vpp)
1832 {
1833 	hammer2_xop_lookup_t *xop;
1834 	hammer2_pfs_t *pmp;
1835 	hammer2_inode_t *ip;
1836 	hammer2_tid_t inum;
1837 	int error;
1838 
1839 	inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1840 
1841 	error = 0;
1842 	pmp = MPTOPMP(mp);
1843 
1844 	/*
1845 	 * Easy if we already have it cached
1846 	 */
1847 	ip = hammer2_inode_lookup(pmp, inum);
1848 	if (ip) {
1849 		hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1850 		*vpp = hammer2_igetv(ip, &error);
1851 		hammer2_inode_unlock(ip);
1852 		hammer2_inode_drop(ip);		/* from lookup */
1853 
1854 		return error;
1855 	}
1856 
1857 	/*
1858 	 * Otherwise we have to find the inode
1859 	 */
1860 	xop = hammer2_xop_alloc(pmp->iroot, 0);
1861 	xop->lhc = inum;
1862 	hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1863 	error = hammer2_xop_collect(&xop->head, 0);
1864 
1865 	if (error == 0) {
1866 		if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1867 			kprintf("vget: no collect error but also no rdata\n");
1868 			kprintf("xop %p\n", xop);
1869 			while ((hammer2_debug & 0x80000) == 0) {
1870 				tsleep(xop, PCATCH, "wait", hz * 10);
1871 			}
1872 			ip = NULL;
1873 		} else {
1874 			ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1875 		}
1876 	}
1877 	hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1878 
1879 	if (ip) {
1880 		*vpp = hammer2_igetv(ip, &error);
1881 		hammer2_inode_unlock(ip);
1882 	} else {
1883 		*vpp = NULL;
1884 		error = ENOENT;
1885 	}
1886 	return (error);
1887 }
1888 
1889 static
1890 int
1891 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1892 {
1893 	hammer2_pfs_t *pmp;
1894 	struct vnode *vp;
1895 	int error;
1896 
1897 	pmp = MPTOPMP(mp);
1898 	if (pmp->iroot == NULL) {
1899 		*vpp = NULL;
1900 		return EINVAL;
1901 	}
1902 
1903 	error = 0;
1904 	hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1905 
1906 	while (pmp->inode_tid == 0) {
1907 		hammer2_xop_ipcluster_t *xop;
1908 		hammer2_inode_meta_t *meta;
1909 
1910 		xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1911 		hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1912 		error = hammer2_xop_collect(&xop->head, 0);
1913 
1914 		if (error == 0) {
1915 			meta = &xop->head.cluster.focus->data->ipdata.meta;
1916 			pmp->iroot->meta = *meta;
1917 			pmp->inode_tid = meta->pfs_inum + 1;
1918 			if (pmp->inode_tid < HAMMER2_INODE_START)
1919 				pmp->inode_tid = HAMMER2_INODE_START;
1920 			pmp->modify_tid =
1921 				xop->head.cluster.focus->bref.modify_tid + 1;
1922 #if 0
1923 			kprintf("PFS: Starting inode %jd\n",
1924 				(intmax_t)pmp->inode_tid);
1925 			kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1926 				pmp->inode_tid, pmp->modify_tid);
1927 #endif
1928 			wakeup(&pmp->iroot);
1929 
1930 			hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1931 
1932 			/*
1933 			 * Prime the mount info.
1934 			 */
1935 			hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1936 			break;
1937 		}
1938 
1939 		/*
1940 		 * Loop, try again
1941 		 */
1942 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1943 		hammer2_inode_unlock(pmp->iroot);
1944 		error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1945 		hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1946 		if (error == EINTR)
1947 			break;
1948 	}
1949 
1950 	if (error) {
1951 		hammer2_inode_unlock(pmp->iroot);
1952 		*vpp = NULL;
1953 	} else {
1954 		vp = hammer2_igetv(pmp->iroot, &error);
1955 		hammer2_inode_unlock(pmp->iroot);
1956 		*vpp = vp;
1957 	}
1958 
1959 	return (error);
1960 }
1961 
1962 /*
1963  * Filesystem status
1964  *
1965  * XXX incorporate ipdata->meta.inode_quota and data_quota
1966  */
1967 static
1968 int
1969 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1970 {
1971 	hammer2_pfs_t *pmp;
1972 	hammer2_dev_t *hmp;
1973 	hammer2_blockref_t bref;
1974 	struct statfs tmp;
1975 	int i;
1976 
1977 	/*
1978 	 * NOTE: iroot might not have validated the cluster yet.
