xref: /dflybsd-src/sys/vfs/hammer2/hammer2_vfsops.c (revision e91e64c7af5788faa55682cd78c0442c83d5d6d5)
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/mountctl.h>
51 #include <sys/dirent.h>
52 #include <sys/uio.h>
53 
54 #include "hammer2.h"
55 #include "hammer2_disk.h"
56 #include "hammer2_mount.h"
57 #include "hammer2_lz4.h"
58 
59 #include "zlib/hammer2_zlib.h"
60 
61 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
62 
63 struct hammer2_sync_info {
64 	int error;
65 	int waitfor;
66 	int pass;
67 };
68 
69 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
70 static struct hammer2_mntlist hammer2_mntlist;
71 
72 struct hammer2_pfslist hammer2_pfslist;
73 struct hammer2_pfslist hammer2_spmplist;
74 struct lock hammer2_mntlk;
75 
76 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
77 int hammer2_debug;
78 int hammer2_xopgroups;
79 long hammer2_debug_inode;
80 int hammer2_cluster_meta_read = 1;	/* physical read-ahead */
81 int hammer2_cluster_data_read = 4;	/* physical read-ahead */
82 int hammer2_cluster_write = 0;		/* physical write clustering */
83 int hammer2_dedup_enable = 1;
84 int hammer2_always_compress = 0;	/* always try to compress */
85 int hammer2_flush_pipe = 100;
86 int hammer2_dio_count;
87 int hammer2_dio_limit = 256;
88 int hammer2_bulkfree_tps = 5000;
89 int hammer2_worker_rmask = 3;
90 long hammer2_chain_allocs;
91 long hammer2_limit_dirty_chains;
92 long hammer2_limit_dirty_inodes;
93 long hammer2_count_modified_chains;
94 long hammer2_iod_file_read;
95 long hammer2_iod_meta_read;
96 long hammer2_iod_indr_read;
97 long hammer2_iod_fmap_read;
98 long hammer2_iod_volu_read;
99 long hammer2_iod_file_write;
100 long hammer2_iod_file_wembed;
101 long hammer2_iod_file_wzero;
102 long hammer2_iod_file_wdedup;
103 long hammer2_iod_meta_write;
104 long hammer2_iod_indr_write;
105 long hammer2_iod_fmap_write;
106 long hammer2_iod_volu_write;
107 static long hammer2_iod_inode_creates;
108 static long hammer2_iod_inode_deletes;
109 
110 long hammer2_process_icrc32;
111 long hammer2_process_xxhash64;
112 
113 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
114 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
115 		"Buffer used for compression.");
116 
117 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
118 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
119 		"Buffer used for decompression.");
120 
121 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
122 
123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
124 	   &hammer2_supported_version, 0, "");
125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
126 	   &hammer2_debug, 0, "");
127 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW,
128 	   &hammer2_debug_inode, 0, "");
129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
130 	   &hammer2_cluster_meta_read, 0, "");
131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
132 	   &hammer2_cluster_data_read, 0, "");
133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
134 	   &hammer2_cluster_write, 0, "");
135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
136 	   &hammer2_dedup_enable, 0, "");
137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
138 	   &hammer2_always_compress, 0, "");
139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
140 	   &hammer2_flush_pipe, 0, "");
141 SYSCTL_INT(_vfs_hammer2, OID_AUTO, worker_rmask, CTLFLAG_RW,
142 	   &hammer2_worker_rmask, 0, "");
143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
144 	   &hammer2_bulkfree_tps, 0, "");
145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
146 	   &hammer2_chain_allocs, 0, "");
147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
148 	   &hammer2_limit_dirty_chains, 0, "");
149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
150 	   &hammer2_limit_dirty_inodes, 0, "");
151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
152 	   &hammer2_count_modified_chains, 0, "");
153 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
154 	   &hammer2_dio_count, 0, "");
155 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
156 	   &hammer2_dio_limit, 0, "");
157 
158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
159 	   &hammer2_iod_file_read, 0, "");
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
161 	   &hammer2_iod_meta_read, 0, "");
162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
163 	   &hammer2_iod_indr_read, 0, "");
164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
165 	   &hammer2_iod_fmap_read, 0, "");
166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
167 	   &hammer2_iod_volu_read, 0, "");
168 
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
170 	   &hammer2_iod_file_write, 0, "");
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
172 	   &hammer2_iod_file_wembed, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
174 	   &hammer2_iod_file_wzero, 0, "");
175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
176 	   &hammer2_iod_file_wdedup, 0, "");
177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
178 	   &hammer2_iod_meta_write, 0, "");
179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
180 	   &hammer2_iod_indr_write, 0, "");
181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
182 	   &hammer2_iod_fmap_write, 0, "");
183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
184 	   &hammer2_iod_volu_write, 0, "");
185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW,
186 	   &hammer2_iod_inode_creates, 0, "");
187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW,
188 	   &hammer2_iod_inode_deletes, 0, "");
189 
190 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
191 	   &hammer2_process_icrc32, 0, "");
192 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
193 	   &hammer2_process_xxhash64, 0, "");
194 
195 static int hammer2_vfs_init(struct vfsconf *conf);
196 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
197 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
198 				struct ucred *cred);
199 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
200 				struct ucred *);
201 static int hammer2_recovery(hammer2_dev_t *hmp);
202 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
203 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
204 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
205 				struct ucred *cred);
206 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
207 				struct ucred *cred);
208 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
209 				struct fid *fhp, struct vnode **vpp);
210 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
211 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
212 				int *exflagsp, struct ucred **credanonp);
213 static int hammer2_vfs_modifying(struct mount *mp);
214 
215 static void hammer2_update_pmps(hammer2_dev_t *hmp);
216 
217 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
218 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
219 				hammer2_dev_t *hmp);
220 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
221 
222 /*
223  * HAMMER2 vfs operations.
224  */
225 static struct vfsops hammer2_vfsops = {
226 	.vfs_flags	= 0,
227 	.vfs_init	= hammer2_vfs_init,
228 	.vfs_uninit	= hammer2_vfs_uninit,
229 	.vfs_sync	= hammer2_vfs_sync,
230 	.vfs_mount	= hammer2_vfs_mount,
231 	.vfs_unmount	= hammer2_vfs_unmount,
232 	.vfs_root 	= hammer2_vfs_root,
233 	.vfs_statfs	= hammer2_vfs_statfs,
234 	.vfs_statvfs	= hammer2_vfs_statvfs,
235 	.vfs_vget	= hammer2_vfs_vget,
236 	.vfs_vptofh	= hammer2_vfs_vptofh,
237 	.vfs_fhtovp	= hammer2_vfs_fhtovp,
238 	.vfs_checkexp	= hammer2_vfs_checkexp,
239 	.vfs_modifying	= hammer2_vfs_modifying
240 };
241 
242 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
243 
244 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
245 MODULE_VERSION(hammer2, 1);
246 
247 static
248 int
249 hammer2_vfs_init(struct vfsconf *conf)
250 {
251 	static struct objcache_malloc_args margs_read;
252 	static struct objcache_malloc_args margs_write;
253 	static struct objcache_malloc_args margs_vop;
254 
255 	int error;
256 
257 	error = 0;
258 	kmalloc_raise_limit(M_HAMMER2, 0);	/* unlimited */
259 
260 	/*
261 	 * hammer2_xopgroups must be even and is most optimal if
262 	 * 2 x ncpus so strategy functions can be queued to the same
263 	 * cpu.
264 	 */
265 	hammer2_xopgroups = HAMMER2_XOPGROUPS_MIN;
266 	if (hammer2_xopgroups < ncpus * 2)
267 		hammer2_xopgroups = ncpus * 2;
268 
269 	/*
270 	 * A large DIO cache is needed to retain dedup enablement masks.
271 	 * The bulkfree code clears related masks as part of the disk block
272 	 * recycling algorithm, preventing it from being used for a later
273 	 * dedup.
274 	 *
275 	 * NOTE: A large buffer cache can actually interfere with dedup
276 	 *	 operation because we dedup based on media physical buffers
277 	 *	 and not logical buffers.  Try to make the DIO case large
278 	 *	 enough to avoid this problem, but also cap it.
279 	 */
280 	hammer2_dio_limit = nbuf * 2;
281 	if (hammer2_dio_limit > 100000)
282 		hammer2_dio_limit = 100000;
283 
284 	if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
285 		error = EINVAL;
286 	if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
287 		error = EINVAL;
288 	if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
289 		error = EINVAL;
290 
291 	if (error)
292 		kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
293 
294 	margs_read.objsize = 65536;
295 	margs_read.mtype = M_HAMMER2_DEBUFFER;
296 
297 	margs_write.objsize = 32768;
298 	margs_write.mtype = M_HAMMER2_CBUFFER;
299 
300 	margs_vop.objsize = sizeof(hammer2_xop_t);
301 	margs_vop.mtype = M_HAMMER2;
302 
303 	/*
304 	 * Note thaht for the XOPS cache we want backing store allocations
305 	 * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
306 	 * confusion), so use the backing store function that does it.  This
307 	 * means that initial XOPS objects are zerod but REUSED objects are
308 	 * not.  So we are responsible for cleaning the object up sufficiently
309 	 * for our needs before objcache_put()ing it back (typically just the
310 	 * FIFO indices).
311 	 */
312 	cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
313 				0, 1, NULL, NULL, NULL,
314 				objcache_malloc_alloc,
315 				objcache_malloc_free,
316 				&margs_read);
317 	cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
318 				0, 1, NULL, NULL, NULL,
319 				objcache_malloc_alloc,
320 				objcache_malloc_free,
321 				&margs_write);
322 	cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
323 				0, 1, NULL, NULL, NULL,
324 				objcache_malloc_alloc_zero,
325 				objcache_malloc_free,
326 				&margs_vop);
327 
328 
329 	lockinit(&hammer2_mntlk, "mntlk", 0, 0);
330 	TAILQ_INIT(&hammer2_mntlist);
331 	TAILQ_INIT(&hammer2_pfslist);
332 	TAILQ_INIT(&hammer2_spmplist);
333 
334 	hammer2_limit_dirty_chains = maxvnodes / 10;
335 	if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
336 		hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
337 	if (hammer2_limit_dirty_chains < 1000)
338 		hammer2_limit_dirty_chains = 1000;
339 
340 	hammer2_limit_dirty_inodes = maxvnodes / 25;
341 	if (hammer2_limit_dirty_inodes < 100)
342 		hammer2_limit_dirty_inodes = 100;
343 	if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
344 		hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
345 
346 	return (error);
347 }
348 
349 static
350 int
351 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
352 {
353 	objcache_destroy(cache_buffer_read);
354 	objcache_destroy(cache_buffer_write);
355 	objcache_destroy(cache_xops);
356 	return 0;
357 }
358 
359 /*
360  * Core PFS allocator.  Used to allocate or reference the pmp structure
361  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
362  * The pmp can be passed in or loaded by this function using the chain and
363  * inode data.
364  *
365  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
366  * transactions.  Note that synchronization does not use this field.
367  * (typically frontend operations and synchronization cannot run on the
368  * same PFS node at the same time).
369  *
370  * XXX check locking
371  */
372 hammer2_pfs_t *
373 hammer2_pfsalloc(hammer2_chain_t *chain,
374 		 const hammer2_inode_data_t *ripdata,
375 		 hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
376 {
377 	hammer2_pfs_t *pmp;
378 	hammer2_inode_t *iroot;
379 	int count;
380 	int i;
381 	int j;
382 
383 	pmp = NULL;
384 
385 	/*
386 	 * Locate or create the PFS based on the cluster id.  If ripdata
387 	 * is NULL this is a spmp which is unique and is always allocated.
388 	 *
389 	 * If the device is mounted in local mode all PFSs are considered
390 	 * independent and not part of any cluster (for debugging only).
391 	 */
392 	if (ripdata) {
393 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
394 			if (force_local != pmp->force_local)
395 				continue;
396 			if (force_local == NULL &&
397 			    bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
398 				 sizeof(pmp->pfs_clid)) == 0) {
399 					break;
400 			} else if (force_local && pmp->pfs_names[0] &&
401 			    strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
402 					break;
403 			}
404 		}
405 	}
406 
407 	if (pmp == NULL) {
408 		pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
409 		pmp->force_local = force_local;
410 		hammer2_trans_manage_init(pmp);
411 		kmalloc_create_obj(&pmp->minode, "HAMMER2-inodes",
412 				   sizeof(struct hammer2_inode));
413 		lockinit(&pmp->lock, "pfslk", 0, 0);
414 		lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
415 		spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
416 		spin_init(&pmp->xop_spin, "h2xop");
417 		spin_init(&pmp->lru_spin, "h2lru");
418 		RB_INIT(&pmp->inum_tree);
419 		TAILQ_INIT(&pmp->syncq);
420 		TAILQ_INIT(&pmp->depq);
421 		TAILQ_INIT(&pmp->lru_list);
422 		spin_init(&pmp->list_spin, "h2pfsalloc_list");
423 
424 		/*
425 		 * Save the last media transaction id for the flusher.  Set
426 		 * initial
427 		 */
428 		if (ripdata) {
429 			pmp->pfs_clid = ripdata->meta.pfs_clid;
430 			TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
431 		} else {
432 			pmp->flags |= HAMMER2_PMPF_SPMP;
433 			TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
434 		}
435 
436 		/*
437 		 * The synchronization thread may start too early, make
438 		 * sure it stays frozen until we are ready to let it go.
439 		 * XXX
440 		 */
441 		/*
442 		pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
443 					 HAMMER2_THREAD_REMASTER;
444 		*/
445 	}
446 
447 	/*
448 	 * Create the PFS's root inode and any missing XOP helper threads.
