xref: /dflybsd-src/sys/vfs/hammer2/hammer2_vfsops.c (revision e9dbfea17a45e13134f19ed8fa22fbe8d11ac99c)
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 #define REPORT_REFS_ERRORS 1	/* XXX remove me */
62 
63 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
64 
65 struct hammer2_sync_info {
66 	int error;
67 	int waitfor;
68 	int pass;
69 };
70 
71 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
72 static struct hammer2_mntlist hammer2_mntlist;
73 
74 struct hammer2_pfslist hammer2_pfslist;
75 struct hammer2_pfslist hammer2_spmplist;
76 struct lock hammer2_mntlk;
77 
78 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
79 int hammer2_debug;
80 int hammer2_xopgroups;
81 long hammer2_debug_inode;
82 int hammer2_cluster_meta_read = 1;	/* physical read-ahead */
83 int hammer2_cluster_data_read = 4;	/* physical read-ahead */
84 int hammer2_cluster_write = 0;		/* physical write clustering */
85 int hammer2_dedup_enable = 1;
86 int hammer2_always_compress = 0;	/* always try to compress */
87 int hammer2_flush_pipe = 100;
88 int hammer2_dio_count;
89 int hammer2_dio_limit = 256;
90 int hammer2_bulkfree_tps = 5000;
91 int hammer2_worker_rmask = 3;
92 long hammer2_chain_allocs;
93 long hammer2_limit_dirty_chains;
94 long hammer2_limit_dirty_inodes;
95 long hammer2_count_modified_chains;
96 long hammer2_iod_file_read;
97 long hammer2_iod_meta_read;
98 long hammer2_iod_indr_read;
99 long hammer2_iod_fmap_read;
100 long hammer2_iod_volu_read;
101 long hammer2_iod_file_write;
102 long hammer2_iod_file_wembed;
103 long hammer2_iod_file_wzero;
104 long hammer2_iod_file_wdedup;
105 long hammer2_iod_meta_write;
106 long hammer2_iod_indr_write;
107 long hammer2_iod_fmap_write;
108 long hammer2_iod_volu_write;
109 static long hammer2_iod_inode_creates;
110 static long hammer2_iod_inode_deletes;
111 
112 long hammer2_process_icrc32;
113 long hammer2_process_xxhash64;
114 
115 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
116 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
117 		"Buffer used for compression.");
118 
119 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
120 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
121 		"Buffer used for decompression.");
122 
123 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
124 
125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
126 	   &hammer2_supported_version, 0, "");
127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
128 	   &hammer2_debug, 0, "");
129 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW,
130 	   &hammer2_debug_inode, 0, "");
131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
132 	   &hammer2_cluster_meta_read, 0, "");
133 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
134 	   &hammer2_cluster_data_read, 0, "");
135 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
136 	   &hammer2_cluster_write, 0, "");
137 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
138 	   &hammer2_dedup_enable, 0, "");
139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
140 	   &hammer2_always_compress, 0, "");
141 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
142 	   &hammer2_flush_pipe, 0, "");
143 SYSCTL_INT(_vfs_hammer2, OID_AUTO, worker_rmask, CTLFLAG_RW,
144 	   &hammer2_worker_rmask, 0, "");
145 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
146 	   &hammer2_bulkfree_tps, 0, "");
147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
148 	   &hammer2_chain_allocs, 0, "");
149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
150 	   &hammer2_limit_dirty_chains, 0, "");
151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
152 	   &hammer2_limit_dirty_inodes, 0, "");
153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
154 	   &hammer2_count_modified_chains, 0, "");
155 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
156 	   &hammer2_dio_count, 0, "");
157 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
158 	   &hammer2_dio_limit, 0, "");
159 
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
161 	   &hammer2_iod_file_read, 0, "");
162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
163 	   &hammer2_iod_meta_read, 0, "");
164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
165 	   &hammer2_iod_indr_read, 0, "");
166 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
167 	   &hammer2_iod_fmap_read, 0, "");
168 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
169 	   &hammer2_iod_volu_read, 0, "");
170 
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
172 	   &hammer2_iod_file_write, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
174 	   &hammer2_iod_file_wembed, 0, "");
175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
176 	   &hammer2_iod_file_wzero, 0, "");
177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
178 	   &hammer2_iod_file_wdedup, 0, "");
179 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
180 	   &hammer2_iod_meta_write, 0, "");
181 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
182 	   &hammer2_iod_indr_write, 0, "");
183 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
184 	   &hammer2_iod_fmap_write, 0, "");
185 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
186 	   &hammer2_iod_volu_write, 0, "");
187 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW,
188 	   &hammer2_iod_inode_creates, 0, "");
189 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW,
190 	   &hammer2_iod_inode_deletes, 0, "");
191 
192 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
193 	   &hammer2_process_icrc32, 0, "");
194 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
195 	   &hammer2_process_xxhash64, 0, "");
196 
197 static int hammer2_vfs_init(struct vfsconf *conf);
198 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
199 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
200 				struct ucred *cred);
201 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
202 				struct ucred *);
203 static int hammer2_recovery(hammer2_dev_t *hmp);
204 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
205 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
206 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
207 				struct ucred *cred);
208 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
209 				struct ucred *cred);
210 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
211 				struct fid *fhp, struct vnode **vpp);
212 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
213 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
214 				int *exflagsp, struct ucred **credanonp);
215 static int hammer2_vfs_modifying(struct mount *mp);
216 
217 static void hammer2_update_pmps(hammer2_dev_t *hmp);
218 
219 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
220 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
221 				hammer2_dev_t *hmp);
222 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
223 
224 /*
225  * HAMMER2 vfs operations.