1979 	 */
1980 	pmp = MPTOPMP(mp);
1981 
1982 	bzero(&tmp, sizeof(tmp));
1983 
1984 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1985 		hmp = pmp->pfs_hmps[i];
1986 		if (hmp == NULL)
1987 			continue;
1988 		if (pmp->iroot->cluster.array[i].chain)
1989 			bref = pmp->iroot->cluster.array[i].chain->bref;
1990 		else
1991 			bzero(&bref, sizeof(bref));
1992 
1993 		tmp.f_files = bref.embed.stats.inode_count;
1994 		tmp.f_ffree = 0;
1995 		tmp.f_blocks = hmp->voldata.allocator_size /
1996 			       mp->mnt_vstat.f_bsize;
1997 		tmp.f_bfree = hmp->voldata.allocator_free /
1998 			      mp->mnt_vstat.f_bsize;
1999 		tmp.f_bavail = tmp.f_bfree;
2000 
2001 		if (cred && cred->cr_uid != 0) {
2002 			uint64_t adj;
2003 
2004 			/* 5% */
2005 			adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2006 			tmp.f_blocks -= adj;
2007 			tmp.f_bfree -= adj;
2008 			tmp.f_bavail -= adj;
2009 		}
2010 
2011 		mp->mnt_stat.f_blocks = tmp.f_blocks;
2012 		mp->mnt_stat.f_bfree = tmp.f_bfree;
2013 		mp->mnt_stat.f_bavail = tmp.f_bavail;
2014 		mp->mnt_stat.f_files = tmp.f_files;
2015 		mp->mnt_stat.f_ffree = tmp.f_ffree;
2016 
2017 		*sbp = mp->mnt_stat;
2018 	}
2019 	return (0);
2020 }
2021 
2022 static
2023 int
2024 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2025 {
2026 	hammer2_pfs_t *pmp;
2027 	hammer2_dev_t *hmp;
2028 	hammer2_blockref_t bref;
2029 	struct statvfs tmp;
2030 	int i;
2031 
2032 	/*
2033 	 * NOTE: iroot might not have validated the cluster yet.
2034 	 */
2035 	pmp = MPTOPMP(mp);
2036 	bzero(&tmp, sizeof(tmp));
2037 
2038 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2039 		hmp = pmp->pfs_hmps[i];
2040 		if (hmp == NULL)
2041 			continue;
2042 		if (pmp->iroot->cluster.array[i].chain)
2043 			bref = pmp->iroot->cluster.array[i].chain->bref;
2044 		else
2045 			bzero(&bref, sizeof(bref));
2046 
2047 		tmp.f_files = bref.embed.stats.inode_count;
2048 		tmp.f_ffree = 0;
2049 		tmp.f_blocks = hmp->voldata.allocator_size /
2050 			       mp->mnt_vstat.f_bsize;
2051 		tmp.f_bfree = hmp->voldata.allocator_free /
2052 			      mp->mnt_vstat.f_bsize;
2053 		tmp.f_bavail = tmp.f_bfree;
2054 
2055 		if (cred && cred->cr_uid != 0) {
2056 			uint64_t adj;
2057 
2058 			/* 5% */
2059 			adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2060 			tmp.f_blocks -= adj;
2061 			tmp.f_bfree -= adj;
2062 			tmp.f_bavail -= adj;
2063 		}
2064 
2065 		mp->mnt_vstat.f_blocks = tmp.f_blocks;
2066 		mp->mnt_vstat.f_bfree = tmp.f_bfree;
2067 		mp->mnt_vstat.f_bavail = tmp.f_bavail;
2068 		mp->mnt_vstat.f_files = tmp.f_files;
2069 		mp->mnt_vstat.f_ffree = tmp.f_ffree;
2070 
2071 		*sbp = mp->mnt_vstat;
2072 	}
2073 	return (0);
2074 }
2075 
2076 /*
2077  * Mount-time recovery (RW mounts)
2078  *
2079  * Updates to the free block table are allowed to lag flushes by one
2080  * transaction.  In case of a crash, then on a fresh mount we must do an
2081  * incremental scan of the last committed transaction id and make sure that
2082  * all related blocks have been marked allocated.
2083  *
2084  * The super-root topology and each PFS has its own transaction id domain,
2085  * so we must track PFS boundary transitions.