449 	 */
450 	if ((iroot = pmp->iroot) == NULL) {
451 		iroot = hammer2_inode_get(pmp, NULL, 1, -1);
452 		if (ripdata)
453 			iroot->meta = ripdata->meta;
454 		pmp->iroot = iroot;
455 		hammer2_inode_ref(iroot);
456 		hammer2_inode_unlock(iroot);
457 	}
458 
459 	/*
460 	 * Stop here if no chain is passed in.
461 	 */
462 	if (chain == NULL)
463 		goto done;
464 
465 	/*
466 	 * When a chain is passed in we must add it to the PFS's root
467 	 * inode, update pmp->pfs_types[], and update the syncronization
468 	 * threads.
469 	 *
470 	 * When forcing local mode, mark the PFS as a MASTER regardless.
471 	 *
472 	 * At the moment empty spots can develop due to removals or failures.
473 	 * Ultimately we want to re-fill these spots but doing so might
474 	 * confused running code. XXX
475 	 */
476 	hammer2_inode_ref(iroot);
477 	hammer2_mtx_ex(&iroot->lock);
478 	j = iroot->cluster.nchains;
479 
480 	if (j == HAMMER2_MAXCLUSTER) {
481 		kprintf("hammer2_pfsalloc: cluster full!\n");
482 		/* XXX fatal error? */
483 	} else {
484 		KKASSERT(chain->pmp == NULL);
485 		chain->pmp = pmp;
486 		hammer2_chain_ref(chain);
487 		iroot->cluster.array[j].chain = chain;
488 		if (force_local)
489 			pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
490 		else
491 			pmp->pfs_types[j] = ripdata->meta.pfs_type;
492 		pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
493 		pmp->pfs_hmps[j] = chain->hmp;
494 		hammer2_spin_ex(&pmp->inum_spin);
495 		pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset;
496 		hammer2_spin_unex(&pmp->inum_spin);
497 
498 		/*
499 		 * If the PFS is already mounted we must account
500 		 * for the mount_count here.
501 		 */
502 		if (pmp->mp)
503 			++chain->hmp->mount_count;
504 
505 		/*
506 		 * May have to fixup dirty chain tracking.  Previous
507 		 * pmp was NULL so nothing to undo.
508 		 */
509 		if (chain->flags & HAMMER2_CHAIN_MODIFIED)
510 			hammer2_pfs_memory_inc(pmp);
511 		++j;
512 	}
513 	iroot->cluster.nchains = j;
514 
515 	/*
516 	 * Update nmasters from any PFS inode which is part of the cluster.
517 	 * It is possible that this will result in a value which is too
518 	 * high.  MASTER PFSs are authoritative for pfs_nmasters and will
519 	 * override this value later on.
520 	 *
521 	 * (This informs us of masters that might not currently be
522 	 *  discoverable by this mount).
523 	 */
524 	if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
525 		pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
526 	}
527 
528 	/*
529 	 * Count visible masters.  Masters are usually added with
530 	 * ripdata->meta.pfs_nmasters set to 1.  This detects when there
531 	 * are more (XXX and must update the master inodes).
532 	 */
533 	count = 0;
534 	for (i = 0; i < iroot->cluster.nchains; ++i) {
535 		if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
536 			++count;
537 	}
538 	if (pmp->pfs_nmasters < count)
539 		pmp->pfs_nmasters = count;
540 
541 	/*
542 	 * Create missing synchronization and support threads.
543 	 *
544 	 * Single-node masters (including snapshots) have nothing to
545 	 * synchronize and do not require this thread.
546 	 *
547 	 * Multi-node masters or any number of soft masters, slaves, copy,
548 	 * or other PFS types need the thread.
549 	 *
550 	 * Each thread is responsible for its particular cluster index.
551 	 * We use independent threads so stalls or mismatches related to
552 	 * any given target do not affect other targets.
553 	 */
554 	for (i = 0; i < iroot->cluster.nchains; ++i) {
555 		/*
556 		 * Single-node masters (including snapshots) have nothing
557 		 * to synchronize and will make direct xops support calls,
558 		 * thus they do not require this thread.
559 		 *
560 		 * Note that there can be thousands of snapshots.  We do not
561 		 * want to create thousands of threads.
562 		 */
563 		if (pmp->pfs_nmasters <= 1 &&
564 		    pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
565 			continue;
566 		}
567 
568 		/*
569 		 * Sync support thread
570 		 */
571 		if (pmp->sync_thrs[i].td == NULL) {
572 			hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
573 					   "h2nod", i, -1,
574 					   hammer2_primary_sync_thread);
575 		}
576 	}
577 
578 	/*
579 	 * Create missing Xop threads
580 	 *
581 	 * NOTE: We create helper threads for all mounted PFSs or any
582 	 *	 PFSs with 2+ nodes (so the sync thread can update them,
583 	 *	 even if not mounted).
584 	 */
585 	if (pmp->mp || iroot->cluster.nchains >= 2)
586 		hammer2_xop_helper_create(pmp);
587 
588 	hammer2_mtx_unlock(&iroot->lock);
589 	hammer2_inode_drop(iroot);
590 done:
591 	return pmp;
592 }
593 
594 /*
595  * Deallocate an element of a probed PFS.  If destroying and this is a
596  * MASTER, adjust nmasters.
597  *
598  * This function does not physically destroy the PFS element in its device
599  * under the super-root  (see hammer2_ioctl_pfs_delete()).
600  */
601 void
602 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
603 {
604 	hammer2_inode_t *iroot;
605 	hammer2_chain_t *chain;
606 	int j;
607 
608 	/*
609 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
610 	 * by the flush code.
611 	 */
612 	iroot = pmp->iroot;
613 	if (iroot) {
614 		/*
615 		 * Stop synchronizing
616 		 *
617 		 * XXX flush after acquiring the iroot lock.
618 		 * XXX clean out the cluster index from all inode structures.
619 		 */
620 		hammer2_thr_delete(&pmp->sync_thrs[clindex]);
621 
622 		/*
623 		 * Remove the cluster index from the group.  If destroying
624 		 * the PFS and this is a master, adjust pfs_nmasters.
625 		 */
626 		hammer2_mtx_ex(&iroot->lock);
627 		chain = iroot->cluster.array[clindex].chain;
628 		iroot->cluster.array[clindex].chain = NULL;
629 
630 		switch(pmp->pfs_types[clindex]) {
631 		case HAMMER2_PFSTYPE_MASTER:
632 			if (destroying && pmp->pfs_nmasters > 0)
633 				--pmp->pfs_nmasters;
634 			/* XXX adjust ripdata->meta.pfs_nmasters */
635 			break;
636 		default:
637 			break;
638 		}
639 		pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
640 
641 		hammer2_mtx_unlock(&iroot->lock);
642 
643 		/*
644 		 * Release the chain.
645 		 */
646 		if (chain) {
647 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
648 			hammer2_chain_drop(chain);
649 		}
650 
651 		/*
652 		 * Terminate all XOP threads for the cluster index.
653 		 */
654 		if (pmp->xop_groups) {
655 			for (j = 0; j < hammer2_xopgroups; ++j) {
656 				hammer2_thr_delete(
657 					&pmp->xop_groups[j].thrs[clindex]);
658 			}
659 		}
660 	}
661 }
662 
663 /*
664  * Destroy a PFS, typically only occurs after the last mount on a device
665  * has gone away.
666  */
667 static void
668 hammer2_pfsfree(hammer2_pfs_t *pmp)
669 {
670 	hammer2_inode_t *iroot;
671 	hammer2_chain_t *chain;
672 	int chains_still_present = 0;
673 	int i;
674 	int j;
675 
676 	/*
677 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
678 	 * by the flush code.
679 	 */
680 	if (pmp->flags & HAMMER2_PMPF_SPMP)
681 		TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
682 	else
683 		TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
684 
685 	/*
686 	 * Cleanup chains remaining on LRU list.
687 	 */
688 	hammer2_spin_ex(&pmp->lru_spin);
689 	while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
690 		KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
691 		atomic_add_int(&pmp->lru_count, -1);
692 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
693 		TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
694 		hammer2_chain_ref(chain);
695 		hammer2_spin_unex(&pmp->lru_spin);
696 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
697 		hammer2_chain_drop(chain);
698 		hammer2_spin_ex(&pmp->lru_spin);
699 	}
700 	hammer2_spin_unex(&pmp->lru_spin);
701 
702 	/*
703 	 * Clean up iroot
704 	 */
705 	iroot = pmp->iroot;
706 	if (iroot) {
707 		for (i = 0; i < iroot->cluster.nchains; ++i) {
708 			hammer2_thr_delete(&pmp->sync_thrs[i]);
709 			if (pmp->xop_groups) {
710 				for (j = 0; j < hammer2_xopgroups; ++j)
711 					hammer2_thr_delete(
712 						&pmp->xop_groups[j].thrs[i]);
713 			}
714 			chain = iroot->cluster.array[i].chain;
715 			if (chain && !RB_EMPTY(&chain->core.rbtree)) {
716 				kprintf("hammer2: Warning pmp %p still "
717 					"has active chains\n", pmp);
718 				chains_still_present = 1;
719 			}
720 		}
721 		KASSERT(iroot->refs == 1,
722 			("PMP->IROOT %p REFS WRONG %d", iroot, iroot->refs));
723 
724 		/* ref for iroot */
725 		hammer2_inode_drop(iroot);
726 		pmp->iroot = NULL;
727 	}
728 
729 	/*
730 	 * Free remaining pmp resources
731 	 */
732 	if (chains_still_present) {
733 		kprintf("hammer2: cannot free pmp %p, still in use\n", pmp);
734 	} else {
735 		kmalloc_destroy_obj(&pmp->minode);
736 		kfree(pmp, M_HAMMER2);
737 	}
738 }
739 
740 /*
741  * Remove all references to hmp from the pfs list.  Any PFS which becomes
742  * empty is terminated and freed.
743  *
744  * XXX inefficient.
745  */
746 static void
747 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
748 {
749 	hammer2_pfs_t *pmp;
750 	hammer2_inode_t *iroot;
751 	hammer2_chain_t *rchain;
752 	int i;
753 	int j;
754 	struct hammer2_pfslist *wlist;
755 
756 	if (which == 0)
757 		wlist = &hammer2_pfslist;
758 	else
759 		wlist = &hammer2_spmplist;
760 again:
761 	TAILQ_FOREACH(pmp, wlist, mntentry) {
762 		if ((iroot = pmp->iroot) == NULL)
763 			continue;
764 
765 		/*
766 		 * Determine if this PFS is affected.  If it is we must
767 		 * freeze all management threads and lock its iroot.
768 		 *
769 		 * Freezing a management thread forces it idle, operations
770 		 * in-progress will be aborted and it will have to start
771 		 * over again when unfrozen, or exit if told to exit.
772 		 */
773 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
774 			if (pmp->pfs_hmps[i] == hmp)
775 				break;
776 		}
777 		if (i == HAMMER2_MAXCLUSTER)
778 			continue;
779 
780 		hammer2_vfs_sync_pmp(pmp, MNT_WAIT);
781 
782 		/*
783 		 * Make sure all synchronization threads are locked
784 		 * down.
785 		 */
786 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
787 			if (pmp->pfs_hmps[i] == NULL)
788 				continue;
789 			hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
790 			if (pmp->xop_groups) {
791 				for (j = 0; j < hammer2_xopgroups; ++j) {
792 					hammer2_thr_freeze_async(
793 						&pmp->xop_groups[j].thrs[i]);
794 				}
795 			}
796 		}
797 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
798 			if (pmp->pfs_hmps[i] == NULL)
799 				continue;
800 			hammer2_thr_freeze(&pmp->sync_thrs[i]);
801 			if (pmp->xop_groups) {
802 				for (j = 0; j < hammer2_xopgroups; ++j) {
803 					hammer2_thr_freeze(
804 						&pmp->xop_groups[j].thrs[i]);
805 				}
806 			}
807 		}
808 
809 		/*
810 		 * Lock the inode and clean out matching chains.
811 		 * Note that we cannot use hammer2_inode_lock_*()
812 		 * here because that would attempt to validate the
813 		 * cluster that we are in the middle of ripping
814 		 * apart.
815 		 *
816 		 * WARNING! We are working directly on the inodes
817 		 *	    embedded cluster.
818 		 */
819 		hammer2_mtx_ex(&iroot->lock);
820 
821 		/*
822 		 * Remove the chain from matching elements of the PFS.
823 		 */
824 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
825 			if (pmp->pfs_hmps[i] != hmp)
826 				continue;
827 			hammer2_thr_delete(&pmp->sync_thrs[i]);
828 			if (pmp->xop_groups) {
829 				for (j = 0; j < hammer2_xopgroups; ++j) {
830 					hammer2_thr_delete(
831 						&pmp->xop_groups[j].thrs[i]);
832 				}
833 			}
834 			rchain = iroot->cluster.array[i].chain;
835 			iroot->cluster.array[i].chain = NULL;
836 			pmp->pfs_types[i] = 0;
837 			if (pmp->pfs_names[i]) {
838 				kfree(pmp->pfs_names[i], M_HAMMER2);
839 				pmp->pfs_names[i] = NULL;
840 			}
841 			if (rchain) {
842 				hammer2_chain_drop(rchain);
843 				/* focus hint */
844 				if (iroot->cluster.focus == rchain)
845 					iroot->cluster.focus = NULL;
846 			}
847 			pmp->pfs_hmps[i] = NULL;
848 		}
849 		hammer2_mtx_unlock(&iroot->lock);
850 
851 		/*
852 		 * Cleanup trailing chains.  Gaps may remain.
853 		 */
854 		for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
855 			if (pmp->pfs_hmps[i])
856 				break;
857 		}
858 		iroot->cluster.nchains = i + 1;
859 
860 		/*
861 		 * If the PMP has no elements remaining we can destroy it.
862 		 * (this will transition management threads from frozen->exit).