226  */
227 static struct vfsops hammer2_vfsops = {
228 	.vfs_flags	= 0,
229 	.vfs_init	= hammer2_vfs_init,
230 	.vfs_uninit	= hammer2_vfs_uninit,
231 	.vfs_sync	= hammer2_vfs_sync,
232 	.vfs_mount	= hammer2_vfs_mount,
233 	.vfs_unmount	= hammer2_vfs_unmount,
234 	.vfs_root 	= hammer2_vfs_root,
235 	.vfs_statfs	= hammer2_vfs_statfs,
236 	.vfs_statvfs	= hammer2_vfs_statvfs,
237 	.vfs_vget	= hammer2_vfs_vget,
238 	.vfs_vptofh	= hammer2_vfs_vptofh,
239 	.vfs_fhtovp	= hammer2_vfs_fhtovp,
240 	.vfs_checkexp	= hammer2_vfs_checkexp,
241 	.vfs_modifying	= hammer2_vfs_modifying
242 };
243 
244 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
245 
246 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
247 MODULE_VERSION(hammer2, 1);
248 
249 static
250 int
251 hammer2_vfs_init(struct vfsconf *conf)
252 {
253 	static struct objcache_malloc_args margs_read;
254 	static struct objcache_malloc_args margs_write;
255 	static struct objcache_malloc_args margs_vop;
256 
257 	int error;
258 
259 	error = 0;
260 	kmalloc_raise_limit(M_HAMMER2, 0);	/* unlimited */
261 
262 	/*
263 	 * hammer2_xopgroups must be even and is most optimal if
264 	 * 2 x ncpus so strategy functions can be queued to the same
265 	 * cpu.
266 	 */
267 	hammer2_xopgroups = HAMMER2_XOPGROUPS_MIN;
268 	if (hammer2_xopgroups < ncpus * 2)
269 		hammer2_xopgroups = ncpus * 2;
270 
271 	/*
272 	 * A large DIO cache is needed to retain dedup enablement masks.
273 	 * The bulkfree code clears related masks as part of the disk block
274 	 * recycling algorithm, preventing it from being used for a later
275 	 * dedup.
276 	 *
277 	 * NOTE: A large buffer cache can actually interfere with dedup
278 	 *	 operation because we dedup based on media physical buffers
279 	 *	 and not logical buffers.  Try to make the DIO case large
280 	 *	 enough to avoid this problem, but also cap it.
281 	 */
282 	hammer2_dio_limit = nbuf * 2;
283 	if (hammer2_dio_limit > 100000)
284 		hammer2_dio_limit = 100000;
285 
286 	if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
287 		error = EINVAL;
288 	if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
289 		error = EINVAL;
290 	if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
291 		error = EINVAL;
292 
293 	if (error)
294 		kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
295 
296 	margs_read.objsize = 65536;
297 	margs_read.mtype = M_HAMMER2_DEBUFFER;
298 
299 	margs_write.objsize = 32768;
300 	margs_write.mtype = M_HAMMER2_CBUFFER;
301 
302 	margs_vop.objsize = sizeof(hammer2_xop_t);
303 	margs_vop.mtype = M_HAMMER2;
304 
305 	/*
306 	 * Note thaht for the XOPS cache we want backing store allocations
307 	 * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
308 	 * confusion), so use the backing store function that does it.  This
309 	 * means that initial XOPS objects are zerod but REUSED objects are
310 	 * not.  So we are responsible for cleaning the object up sufficiently
311 	 * for our needs before objcache_put()ing it back (typically just the
312 	 * FIFO indices).
313 	 */
314 	cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
315 				0, 1, NULL, NULL, NULL,
316 				objcache_malloc_alloc,
317 				objcache_malloc_free,
318 				&margs_read);
319 	cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
320 				0, 1, NULL, NULL, NULL,
321 				objcache_malloc_alloc,
322 				objcache_malloc_free,
323 				&margs_write);
324 	cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
325 				0, 1, NULL, NULL, NULL,
326 				objcache_malloc_alloc_zero,
327 				objcache_malloc_free,
328 				&margs_vop);
329 
330 
331 	lockinit(&hammer2_mntlk, "mntlk", 0, 0);
332 	TAILQ_INIT(&hammer2_mntlist);
333 	TAILQ_INIT(&hammer2_pfslist);
334 	TAILQ_INIT(&hammer2_spmplist);
335 
336 	hammer2_limit_dirty_chains = maxvnodes / 10;
337 	if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
338 		hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
339 	if (hammer2_limit_dirty_chains < 1000)
340 		hammer2_limit_dirty_chains = 1000;
341 
342 	hammer2_limit_dirty_inodes = maxvnodes / 25;
343 	if (hammer2_limit_dirty_inodes < 100)
344 		hammer2_limit_dirty_inodes = 100;
345 	if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
346 		hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
347 
348 	return (error);
349 }
350 
351 static
352 int
353 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
354 {
355 	objcache_destroy(cache_buffer_read);
356 	objcache_destroy(cache_buffer_write);
357 	objcache_destroy(cache_xops);
358 	return 0;
359 }
360 
361 /*
362  * Core PFS allocator.  Used to allocate or reference the pmp structure
363  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
364  * The pmp can be passed in or loaded by this function using the chain and
365  * inode data.
366  *
367  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
368  * transactions.  Note that synchronization does not use this field.
369  * (typically frontend operations and synchronization cannot run on the
370  * same PFS node at the same time).
371  *
372  * XXX check locking
373  */
374 hammer2_pfs_t *
375 hammer2_pfsalloc(hammer2_chain_t *chain,
376 		 const hammer2_inode_data_t *ripdata,
377 		 hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
378 {
379 	hammer2_pfs_t *pmp;
380 	hammer2_inode_t *iroot;
381 	int count;
382 	int i;
383 	int j;
384 
385 	pmp = NULL;
386 
387 	/*
388 	 * Locate or create the PFS based on the cluster id.  If ripdata
389 	 * is NULL this is a spmp which is unique and is always allocated.
390 	 *
391 	 * If the device is mounted in local mode all PFSs are considered
392 	 * independent and not part of any cluster (for debugging only).
393 	 */
394 	if (ripdata) {
395 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
396 			if (force_local != pmp->force_local)
397 				continue;
398 			if (force_local == NULL &&
399 			    bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
400 				 sizeof(pmp->pfs_clid)) == 0) {
401 					break;
402 			} else if (force_local && pmp->pfs_names[0] &&
403 			    strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
404 					break;
405 			}
406 		}
407 	}
408 
409 	if (pmp == NULL) {
410 		pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
411 		pmp->force_local = force_local;
412 		hammer2_trans_manage_init(pmp);
413 		kmalloc_create_obj(&pmp->minode, "HAMMER2-inodes",
414 				   sizeof(struct hammer2_inode));
415 		lockinit(&pmp->lock, "pfslk", 0, 0);
416 		lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
417 		spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
418 		spin_init(&pmp->xop_spin, "h2xop");
419 		spin_init(&pmp->lru_spin, "h2lru");
420 		RB_INIT(&pmp->inum_tree);
421 		TAILQ_INIT(&pmp->syncq);
422 		TAILQ_INIT(&pmp->depq);
423 		TAILQ_INIT(&pmp->lru_list);
424 		spin_init(&pmp->list_spin, "h2pfsalloc_list");
425 
426 		/*
427 		 * Save the last media transaction id for the flusher.  Set
428 		 * initial
429 		 */
430 		if (ripdata) {
431 			pmp->pfs_clid = ripdata->meta.pfs_clid;
432 			TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
433 		} else {
434 			pmp->flags |= HAMMER2_PMPF_SPMP;
435 			TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
436 		}
437 
438 		/*
439 		 * The synchronization thread may start too early, make
440 		 * sure it stays frozen until we are ready to let it go.