2086  */
2087 struct hammer2_recovery_elm {
2088 	TAILQ_ENTRY(hammer2_recovery_elm) entry;
2089 	hammer2_chain_t *chain;
2090 	hammer2_tid_t sync_tid;
2091 };
2092 
2093 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2094 
2095 struct hammer2_recovery_info {
2096 	struct hammer2_recovery_list list;
2097 	hammer2_tid_t	mtid;
2098 	int	depth;
2099 };
2100 
2101 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2102 			hammer2_chain_t *parent,
2103 			struct hammer2_recovery_info *info,
2104 			hammer2_tid_t sync_tid);
2105 
2106 #define HAMMER2_RECOVERY_MAXDEPTH	10
2107 
2108 static
2109 int
2110 hammer2_recovery(hammer2_dev_t *hmp)
2111 {
2112 	struct hammer2_recovery_info info;
2113 	struct hammer2_recovery_elm *elm;
2114 	hammer2_chain_t *parent;
2115 	hammer2_tid_t sync_tid;
2116 	hammer2_tid_t mirror_tid;
2117 	int error;
2118 
2119 	hammer2_trans_init(hmp->spmp, 0);
2120 
2121 	sync_tid = hmp->voldata.freemap_tid;
2122 	mirror_tid = hmp->voldata.mirror_tid;
2123 
2124 	kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2125 	if (sync_tid >= mirror_tid) {
2126 		kprintf(" no recovery needed\n");
2127 	} else {
2128 		kprintf(" freemap recovery %016jx-%016jx\n",
2129 			sync_tid + 1, mirror_tid);
2130 	}
2131 
2132 	TAILQ_INIT(&info.list);
2133 	info.depth = 0;
2134 	parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2135 	error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2136 	hammer2_chain_lookup_done(parent);
2137 
2138 	while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2139 		TAILQ_REMOVE(&info.list, elm, entry);
2140 		parent = elm->chain;
2141 		sync_tid = elm->sync_tid;
2142 		kfree(elm, M_HAMMER2);
2143 
2144 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2145 		error |= hammer2_recovery_scan(hmp, parent, &info,
2146 					      hmp->voldata.freemap_tid);
2147 		hammer2_chain_unlock(parent);
2148 		hammer2_chain_drop(parent);	/* drop elm->chain ref */
2149 	}
2150 
2151 	hammer2_trans_done(hmp->spmp);
2152 
2153 	return error;
2154 }
2155 
2156 static
2157 int
2158 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2159 		      struct hammer2_recovery_info *info,
2160 		      hammer2_tid_t sync_tid)
2161 {
2162 	const hammer2_inode_data_t *ripdata;
2163 	hammer2_chain_t *chain;
2164 	hammer2_blockref_t bref;
2165 	int tmp_error;
2166 	int rup_error;
2167 	int error;
2168 	int first;
2169 
2170 	/*
2171 	 * Adjust freemap to ensure that the block(s) are marked allocated.
2172 	 */
2173 	if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2174 		hammer2_freemap_adjust(hmp, &parent->bref,
2175 				       HAMMER2_FREEMAP_DORECOVER);
2176 	}
2177 
2178 	/*
2179 	 * Check type for recursive scan
2180 	 */
2181 	switch(parent->bref.type) {
2182 	case HAMMER2_BREF_TYPE_VOLUME:
2183 		/* data already instantiated */
2184 		break;
2185 	case HAMMER2_BREF_TYPE_INODE:
2186 		/*
2187 		 * Must instantiate data for DIRECTDATA test and also
2188 		 * for recursion.
2189 		 */
2190 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2191 		ripdata = &hammer2_chain_rdata(parent)->ipdata;
2192 		if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2193 			/* not applicable to recovery scan */
2194 			hammer2_chain_unlock(parent);
2195 			return 0;
2196 		}
2197 		hammer2_chain_unlock(parent);
2198 		break;
2199 	case HAMMER2_BREF_TYPE_INDIRECT:
2200 		/*
2201 		 * Must instantiate data for recursion
2202 		 */
2203 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2204 		hammer2_chain_unlock(parent);
2205 		break;
2206 	case HAMMER2_BREF_TYPE_DIRENT:
2207 	case HAMMER2_BREF_TYPE_DATA:
2208 	case HAMMER2_BREF_TYPE_FREEMAP:
2209 	case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2210 	case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2211 		/* not applicable to recovery scan */
2212 		return 0;
2213 		break;
2214 	default:
2215 		return HAMMER2_ERROR_BADBREF;
2216 	}
2217 
2218 	/*
2219 	 * Defer operation if depth limit reached or if we are crossing a
2220 	 * PFS boundary.
2221 	 */
2222 	if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2223 		struct hammer2_recovery_elm *elm;
2224 
2225 		elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2226 		elm->chain = parent;
2227 		elm->sync_tid = sync_tid;
2228 		hammer2_chain_ref(parent);
2229 		TAILQ_INSERT_TAIL(&info->list, elm, entry);
2230 		/* unlocked by caller */
2231 
2232 		return(0);
2233 	}
2234 
2235 
2236 	/*
2237 	 * Recursive scan of the last flushed transaction only.  We are
2238 	 * doing this without pmp assignments so don't leave the chains
2239 	 * hanging around after we are done with them.