863 		 */
864 		if (iroot->cluster.nchains == 0) {
865 			/*
866 			 * If this was the hmp's spmp, we need to clean
867 			 * a little more stuff out.
868 			 */
869 			if (hmp->spmp == pmp) {
870 				hmp->spmp = NULL;
871 				hmp->vchain.pmp = NULL;
872 				hmp->fchain.pmp = NULL;
873 			}
874 
875 			/*
876 			 * Free the pmp and restart the loop
877 			 */
878 			KKASSERT(TAILQ_EMPTY(&pmp->syncq));
879 			KKASSERT(TAILQ_EMPTY(&pmp->depq));
880 			hammer2_pfsfree(pmp);
881 			goto again;
882 		}
883 
884 		/*
885 		 * If elements still remain we need to set the REMASTER
886 		 * flag and unfreeze it.
887 		 */
888 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
889 			if (pmp->pfs_hmps[i] == NULL)
890 				continue;
891 			hammer2_thr_remaster(&pmp->sync_thrs[i]);
892 			hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
893 			if (pmp->xop_groups) {
894 				for (j = 0; j < hammer2_xopgroups; ++j) {
895 					hammer2_thr_remaster(
896 						&pmp->xop_groups[j].thrs[i]);
897 					hammer2_thr_unfreeze(
898 						&pmp->xop_groups[j].thrs[i]);
899 				}
900 			}
901 		}
902 	}
903 }
904 
905 /*
906  * Mount or remount HAMMER2 fileystem from physical media
907  *
908  *	mountroot
909  *		mp		mount point structure
910  *		path		NULL
911  *		data		<unused>
912  *		cred		<unused>
913  *
914  *	mount
915  *		mp		mount point structure
916  *		path		path to mount point
917  *		data		pointer to argument structure in user space
918  *			volume	volume path (device@LABEL form)
919  *			hflags	user mount flags
920  *		cred		user credentials
921  *
922  * RETURNS:	0	Success
923  *		!0	error number
924  */
925 static
926 int
927 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
928 		  struct ucred *cred)
929 {
930 	struct hammer2_mount_info info;
931 	hammer2_pfs_t *pmp;
932 	hammer2_pfs_t *spmp;
933 	hammer2_dev_t *hmp, *hmp_tmp;
934 	hammer2_dev_t *force_local;
935 	hammer2_key_t key_next;
936 	hammer2_key_t key_dummy;
937 	hammer2_key_t lhc;
938 	hammer2_chain_t *parent;
939 	hammer2_chain_t *chain;
940 	const hammer2_inode_data_t *ripdata;
941 	hammer2_blockref_t bref;
942 	hammer2_devvp_list_t devvpl;
943 	hammer2_devvp_t *e, *e_tmp;
944 	struct file *fp;
945 	char devstr[MNAMELEN];
946 	size_t size;
947 	size_t done;
948 	char *dev;
949 	char *label;
950 	int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
951 	int error;
952 	int i;
953 
954 	hmp = NULL;
955 	pmp = NULL;
956 	dev = NULL;
957 	label = NULL;
958 	bzero(&info, sizeof(info));
959 
960 	if (path) {
961 		/*
962 		 * Non-root mount or updating a mount
963 		 */
964 		error = copyin(data, &info, sizeof(info));
965 		if (error)
966 			return (error);
967 	}
968 
969 	if (mp->mnt_flag & MNT_UPDATE) {
970 		/*
971 		 * Update mount.  Note that pmp->iroot->cluster is
972 		 * an inode-embedded cluster and thus cannot be
973 		 * directly locked.
974 		 *
975 		 * XXX HAMMER2 needs to implement NFS export via
976 		 *     mountctl.
977 		 */
978 		hammer2_cluster_t *cluster;
979 
980 		pmp = MPTOPMP(mp);
981 		pmp->hflags = info.hflags;
982 		cluster = &pmp->iroot->cluster;
983 		for (i = 0; i < cluster->nchains; ++i) {
984 			if (cluster->array[i].chain == NULL)
985 				continue;
986 			hmp = cluster->array[i].chain->hmp;
987 			error = hammer2_remount(hmp, mp, path, cred);
988 			if (error)
989 				break;
990 		}
991 
992 		return error;
993 	}
994 
995 	if (path == NULL) {
996 		/*
997 		 * Root mount
998 		 */
999 		info.cluster_fd = -1;
1000 		ksnprintf(devstr, sizeof(devstr), "%s",
1001 			  mp->mnt_stat.f_mntfromname);
1002 		done = strlen(devstr) + 1;
1003 		kprintf("hammer2_mount: root devstr=\"%s\"\n", devstr);
1004 	} else {
1005 		error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
1006 		if (error)
1007 			return (error);
1008 		kprintf("hammer2_mount: devstr=\"%s\"\n", devstr);
1009 	}
1010 
1011 	/*
1012 	 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
1013 	 * if no label specified, based on the partition id.  Error out if no
1014 	 * label or device (with partition id) is specified.  This is strictly
1015 	 * a convenience to match the default label created by newfs_hammer2,
1016 	 * our preference is that a label always be specified.
1017 	 *
1018 	 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
1019 	 *	 that does not specify a device, as long as some H2 label
1020 	 *	 has already been mounted from that device.  This makes
1021 	 *	 mounting snapshots a lot easier.
1022 	 */
1023 	dev = devstr;
1024 	label = strchr(devstr, '@');
1025 	if (label && ((label + 1) - dev) > done) {
1026 		kprintf("hammer2_mount: bad label %s/%zd\n", devstr, done);
1027 		return (EINVAL);
1028 	}
1029 	if (label == NULL || label[1] == 0) {
1030 		char slice;
1031 
1032 		if (label == NULL)
1033 			label = devstr + strlen(devstr);
1034 		else
1035 			*label = '\0';		/* clean up trailing @ */
1036 
1037 		slice = label[-1];
1038 		switch(slice) {
1039 		case 'a':
1040 			label = "BOOT";
1041 			break;
1042 		case 'd':
1043 			label = "ROOT";
1044 			break;
1045 		default:
1046 			label = "DATA";
1047 			break;
1048 		}
1049 	} else {
1050 		*label = '\0';
1051 		label++;
1052 	}
1053 
1054 	kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
1055 		dev, label, ronly);
1056 
1057 	/*
1058 	 * Initialize all device vnodes.
1059 	 */
1060 	TAILQ_INIT(&devvpl);
1061 	error = hammer2_init_devvp(dev, path == NULL, &devvpl);
1062 	if (error) {
1063 		kprintf("hammer2: failed to initialize devvp in %s\n", dev);
1064 		hammer2_cleanup_devvp(&devvpl);
1065 		return error;
1066 	}
1067 
1068 	/*
1069 	 * Determine if the device has already been mounted.  After this
1070 	 * check hmp will be non-NULL if we are doing the second or more
1071 	 * hammer2 mounts from the same device.
1072 	 */
1073 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1074 	if (!TAILQ_EMPTY(&devvpl)) {
1075 		/*
1076 		 * Match the device.  Due to the way devfs works,
1077 		 * we may not be able to directly match the vnode pointer,
1078 		 * so also check to see if the underlying device matches.
1079 		 */
1080 		TAILQ_FOREACH(hmp_tmp, &hammer2_mntlist, mntentry) {
1081 			TAILQ_FOREACH(e_tmp, &hmp_tmp->devvpl, entry) {
1082 				int devvp_found = 0;
1083 				TAILQ_FOREACH(e, &devvpl, entry) {
1084 					KKASSERT(e->devvp);
1085 					if (e_tmp->devvp == e->devvp)
1086 						devvp_found = 1;
1087 					if (e_tmp->devvp->v_rdev &&
1088 					    e_tmp->devvp->v_rdev == e->devvp->v_rdev)
1089 						devvp_found = 1;
1090 				}
1091 				if (!devvp_found)
1092 					goto next_hmp;
1093 			}
1094 			hmp = hmp_tmp;
1095 			kprintf("hammer2_mount: hmp=%p matched\n", hmp);
1096 			break;
1097 next_hmp:
1098 			continue;
1099 		}
1100 
1101 		/*
1102 		 * If no match this may be a fresh H2 mount, make sure
1103 		 * the device is not mounted on anything else.
1104 		 */
1105 		if (hmp == NULL) {
1106 			TAILQ_FOREACH(e, &devvpl, entry) {
1107 				struct vnode *devvp = e->devvp;
1108 				KKASSERT(devvp);
1109 				error = vfs_mountedon(devvp);
1110 				if (error) {
1111 					kprintf("hammer2_mount: %s mounted %d\n",
1112 						e->path, error);
1113 					hammer2_cleanup_devvp(&devvpl);
1114 					lockmgr(&hammer2_mntlk, LK_RELEASE);
1115 					return error;
1116 				}
1117 			}
1118 		}
1119 	} else {
1120 		/*
1121 		 * Match the label to a pmp already probed.
1122 		 */
1123 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1124 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1125 				if (pmp->pfs_names[i] &&
1126 				    strcmp(pmp->pfs_names[i], label) == 0) {
1127 					hmp = pmp->pfs_hmps[i];
1128 					break;
1129 				}
1130 			}
1131 			if (hmp)
1132 				break;
1133 		}
1134 		if (hmp == NULL) {
1135 			kprintf("hammer2_mount: PFS label \"%s\" not found\n",
1136 				label);
1137 			hammer2_cleanup_devvp(&devvpl);
1138 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1139 			return ENOENT;
1140 		}
1141 	}
1142 
1143 	/*
1144 	 * Open the device if this isn't a secondary mount and construct
1145 	 * the H2 device mount (hmp).
1146 	 */
1147 	if (hmp == NULL) {
1148 		hammer2_chain_t *schain;
1149 		hammer2_xop_head_t xop;
1150 
1151 		/*
1152 		 * Now open the device
1153 		 */
1154 		KKASSERT(!TAILQ_EMPTY(&devvpl));
1155 		if (error == 0) {
1156 			error = hammer2_open_devvp(&devvpl, ronly);
1157 			if (error) {
1158 				hammer2_close_devvp(&devvpl, ronly);
1159 				hammer2_cleanup_devvp(&devvpl);
1160 				lockmgr(&hammer2_mntlk, LK_RELEASE);
1161 				return error;
1162 			}
1163 		}
1164 
1165 		/*
1166 		 * Construct volumes and link with device vnodes.
1167 		 */
1168 		hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1169 		hmp->devvp = NULL;
1170 		error = hammer2_init_volumes(mp, &devvpl, hmp->volumes,
1171 					     &hmp->voldata, &hmp->devvp);
1172 		if (error) {
1173 			hammer2_close_devvp(&devvpl, ronly);
1174 			hammer2_cleanup_devvp(&devvpl);
1175 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1176 			kfree(hmp, M_HAMMER2);
1177 			return error;
1178 		}
1179 		if (!hmp->devvp) {
1180 			kprintf("hammer2: failed to initialize root volume\n");
1181 			hammer2_unmount_helper(mp, NULL, hmp);
1182 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1183 			hammer2_vfs_unmount(mp, MNT_FORCE);
1184 			return EINVAL;
1185 		}
1186 
1187 		ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1188 		hmp->ronly = ronly;
1189 		hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1190 		kmalloc_create_obj(&hmp->mchain, "HAMMER2-chains",
1191 				   sizeof(struct hammer2_chain));
1192 		kmalloc_create_obj(&hmp->mio, "HAMMER2-dio",
1193 				   sizeof(struct hammer2_io));
1194 		kmalloc_create(&hmp->mmsg, "HAMMER2-msg");
1195 		TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1196 		RB_INIT(&hmp->iotree);
1197 		spin_init(&hmp->io_spin, "h2mount_io");
1198 		spin_init(&hmp->list_spin, "h2mount_list");
1199 
1200 		lockinit(&hmp->vollk, "h2vol", 0, 0);
1201 		lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1202 		lockinit(&hmp->bflock, "h2bflk", 0, 0);
1203 
1204 		/*
1205 		 * vchain setup. vchain.data is embedded.
1206 		 * vchain.refs is initialized and will never drop to 0.
1207 		 *
1208 		 * NOTE! voldata is not yet loaded.
1209 		 */
1210 		hmp->vchain.hmp = hmp;
1211 		hmp->vchain.refs = 1;
1212 		hmp->vchain.data = (void *)&hmp->voldata;
1213 		hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1214 		hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1215 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1216 		hammer2_chain_core_init(&hmp->vchain);
1217 
1218 		/*
1219 		 * fchain setup.  fchain.data is embedded.
1220 		 * fchain.refs is initialized and will never drop to 0.
1221 		 *
1222 		 * The data is not used but needs to be initialized to
1223 		 * pass assertion muster.  We use this chain primarily
1224 		 * as a placeholder for the freemap's top-level radix tree
1225 		 * so it does not interfere with the volume's topology
1226 		 * radix tree.
1227 		 */
1228 		hmp->fchain.hmp = hmp;
1229 		hmp->fchain.refs = 1;
1230 		hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1231 		hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1232 		hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1233 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1234 		hmp->fchain.bref.methods =
1235 			HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1236 			HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1237 		hammer2_chain_core_init(&hmp->fchain);
1238 
1239 		/*
1240 		 * Initialize volume header related fields.
1241 		 */
1242 		KKASSERT(hmp->voldata.magic == HAMMER2_VOLUME_ID_HBO ||
1243 			 hmp->voldata.magic == HAMMER2_VOLUME_ID_ABO);
1244 		hmp->volhdrno = error;
1245 		hmp->volsync = hmp->voldata;
1246 		hmp->free_reserved = hmp->voldata.allocator_size / 20;
1247 		/*
1248 		 * Must use hmp instead of volume header for these two
1249 		 * in order to handle volume versions transparently.