441 		 * XXX
442 		 */
443 		/*
444 		pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
445 					 HAMMER2_THREAD_REMASTER;
446 		*/
447 	}
448 
449 	/*
450 	 * Create the PFS's root inode and any missing XOP helper threads.
451 	 */
452 	if ((iroot = pmp->iroot) == NULL) {
453 		iroot = hammer2_inode_get(pmp, NULL, 1, -1);
454 		if (ripdata)
455 			iroot->meta = ripdata->meta;
456 		pmp->iroot = iroot;
457 		hammer2_inode_ref(iroot);
458 		hammer2_inode_unlock(iroot);
459 	}
460 
461 	/*
462 	 * Stop here if no chain is passed in.
463 	 */
464 	if (chain == NULL)
465 		goto done;
466 
467 	/*
468 	 * When a chain is passed in we must add it to the PFS's root
469 	 * inode, update pmp->pfs_types[], and update the syncronization
470 	 * threads.
471 	 *
472 	 * When forcing local mode, mark the PFS as a MASTER regardless.
473 	 *
474 	 * At the moment empty spots can develop due to removals or failures.
475 	 * Ultimately we want to re-fill these spots but doing so might
476 	 * confused running code. XXX
477 	 */
478 	hammer2_inode_ref(iroot);
479 	hammer2_mtx_ex(&iroot->lock);
480 	j = iroot->cluster.nchains;
481 
482 	if (j == HAMMER2_MAXCLUSTER) {
483 		kprintf("hammer2_pfsalloc: cluster full!\n");
484 		/* XXX fatal error? */
485 	} else {
486 		KKASSERT(chain->pmp == NULL);
487 		chain->pmp = pmp;
488 		hammer2_chain_ref(chain);
489 		iroot->cluster.array[j].chain = chain;
490 		if (force_local)
491 			pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
492 		else
493 			pmp->pfs_types[j] = ripdata->meta.pfs_type;
494 		pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
495 		pmp->pfs_hmps[j] = chain->hmp;
496 		hammer2_spin_ex(&pmp->inum_spin);
497 		pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset;
498 		hammer2_spin_unex(&pmp->inum_spin);
499 
500 		/*
501 		 * If the PFS is already mounted we must account
502 		 * for the mount_count here.
503 		 */
504 		if (pmp->mp)
505 			++chain->hmp->mount_count;
506 
507 		/*
508 		 * May have to fixup dirty chain tracking.  Previous
509 		 * pmp was NULL so nothing to undo.
510 		 */
511 		if (chain->flags & HAMMER2_CHAIN_MODIFIED)
512 			hammer2_pfs_memory_inc(pmp);
513 		++j;
514 	}
515 	iroot->cluster.nchains = j;
516 
517 	/*
518 	 * Update nmasters from any PFS inode which is part of the cluster.
519 	 * It is possible that this will result in a value which is too
520 	 * high.  MASTER PFSs are authoritative for pfs_nmasters and will
521 	 * override this value later on.
522 	 *
523 	 * (This informs us of masters that might not currently be
524 	 *  discoverable by this mount).
525 	 */
526 	if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
527 		pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
528 	}
529 
530 	/*
531 	 * Count visible masters.  Masters are usually added with
532 	 * ripdata->meta.pfs_nmasters set to 1.  This detects when there
533 	 * are more (XXX and must update the master inodes).
534 	 */
535 	count = 0;
536 	for (i = 0; i < iroot->cluster.nchains; ++i) {
537 		if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
538 			++count;
539 	}
540 	if (pmp->pfs_nmasters < count)
541 		pmp->pfs_nmasters = count;
542 
543 	/*
544 	 * Create missing synchronization and support threads.
545 	 *
546 	 * Single-node masters (including snapshots) have nothing to
547 	 * synchronize and do not require this thread.
548 	 *
549 	 * Multi-node masters or any number of soft masters, slaves, copy,
550 	 * or other PFS types need the thread.
551 	 *
552 	 * Each thread is responsible for its particular cluster index.
553 	 * We use independent threads so stalls or mismatches related to
554 	 * any given target do not affect other targets.
555 	 */
556 	for (i = 0; i < iroot->cluster.nchains; ++i) {
557 		/*
558 		 * Single-node masters (including snapshots) have nothing
559 		 * to synchronize and will make direct xops support calls,
560 		 * thus they do not require this thread.
561 		 *
562 		 * Note that there can be thousands of snapshots.  We do not
563 		 * want to create thousands of threads.
564 		 */
565 		if (pmp->pfs_nmasters <= 1 &&
566 		    pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
567 			continue;
568 		}
569 
570 		/*
571 		 * Sync support thread
572 		 */
573 		if (pmp->sync_thrs[i].td == NULL) {
574 			hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
575 					   "h2nod", i, -1,
576 					   hammer2_primary_sync_thread);
577 		}
578 	}
579 
580 	/*
581 	 * Create missing Xop threads
582 	 *
583 	 * NOTE: We create helper threads for all mounted PFSs or any
584 	 *	 PFSs with 2+ nodes (so the sync thread can update them,
585 	 *	 even if not mounted).
586 	 */
587 	if (pmp->mp || iroot->cluster.nchains >= 2)
588 		hammer2_xop_helper_create(pmp);
589 
590 	hammer2_mtx_unlock(&iroot->lock);
591 	hammer2_inode_drop(iroot);
592 done:
593 	return pmp;
594 }
595 
596 /*
597  * Deallocate an element of a probed PFS.  If destroying and this is a
598  * MASTER, adjust nmasters.