2240 	 *
2241 	 * error	Cumulative error this level only
2242 	 * rup_error	Cumulative error for recursion
2243 	 * tmp_error	Specific non-cumulative recursion error
2244 	 */
2245 	chain = NULL;
2246 	first = 1;
2247 	rup_error = 0;
2248 	error = 0;
2249 
2250 	for (;;) {
2251 		error |= hammer2_chain_scan(parent, &chain, &bref,
2252 					    &first,
2253 					    HAMMER2_LOOKUP_NODATA);
2254 
2255 		/*
2256 		 * Problem during scan or EOF
2257 		 */
2258 		if (error)
2259 			break;
2260 
2261 		/*
2262 		 * If this is a leaf
2263 		 */
2264 		if (chain == NULL) {
2265 			if (bref.mirror_tid > sync_tid) {
2266 				hammer2_freemap_adjust(hmp, &bref,
2267 						     HAMMER2_FREEMAP_DORECOVER);
2268 			}
2269 			continue;
2270 		}
2271 
2272 		/*
2273 		 * This may or may not be a recursive node.
2274 		 */
2275 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2276 		if (bref.mirror_tid > sync_tid) {
2277 			++info->depth;
2278 			tmp_error = hammer2_recovery_scan(hmp, chain,
2279 							   info, sync_tid);
2280 			--info->depth;
2281 		} else {
2282 			tmp_error = 0;
2283 		}
2284 
2285 		/*
2286 		 * Flush the recovery at the PFS boundary to stage it for
2287 		 * the final flush of the super-root topology.
2288 		 */
2289 		if (tmp_error == 0 &&
2290 		    (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2291 		    (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2292 			hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2293 					     HAMMER2_FLUSH_ALL);
2294 		}
2295 		rup_error |= tmp_error;
2296 	}
2297 	return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2298 }
2299 
2300 /*
2301  * This fixes up an error introduced in earlier H2 implementations where
2302  * moving a PFS inode into an indirect block wound up causing the
2303  * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2304  */
2305 static
2306 int
2307 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2308 {
2309 	const hammer2_inode_data_t *ripdata;
2310 	hammer2_chain_t *parent;
2311 	hammer2_chain_t *chain;
2312 	hammer2_key_t key_next;
2313 	hammer2_pfs_t *spmp;
2314 	int error;
2315 
2316 	error = 0;
2317 
2318 	/*
2319 	 * Lookup mount point under the media-localized super-root.
2320 	 *
2321 	 * cluster->pmp will incorrectly point to spmp and must be fixed
2322 	 * up later on.
2323 	 */
2324 	spmp = hmp->spmp;
2325 	hammer2_inode_lock(spmp->iroot, 0);
2326 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2327 	chain = hammer2_chain_lookup(&parent, &key_next,
2328 					 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2329 					 &error, 0);
2330 	while (chain) {
2331 		if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2332 			continue;
2333 		if (chain->error) {
2334 			kprintf("I/O error scanning PFS labels\n");
2335 			error |= chain->error;
2336 		} else if ((chain->bref.flags &
2337 			    HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2338 			int error2;
2339 
2340 			ripdata = &chain->data->ipdata;
2341 			hammer2_trans_init(hmp->spmp, 0);
2342 			error2 = hammer2_chain_modify(chain,
2343 						      chain->bref.modify_tid,
2344 						      0, 0);
2345 			if (error2 == 0) {
2346 				kprintf("hammer2: Correct mis-flagged PFS %s\n",
2347 					ripdata->filename);
2348 				chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2349 			} else {
2350 				error |= error2;
2351 			}
2352 			hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2353 					     HAMMER2_FLUSH_ALL);
2354 			hammer2_trans_done(hmp->spmp);
2355 		}
2356 		chain = hammer2_chain_next(&parent, chain, &key_next,
2357 					   key_next, HAMMER2_KEY_MAX,
2358 					   &error, 0);
2359 	}
2360 	if (parent) {
2361 		hammer2_chain_unlock(parent);
2362 		hammer2_chain_drop(parent);
2363 	}
2364 	hammer2_inode_unlock(spmp->iroot);
2365 
2366 	return error;
2367 }
2368 
2369 /*
2370  * Sync a mount point; this is called periodically on a per-mount basis from
2371  * the filesystem syncer, and whenever a user issues a sync.
2372  */
2373 int
2374 hammer2_vfs_sync(struct mount *mp, int waitfor)
2375 {
2376 	hammer2_xop_flush_t *xop;
2377 	struct hammer2_sync_info info;
2378 	hammer2_inode_t *iroot;
2379 	hammer2_pfs_t *pmp;
2380 	int flags;
2381 	int error;
2382 
2383 	pmp = MPTOPMP(mp);
2384 	iroot = pmp->iroot;
2385 	KKASSERT(iroot);
2386 	KKASSERT(iroot->pmp == pmp);
2387 
2388 	/*
2389 	 * We can't acquire locks on existing vnodes while in a transaction
2390 	 * without risking a deadlock.  This assumes that vfsync() can be
2391 	 * called without the vnode locked (which it can in DragonFly).
2392 	 * Otherwise we'd have to implement a multi-pass or flag the lock
2393 	 * failures and retry.
2394 	 *
2395 	 * The reclamation code interlocks with the sync list's token
2396 	 * (by removing the vnode from the scan list) before unlocking
2397 	 * the inode, giving us time to ref the inode.