1250 		 */
1251 		if (hmp->voldata.version >= HAMMER2_VOL_VERSION_MULTI_VOLUMES) {
1252 			hmp->nvolumes = hmp->voldata.nvolumes;
1253 			hmp->total_size = hmp->voldata.total_size;
1254 		} else {
1255 			hmp->nvolumes = 1;
1256 			hmp->total_size = hmp->voldata.volu_size;
1257 		}
1258 		KKASSERT(hmp->nvolumes > 0);
1259 
1260 		/*
1261 		 * Move devvpl entries to hmp.
1262 		 */
1263 		TAILQ_INIT(&hmp->devvpl);
1264 		while ((e = TAILQ_FIRST(&devvpl)) != NULL) {
1265 			TAILQ_REMOVE(&devvpl, e, entry);
1266 			TAILQ_INSERT_TAIL(&hmp->devvpl, e, entry);
1267 		}
1268 		KKASSERT(TAILQ_EMPTY(&devvpl));
1269 		KKASSERT(!TAILQ_EMPTY(&hmp->devvpl));
1270 
1271 		/*
1272 		 * Really important to get these right or the flush and
1273 		 * teardown code will get confused.
1274 		 */
1275 		hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1276 		spmp = hmp->spmp;
1277 		spmp->pfs_hmps[0] = hmp;
1278 
1279 		/*
1280 		 * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1281 		 * is inherited from the volume header.
1282 		 */
1283 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1284 		hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1285 		hmp->vchain.pmp = spmp;
1286 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1287 		hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1288 		hmp->fchain.pmp = spmp;
1289 
1290 		/*
1291 		 * First locate the super-root inode, which is key 0
1292 		 * relative to the volume header's blockset.
1293 		 *
1294 		 * Then locate the root inode by scanning the directory keyspace
1295 		 * represented by the label.
1296 		 */
1297 		parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1298 		schain = hammer2_chain_lookup(&parent, &key_dummy,
1299 				      HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1300 				      &error, 0);
1301 		hammer2_chain_lookup_done(parent);
1302 		if (schain == NULL) {
1303 			kprintf("hammer2_mount: invalid super-root\n");
1304 			hammer2_unmount_helper(mp, NULL, hmp);
1305 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1306 			hammer2_vfs_unmount(mp, MNT_FORCE);
1307 			return EINVAL;
1308 		}
1309 		if (schain->error) {
1310 			kprintf("hammer2_mount: error %s reading super-root\n",
1311 				hammer2_error_str(schain->error));
1312 			hammer2_chain_unlock(schain);
1313 			hammer2_chain_drop(schain);
1314 			schain = NULL;
1315 			hammer2_unmount_helper(mp, NULL, hmp);
1316 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1317 			hammer2_vfs_unmount(mp, MNT_FORCE);
1318 			return EINVAL;
1319 		}
1320 
1321 		/*
1322 		 * The super-root always uses an inode_tid of 1 when
1323 		 * creating PFSs.
1324 		 */
1325 		spmp->inode_tid = 1;
1326 		spmp->modify_tid = schain->bref.modify_tid + 1;
1327 
1328 		/*
1329 		 * Sanity-check schain's pmp and finish initialization.
1330 		 * Any chain belonging to the super-root topology should
1331 		 * have a NULL pmp (not even set to spmp).
1332 		 */
1333 		ripdata = &schain->data->ipdata;
1334 		KKASSERT(schain->pmp == NULL);
1335 		spmp->pfs_clid = ripdata->meta.pfs_clid;
1336 
1337 		/*
1338 		 * Replace the dummy spmp->iroot with a real one.  It's
1339 		 * easier to just do a wholesale replacement than to try
1340 		 * to update the chain and fixup the iroot fields.
1341 		 *
1342 		 * The returned inode is locked with the supplied cluster.
1343 		 */
1344 		hammer2_dummy_xop_from_chain(&xop, schain);
1345 		hammer2_inode_drop(spmp->iroot);
1346 		spmp->iroot = NULL;
1347 		spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1);
1348 		spmp->spmp_hmp = hmp;
1349 		spmp->pfs_types[0] = ripdata->meta.pfs_type;
1350 		spmp->pfs_hmps[0] = hmp;
1351 		hammer2_inode_ref(spmp->iroot);
1352 		hammer2_inode_unlock(spmp->iroot);
1353 		hammer2_cluster_unlock(&xop.cluster);
1354 		hammer2_chain_drop(schain);
1355 		/* do not call hammer2_cluster_drop() on an embedded cluster */
1356 		schain = NULL;	/* now invalid */
1357 		/* leave spmp->iroot with one ref */
1358 
1359 		if (!hmp->ronly) {
1360 			error = hammer2_recovery(hmp);
1361 			if (error == 0)
1362 				error |= hammer2_fixup_pfses(hmp);
1363 			/* XXX do something with error */
1364 		}
1365 		hammer2_update_pmps(hmp);
1366 		hammer2_iocom_init(hmp);
1367 		hammer2_bulkfree_init(hmp);
1368 
1369 		/*
1370 		 * Ref the cluster management messaging descriptor.  The mount
1371 		 * program deals with the other end of the communications pipe.
1372 		 *
1373 		 * Root mounts typically do not supply one.
1374 		 */
1375 		if (info.cluster_fd >= 0) {
1376 			fp = holdfp(curthread, info.cluster_fd, -1);
1377 			if (fp) {
1378 				hammer2_cluster_reconnect(hmp, fp);
1379 			} else {
1380 				kprintf("hammer2_mount: bad cluster_fd!\n");
1381 			}
1382 		}
1383 	} else {
1384 		spmp = hmp->spmp;
1385 		if (info.hflags & HMNT2_DEVFLAGS) {
1386 			kprintf("hammer2_mount: Warning: mount flags pertaining "
1387 				"to the whole device may only be specified "
1388 				"on the first mount of the device: %08x\n",
1389 				info.hflags & HMNT2_DEVFLAGS);
1390 		}
1391 	}
1392 
1393 	/*
1394 	 * Force local mount (disassociate all PFSs from their clusters).
1395 	 * Used primarily for debugging.
1396 	 */
1397 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1398 
1399 	/*
1400 	 * Lookup the mount point under the media-localized super-root.
1401 	 * Scanning hammer2_pfslist doesn't help us because it represents
1402 	 * PFS cluster ids which can aggregate several named PFSs together.
1403 	 *
1404 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1405 	 * up later on.
1406 	 */
1407 	hammer2_inode_lock(spmp->iroot, 0);
1408 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1409 	lhc = hammer2_dirhash(label, strlen(label));
1410 	chain = hammer2_chain_lookup(&parent, &key_next,
1411 				     lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1412 				     &error, 0);
1413 	while (chain) {
1414 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1415 		    strcmp(label, chain->data->ipdata.filename) == 0) {
1416 			break;
1417 		}
1418 		chain = hammer2_chain_next(&parent, chain, &key_next,
1419 					    key_next,
1420 					    lhc + HAMMER2_DIRHASH_LOMASK,
1421 					    &error, 0);
1422 	}
1423 	if (parent) {
1424 		hammer2_chain_unlock(parent);
1425 		hammer2_chain_drop(parent);
1426 	}
1427 	hammer2_inode_unlock(spmp->iroot);
1428 
1429 	/*
1430 	 * PFS could not be found?
1431 	 */
1432 	if (chain == NULL) {
1433 		hammer2_unmount_helper(mp, NULL, hmp);
1434 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1435 		hammer2_vfs_unmount(mp, MNT_FORCE);
1436 
1437 		if (error) {
1438 			kprintf("hammer2_mount: PFS label I/O error\n");
1439 			return EINVAL;
1440 		} else {
1441 			kprintf("hammer2_mount: PFS label \"%s\" not found\n",
1442 				label);
1443 			return ENOENT;
1444 		}
1445 	}
1446 
1447 	/*
1448 	 * Acquire the pmp structure (it should have already been allocated
1449 	 * via hammer2_update_pmps() so do not pass cluster in to add to
1450 	 * available chains).
1451 	 *
1452 	 * Check if the cluster has already been mounted.  A cluster can
1453 	 * only be mounted once, use null mounts to mount additional copies.
1454 	 */
1455 	if (chain->error) {
1456 		kprintf("hammer2_mount: PFS label I/O error\n");
1457 	} else {
1458 		ripdata = &chain->data->ipdata;
1459 		bref = chain->bref;
1460 		pmp = hammer2_pfsalloc(NULL, ripdata,
1461 				       bref.modify_tid, force_local);
1462 	}
1463 	hammer2_chain_unlock(chain);
1464 	hammer2_chain_drop(chain);
1465 
1466 	/*
1467 	 * Finish the mount
1468 	 */
1469         kprintf("hammer2_mount: hmp=%p pmp=%p\n", hmp, pmp);
1470 
1471 	if (pmp->mp) {
1472 		kprintf("hammer2_mount: PFS already mounted!\n");
1473 		hammer2_unmount_helper(mp, NULL, hmp);
1474 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1475 		hammer2_vfs_unmount(mp, MNT_FORCE);
1476 
1477 		return EBUSY;
1478 	}
1479 
1480 	pmp->hflags = info.hflags;
1481         mp->mnt_flag |= MNT_LOCAL;
1482         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1483         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1484 
1485         /*
1486          * required mount structure initializations
1487          */
1488         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1489         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1490 
1491         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1492         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1493 
1494         /*
1495          * Optional fields
1496          */
1497         mp->mnt_iosize_max = MAXPHYS;
1498 
1499 	/*
1500 	 * Connect up mount pointers.
1501 	 */
1502 	hammer2_mount_helper(mp, pmp);
1503         lockmgr(&hammer2_mntlk, LK_RELEASE);
1504 
1505 	/*
1506 	 * Finish setup
1507 	 */
1508 	vfs_getnewfsid(mp);
1509 	vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1510 	vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1511 	vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1512 
1513 	if (path) {
1514 		copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1515 			  MNAMELEN - 1, &size);
1516 		bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1517 	} /* else root mount, already in there */
1518 
1519 	bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1520 	if (path) {
1521 		copyinstr(path, mp->mnt_stat.f_mntonname,
1522 			  sizeof(mp->mnt_stat.f_mntonname) - 1,
1523 			  &size);
1524 	} else {
1525 		/* root mount */
1526 		mp->mnt_stat.f_mntonname[0] = '/';
1527 	}
1528 
1529 	/*
1530 	 * Initial statfs to prime mnt_stat.
1531 	 */
1532 	hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1533 
1534 	return 0;
1535 }
1536 
1537 /*
1538  * Scan PFSs under the super-root and create hammer2_pfs structures.
1539  */
1540 static
1541 void
1542 hammer2_update_pmps(hammer2_dev_t *hmp)
1543 {
1544 	const hammer2_inode_data_t *ripdata;
1545 	hammer2_chain_t *parent;
1546 	hammer2_chain_t *chain;
1547 	hammer2_blockref_t bref;
1548 	hammer2_dev_t *force_local;
1549 	hammer2_pfs_t *spmp;
1550 	hammer2_pfs_t *pmp;
1551 	hammer2_key_t key_next;
1552 	int error;
1553 
1554 	/*
1555 	 * Force local mount (disassociate all PFSs from their clusters).
1556 	 * Used primarily for debugging.
1557 	 */
1558 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1559 
1560 	/*
1561 	 * Lookup mount point under the media-localized super-root.
1562 	 *
1563 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1564 	 * up later on.
1565 	 */
1566 	spmp = hmp->spmp;
1567 	hammer2_inode_lock(spmp->iroot, 0);
1568 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1569 	chain = hammer2_chain_lookup(&parent, &key_next,
1570 					 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1571 					 &error, 0);
1572 	while (chain) {
1573 		if (chain->error) {
1574 			kprintf("I/O error scanning PFS labels\n");
1575 		} else if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
1576 			kprintf("Non inode chain type %d under super-root\n",
1577 				chain->bref.type);
1578 		} else {
1579 			ripdata = &chain->data->ipdata;
1580 			bref = chain->bref;
1581 			pmp = hammer2_pfsalloc(chain, ripdata,
1582 					       bref.modify_tid, force_local);
1583 		}
1584 		chain = hammer2_chain_next(&parent, chain, &key_next,
1585 					   key_next, HAMMER2_KEY_MAX,
1586 					   &error, 0);
1587 	}
1588 	if (parent) {
1589 		hammer2_chain_unlock(parent);
1590 		hammer2_chain_drop(parent);
1591 	}
1592 	hammer2_inode_unlock(spmp->iroot);
1593 }
1594 
1595 static
1596 int
1597 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1598 		struct ucred *cred)
1599 {
1600 	hammer2_volume_t *vol;
1601 	struct vnode *devvp;
1602 	int i, error, result = 0;
1603 
1604 	if (!(hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)))
1605 		return 0;
1606 
1607 	for (i = 0; i < hmp->nvolumes; ++i) {
1608 		vol = &hmp->volumes[i];
1609 		devvp = vol->dev->devvp;
1610 		KKASSERT(devvp);
1611 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1612 		VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1613 		vn_unlock(devvp);
1614 		error = 0;
1615 		if (vol->id == HAMMER2_ROOT_VOLUME) {
1616 			error = hammer2_recovery(hmp);
1617 			if (error == 0)
1618 				error |= hammer2_fixup_pfses(hmp);
1619 		}
1620 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1621 		if (error == 0) {
1622 			VOP_CLOSE(devvp, FREAD, NULL);
1623 		} else {
1624 			VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1625 		}
1626 		vn_unlock(devvp);
1627 		result |= error;
1628 	}
1629 	if (result == 0) {
1630 		kprintf("hammer2: enable read/write\n");
1631 		hmp->ronly = 0;
1632 	}
1633 
1634 	return result;
1635 }
1636 
1637 static
1638 int
1639 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1640 {
1641 	hammer2_pfs_t *pmp;
1642 	int flags;
1643 	int error = 0;
1644 
1645 	pmp = MPTOPMP(mp);
1646 
1647 	if (pmp == NULL)
1648 		return(0);
1649 
1650 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1651 
1652 	/*
1653 	 * If mount initialization proceeded far enough we must flush
1654 	 * its vnodes and sync the underlying mount points.  Three syncs
1655 	 * are required to fully flush the filesystem (freemap updates lag
1656 	 * by one flush, and one extra for safety).