599  *
600  * This function does not physically destroy the PFS element in its device
601  * under the super-root  (see hammer2_ioctl_pfs_delete()).
602  */
603 void
604 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
605 {
606 	hammer2_inode_t *iroot;
607 	hammer2_chain_t *chain;
608 	int j;
609 
610 	/*
611 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
612 	 * by the flush code.
613 	 */
614 	iroot = pmp->iroot;
615 	if (iroot) {
616 		/*
617 		 * Stop synchronizing
618 		 *
619 		 * XXX flush after acquiring the iroot lock.
620 		 * XXX clean out the cluster index from all inode structures.
621 		 */
622 		hammer2_thr_delete(&pmp->sync_thrs[clindex]);
623 
624 		/*
625 		 * Remove the cluster index from the group.  If destroying
626 		 * the PFS and this is a master, adjust pfs_nmasters.
627 		 */
628 		hammer2_mtx_ex(&iroot->lock);
629 		chain = iroot->cluster.array[clindex].chain;
630 		iroot->cluster.array[clindex].chain = NULL;
631 
632 		switch(pmp->pfs_types[clindex]) {
633 		case HAMMER2_PFSTYPE_MASTER:
634 			if (destroying && pmp->pfs_nmasters > 0)
635 				--pmp->pfs_nmasters;
636 			/* XXX adjust ripdata->meta.pfs_nmasters */
637 			break;
638 		default:
639 			break;
640 		}
641 		pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
642 
643 		hammer2_mtx_unlock(&iroot->lock);
644 
645 		/*
646 		 * Release the chain.
647 		 */
648 		if (chain) {
649 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
650 			hammer2_chain_drop(chain);
651 		}
652 
653 		/*
654 		 * Terminate all XOP threads for the cluster index.
655 		 */
656 		if (pmp->xop_groups) {
657 			for (j = 0; j < hammer2_xopgroups; ++j) {
658 				hammer2_thr_delete(
659 					&pmp->xop_groups[j].thrs[clindex]);
660 			}
661 		}
662 	}
663 }
664 
665 /*
666  * Destroy a PFS, typically only occurs after the last mount on a device
667  * has gone away.
668  */
669 static void
670 hammer2_pfsfree(hammer2_pfs_t *pmp)
671 {
672 	hammer2_inode_t *iroot;
673 	hammer2_chain_t *chain;
674 	int chains_still_present = 0;
675 	int i;
676 	int j;
677 
678 	/*
679 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
680 	 * by the flush code.
681 	 */
682 	if (pmp->flags & HAMMER2_PMPF_SPMP)
683 		TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
684 	else
685 		TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
686 
687 	/*
688 	 * Cleanup chains remaining on LRU list.
689 	 */
690 	hammer2_spin_ex(&pmp->lru_spin);
691 	while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
692 		KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
693 		atomic_add_int(&pmp->lru_count, -1);
694 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
695 		TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
696 		hammer2_chain_ref(chain);
697 		hammer2_spin_unex(&pmp->lru_spin);
698 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
699 		hammer2_chain_drop(chain);
700 		hammer2_spin_ex(&pmp->lru_spin);
701 	}
702 	hammer2_spin_unex(&pmp->lru_spin);
703 
704 	/*
705 	 * Clean up iroot
706 	 */
707 	iroot = pmp->iroot;
708 	if (iroot) {
709 		for (i = 0; i < iroot->cluster.nchains; ++i) {
710 			hammer2_thr_delete(&pmp->sync_thrs[i]);
711 			if (pmp->xop_groups) {
712 				for (j = 0; j < hammer2_xopgroups; ++j)
713 					hammer2_thr_delete(
714 						&pmp->xop_groups[j].thrs[i]);
715 			}
716 			chain = iroot->cluster.array[i].chain;
717 			if (chain && !RB_EMPTY(&chain->core.rbtree)) {
718 				kprintf("hammer2: Warning pmp %p still "
719 					"has active chains\n", pmp);
720 				chains_still_present = 1;
721 			}
722 		}
723 #if REPORT_REFS_ERRORS
724 		if (iroot->refs != 1)
725 			kprintf("PMP->IROOT %p REFS WRONG %d\n",
726 				iroot, iroot->refs);
727 #else
728 		KKASSERT(iroot->refs == 1);
729 #endif
730 		/* ref for iroot */
731 		hammer2_inode_drop(iroot);
732 		pmp->iroot = NULL;
733 	}
734 
735 	/*
736 	 * Free remaining pmp resources
737 	 */
738 	if (chains_still_present) {
739 		kprintf("hammer2: cannot free pmp %p, still in use\n", pmp);
740 	} else {
741 		kmalloc_destroy_obj(&pmp->minode);
742 		kfree(pmp, M_HAMMER2);
743 	}
744 }
745 
746 /*
747  * Remove all references to hmp from the pfs list.  Any PFS which becomes
748  * empty is terminated and freed.
749  *
750  * XXX inefficient.
751  */
752 static void
753 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
754 {
755 	hammer2_pfs_t *pmp;
756 	hammer2_inode_t *iroot;
757 	hammer2_chain_t *rchain;
758 	int i;
759 	int j;
760 	struct hammer2_pfslist *wlist;
761 
762 	if (which == 0)
763 		wlist = &hammer2_pfslist;
764 	else
765 		wlist = &hammer2_spmplist;
766 again:
767 	TAILQ_FOREACH(pmp, wlist, mntentry) {
768 		if ((iroot = pmp->iroot) == NULL)
769 			continue;
770 
771 		/*
772 		 * Determine if this PFS is affected.  If it is we must
773 		 * freeze all management threads and lock its iroot.
774 		 *
775 		 * Freezing a management thread forces it idle, operations
776 		 * in-progress will be aborted and it will have to start
777 		 * over again when unfrozen, or exit if told to exit.
778 		 */
779 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
780 			if (pmp->pfs_hmps[i] == hmp)
781 				break;
782 		}
783 		if (i == HAMMER2_MAXCLUSTER)
784 			continue;
785 
786 		hammer2_vfs_sync_pmp(pmp, MNT_WAIT);
787 
788 		/*
789 		 * Make sure all synchronization threads are locked
790 		 * down.