2398 	 */
2399 	/*flags = VMSC_GETVP;*/
2400 	flags = 0;
2401 	if (waitfor & MNT_LAZY)
2402 		flags |= VMSC_ONEPASS;
2403 
2404 	/*
2405 	 * Flush vnodes individually using a normal transaction to avoid
2406 	 * stalling any concurrent operations.  This will flush the related
2407 	 * buffer cache buffers and inodes to the media.
2408 	 *
2409 	 * For efficiency do an async pass before making sure with a
2410 	 * synchronous pass on all related buffer cache buffers.
2411 	 */
2412 	hammer2_trans_init(pmp, 0);
2413 
2414 	info.error = 0;
2415 
2416 	info.waitfor = MNT_NOWAIT;
2417 	info.pass = 1;
2418 	vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2419 
2420 	/*
2421 	 * Now do two passes making sure we get everything.  The first pass
2422 	 * vfsync()s dirty vnodes.  The second pass waits for their I/O's
2423 	 * to finish and cleans up the dirty flag on the vnode.
2424 	 */
2425 	info.pass = 1;
2426 	info.waitfor = MNT_WAIT;
2427 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2428 
2429 	info.pass = 2;
2430 	info.waitfor = MNT_WAIT;
2431 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2432 
2433 	/*
2434 	 * We must also run the sideq to handle any disconnected inodes
2435 	 * as the vnode scan will not see these.
2436 	 */
2437 	hammer2_inode_run_sideq(pmp, 1);
2438 	hammer2_trans_done(pmp);
2439 
2440 	/*
2441 	 * Start our flush transaction and flush the root topology down to
2442 	 * the inodes, but not the inodes themselves (which we already flushed
2443 	 * above).  Any concurrent activity effecting inode contents will not
2444 	 *
2445 	 * The flush sequence will
2446 	 *
2447 	 * NOTE! It is still possible for the paging code to push pages
2448 	 *	 out via a UIO_NOCOPY hammer2_vop_write() during the main
2449 	 *	 flush.
2450 	 */
2451 	hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2452 
2453 	/*
2454 	 * sync dirty vnodes again while in the flush transaction.  This is
2455 	 * currently an expensive shim to makre sure the logical topology is
2456 	 * completely consistent before we flush the volume header.
2457 	 */
2458 	info.pass = 1;
2459 	info.waitfor = MNT_WAIT;
2460 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2461 
2462 	info.pass = 2;
2463 	info.waitfor = MNT_WAIT;
2464 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2465 
2466 	/*
2467 	 * Use the XOP interface to concurrently flush all nodes to
2468 	 * synchronize the PFSROOT subtopology to the media.  A standard
2469 	 * end-of-scan ENOENT error indicates cluster sufficiency.
2470 	 *
2471 	 * Note that this flush will not be visible on crash recovery until
2472 	 * we flush the super-root topology in the next loop.
2473 	 *
2474 	 * XXX For now wait for all flushes to complete.
2475 	 */
2476 	if (iroot) {
2477 		/*
2478 		 * If unmounting try to flush everything including any
2479 		 * sub-trees under inodes, just in case there is dangling
2480 		 * modified data, as a safety.  Otherwise just flush up to
2481 		 * the inodes in this stage.
2482 		 */
2483 		if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2484 			xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2485 						       HAMMER2_XOP_VOLHDR);
2486 		} else {
2487 			xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING |
2488 						       HAMMER2_XOP_INODE_STOP |
2489 						       HAMMER2_XOP_VOLHDR);
2490 		}
2491 		hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2492 		error = hammer2_xop_collect(&xop->head,
2493 					    HAMMER2_XOP_COLLECT_WAITALL);
2494 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2495 		if (error == HAMMER2_ERROR_ENOENT)
2496 			error = 0;
2497 		else
2498 			error = hammer2_error_to_errno(error);
2499 	} else {
2500 		error = 0;
2501 	}
2502 	hammer2_trans_done(pmp);
2503 
2504 	return (error);
2505 }
2506 
2507 /*
2508  * Sync passes.
2509  *
2510  * Note that we ignore the tranasction mtid we got above.  Instead,
2511  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2512  * transactions.
2513  *
2514  * WARNING! The frontend might be waiting on chnmem (limit_dirty_chains)
2515  * while holding a vnode locked.  When this situation occurs we cannot
2516  * safely test whether it is ok to clear the dirty bit on the vnode.
2517  * However, we can still flush the inode's topology.
2518  */
2519 static int
2520 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2521 {
2522 	struct hammer2_sync_info *info = data;
2523 	hammer2_inode_t *ip;
2524 	int error;
2525 
2526 	/*
2527 	 * Degenerate cases.  Note that ip == NULL typically means the
2528 	 * syncer vnode itself and we don't want to vclrisdirty() in that
2529 	 * situation.