1657 	 */
1658 	if (mntflags & MNT_FORCE)
1659 		flags = FORCECLOSE;
1660 	else
1661 		flags = 0;
1662 	if (pmp->iroot) {
1663 		error = vflush(mp, 0, flags);
1664 		if (error)
1665 			goto failed;
1666 		hammer2_vfs_sync(mp, MNT_WAIT);
1667 		hammer2_vfs_sync(mp, MNT_WAIT);
1668 		hammer2_vfs_sync(mp, MNT_WAIT);
1669 	}
1670 
1671 	/*
1672 	 * Cleanup the frontend support XOPS threads
1673 	 */
1674 	hammer2_xop_helper_cleanup(pmp);
1675 
1676 	if (pmp->mp)
1677 		hammer2_unmount_helper(mp, pmp, NULL);
1678 
1679 	error = 0;
1680 failed:
1681 	lockmgr(&hammer2_mntlk, LK_RELEASE);
1682 
1683 	return (error);
1684 }
1685 
1686 /*
1687  * Mount helper, hook the system mount into our PFS.
1688  * The mount lock is held.
1689  *
1690  * We must bump the mount_count on related devices for any
1691  * mounted PFSs.
1692  */
1693 static
1694 void
1695 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1696 {
1697 	hammer2_cluster_t *cluster;
1698 	hammer2_chain_t *rchain;
1699 	int i;
1700 
1701         mp->mnt_data = (qaddr_t)pmp;
1702 	pmp->mp = mp;
1703 
1704 	/*
1705 	 * After pmp->mp is set we have to adjust hmp->mount_count.
1706 	 */
1707 	cluster = &pmp->iroot->cluster;
1708 	for (i = 0; i < cluster->nchains; ++i) {
1709 		rchain = cluster->array[i].chain;
1710 		if (rchain == NULL)
1711 			continue;
1712 		++rchain->hmp->mount_count;
1713 	}
1714 
1715 	/*
1716 	 * Create missing Xop threads
1717 	 */
1718 	hammer2_xop_helper_create(pmp);
1719 }
1720 
1721 /*
1722  * Mount helper, unhook the system mount from our PFS.
1723  * The mount lock is held.
1724  *
1725  * If hmp is supplied a mount responsible for being the first to open
1726  * the block device failed and the block device and all PFSs using the
1727  * block device must be cleaned up.
1728  *
1729  * If pmp is supplied multiple devices might be backing the PFS and each
1730  * must be disconnected.  This might not be the last PFS using some of the
1731  * underlying devices.  Also, we have to adjust our hmp->mount_count
1732  * accounting for the devices backing the pmp which is now undergoing an
1733  * unmount.
1734  */
1735 static
1736 void
1737 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1738 {
1739 	hammer2_cluster_t *cluster;
1740 	hammer2_chain_t *rchain;
1741 	int dumpcnt;
1742 	int i;
1743 
1744 	/*
1745 	 * If no device supplied this is a high-level unmount and we have to
1746 	 * to disconnect the mount, adjust mount_count, and locate devices
1747 	 * that might now have no mounts.
1748 	 */
1749 	if (pmp) {
1750 		KKASSERT(hmp == NULL);
1751 		KKASSERT(MPTOPMP(mp) == pmp);
1752 		pmp->mp = NULL;
1753 		mp->mnt_data = NULL;
1754 
1755 		/*
1756 		 * After pmp->mp is cleared we have to account for
1757 		 * mount_count.
1758 		 */
1759 		cluster = &pmp->iroot->cluster;
1760 		for (i = 0; i < cluster->nchains; ++i) {
1761 			rchain = cluster->array[i].chain;
1762 			if (rchain == NULL)
1763 				continue;
1764 			--rchain->hmp->mount_count;
1765 			/* scrapping hmp now may invalidate the pmp */
1766 		}
1767 again:
1768 		TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1769 			if (hmp->mount_count == 0) {
1770 				hammer2_unmount_helper(NULL, NULL, hmp);
1771 				goto again;
1772 			}
1773 		}
1774 		return;
1775 	}
1776 
1777 	/*
1778 	 * Try to terminate the block device.  We can't terminate it if
1779 	 * there are still PFSs referencing it.
1780 	 */
1781 	if (hmp->mount_count)
1782 		return;
1783 
1784 	/*
1785 	 * Decomission the network before we start messing with the
1786 	 * device and PFS.
1787 	 */
1788 	hammer2_iocom_uninit(hmp);
1789 
1790 	hammer2_bulkfree_uninit(hmp);
1791 	hammer2_pfsfree_scan(hmp, 0);
1792 
1793 	/*
1794 	 * Cycle the volume data lock as a safety (probably not needed any
1795 	 * more).  To ensure everything is out we need to flush at least
1796 	 * three times.  (1) The running of the sideq can dirty the
1797 	 * filesystem, (2) A normal flush can dirty the freemap, and
1798 	 * (3) ensure that the freemap is fully synchronized.
1799 	 *
1800 	 * The next mount's recovery scan can clean everything up but we want
1801 	 * to leave the filesystem in a 100% clean state on a normal unmount.
1802 	 */
1803 #if 0
1804 	hammer2_voldata_lock(hmp);
1805 	hammer2_voldata_unlock(hmp);
1806 #endif
1807 
1808 	/*
1809 	 * Flush whatever is left.  Unmounted but modified PFS's might still
1810 	 * have some dirty chains on them.
1811 	 */
1812 	hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1813 	hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1814 
1815 	if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1816 		hammer2_voldata_modify(hmp);
1817 		hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1818 					    HAMMER2_FLUSH_ALL);
1819 	}
1820 	hammer2_chain_unlock(&hmp->fchain);
1821 
1822 	if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1823 		hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1824 					    HAMMER2_FLUSH_ALL);
1825 	}
1826 	hammer2_chain_unlock(&hmp->vchain);
1827 
1828 	if ((hmp->vchain.flags | hmp->fchain.flags) &
1829 	    HAMMER2_CHAIN_FLUSH_MASK) {
1830 		kprintf("hammer2_unmount: chains left over after final sync\n");
1831 		kprintf("    vchain %08x\n", hmp->vchain.flags);
1832 		kprintf("    fchain %08x\n", hmp->fchain.flags);
1833 
1834 		if (hammer2_debug & 0x0010)
1835 			Debugger("entered debugger");
1836 	}
1837 
1838 	hammer2_pfsfree_scan(hmp, 1);
1839 
1840 	KKASSERT(hmp->spmp == NULL);
1841 
1842 	/*
1843 	 * Finish up with the device vnode
1844 	 */
1845 	if (!TAILQ_EMPTY(&hmp->devvpl)) {
1846 		hammer2_close_devvp(&hmp->devvpl, hmp->ronly);
1847 		hammer2_cleanup_devvp(&hmp->devvpl);
1848 	}
1849 	KKASSERT(TAILQ_EMPTY(&hmp->devvpl));
1850 
1851 	/*
1852 	 * Clear vchain/fchain flags that might prevent final cleanup
1853 	 * of these chains.
1854 	 */
1855 	if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1856 		atomic_add_long(&hammer2_count_modified_chains, -1);
1857 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1858 		hammer2_pfs_memory_wakeup(hmp->vchain.pmp, -1);
1859 	}
1860 	if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1861 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1862 	}
1863 
1864 	if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1865 		atomic_add_long(&hammer2_count_modified_chains, -1);
1866 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1867 		hammer2_pfs_memory_wakeup(hmp->fchain.pmp, -1);
1868 	}
1869 	if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1870 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1871 	}
1872 
1873 	/*
1874 	 * Final drop of embedded freemap root chain to
1875 	 * clean up fchain.core (fchain structure is not
1876 	 * flagged ALLOCATED so it is cleaned out and then
1877 	 * left to rot).
1878 	 */
1879 	hammer2_chain_drop(&hmp->fchain);
1880 
1881 	/*
1882 	 * Final drop of embedded volume root chain to clean
1883 	 * up vchain.core (vchain structure is not flagged
1884 	 * ALLOCATED so it is cleaned out and then left to
1885 	 * rot).
1886 	 */
1887 	dumpcnt = 50;
1888 	hammer2_dump_chain(&hmp->vchain, 0, 0, &dumpcnt, 'v', (u_int)-1);
1889 	dumpcnt = 50;
1890 	hammer2_dump_chain(&hmp->fchain, 0, 0, &dumpcnt, 'f', (u_int)-1);
1891 
1892 	hammer2_chain_drop(&hmp->vchain);
1893 
1894 	hammer2_io_cleanup(hmp, &hmp->iotree);
1895 	if (hmp->iofree_count) {
1896 		kprintf("io_cleanup: %d I/O's left hanging\n",
1897 			hmp->iofree_count);
1898 	}
1899 
1900 	TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1901 	kmalloc_destroy_obj(&hmp->mchain);
1902 	kmalloc_destroy_obj(&hmp->mio);
1903 	kmalloc_destroy(&hmp->mmsg);
1904 	kfree(hmp, M_HAMMER2);
1905 }
1906 
1907 int
1908 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1909 		 ino_t ino, struct vnode **vpp)
1910 {
1911 	hammer2_xop_lookup_t *xop;
1912 	hammer2_pfs_t *pmp;
1913 	hammer2_inode_t *ip;
1914 	hammer2_tid_t inum;
1915 	int error;
1916 
1917 	inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1918 
1919 	error = 0;
1920 	pmp = MPTOPMP(mp);
1921 
1922 	/*
1923 	 * Easy if we already have it cached
1924 	 */
1925 	ip = hammer2_inode_lookup(pmp, inum);
1926 	if (ip) {
1927 		hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1928 		*vpp = hammer2_igetv(ip, &error);
1929 		hammer2_inode_unlock(ip);
1930 		hammer2_inode_drop(ip);		/* from lookup */
1931 
1932 		return error;
1933 	}
1934 
1935 	/*
1936 	 * Otherwise we have to find the inode
1937 	 */
1938 	xop = hammer2_xop_alloc(pmp->iroot, 0);
1939 	xop->lhc = inum;
1940 	hammer2_xop_start(&xop->head, &hammer2_lookup_desc);
1941 	error = hammer2_xop_collect(&xop->head, 0);
1942 
1943 	if (error == 0)
1944 		ip = hammer2_inode_get(pmp, &xop->head, -1, -1);
1945 	hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1946 
1947 	if (ip) {
1948 		*vpp = hammer2_igetv(ip, &error);
1949 		hammer2_inode_unlock(ip);
1950 	} else {
1951 		*vpp = NULL;
1952 		error = ENOENT;
1953 	}
1954 	return (error);
1955 }
1956 
1957 static
1958 int
1959 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1960 {
1961 	hammer2_pfs_t *pmp;
1962 	struct vnode *vp;
1963 	int error;
1964 
1965 	pmp = MPTOPMP(mp);
1966 	if (pmp->iroot == NULL) {
1967 		kprintf("hammer2 (%s): no root inode\n",
1968 			mp->mnt_stat.f_mntfromname);
1969 		*vpp = NULL;
1970 		return EINVAL;
1971 	}
1972 
1973 	error = 0;
1974 	hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1975 
1976 	while (pmp->inode_tid == 0) {
1977 		hammer2_xop_ipcluster_t *xop;
1978 		const hammer2_inode_meta_t *meta;
1979 
1980 		xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1981 		hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc);
1982 		error = hammer2_xop_collect(&xop->head, 0);
1983 
1984 		if (error == 0) {
1985 			meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1986 			pmp->iroot->meta = *meta;
1987 			pmp->inode_tid = meta->pfs_inum + 1;
1988 			hammer2_xop_pdata(&xop->head);
1989 			/* meta invalid */
1990 
1991 			if (pmp->inode_tid < HAMMER2_INODE_START)
1992 				pmp->inode_tid = HAMMER2_INODE_START;
1993 			pmp->modify_tid =
1994 				xop->head.cluster.focus->bref.modify_tid + 1;
1995 #if 0
1996 			kprintf("PFS: Starting inode %jd\n",
1997 				(intmax_t)pmp->inode_tid);
1998 			kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1999 				pmp->inode_tid, pmp->modify_tid);
2000 #endif
2001 			wakeup(&pmp->iroot);
2002 
2003 			hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2004 
2005 			/*
2006 			 * Prime the mount info.
2007 			 */
2008 			hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
2009 			break;
2010 		}
2011 
2012 		/*
2013 		 * Loop, try again
2014 		 */
2015 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2016 		hammer2_inode_unlock(pmp->iroot);
2017 		error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
2018 		hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
2019 		if (error == EINTR)
2020 			break;
2021 	}
2022 
2023 	if (error) {
2024 		hammer2_inode_unlock(pmp->iroot);
2025 		*vpp = NULL;
2026 	} else {
2027 		vp = hammer2_igetv(pmp->iroot, &error);
2028 		hammer2_inode_unlock(pmp->iroot);
2029 		*vpp = vp;
2030 	}
2031 
2032 	return (error);
2033 }
2034 
2035 /*
2036  * Filesystem status
2037  *
2038  * XXX incorporate ipdata->meta.inode_quota and data_quota
2039  */
2040 static
2041 int
2042 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
2043 {
2044 	hammer2_pfs_t *pmp;
2045 	hammer2_dev_t *hmp;
2046 	hammer2_blockref_t bref;
2047 	struct statfs tmp;
2048 	int i;
2049 
2050 	/*
2051 	 * NOTE: iroot might not have validated the cluster yet.