791 		 */
792 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
793 			if (pmp->pfs_hmps[i] == NULL)
794 				continue;
795 			hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
796 			if (pmp->xop_groups) {
797 				for (j = 0; j < hammer2_xopgroups; ++j) {
798 					hammer2_thr_freeze_async(
799 						&pmp->xop_groups[j].thrs[i]);
800 				}
801 			}
802 		}
803 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
804 			if (pmp->pfs_hmps[i] == NULL)
805 				continue;
806 			hammer2_thr_freeze(&pmp->sync_thrs[i]);
807 			if (pmp->xop_groups) {
808 				for (j = 0; j < hammer2_xopgroups; ++j) {
809 					hammer2_thr_freeze(
810 						&pmp->xop_groups[j].thrs[i]);
811 				}
812 			}
813 		}
814 
815 		/*
816 		 * Lock the inode and clean out matching chains.
817 		 * Note that we cannot use hammer2_inode_lock_*()
818 		 * here because that would attempt to validate the
819 		 * cluster that we are in the middle of ripping
820 		 * apart.
821 		 *
822 		 * WARNING! We are working directly on the inodes
823 		 *	    embedded cluster.
824 		 */
825 		hammer2_mtx_ex(&iroot->lock);
826 
827 		/*
828 		 * Remove the chain from matching elements of the PFS.
829 		 */
830 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
831 			if (pmp->pfs_hmps[i] != hmp)
832 				continue;
833 			hammer2_thr_delete(&pmp->sync_thrs[i]);
834 			if (pmp->xop_groups) {
835 				for (j = 0; j < hammer2_xopgroups; ++j) {
836 					hammer2_thr_delete(
837 						&pmp->xop_groups[j].thrs[i]);
838 				}
839 			}
840 			rchain = iroot->cluster.array[i].chain;
841 			iroot->cluster.array[i].chain = NULL;
842 			pmp->pfs_types[i] = 0;
843 			if (pmp->pfs_names[i]) {
844 				kfree(pmp->pfs_names[i], M_HAMMER2);
845 				pmp->pfs_names[i] = NULL;
846 			}
847 			if (rchain) {
848 				hammer2_chain_drop(rchain);
849 				/* focus hint */
850 				if (iroot->cluster.focus == rchain)
851 					iroot->cluster.focus = NULL;
852 			}
853 			pmp->pfs_hmps[i] = NULL;
854 		}
855 		hammer2_mtx_unlock(&iroot->lock);
856 
857 		/*
858 		 * Cleanup trailing chains.  Gaps may remain.
859 		 */
860 		for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
861 			if (pmp->pfs_hmps[i])
862 				break;
863 		}
864 		iroot->cluster.nchains = i + 1;
865 
866 		/*
867 		 * If the PMP has no elements remaining we can destroy it.
868 		 * (this will transition management threads from frozen->exit).
869 		 */
870 		if (iroot->cluster.nchains == 0) {
871 			/*
872 			 * If this was the hmp's spmp, we need to clean
873 			 * a little more stuff out.
874 			 */
875 			if (hmp->spmp == pmp) {
876 				hmp->spmp = NULL;
877 				hmp->vchain.pmp = NULL;
878 				hmp->fchain.pmp = NULL;
879 			}
880 
881 			/*
882 			 * Free the pmp and restart the loop
883 			 */
884 			KKASSERT(TAILQ_EMPTY(&pmp->syncq));
885 			KKASSERT(TAILQ_EMPTY(&pmp->depq));
886 			hammer2_pfsfree(pmp);
887 			goto again;
888 		}
889 
890 		/*
891 		 * If elements still remain we need to set the REMASTER
892 		 * flag and unfreeze it.
893 		 */
894 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
895 			if (pmp->pfs_hmps[i] == NULL)
896 				continue;
897 			hammer2_thr_remaster(&pmp->sync_thrs[i]);
898 			hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
899 			if (pmp->xop_groups) {
900 				for (j = 0; j < hammer2_xopgroups; ++j) {
901 					hammer2_thr_remaster(
902 						&pmp->xop_groups[j].thrs[i]);
903 					hammer2_thr_unfreeze(
904 						&pmp->xop_groups[j].thrs[i]);
905 				}
906 			}
907 		}
908 	}
909 }
910 
911 /*
912  * Mount or remount HAMMER2 fileystem from physical media
913  *
914  *	mountroot
915  *		mp		mount point structure
916  *		path		NULL
917  *		data		<unused>
918  *		cred		<unused>
919  *
920  *	mount
921  *		mp		mount point structure
922  *		path		path to mount point
923  *		data		pointer to argument structure in user space
924  *			volume	volume path (device@LABEL form)
925  *			hflags	user mount flags
926  *		cred		user credentials
927  *
928  * RETURNS:	0	Success
929  *		!0	error number
930  */
931 static
932 int
933 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
934 		  struct ucred *cred)
935 {
936 	struct hammer2_mount_info info;
937 	hammer2_pfs_t *pmp;
938 	hammer2_pfs_t *spmp;
939 	hammer2_dev_t *hmp, *hmp_tmp;
940 	hammer2_dev_t *force_local;
941 	hammer2_key_t key_next;
942 	hammer2_key_t key_dummy;
943 	hammer2_key_t lhc;
944 	hammer2_chain_t *parent;
945 	hammer2_chain_t *chain;
946 	const hammer2_inode_data_t *ripdata;
947 	hammer2_blockref_t bref;
948 	hammer2_devvp_list_t devvpl;
949 	hammer2_devvp_t *e, *e_tmp;
950 	struct file *fp;
951 	char devstr[MNAMELEN];
952 	size_t size;
953 	size_t done;
954 	char *dev;
955 	char *label;
956 	int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
957 	int error;
958 	int i;
959 
960 	hmp = NULL;
961 	pmp = NULL;
962 	dev = NULL;
963 	label = NULL;
964 	bzero(&info, sizeof(info));
965 
966 	if (path) {
967 		/*
968 		 * Non-root mount or updating a mount
969 		 */
970 		error = copyin(data, &info, sizeof(info));
971 		if (error)
972 			return (error);
973 	}
974 
975 	if (mp->mnt_flag & MNT_UPDATE) {
976 		/*
977 		 * Update mount.  Note that pmp->iroot->cluster is
978 		 * an inode-embedded cluster and thus cannot be
979 		 * directly locked.
980 		 *
981 		 * XXX HAMMER2 needs to implement NFS export via
982 		 *     mountctl.