2530 	 */
2531 	ip = VTOI(vp);
2532 	if (ip == NULL) {
2533 		return(0);
2534 	}
2535 	if (vp->v_type == VNON || vp->v_type == VBAD) {
2536 		vclrisdirty(vp);
2537 		return(0);
2538 	}
2539 
2540 	/*
2541 	 * Synchronize the buffer cche and inode meta-data to the backing
2542 	 * chain topology.
2543 	 *
2544 	 * vfsync is not necessarily synchronous, so it is best NOT to try
2545 	 * to flush the backing topology to media at this point.
2546 	 */
2547 	hammer2_inode_ref(ip);
2548 	if ((ip->flags & (HAMMER2_INODE_RESIZED|HAMMER2_INODE_MODIFIED)) ||
2549 	    !RB_EMPTY(&vp->v_rbdirty_tree)) {
2550 		if (info->pass == 1)
2551 			vfsync(vp, info->waitfor, 1, NULL, NULL);
2552 		else
2553 			bio_track_wait(&vp->v_track_write, 0, 0);
2554 	}
2555 	if (info->pass == 2 && (vp->v_flag & VISDIRTY)) {
2556 		/*
2557 		 * v_token is needed to interlock v_rbdirty_tree.
2558 		 */
2559 		lwkt_gettoken(&vp->v_token);
2560 		hammer2_inode_lock(ip, 0);
2561 		hammer2_inode_chain_sync(ip);
2562 		hammer2_inode_chain_flush(ip);
2563 		if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2564 				  HAMMER2_INODE_RESIZED |
2565 				  HAMMER2_INODE_DIRTYDATA)) == 0 &&
2566 		    RB_EMPTY(&vp->v_rbdirty_tree) &&
2567 		    !bio_track_active(&vp->v_track_write)) {
2568 			vclrisdirty(vp);
2569 		}
2570 		hammer2_inode_unlock(ip);
2571 		lwkt_reltoken(&vp->v_token);
2572 	}
2573 	hammer2_inode_drop(ip);
2574 #if 1
2575 	error = 0;
2576 	if (error)
2577 		info->error = error;
2578 #endif
2579 	return(0);
2580 }
2581 
2582 static
2583 int
2584 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2585 {
2586 	hammer2_inode_t *ip;
2587 
2588 	KKASSERT(MAXFIDSZ >= 16);
2589 	ip = VTOI(vp);
2590 	fhp->fid_len = offsetof(struct fid, fid_data[16]);
2591 	fhp->fid_ext = 0;
2592 	((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2593 	((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2594 
2595 	return 0;
2596 }
2597 
2598 static
2599 int
2600 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2601 	       struct fid *fhp, struct vnode **vpp)
2602 {
2603 	hammer2_pfs_t *pmp;
2604 	hammer2_tid_t inum;
2605 	int error;
2606 
2607 	pmp = MPTOPMP(mp);
2608 	inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2609 	if (vpp) {
2610 		if (inum == 1)
2611 			error = hammer2_vfs_root(mp, vpp);
2612 		else
2613 			error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2614 	} else {
2615 		error = 0;
2616 	}
2617 	if (error)
2618 		kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2619 	return error;
2620 }
2621 
2622 static
2623 int
2624 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2625 		 int *exflagsp, struct ucred **credanonp)
2626 {
2627 	hammer2_pfs_t *pmp;
2628 	struct netcred *np;
2629 	int error;
2630 
2631 	pmp = MPTOPMP(mp);
2632 	np = vfs_export_lookup(mp, &pmp->export, nam);
2633 	if (np) {
2634 		*exflagsp = np->netc_exflags;
2635 		*credanonp = &np->netc_anon;
2636 		error = 0;
2637 	} else {
2638 		error = EACCES;
2639 	}
2640 	return error;
2641 }
2642 
2643 /*
2644  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2645  * header into the HMP
2646  *
2647  * XXX read four volhdrs and use the one with the highest TID whos CRC
2648  *     matches.
2649  *
2650  * XXX check iCRCs.
2651  *
2652  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2653  *     nonexistant locations.
2654  *
2655  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2656  */
2657 static
2658 int
2659 hammer2_install_volume_header(hammer2_dev_t *hmp)
2660 {
2661 	hammer2_volume_data_t *vd;
2662 	struct buf *bp;
2663 	hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2664 	int error_reported;
2665 	int error;
2666 	int valid;
2667 	int i;
2668 
2669 	error_reported = 0;
2670 	error = 0;
2671 	valid = 0;
2672 	bp = NULL;
2673 
2674 	/*
2675 	 * There are up to 4 copies of the volume header (syncs iterate
2676 	 * between them so there is no single master).  We don't trust the
2677 	 * volu_size field so we don't know precisely how large the filesystem
2678 	 * is, so depend on the OS to return an error if we go beyond the
2679 	 * block device's EOF.