2052 	 */
2053 	pmp = MPTOPMP(mp);
2054 
2055 	bzero(&tmp, sizeof(tmp));
2056 
2057 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2058 		hmp = pmp->pfs_hmps[i];
2059 		if (hmp == NULL)
2060 			continue;
2061 		if (pmp->iroot->cluster.array[i].chain)
2062 			bref = pmp->iroot->cluster.array[i].chain->bref;
2063 		else
2064 			bzero(&bref, sizeof(bref));
2065 
2066 		tmp.f_files = bref.embed.stats.inode_count;
2067 		tmp.f_ffree = 0;
2068 		tmp.f_blocks = hmp->voldata.allocator_size /
2069 			       mp->mnt_vstat.f_bsize;
2070 		tmp.f_bfree = hmp->voldata.allocator_free /
2071 			      mp->mnt_vstat.f_bsize;
2072 		tmp.f_bavail = tmp.f_bfree;
2073 
2074 		if (cred && cred->cr_uid != 0) {
2075 			uint64_t adj;
2076 
2077 			/* 5% */
2078 			adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2079 			tmp.f_blocks -= adj;
2080 			tmp.f_bfree -= adj;
2081 			tmp.f_bavail -= adj;
2082 		}
2083 
2084 		mp->mnt_stat.f_blocks = tmp.f_blocks;
2085 		mp->mnt_stat.f_bfree = tmp.f_bfree;
2086 		mp->mnt_stat.f_bavail = tmp.f_bavail;
2087 		mp->mnt_stat.f_files = tmp.f_files;
2088 		mp->mnt_stat.f_ffree = tmp.f_ffree;
2089 
2090 		*sbp = mp->mnt_stat;
2091 	}
2092 	return (0);
2093 }
2094 
2095 static
2096 int
2097 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2098 {
2099 	hammer2_pfs_t *pmp;
2100 	hammer2_dev_t *hmp;
2101 	hammer2_blockref_t bref;
2102 	struct statvfs tmp;
2103 	int i;
2104 
2105 	/*
2106 	 * NOTE: iroot might not have validated the cluster yet.
2107 	 */
2108 	pmp = MPTOPMP(mp);
2109 	bzero(&tmp, sizeof(tmp));
2110 
2111 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2112 		hmp = pmp->pfs_hmps[i];
2113 		if (hmp == NULL)
2114 			continue;
2115 		if (pmp->iroot->cluster.array[i].chain)
2116 			bref = pmp->iroot->cluster.array[i].chain->bref;
2117 		else
2118 			bzero(&bref, sizeof(bref));
2119 
2120 		tmp.f_files = bref.embed.stats.inode_count;
2121 		tmp.f_ffree = 0;
2122 		tmp.f_blocks = hmp->voldata.allocator_size /
2123 			       mp->mnt_vstat.f_bsize;
2124 		tmp.f_bfree = hmp->voldata.allocator_free /
2125 			      mp->mnt_vstat.f_bsize;
2126 		tmp.f_bavail = tmp.f_bfree;
2127 
2128 		if (cred && cred->cr_uid != 0) {
2129 			uint64_t adj;
2130 
2131 			/* 5% */
2132 			adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2133 			tmp.f_blocks -= adj;
2134 			tmp.f_bfree -= adj;
2135 			tmp.f_bavail -= adj;
2136 		}
2137 
2138 		mp->mnt_vstat.f_blocks = tmp.f_blocks;
2139 		mp->mnt_vstat.f_bfree = tmp.f_bfree;
2140 		mp->mnt_vstat.f_bavail = tmp.f_bavail;
2141 		mp->mnt_vstat.f_files = tmp.f_files;
2142 		mp->mnt_vstat.f_ffree = tmp.f_ffree;
2143 
2144 		*sbp = mp->mnt_vstat;
2145 	}
2146 	return (0);
2147 }
2148 
2149 /*
2150  * Mount-time recovery (RW mounts)
2151  *
2152  * Updates to the free block table are allowed to lag flushes by one
2153  * transaction.  In case of a crash, then on a fresh mount we must do an
2154  * incremental scan of the last committed transaction id and make sure that
2155  * all related blocks have been marked allocated.
2156  */
2157 struct hammer2_recovery_elm {
2158 	TAILQ_ENTRY(hammer2_recovery_elm) entry;
2159 	hammer2_chain_t *chain;
2160 	hammer2_tid_t sync_tid;
2161 };
2162 
2163 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2164 
2165 struct hammer2_recovery_info {
2166 	struct hammer2_recovery_list list;
2167 	hammer2_tid_t	mtid;
2168 	int	depth;
2169 };
2170 
2171 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2172 			hammer2_chain_t *parent,
2173 			struct hammer2_recovery_info *info,
2174 			hammer2_tid_t sync_tid);
2175 
2176 #define HAMMER2_RECOVERY_MAXDEPTH	10
2177 
2178 static
2179 int
2180 hammer2_recovery(hammer2_dev_t *hmp)
2181 {
2182 	struct hammer2_recovery_info info;
2183 	struct hammer2_recovery_elm *elm;
2184 	hammer2_chain_t *parent;
2185 	hammer2_tid_t sync_tid;
2186 	hammer2_tid_t mirror_tid;
2187 	int error;
2188 
2189 	hammer2_trans_init(hmp->spmp, 0);
2190 
2191 	sync_tid = hmp->voldata.freemap_tid;
2192 	mirror_tid = hmp->voldata.mirror_tid;
2193 
2194 	kprintf("hammer2_mount: \"%s\": ", hmp->devrepname);
2195 	if (sync_tid >= mirror_tid) {
2196 		kprintf("no recovery needed\n");
2197 	} else {
2198 		kprintf("freemap recovery %016jx-%016jx\n",
2199 			sync_tid + 1, mirror_tid);
2200 	}
2201 
2202 	TAILQ_INIT(&info.list);
2203 	info.depth = 0;
2204 	parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2205 	error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2206 	hammer2_chain_lookup_done(parent);
2207 
2208 	while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2209 		TAILQ_REMOVE(&info.list, elm, entry);
2210 		parent = elm->chain;
2211 		sync_tid = elm->sync_tid;
2212 		kfree(elm, M_HAMMER2);
2213 
2214 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2215 		error |= hammer2_recovery_scan(hmp, parent, &info,
2216 					      hmp->voldata.freemap_tid);
2217 		hammer2_chain_unlock(parent);
2218 		hammer2_chain_drop(parent);	/* drop elm->chain ref */
2219 	}
2220 
2221 	hammer2_trans_done(hmp->spmp, 0);
2222 
2223 	return error;
2224 }
2225 
2226 static
2227 int
2228 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2229 		      struct hammer2_recovery_info *info,
2230 		      hammer2_tid_t sync_tid)
2231 {
2232 	const hammer2_inode_data_t *ripdata;
2233 	hammer2_chain_t *chain;
2234 	hammer2_blockref_t bref;
2235 	int tmp_error;
2236 	int rup_error;
2237 	int error;
2238 	int first;
2239 
2240 	/*
2241 	 * Adjust freemap to ensure that the block(s) are marked allocated.
2242 	 */
2243 	if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2244 		hammer2_freemap_adjust(hmp, &parent->bref,
2245 				       HAMMER2_FREEMAP_DORECOVER);
2246 	}
2247 
2248 	/*
2249 	 * Check type for recursive scan
2250 	 */
2251 	switch(parent->bref.type) {
2252 	case HAMMER2_BREF_TYPE_VOLUME:
2253 		/* data already instantiated */
2254 		break;
2255 	case HAMMER2_BREF_TYPE_INODE:
2256 		/*
2257 		 * Must instantiate data for DIRECTDATA test and also
2258 		 * for recursion.
2259 		 */
2260 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2261 		ripdata = &parent->data->ipdata;
2262 		if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2263 			/* not applicable to recovery scan */
2264 			hammer2_chain_unlock(parent);
2265 			return 0;
2266 		}
2267 		hammer2_chain_unlock(parent);
2268 		break;
2269 	case HAMMER2_BREF_TYPE_INDIRECT:
2270 		/*
2271 		 * Must instantiate data for recursion
2272 		 */
2273 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2274 		hammer2_chain_unlock(parent);
2275 		break;
2276 	case HAMMER2_BREF_TYPE_DIRENT:
2277 	case HAMMER2_BREF_TYPE_DATA:
2278 	case HAMMER2_BREF_TYPE_FREEMAP:
2279 	case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2280 	case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2281 		/* not applicable to recovery scan */
2282 		return 0;
2283 		break;
2284 	default:
2285 		return HAMMER2_ERROR_BADBREF;
2286 	}
2287 
2288 	/*
2289 	 * Defer operation if depth limit reached.
2290 	 */
2291 	if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2292 		struct hammer2_recovery_elm *elm;
2293 
2294 		elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2295 		elm->chain = parent;
2296 		elm->sync_tid = sync_tid;
2297 		hammer2_chain_ref(parent);
2298 		TAILQ_INSERT_TAIL(&info->list, elm, entry);
2299 		/* unlocked by caller */
2300 
2301 		return(0);
2302 	}
2303 
2304 
2305 	/*
2306 	 * Recursive scan of the last flushed transaction only.  We are
2307 	 * doing this without pmp assignments so don't leave the chains
2308 	 * hanging around after we are done with them.
2309 	 *
2310 	 * error	Cumulative error this level only
2311 	 * rup_error	Cumulative error for recursion
2312 	 * tmp_error	Specific non-cumulative recursion error
2313 	 */
2314 	chain = NULL;
2315 	first = 1;
2316 	rup_error = 0;
2317 	error = 0;
2318 
2319 	for (;;) {
2320 		error |= hammer2_chain_scan(parent, &chain, &bref,
2321 					    &first,
2322 					    HAMMER2_LOOKUP_NODATA);
2323 
2324 		/*
2325 		 * Problem during scan or EOF
2326 		 */
2327 		if (error)
2328 			break;
2329 
2330 		/*
2331 		 * If this is a leaf
2332 		 */
2333 		if (chain == NULL) {
2334 			if (bref.mirror_tid > sync_tid) {
2335 				hammer2_freemap_adjust(hmp, &bref,
2336 						     HAMMER2_FREEMAP_DORECOVER);
2337 			}
2338 			continue;
2339 		}
2340 
2341 		/*
2342 		 * This may or may not be a recursive node.
2343 		 */
2344 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2345 		if (bref.mirror_tid > sync_tid) {
2346 			++info->depth;
2347 			tmp_error = hammer2_recovery_scan(hmp, chain,
2348 							   info, sync_tid);
2349 			--info->depth;
2350 		} else {
2351 			tmp_error = 0;
2352 		}
2353 
2354 		/*
2355 		 * Flush the recovery at the PFS boundary to stage it for
2356 		 * the final flush of the super-root topology.
2357 		 */
2358 		if (tmp_error == 0 &&
2359 		    (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2360 		    (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2361 			hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2362 					     HAMMER2_FLUSH_ALL);
2363 		}
2364 		rup_error |= tmp_error;
2365 	}
2366 	return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2367 }
2368 
2369 /*
2370  * This fixes up an error introduced in earlier H2 implementations where
2371  * moving a PFS inode into an indirect block wound up causing the
2372  * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2373  */
2374 static
2375 int
2376 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2377 {
2378 	const hammer2_inode_data_t *ripdata;
2379 	hammer2_chain_t *parent;
2380 	hammer2_chain_t *chain;
2381 	hammer2_key_t key_next;
2382 	hammer2_pfs_t *spmp;
2383 	int error;
2384 
2385 	error = 0;
2386 
2387 	/*
2388 	 * Lookup mount point under the media-localized super-root.
2389 	 *
2390 	 * cluster->pmp will incorrectly point to spmp and must be fixed
2391 	 * up later on.
2392 	 */
2393 	spmp = hmp->spmp;
2394 	hammer2_inode_lock(spmp->iroot, 0);
2395 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2396 	chain = hammer2_chain_lookup(&parent, &key_next,
2397 					 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2398 					 &error, 0);
2399 	while (chain) {
2400 		if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2401 			continue;
2402 		if (chain->error) {
2403 			kprintf("I/O error scanning PFS labels\n");
2404 			error |= chain->error;
2405 		} else if ((chain->bref.flags &
2406 			    HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2407 			int error2;
2408 
2409 			ripdata = &chain->data->ipdata;
2410 			hammer2_trans_init(hmp->spmp, 0);
2411 			error2 = hammer2_chain_modify(chain,
2412 						      chain->bref.modify_tid,
2413 						      0, 0);
2414 			if (error2 == 0) {
2415 				kprintf("hammer2: Correct mis-flagged PFS %s\n",
2416 					ripdata->filename);
2417 				chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2418 			} else {
2419 				error |= error2;
2420 			}
2421 			hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2422 					     HAMMER2_FLUSH_ALL);
2423 			hammer2_trans_done(hmp->spmp, 0);
2424 		}
2425 		chain = hammer2_chain_next(&parent, chain, &key_next,
2426 					   key_next, HAMMER2_KEY_MAX,
2427 					   &error, 0);
2428 	}
2429 	if (parent) {
2430 		hammer2_chain_unlock(parent);
2431 		hammer2_chain_drop(parent);
2432 	}
2433 	hammer2_inode_unlock(spmp->iroot);
2434 
2435 	return error;
2436 }
2437 
2438 /*
2439  * Sync a mount point; this is called periodically on a per-mount basis from
2440  * the filesystem syncer, and whenever a user issues a sync.
2441  */
2442 int
2443 hammer2_vfs_sync(struct mount *mp, int waitfor)
2444 {
2445 	int error;
2446 
2447 	error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor);
2448 
2449 	return error;
2450 }
2451 
2452 /*
2453  * Because frontend operations lock vnodes before we get a chance to
2454  * lock the related inode, we can't just acquire a vnode lock without
2455  * risking a deadlock.  The frontend may be holding a vnode lock while
2456  * also blocked on our SYNCQ flag while trying to get the inode lock.