983 		 */
984 		hammer2_cluster_t *cluster;
985 
986 		pmp = MPTOPMP(mp);
987 		pmp->hflags = info.hflags;
988 		cluster = &pmp->iroot->cluster;
989 		for (i = 0; i < cluster->nchains; ++i) {
990 			if (cluster->array[i].chain == NULL)
991 				continue;
992 			hmp = cluster->array[i].chain->hmp;
993 			error = hammer2_remount(hmp, mp, path, cred);
994 			if (error)
995 				break;
996 		}
997 
998 		return error;
999 	}
1000 
1001 	if (path == NULL) {
1002 		/*
1003 		 * Root mount
1004 		 */
1005 		info.cluster_fd = -1;
1006 		ksnprintf(devstr, sizeof(devstr), "%s",
1007 			  mp->mnt_stat.f_mntfromname);
1008 		done = strlen(devstr) + 1;
1009 		kprintf("hammer2_mount: root devstr=\"%s\"\n", devstr);
1010 	} else {
1011 		error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
1012 		if (error)
1013 			return (error);
1014 		kprintf("hammer2_mount: devstr=\"%s\"\n", devstr);
1015 	}
1016 
1017 	/*
1018 	 * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
1019 	 * if no label specified, based on the partition id.  Error out if no
1020 	 * label or device (with partition id) is specified.  This is strictly
1021 	 * a convenience to match the default label created by newfs_hammer2,
1022 	 * our preference is that a label always be specified.
1023 	 *
1024 	 * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
1025 	 *	 that does not specify a device, as long as some H2 label
1026 	 *	 has already been mounted from that device.  This makes
1027 	 *	 mounting snapshots a lot easier.
1028 	 */
1029 	dev = devstr;
1030 	label = strchr(devstr, '@');
1031 	if (label && ((label + 1) - dev) > done) {
1032 		kprintf("hammer2_mount: bad label %s/%zd\n", devstr, done);
1033 		return (EINVAL);
1034 	}
1035 	if (label == NULL || label[1] == 0) {
1036 		char slice;
1037 
1038 		if (label == NULL)
1039 			label = devstr + strlen(devstr);
1040 		else
1041 			*label = '\0';		/* clean up trailing @ */
1042 
1043 		slice = label[-1];
1044 		switch(slice) {
1045 		case 'a':
1046 			label = "BOOT";
1047 			break;
1048 		case 'd':
1049 			label = "ROOT";
1050 			break;
1051 		default:
1052 			label = "DATA";
1053 			break;
1054 		}
1055 	} else {
1056 		*label = '\0';
1057 		label++;
1058 	}
1059 
1060 	kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
1061 		dev, label, ronly);
1062 
1063 	/*
1064 	 * Initialize all device vnodes.
1065 	 */
1066 	TAILQ_INIT(&devvpl);
1067 	error = hammer2_init_devvp(dev, path == NULL, &devvpl);
1068 	if (error) {
1069 		kprintf("hammer2: failed to initialize devvp in %s\n", dev);
1070 		hammer2_cleanup_devvp(&devvpl);
1071 		return error;
1072 	}
1073 
1074 	/*
1075 	 * Determine if the device has already been mounted.  After this
1076 	 * check hmp will be non-NULL if we are doing the second or more
1077 	 * hammer2 mounts from the same device.
1078 	 */
1079 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1080 	if (!TAILQ_EMPTY(&devvpl)) {
1081 		/*
1082 		 * Match the device.  Due to the way devfs works,
1083 		 * we may not be able to directly match the vnode pointer,
1084 		 * so also check to see if the underlying device matches.
1085 		 */
1086 		TAILQ_FOREACH(hmp_tmp, &hammer2_mntlist, mntentry) {
1087 			TAILQ_FOREACH(e_tmp, &hmp_tmp->devvpl, entry) {
1088 				int devvp_found = 0;
1089 				TAILQ_FOREACH(e, &devvpl, entry) {
1090 					KKASSERT(e->devvp);
1091 					if (e_tmp->devvp == e->devvp)
1092 						devvp_found = 1;
1093 					if (e_tmp->devvp->v_rdev &&
1094 					    e_tmp->devvp->v_rdev == e->devvp->v_rdev)
1095 						devvp_found = 1;
1096 				}
1097 				if (!devvp_found)
1098 					goto next_hmp;
1099 			}
1100 			hmp = hmp_tmp;
1101 			kprintf("hammer2_mount: hmp=%p matched\n", hmp);
1102 			break;
1103 next_hmp:
1104 			continue;
1105 		}
1106 
1107 		/*
1108 		 * If no match this may be a fresh H2 mount, make sure
1109 		 * the device is not mounted on anything else.
1110 		 */
1111 		if (hmp == NULL) {
1112 			TAILQ_FOREACH(e, &devvpl, entry) {
1113 				struct vnode *devvp = e->devvp;
1114 				KKASSERT(devvp);
1115 				error = vfs_mountedon(devvp);
1116 				if (error) {
1117 					kprintf("hammer2_mount: %s mounted %d\n",
1118 						e->path, error);
1119 					hammer2_cleanup_devvp(&devvpl);
1120 					lockmgr(&hammer2_mntlk, LK_RELEASE);
1121 					return error;
1122 				}
1123 			}
1124 		}
1125 	} else {
1126 		/*
1127 		 * Match the label to a pmp already probed.
1128 		 */
1129 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1130 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1131 				if (pmp->pfs_names[i] &&
1132 				    strcmp(pmp->pfs_names[i], label) == 0) {
1133 					hmp = pmp->pfs_hmps[i];
1134 					break;
1135 				}
1136 			}
1137 			if (hmp)
1138 				break;
1139 		}
1140 		if (hmp == NULL) {
1141 			kprintf("hammer2_mount: PFS label \"%s\" not found\n",
1142 				label);
1143 			hammer2_cleanup_devvp(&devvpl);
1144 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1145 			return ENOENT;
1146 		}
1147 	}
1148 
1149 	/*
1150 	 * Open the device if this isn't a secondary mount and construct
1151 	 * the H2 device mount (hmp).