2680 	 */
2681 	for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2682 		error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2683 			      HAMMER2_VOLUME_BYTES, &bp);
2684 		if (error) {
2685 			brelse(bp);
2686 			bp = NULL;
2687 			continue;
2688 		}
2689 
2690 		vd = (struct hammer2_volume_data *) bp->b_data;
2691 		if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2692 		    (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2693 			brelse(bp);
2694 			bp = NULL;
2695 			continue;
2696 		}
2697 
2698 		if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2699 			/* XXX: Reversed-endianness filesystem */
2700 			kprintf("hammer2: reverse-endian filesystem detected");
2701 			brelse(bp);
2702 			bp = NULL;
2703 			continue;
2704 		}
2705 
2706 		crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2707 		crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2708 				      HAMMER2_VOLUME_ICRC0_SIZE);
2709 		bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2710 		bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2711 				       HAMMER2_VOLUME_ICRC1_SIZE);
2712 		if ((crc0 != crc) || (bcrc0 != bcrc)) {
2713 			kprintf("hammer2 volume header crc "
2714 				"mismatch copy #%d %08x/%08x\n",
2715 				i, crc0, crc);
2716 			error_reported = 1;
2717 			brelse(bp);
2718 			bp = NULL;
2719 			continue;
2720 		}
2721 		if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2722 			valid = 1;
2723 			hmp->voldata = *vd;
2724 			hmp->volhdrno = i;
2725 		}
2726 		brelse(bp);
2727 		bp = NULL;
2728 	}
2729 	if (valid) {
2730 		hmp->volsync = hmp->voldata;
2731 		hmp->free_reserved = hmp->voldata.allocator_size / 20;
2732 		error = 0;
2733 		if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2734 			kprintf("hammer2: using volume header #%d\n",
2735 				hmp->volhdrno);
2736 		}
2737 	} else {
2738 		error = EINVAL;
2739 		kprintf("hammer2: no valid volume headers found!\n");
2740 	}
2741 	return (error);
2742 }
2743 
2744 /*
2745  * This handles hysteresis on regular file flushes.  Because the BIOs are
2746  * routed to a thread it is possible for an excessive number to build up
2747  * and cause long front-end stalls long before the runningbuffspace limit
2748  * is hit, so we implement hammer2_flush_pipe to control the
2749  * hysteresis.
2750  *
2751  * This is a particular problem when compression is used.
2752  */
2753 void
2754 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2755 {
2756 	atomic_add_int(&pmp->count_lwinprog, 1);
2757 }
2758 
2759 void
2760 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2761 {
2762 	int lwinprog;
2763 
2764 	lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2765 	if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2766 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2767 		atomic_clear_int(&pmp->count_lwinprog,
2768 				 HAMMER2_LWINPROG_WAITING);
2769 		wakeup(&pmp->count_lwinprog);
2770 	}
2771 	if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2772 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2773 		atomic_clear_int(&pmp->count_lwinprog,
2774 				 HAMMER2_LWINPROG_WAITING0);
2775 		wakeup(&pmp->count_lwinprog);
2776 	}
2777 }
2778 
2779 void
2780 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2781 {
2782 	int lwinprog;
2783 	int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2784 				    HAMMER2_LWINPROG_WAITING0;
2785 
2786 	for (;;) {
2787 		lwinprog = pmp->count_lwinprog;
2788 		cpu_ccfence();
2789 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2790 			break;
2791 		tsleep_interlock(&pmp->count_lwinprog, 0);
2792 		atomic_set_int(&pmp->count_lwinprog, lwflag);
2793 		lwinprog = pmp->count_lwinprog;
2794 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2795 			break;
2796 		tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2797 	}
2798 }
2799 
2800 /*
2801  * Manage excessive memory resource use for chain and related
2802  * structures.
2803  */
2804 void
2805 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2806 {
2807 	uint32_t waiting;
2808 	uint32_t count;
2809 	uint32_t limit;
2810 #if 0
2811 	static int zzticks;
2812 #endif
2813 
2814 	/*
2815 	 * Atomic check condition and wait.  Also do an early speedup of
2816 	 * the syncer to try to avoid hitting the wait.
2817 	 */
2818 	for (;;) {
2819 		waiting = pmp->inmem_dirty_chains;
2820 		cpu_ccfence();
2821 		count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2822 
2823 		limit = pmp->mp->mnt_nvnodelistsize / 10;
2824 		if (limit < hammer2_limit_dirty_chains)
2825 			limit = hammer2_limit_dirty_chains;
2826 		if (limit < 1000)
2827 			limit = 1000;
2828 
2829 #if 0
2830 		if ((int)(ticks - zzticks) > hz) {
2831 			zzticks = ticks;
2832 			kprintf("count %ld %ld\n", count, limit);
2833 		}
2834 #endif
2835 
2836 		/*
2837 		 * Block if there are too many dirty chains present, wait
2838 		 * for the flush to clean some out.