2457  *
2458  * To deal with this situation we can check the vnode lock situation
2459  * after locking the inode and perform a work-around.
2460  */
2461 int
2462 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor)
2463 {
2464 	struct mount *mp;
2465 	/*hammer2_xop_flush_t *xop;*/
2466 	/*struct hammer2_sync_info info;*/
2467 	hammer2_inode_t *ip;
2468 	hammer2_depend_t *depend;
2469 	hammer2_depend_t *depend_next;
2470 	struct vnode *vp;
2471 	uint32_t pass2;
2472 	int error;
2473 	int wakecount;
2474 	int dorestart;
2475 
2476 	mp = pmp->mp;
2477 
2478 	/*
2479 	 * Move all inodes on sideq to syncq.  This will clear sideq.
2480 	 * This should represent all flushable inodes.  These inodes
2481 	 * will already have refs due to being on syncq or sideq.  We
2482 	 * must do this all at once with the spinlock held to ensure that
2483 	 * all inode dependencies are part of the same flush.
2484 	 *
2485 	 * We should be able to do this asynchronously from frontend
2486 	 * operations because we will be locking the inodes later on
2487 	 * to actually flush them, and that will partition any frontend
2488 	 * op using the same inode.  Either it has already locked the
2489 	 * inode and we will block, or it has not yet locked the inode
2490 	 * and it will block until we are finished flushing that inode.
2491 	 *
2492 	 * When restarting, only move the inodes flagged as PASS2 from
2493 	 * SIDEQ to SYNCQ.  PASS2 propagation by inode_lock4() and
2494 	 * inode_depend() are atomic with the spin-lock.
2495 	 */
2496 	hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2497 #ifdef HAMMER2_DEBUG_SYNC
2498 	kprintf("FILESYSTEM SYNC BOUNDARY\n");
2499 #endif
2500 	dorestart = 0;
2501 
2502 	/*
2503 	 * Move inodes from depq to syncq, releasing the related
2504 	 * depend structures.
2505 	 */
2506 restart:
2507 #ifdef HAMMER2_DEBUG_SYNC
2508 	kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart);
2509 #endif
2510 	hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/);
2511 	hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN);
2512 
2513 	/*
2514 	 * Move inodes from depq to syncq.  When restarting, only depq's
2515 	 * marked pass2 are moved.
2516 	 */
2517 	hammer2_spin_ex(&pmp->list_spin);
2518 	depend_next = TAILQ_FIRST(&pmp->depq);
2519 	wakecount = 0;
2520 
2521 	while ((depend = depend_next) != NULL) {
2522 		depend_next = TAILQ_NEXT(depend, entry);
2523 		if (dorestart && depend->pass2 == 0)
2524 			continue;
2525 		TAILQ_FOREACH(ip, &depend->sideq, entry) {
2526 			KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ);
2527 			atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ);
2528 			atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ);
2529 			ip->depend = NULL;
2530 		}
2531 
2532 		/*
2533 		 * NOTE: pmp->sideq_count includes both sideq and syncq
2534 		 */
2535 		TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry);
2536 
2537 		depend->count = 0;
2538 		depend->pass2 = 0;
2539 		TAILQ_REMOVE(&pmp->depq, depend, entry);
2540 	}
2541 
2542 	hammer2_spin_unex(&pmp->list_spin);
2543 	hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/
2544 				      HAMMER2_TRANS_WAITING);
2545 	dorestart = 0;
2546 
2547 	/*
2548 	 * sideq_count may have dropped enough to allow us to unstall
2549 	 * the frontend.
2550 	 */
2551 	hammer2_pfs_memory_wakeup(pmp, 0);
2552 
2553 	/*
2554 	 * Now run through all inodes on syncq.
2555 	 *
2556 	 * Flush transactions only interlock with other flush transactions.
2557 	 * Any conflicting frontend operations will block on the inode, but
2558 	 * may hold a vnode lock while doing so.
2559 	 */
2560 	hammer2_spin_ex(&pmp->list_spin);
2561 	while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) {
2562 		/*
2563 		 * Remove the inode from the SYNCQ, transfer the syncq ref
2564 		 * to us.  We must clear SYNCQ to allow any potential
2565 		 * front-end deadlock to proceed.  We must set PASS2 so
2566 		 * the dependency code knows what to do.
2567 		 */
2568 		pass2 = ip->flags;
2569 		cpu_ccfence();
2570 		if (atomic_cmpset_int(&ip->flags,
2571 			      pass2,
2572 			      (pass2 & ~(HAMMER2_INODE_SYNCQ |
2573 					 HAMMER2_INODE_SYNCQ_WAKEUP)) |
2574 			      HAMMER2_INODE_SYNCQ_PASS2) == 0) {
2575 			continue;
2576 		}
2577 		TAILQ_REMOVE(&pmp->syncq, ip, entry);
2578 		--pmp->sideq_count;
2579 		hammer2_spin_unex(&pmp->list_spin);
2580 
2581 		/*
2582 		 * Tickle anyone waiting on ip->flags or the hysteresis
2583 		 * on the dirty inode count.
2584 		 */
2585 		if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP)
2586 			wakeup(&ip->flags);
2587 		if (++wakecount >= hammer2_limit_dirty_inodes / 20 + 1) {
2588 			wakecount = 0;
2589 			hammer2_pfs_memory_wakeup(pmp, 0);
2590 		}
2591 
2592 		/*
2593 		 * Relock the inode, and we inherit a ref from the above.
2594 		 * We will check for a race after we acquire the vnode.
2595 		 */
2596 		hammer2_mtx_ex(&ip->lock);
2597 
2598 		/*
2599 		 * We need the vp in order to vfsync() dirty buffers, so if
2600 		 * one isn't attached we can skip it.
2601 		 *
2602 		 * Ordering the inode lock and then the vnode lock has the
2603 		 * potential to deadlock.  If we had left SYNCQ set that could
2604 		 * also deadlock us against the frontend even if we don't hold
2605 		 * any locks, but the latter is not a problem now since we
2606 		 * cleared it.  igetv will temporarily release the inode lock
2607 		 * in a safe manner to work-around the deadlock.
2608 		 *
2609 		 * Unfortunately it is still possible to deadlock when the
2610 		 * frontend obtains multiple inode locks, because all the
2611 		 * related vnodes are already locked (nor can the vnode locks
2612 		 * be released and reacquired without messing up RECLAIM and
2613 		 * INACTIVE sequencing).
2614 		 *
2615 		 * The solution for now is to move the vp back onto SIDEQ
2616 		 * and set dorestart, which will restart the flush after we
2617 		 * exhaust the current SYNCQ.  Note that additional
2618 		 * dependencies may build up, so we definitely need to move
2619 		 * the whole SIDEQ back to SYNCQ when we restart.
2620 		 */
2621 		vp = ip->vp;
2622 		if (vp) {
2623 			if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) {
2624 				/*
2625 				 * Failed to get the vnode, requeue the inode
2626 				 * (PASS2 is already set so it will be found
2627 				 * again on the restart).
2628 				 *
2629 				 * Then unlock, possibly sleep, and retry
2630 				 * later.  We sleep if PASS2 was *previously*
2631 				 * set, before we set it again above.
2632 				 */
2633 				vp = NULL;
2634 				dorestart = 1;
2635 #ifdef HAMMER2_DEBUG_SYNC
2636 				kprintf("inum %ld (sync delayed by vnode)\n",
2637 					(long)ip->meta.inum);
2638 #endif
2639 				hammer2_inode_delayed_sideq(ip);
2640 
2641 				hammer2_mtx_unlock(&ip->lock);
2642 				hammer2_inode_drop(ip);
2643 
2644 				if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) {
2645 					tsleep(&dorestart, 0, "h2syndel", 2);
2646 				}
2647 				hammer2_spin_ex(&pmp->list_spin);
2648 				continue;
2649 			}
2650 		} else {
2651 			vp = NULL;
2652 		}
2653 
2654 		/*
2655 		 * If the inode wound up on a SIDEQ again it will already be
2656 		 * prepped for another PASS2.  In this situation if we flush
2657 		 * it now we will just wind up flushing it again in the same
2658 		 * syncer run, so we might as well not flush it now.
2659 		 */
2660 		if (ip->flags & HAMMER2_INODE_SIDEQ) {
2661 			hammer2_mtx_unlock(&ip->lock);
2662 			hammer2_inode_drop(ip);
2663 			if (vp)
2664 				vput(vp);
2665 			dorestart = 1;
2666 			hammer2_spin_ex(&pmp->list_spin);
2667 			continue;
2668 		}
2669 
2670 		/*
2671 		 * Ok we have the inode exclusively locked and if vp is
2672 		 * not NULL that will also be exclusively locked.  Do the
2673 		 * meat of the flush.
2674 		 *
2675 		 * vp token needed for v_rbdirty_tree check / vclrisdirty
2676 		 * sequencing.  Though we hold the vnode exclusively so
2677 		 * we shouldn't need to hold the token also in this case.
2678 		 */
2679 		if (vp) {
2680 			vfsync(vp, MNT_WAIT, 1, NULL, NULL);
2681 			bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */
2682 		}
2683 
2684 		/*
2685 		 * If the inode has not yet been inserted into the tree
2686 		 * we must do so.  Then sync and flush it.  The flush should
2687 		 * update the parent.
2688 		 */
2689 		if (ip->flags & HAMMER2_INODE_DELETING) {
2690 #ifdef HAMMER2_DEBUG_SYNC
2691 			kprintf("inum %ld destroy\n", (long)ip->meta.inum);
2692 #endif
2693 			hammer2_inode_chain_des(ip);
2694 			atomic_add_long(&hammer2_iod_inode_deletes, 1);
2695 		} else if (ip->flags & HAMMER2_INODE_CREATING) {
2696 #ifdef HAMMER2_DEBUG_SYNC
2697 			kprintf("inum %ld insert\n", (long)ip->meta.inum);
2698 #endif
2699 			hammer2_inode_chain_ins(ip);
2700 			atomic_add_long(&hammer2_iod_inode_creates, 1);
2701 		}
2702 #ifdef HAMMER2_DEBUG_SYNC
2703 		kprintf("inum %ld chain-sync\n", (long)ip->meta.inum);
2704 #endif
2705 
2706 		/*
2707 		 * Because I kinda messed up the design and index the inodes
2708 		 * under the root inode, along side the directory entries,
2709 		 * we can't flush the inode index under the iroot until the
2710 		 * end.  If we do it now we might miss effects created by
2711 		 * other inodes on the SYNCQ.
2712 		 *
2713 		 * Do a normal (non-FSSYNC) flush instead, which allows the
2714 		 * vnode code to work the same.  We don't want to force iroot
2715 		 * back onto the SIDEQ, and we also don't want the flush code
2716 		 * to update pfs_iroot_blocksets until the final flush later.
2717 		 *
2718 		 * XXX at the moment this will likely result in a double-flush
2719 		 * of the iroot chain.
2720 		 */
2721 		hammer2_inode_chain_sync(ip);
2722 		if (ip == pmp->iroot) {
2723 			hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP);
2724 		} else {
2725 			hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2726 						      HAMMER2_XOP_FSSYNC);
2727 		}
2728 		if (vp) {
2729 			lwkt_gettoken(&vp->v_token);
2730 			if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2731 					  HAMMER2_INODE_RESIZED |
2732 					  HAMMER2_INODE_DIRTYDATA)) == 0 &&
2733 			    RB_EMPTY(&vp->v_rbdirty_tree) &&
2734 			    !bio_track_active(&vp->v_track_write)) {
2735 				vclrisdirty(vp);
2736 			} else {
2737 				hammer2_inode_delayed_sideq(ip);
2738 			}
2739 			lwkt_reltoken(&vp->v_token);
2740 			vput(vp);
2741 			vp = NULL;	/* safety */
2742 		}
2743 		atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2);
2744 		hammer2_inode_unlock(ip);	/* unlock+drop */
2745 		/* ip pointer invalid */
2746 
2747 		/*
2748 		 * If the inode got dirted after we dropped our locks,
2749 		 * it will have already been moved back to the SIDEQ.
2750 		 */
2751 		hammer2_spin_ex(&pmp->list_spin);
2752 	}
2753 	hammer2_spin_unex(&pmp->list_spin);
2754 	hammer2_pfs_memory_wakeup(pmp, 0);
2755 
2756 	if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) {
2757 #ifdef HAMMER2_DEBUG_SYNC
2758 		kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n");
2759 		/*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/
2760 #endif
2761 		dorestart = 1;
2762 		goto restart;
2763 	}
2764 #ifdef HAMMER2_DEBUG_SYNC
2765 	kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n");
2766 	/*tsleep(&dorestart, 0, "h2STG2", hz*20);*/
2767 #endif
2768 
2769 	/*
2770 	 * We have to flush the PFS root last, even if it does not appear to
2771 	 * be dirty, because all the inodes in the PFS are indexed under it.
2772 	 * The normal flushing of iroot above would only occur if directory
2773 	 * entries under the root were changed.
2774 	 *
2775 	 * Specifying VOLHDR will cause an additionl flush of hmp->spmp
2776 	 * for the media making up the cluster.
2777 	 */
2778 	if ((ip = pmp->iroot) != NULL) {
2779 		hammer2_inode_ref(ip);
2780 		hammer2_mtx_ex(&ip->lock);
2781 		hammer2_inode_chain_sync(ip);
2782 		hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2783 					      HAMMER2_XOP_FSSYNC |
2784 					      HAMMER2_XOP_VOLHDR);
2785 		hammer2_inode_unlock(ip);	/* unlock+drop */
2786 	}
2787 #ifdef HAMMER2_DEBUG_SYNC
2788 	kprintf("FILESYSTEM SYNC STAGE 2 DONE\n");
2789 #endif
2790 
2791 	/*
2792 	 * device bioq sync
2793 	 */
2794 	hammer2_bioq_sync(pmp);
2795 
2796 #if 0
2797 	/*
2798 	 * Generally speaking we now want to flush the media topology from
2799 	 * the iroot through to the inodes.  The flush stops at any inode
2800 	 * boundary, which allows the frontend to continue running concurrent
2801 	 * modifying operations on inodes (including kernel flushes of
2802 	 * buffers) without interfering with the main sync.