1152 	 */
1153 	if (hmp == NULL) {
1154 		hammer2_chain_t *schain;
1155 		hammer2_xid_t xid;
1156 		hammer2_xop_head_t xop;
1157 
1158 		/*
1159 		 * Now open the device
1160 		 */
1161 		KKASSERT(!TAILQ_EMPTY(&devvpl));
1162 		if (error == 0) {
1163 			error = hammer2_open_devvp(&devvpl, ronly);
1164 			if (error) {
1165 				hammer2_close_devvp(&devvpl, ronly);
1166 				hammer2_cleanup_devvp(&devvpl);
1167 				lockmgr(&hammer2_mntlk, LK_RELEASE);
1168 				return error;
1169 			}
1170 		}
1171 
1172 		/*
1173 		 * Construct volumes and link with device vnodes.
1174 		 */
1175 		hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1176 		hmp->devvp = NULL;
1177 		error = hammer2_init_volumes(mp, &devvpl, hmp->volumes,
1178 					     &hmp->voldata, &hmp->devvp);
1179 		if (error) {
1180 			hammer2_close_devvp(&devvpl, ronly);
1181 			hammer2_cleanup_devvp(&devvpl);
1182 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1183 			kfree(hmp, M_HAMMER2);
1184 			return error;
1185 		}
1186 		if (!hmp->devvp) {
1187 			kprintf("hammer2: failed to initialize root volume\n");
1188 			hammer2_unmount_helper(mp, NULL, hmp);
1189 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1190 			hammer2_vfs_unmount(mp, MNT_FORCE);
1191 			return EINVAL;
1192 		}
1193 
1194 		ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1195 		hmp->ronly = ronly;
1196 		hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1197 		kmalloc_create_obj(&hmp->mchain, "HAMMER2-chains",
1198 				   sizeof(struct hammer2_chain));
1199 		kmalloc_create(&hmp->mmsg, "HAMMER2-msg");
1200 		TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1201 		RB_INIT(&hmp->iotree);
1202 		spin_init(&hmp->io_spin, "h2mount_io");
1203 		spin_init(&hmp->list_spin, "h2mount_list");
1204 
1205 		lockinit(&hmp->vollk, "h2vol", 0, 0);
1206 		lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1207 		lockinit(&hmp->bflock, "h2bflk", 0, 0);
1208 
1209 		/*
1210 		 * vchain setup. vchain.data is embedded.
1211 		 * vchain.refs is initialized and will never drop to 0.
1212 		 *
1213 		 * NOTE! voldata is not yet loaded.
1214 		 */
1215 		hmp->vchain.hmp = hmp;
1216 		hmp->vchain.refs = 1;
1217 		hmp->vchain.data = (void *)&hmp->voldata;
1218 		hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1219 		hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1220 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1221 		hammer2_chain_core_init(&hmp->vchain);
1222 
1223 		/*
1224 		 * fchain setup.  fchain.data is embedded.
1225 		 * fchain.refs is initialized and will never drop to 0.
1226 		 *
1227 		 * The data is not used but needs to be initialized to
1228 		 * pass assertion muster.  We use this chain primarily
1229 		 * as a placeholder for the freemap's top-level radix tree
1230 		 * so it does not interfere with the volume's topology
1231 		 * radix tree.
1232 		 */
1233 		hmp->fchain.hmp = hmp;
1234 		hmp->fchain.refs = 1;
1235 		hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1236 		hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1237 		hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1238 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1239 		hmp->fchain.bref.methods =
1240 			HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1241 			HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1242 		hammer2_chain_core_init(&hmp->fchain);
1243 
1244 		/*
1245 		 * Initialize volume header related fields.
1246 		 */
1247 		KKASSERT(hmp->voldata.magic == HAMMER2_VOLUME_ID_HBO ||
1248 			 hmp->voldata.magic == HAMMER2_VOLUME_ID_ABO);
1249 		hmp->volhdrno = error;
1250 		hmp->volsync = hmp->voldata;
1251 		hmp->free_reserved = hmp->voldata.allocator_size / 20;
1252 		/*
1253 		 * Must use hmp instead of volume header for these two
1254 		 * in order to handle volume versions transparently.
1255 		 */
1256 		if (hmp->voldata.version >= HAMMER2_VOL_VERSION_MULTI_VOLUMES) {
1257 			hmp->nvolumes = hmp->voldata.nvolumes;
1258 			hmp->total_size = hmp->voldata.total_size;
1259 		} else {
1260 			hmp->nvolumes = 1;
1261 			hmp->total_size = hmp->voldata.volu_size;
1262 		}
1263 		KKASSERT(hmp->nvolumes > 0);
1264 
1265 		/*
1266 		 * Move devvpl entries to hmp.
1267 		 */
1268 		TAILQ_INIT(&hmp->devvpl);
1269 		while ((e = TAILQ_FIRST(&devvpl)) != NULL) {
1270 			TAILQ_REMOVE(&devvpl, e, entry);
1271 			TAILQ_INSERT_TAIL(&hmp->devvpl, e, entry);
1272 		}
1273 		KKASSERT(TAILQ_EMPTY(&devvpl));
1274 		KKASSERT(!TAILQ_EMPTY(&hmp->devvpl));
1275 
1276 		/*
1277 		 * Really important to get these right or the flush and
1278 		 * teardown code will get confused.
1279 		 */
1280 		hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1281 		spmp = hmp->spmp;
1282 		spmp->pfs_hmps[0] = hmp;
1283 
1284 		/*
1285 		 * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1286 		 * is inherited from the volume header.
1287 		 */
1288 		xid = 0;
1289 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1290 		hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1291 		hmp->vchain.pmp = spmp;
1292 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1293 		hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1294 		hmp->fchain.pmp = spmp;
1295 
1296 		/*
1297 		 * First locate the super-root inode, which is key 0
1298 		 * relative to the volume header's blockset.
1299 		 *
1300 		 * Then locate the root inode by scanning the directory keyspace
1301 		 * represented by the label.
1302 		 */
1303 		parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1304 		schain = hammer2_chain_lookup(&parent, &key_dummy,
1305 				      HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1306 				      &error, 0);
1307 		hammer2_chain_lookup_done(parent);
1308 		if (schain == NULL) {
1309 			kprintf("hammer2_mount: invalid super-root\n");
1310 			hammer2_unmount_helper(mp, NULL, hmp);
1311 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1312 			hammer2_vfs_unmount(mp, MNT_FORCE);
1313 			return EINVAL;
1314 		}
1315 		if (schain->error) {
1316 			kprintf("hammer2_mount: error %s reading super-root\n",
1317 				hammer2_error_str(schain->error));
1318 			hammer2_chain_unlock(schain);
1319 			hammer2_chain_drop(schain);
1320 			schain = NULL;
1321 			hammer2_unmount_helper(mp, NULL, hmp);
1322 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1323 			hammer2_vfs_unmount(mp, MNT_FORCE);
1324 			return EINVAL;
1325 		}
1326 
1327 		/*
1328 		 * The super-root always uses an inode_tid of 1 when
1329 		 * creating PFSs.