2839 		 */
2840 		if (count > limit) {
2841 			tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2842 			if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2843 					       waiting,
2844 				       waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2845 				speedup_syncer(pmp->mp);
2846 				tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2847 				       "chnmem", hz);
2848 			}
2849 			continue;	/* loop on success or fail */
2850 		}
2851 
2852 		/*
2853 		 * Try to start an early flush before we are forced to block.
2854 		 */
2855 		if (count > limit * 5 / 10)
2856 			speedup_syncer(pmp->mp);
2857 		break;
2858 	}
2859 }
2860 
2861 void
2862 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2863 {
2864 	if (pmp) {
2865 		atomic_add_int(&pmp->inmem_dirty_chains, 1);
2866 	}
2867 }
2868 
2869 void
2870 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2871 {
2872 	uint32_t waiting;
2873 
2874 	if (pmp) {
2875 		waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
2876 		/* don't need --waiting to test flag */
2877 		if (waiting & HAMMER2_DIRTYCHAIN_WAITING) {
2878 			atomic_clear_int(&pmp->inmem_dirty_chains,
2879 					 HAMMER2_DIRTYCHAIN_WAITING);
2880 			wakeup(&pmp->inmem_dirty_chains);
2881 		}
2882 	}
2883 }
2884 
2885 /*
2886  * Returns 0 if the filesystem has tons of free space
2887  * Returns 1 if the filesystem has less than 10% remaining
2888  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
2889  */
2890 int
2891 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
2892 {
2893 	hammer2_pfs_t *pmp;
2894 	hammer2_dev_t *hmp;
2895 	hammer2_off_t free_reserved;
2896 	hammer2_off_t free_nominal;
2897 	int i;
2898 
2899 	pmp = ip->pmp;
2900 
2901 	if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
2902 		free_reserved = HAMMER2_SEGSIZE;
2903 		free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
2904 		for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2905 			hmp = pmp->pfs_hmps[i];
2906 			if (hmp == NULL)
2907 				continue;
2908 			if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
2909 			    pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
2910 				continue;
2911 
2912 			if (free_nominal > hmp->voldata.allocator_free)
2913 				free_nominal = hmp->voldata.allocator_free;
2914 			if (free_reserved < hmp->free_reserved)
2915 				free_reserved = hmp->free_reserved;
2916 		}
2917 
2918 		/*
2919 		 * SMP races ok
2920 		 */
2921 		pmp->free_reserved = free_reserved;
2922 		pmp->free_nominal = free_nominal;
2923 		pmp->free_ticks = ticks;
2924 	} else {
2925 		free_reserved = pmp->free_reserved;
2926 		free_nominal = pmp->free_nominal;
2927 	}
2928 	if (cred && cred->cr_uid != 0) {
2929 		if ((int64_t)(free_nominal - bytes) <
2930 		    (int64_t)free_reserved) {
2931 			return 2;
2932 		}
2933 	} else {
2934 		if ((int64_t)(free_nominal - bytes) <
2935 		    (int64_t)free_reserved / 2) {
2936 			return 2;
2937 		}
2938 	}
2939 	if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
2940 		return 1;
2941 	return 0;
2942 }
2943 
2944 /*
2945  * Debugging
2946  */
2947 void
2948 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
2949 		   u_int flags)
2950 {
2951 	hammer2_chain_t *scan;
2952 	hammer2_chain_t *parent;
2953 
2954 	--*countp;
2955 	if (*countp == 0) {
2956 		kprintf("%*.*s...\n", tab, tab, "");
2957 		return;
2958 	}
2959 	if (*countp < 0)
2960 		return;
2961 	kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2962 		tab, tab, "", pfx,
2963 		chain, chain->bref.type,
2964 		chain->bref.key, chain->bref.keybits,
2965 		chain->bref.mirror_tid);
2966 
2967 	kprintf("%*.*s      [%08x] (%s) refs=%d",
2968 		tab, tab, "",
2969 		chain->flags,
2970 		((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2971 		chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2972 		chain->refs);
2973 
2974 	parent = chain->parent;
2975 	if (parent)
2976 		kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2977 			tab, tab, "",
2978 			parent, parent->flags, parent->refs);
2979 	if (RB_EMPTY(&chain->core.rbtree)) {
2980 		kprintf("\n");
2981 	} else {
2982 		kprintf(" {\n");
2983 		RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
2984 			if ((scan->flags & flags) || flags == (u_int)-1) {
2985 				hammer2_dump_chain(scan, tab + 4, countp, 'a',
2986 						   flags);
2987 			}
2988 		}
2989 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2990 			kprintf("%*.*s}(%s)\n", tab, tab, "",
2991 				chain->data->ipdata.filename);
2992 		else
2993 			kprintf("%*.*s}\n", tab, tab, "");
2994 	}
2995 }
2996