2803 	 *
2804 	 * Use the XOP interface to concurrently flush all nodes to
2805 	 * synchronize the PFSROOT subtopology to the media.  A standard
2806 	 * end-of-scan ENOENT error indicates cluster sufficiency.
2807 	 *
2808 	 * Note that this flush will not be visible on crash recovery until
2809 	 * we flush the super-root topology in the next loop.
2810 	 *
2811 	 * XXX For now wait for all flushes to complete.
2812 	 */
2813 	if (mp && (ip = pmp->iroot) != NULL) {
2814 		/*
2815 		 * If unmounting try to flush everything including any
2816 		 * sub-trees under inodes, just in case there is dangling
2817 		 * modified data, as a safety.  Otherwise just flush up to
2818 		 * the inodes in this stage.
2819 		 */
2820 		kprintf("MP & IROOT\n");
2821 #ifdef HAMMER2_DEBUG_SYNC
2822 		kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n");
2823 #endif
2824 		if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2825 			xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2826 						    HAMMER2_XOP_VOLHDR |
2827 						    HAMMER2_XOP_FSSYNC |
2828 						    HAMMER2_XOP_INODE_STOP);
2829 		} else {
2830 			xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2831 						    HAMMER2_XOP_INODE_STOP |
2832 						    HAMMER2_XOP_VOLHDR |
2833 						    HAMMER2_XOP_FSSYNC |
2834 						    HAMMER2_XOP_INODE_STOP);
2835 		}
2836 		hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc);
2837 		error = hammer2_xop_collect(&xop->head,
2838 					    HAMMER2_XOP_COLLECT_WAITALL);
2839 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2840 #ifdef HAMMER2_DEBUG_SYNC
2841 		kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n");
2842 #endif
2843 		if (error == HAMMER2_ERROR_ENOENT)
2844 			error = 0;
2845 		else
2846 			error = hammer2_error_to_errno(error);
2847 	} else {
2848 		error = 0;
2849 	}
2850 #endif
2851 	error = 0;	/* XXX */
2852 	hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH);
2853 
2854 	return (error);
2855 }
2856 
2857 static
2858 int
2859 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2860 {
2861 	hammer2_inode_t *ip;
2862 
2863 	KKASSERT(MAXFIDSZ >= 16);
2864 	ip = VTOI(vp);
2865 	fhp->fid_len = offsetof(struct fid, fid_data[16]);
2866 	fhp->fid_ext = 0;
2867 	((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2868 	((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2869 
2870 	return 0;
2871 }
2872 
2873 static
2874 int
2875 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2876 	       struct fid *fhp, struct vnode **vpp)
2877 {
2878 	hammer2_pfs_t *pmp;
2879 	hammer2_tid_t inum;
2880 	int error;
2881 
2882 	pmp = MPTOPMP(mp);
2883 	inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2884 	if (vpp) {
2885 		if (inum == 1)
2886 			error = hammer2_vfs_root(mp, vpp);
2887 		else
2888 			error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2889 	} else {
2890 		error = 0;
2891 	}
2892 	if (error)
2893 		kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2894 	return error;
2895 }
2896 
2897 static
2898 int
2899 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2900 		 int *exflagsp, struct ucred **credanonp)
2901 {
2902 	hammer2_pfs_t *pmp;
2903 	struct netcred *np;
2904 	int error;
2905 
2906 	pmp = MPTOPMP(mp);
2907 	np = vfs_export_lookup(mp, &pmp->export, nam);
2908 	if (np) {
2909 		*exflagsp = np->netc_exflags;
2910 		*credanonp = &np->netc_anon;
2911 		error = 0;
2912 	} else {
2913 		error = EACCES;
2914 	}
2915 	return error;
2916 }
2917 
2918 /*
2919  * This handles hysteresis on regular file flushes.  Because the BIOs are
2920  * routed to a thread it is possible for an excessive number to build up
2921  * and cause long front-end stalls long before the runningbuffspace limit
2922  * is hit, so we implement hammer2_flush_pipe to control the
2923  * hysteresis.
2924  *
2925  * This is a particular problem when compression is used.
2926  */
2927 void
2928 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2929 {
2930 	atomic_add_int(&pmp->count_lwinprog, 1);
2931 }
2932 
2933 void
2934 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2935 {
2936 	int lwinprog;
2937 
2938 	lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2939 	if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2940 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2941 		atomic_clear_int(&pmp->count_lwinprog,
2942 				 HAMMER2_LWINPROG_WAITING);
2943 		wakeup(&pmp->count_lwinprog);
2944 	}
2945 	if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2946 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2947 		atomic_clear_int(&pmp->count_lwinprog,
2948 				 HAMMER2_LWINPROG_WAITING0);
2949 		wakeup(&pmp->count_lwinprog);
2950 	}
2951 }
2952 
2953 void
2954 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2955 {
2956 	int lwinprog;
2957 	int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2958 				    HAMMER2_LWINPROG_WAITING0;
2959 
2960 	for (;;) {
2961 		lwinprog = pmp->count_lwinprog;
2962 		cpu_ccfence();
2963 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2964 			break;
2965 		tsleep_interlock(&pmp->count_lwinprog, 0);
2966 		atomic_set_int(&pmp->count_lwinprog, lwflag);
2967 		lwinprog = pmp->count_lwinprog;
2968 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2969 			break;
2970 		tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2971 	}
2972 }
2973 
2974 /*
2975  * It is possible for an excessive number of dirty chains or dirty inodes
2976  * to build up.  When this occurs we start an asynchronous filesystem sync.
2977  * If the level continues to build up, we stall, waiting for it to drop,
2978  * with some hysteresis.
2979  *
2980  * This relies on the kernel calling hammer2_vfs_modifying() prior to
2981  * obtaining any vnode locks before making a modifying VOP call.
2982  */
2983 static int
2984 hammer2_vfs_modifying(struct mount *mp)
2985 {
2986 	if (mp->mnt_flag & MNT_RDONLY)
2987 		return EROFS;
2988 	hammer2_pfs_memory_wait(MPTOPMP(mp));
2989 
2990 	return 0;
2991 }
2992 
2993 /*
2994  * Initiate an asynchronous filesystem sync and, with hysteresis,
2995  * stall if the internal data structure count becomes too bloated.
2996  */
2997 void
2998 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2999 {
3000 	uint32_t waiting;
3001 	int pcatch;
3002 	int error;
3003 
3004 	if (pmp == NULL || pmp->mp == NULL)
3005 		return;
3006 
3007 	for (;;) {
3008 		waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK;
3009 		cpu_ccfence();
3010 
3011 		/*
3012 		 * Start the syncer running at 1/2 the limit
3013 		 */
3014 		if (waiting > hammer2_limit_dirty_chains / 2 ||
3015 		    pmp->sideq_count > hammer2_limit_dirty_inodes / 2) {
3016 			trigger_syncer(pmp->mp);
3017 		}
3018 
3019 		/*
3020 		 * Stall at the limit waiting for the counts to drop.
3021 		 * This code will typically be woken up once the count
3022 		 * drops below 3/4 the limit, or in one second.
3023 		 */
3024 		if (waiting < hammer2_limit_dirty_chains &&
3025 		    pmp->sideq_count < hammer2_limit_dirty_inodes) {
3026 			break;
3027 		}
3028 
3029 		pcatch = curthread->td_proc ? PCATCH : 0;
3030 
3031 		tsleep_interlock(&pmp->inmem_dirty_chains, pcatch);
3032 		atomic_set_int(&pmp->inmem_dirty_chains,
3033 			       HAMMER2_DIRTYCHAIN_WAITING);
3034 		if (waiting < hammer2_limit_dirty_chains &&
3035 		    pmp->sideq_count < hammer2_limit_dirty_inodes) {
3036 			break;
3037 		}
3038 		trigger_syncer(pmp->mp);
3039 		error = tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED | pcatch,
3040 			       "h2memw", hz);
3041 		if (error == ERESTART)
3042 			break;
3043 	}
3044 }
3045 
3046 /*
3047  * Wake up any stalled frontend ops waiting, with hysteresis, using
3048  * 2/3 of the limit.
3049  */
3050 void
3051 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp, int count)
3052 {
3053 	uint32_t waiting;
3054 
3055 	if (pmp) {
3056 		waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, count);
3057 		/* don't need --waiting to test flag */
3058 
3059 		if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) &&
3060 		    (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <=
3061 		    hammer2_limit_dirty_chains * 2 / 3 &&
3062 		    pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) {
3063 			atomic_clear_int(&pmp->inmem_dirty_chains,
3064 					 HAMMER2_DIRTYCHAIN_WAITING);
3065 			wakeup(&pmp->inmem_dirty_chains);
3066 		}
3067 	}
3068 }
3069 
3070 void
3071 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
3072 {
3073 	if (pmp) {
3074 		atomic_add_int(&pmp->inmem_dirty_chains, 1);
3075 	}
3076 }
3077 
3078 /*
3079  * Volume header data locks
3080  */
3081 void
3082 hammer2_voldata_lock(hammer2_dev_t *hmp)
3083 {
3084 	lockmgr(&hmp->vollk, LK_EXCLUSIVE);
3085 }
3086 
3087 void
3088 hammer2_voldata_unlock(hammer2_dev_t *hmp)
3089 {
3090 	lockmgr(&hmp->vollk, LK_RELEASE);
3091 }
3092 
3093 void
3094 hammer2_voldata_modify(hammer2_dev_t *hmp)
3095 {
3096 	if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) {
3097 		atomic_add_long(&hammer2_count_modified_chains, 1);
3098 		atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
3099 		hammer2_pfs_memory_inc(hmp->vchain.pmp);
3100 	}
3101 }
3102 
3103 /*
3104  * Returns 0 if the filesystem has tons of free space
3105  * Returns 1 if the filesystem has less than 10% remaining
3106  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
3107  */
3108 int
3109 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
3110 {
3111 	hammer2_pfs_t *pmp;
3112 	hammer2_dev_t *hmp;
3113 	hammer2_off_t free_reserved;
3114 	hammer2_off_t free_nominal;
3115 	int i;
3116 
3117 	pmp = ip->pmp;
3118 
3119 	if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
3120 		free_reserved = HAMMER2_SEGSIZE;
3121 		free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
3122 		for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
3123 			hmp = pmp->pfs_hmps[i];
3124 			if (hmp == NULL)
3125 				continue;
3126 			if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
3127 			    pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
3128 				continue;
3129 
3130 			if (free_nominal > hmp->voldata.allocator_free)
3131 				free_nominal = hmp->voldata.allocator_free;
3132 			if (free_reserved < hmp->free_reserved)
3133 				free_reserved = hmp->free_reserved;
3134 		}
3135 
3136 		/*
3137 		 * SMP races ok
3138 		 */
3139 		pmp->free_reserved = free_reserved;
3140 		pmp->free_nominal = free_nominal;
3141 		pmp->free_ticks = ticks;
3142 	} else {
3143 		free_reserved = pmp->free_reserved;
3144 		free_nominal = pmp->free_nominal;
3145 	}
3146 	if (cred && cred->cr_uid != 0) {
3147 		if ((int64_t)(free_nominal - bytes) <
3148 		    (int64_t)free_reserved) {
3149 			return 2;
3150 		}
3151 	} else {
3152 		if ((int64_t)(free_nominal - bytes) <
3153 		    (int64_t)free_reserved / 2) {
3154 			return 2;
3155 		}
3156 	}
3157 	if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
3158 		return 1;
3159 	return 0;
3160 }
3161 
3162 /*
3163  * Debugging
3164  */
3165 void
3166 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int bi, int *countp,
3167 		   char pfx, u_int flags)
3168 {
3169 	hammer2_chain_t *scan;
3170 	hammer2_chain_t *parent;
3171 
3172 	--*countp;
3173 	if (*countp == 0) {
3174 		kprintf("%*.*s...\n", tab, tab, "");
3175 		return;
3176 	}
3177 	if (*countp < 0)
3178 		return;
3179 	kprintf("%*.*s%c-chain %p %s.%-3d %016jx %016jx/%-2d mir=%016jx\n",
3180 		tab, tab, "", pfx, chain,
3181 		hammer2_bref_type_str(chain->bref.type), bi,
3182 		chain->bref.data_off, chain->bref.key, chain->bref.keybits,
3183 		chain->bref.mirror_tid);
3184 
3185 	kprintf("%*.*s      [%08x] (%s) refs=%d",
3186 		tab, tab, "",
3187 		chain->flags,
3188 		((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3189 		chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
3190 		chain->refs);
3191 
3192 	parent = chain->parent;
3193 	if (parent)
3194 		kprintf("\n%*.*s      p=%p [pflags %08x prefs %d]",
3195 			tab, tab, "",
3196 			parent, parent->flags, parent->refs);
3197 	if (RB_EMPTY(&chain->core.rbtree)) {
3198 		kprintf("\n");
3199 	} else {
3200 		int bi = 0;
3201 		kprintf(" {\n");
3202 		RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3203 			if ((scan->flags & flags) || flags == (u_int)-1) {
3204 				hammer2_dump_chain(scan, tab + 4, bi, countp,
3205 						   'a', flags);
3206 			}
3207 			bi++;
3208 		}
3209 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3210 			kprintf("%*.*s}(%s)\n", tab, tab, "",
3211 				chain->data->ipdata.filename);
3212 		else
3213 			kprintf("%*.*s}\n", tab, tab, "");
3214 	}
3215 }
3216