1330 		 */
1331 		spmp->inode_tid = 1;
1332 		spmp->modify_tid = schain->bref.modify_tid + 1;
1333 
1334 		/*
1335 		 * Sanity-check schain's pmp and finish initialization.
1336 		 * Any chain belonging to the super-root topology should
1337 		 * have a NULL pmp (not even set to spmp).
1338 		 */
1339 		ripdata = &schain->data->ipdata;
1340 		KKASSERT(schain->pmp == NULL);
1341 		spmp->pfs_clid = ripdata->meta.pfs_clid;
1342 
1343 		/*
1344 		 * Replace the dummy spmp->iroot with a real one.  It's
1345 		 * easier to just do a wholesale replacement than to try
1346 		 * to update the chain and fixup the iroot fields.
1347 		 *
1348 		 * The returned inode is locked with the supplied cluster.
1349 		 */
1350 		hammer2_dummy_xop_from_chain(&xop, schain);
1351 		hammer2_inode_drop(spmp->iroot);
1352 		spmp->iroot = NULL;
1353 		spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1);
1354 		spmp->spmp_hmp = hmp;
1355 		spmp->pfs_types[0] = ripdata->meta.pfs_type;
1356 		spmp->pfs_hmps[0] = hmp;
1357 		hammer2_inode_ref(spmp->iroot);
1358 		hammer2_inode_unlock(spmp->iroot);
1359 		hammer2_cluster_unlock(&xop.cluster);
1360 		hammer2_chain_drop(schain);
1361 		/* do not call hammer2_cluster_drop() on an embedded cluster */
1362 		schain = NULL;	/* now invalid */
1363 		/* leave spmp->iroot with one ref */
1364 
1365 		if (!hmp->ronly) {
1366 			error = hammer2_recovery(hmp);
1367 			if (error == 0)
1368 				error |= hammer2_fixup_pfses(hmp);
1369 			/* XXX do something with error */
1370 		}
1371 		hammer2_update_pmps(hmp);
1372 		hammer2_iocom_init(hmp);
1373 		hammer2_bulkfree_init(hmp);
1374 
1375 		/*
1376 		 * Ref the cluster management messaging descriptor.  The mount
1377 		 * program deals with the other end of the communications pipe.
1378 		 *
1379 		 * Root mounts typically do not supply one.
1380 		 */
1381 		if (info.cluster_fd >= 0) {
1382 			fp = holdfp(curthread, info.cluster_fd, -1);
1383 			if (fp) {
1384 				hammer2_cluster_reconnect(hmp, fp);
1385 			} else {
1386 				kprintf("hammer2_mount: bad cluster_fd!\n");
1387 			}
1388 		}
1389 	} else {
1390 		spmp = hmp->spmp;
1391 		if (info.hflags & HMNT2_DEVFLAGS) {
1392 			kprintf("hammer2_mount: Warning: mount flags pertaining "
1393 				"to the whole device may only be specified "
1394 				"on the first mount of the device: %08x\n",
1395 				info.hflags & HMNT2_DEVFLAGS);
1396 		}
1397 	}
1398 
1399 	/*
1400 	 * Force local mount (disassociate all PFSs from their clusters).
1401 	 * Used primarily for debugging.
1402 	 */
1403 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1404 
1405 	/*
1406 	 * Lookup the mount point under the media-localized super-root.
1407 	 * Scanning hammer2_pfslist doesn't help us because it represents
1408 	 * PFS cluster ids which can aggregate several named PFSs together.
1409 	 *
1410 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1411 	 * up later on.
1412 	 */
1413 	hammer2_inode_lock(spmp->iroot, 0);
1414 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1415 	lhc = hammer2_dirhash(label, strlen(label));
1416 	chain = hammer2_chain_lookup(&parent, &key_next,
1417 				     lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1418 				     &error, 0);
1419 	while (chain) {
1420 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1421 		    strcmp(label, chain->data->ipdata.filename) == 0) {
1422 			break;
1423 		}
1424 		chain = hammer2_chain_next(&parent, chain, &key_next,
1425 					    key_next,
1426 					    lhc + HAMMER2_DIRHASH_LOMASK,
1427 					    &error, 0);
1428 	}
1429 	if (parent) {
1430 		hammer2_chain_unlock(parent);
1431 		hammer2_chain_drop(parent);
1432 	}
1433 	hammer2_inode_unlock(spmp->iroot);
1434 
1435 	/*
1436 	 * PFS could not be found?
1437 	 */
1438 	if (chain == NULL) {
1439 		hammer2_unmount_helper(mp, NULL, hmp);
1440 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1441 		hammer2_vfs_unmount(mp, MNT_FORCE);
1442 
1443 		if (error) {
1444 			kprintf("hammer2_mount: PFS label I/O error\n");
1445 			return EINVAL;
1446 		} else {
1447 			kprintf("hammer2_mount: PFS label \"%s\" not found\n",
1448 				label);
1449 			return ENOENT;
1450 		}
1451 	}
1452 
1453 	/*
1454 	 * Acquire the pmp structure (it should have already been allocated
1455 	 * via hammer2_update_pmps() so do not pass cluster in to add to
1456 	 * available chains).
1457 	 *
1458 	 * Check if the cluster has already been mounted.  A cluster can
1459 	 * only be mounted once, use null mounts to mount additional copies.
1460 	 */
1461 	if (chain->error) {
1462 		kprintf("hammer2_mount: PFS label I/O error\n");
1463 	} else {
1464 		ripdata = &chain->data->ipdata;
1465 		bref = chain->bref;
1466 		pmp = hammer2_pfsalloc(NULL, ripdata,
1467 				       bref.modify_tid, force_local);
1468 	}
1469 	hammer2_chain_unlock(chain);
1470 	hammer2_chain_drop(chain);
1471 
1472 	/*
1473 	 * Finish the mount
1474 	 */
1475         kprintf("hammer2_mount: hmp=%p pmp=%p\n", hmp, pmp);
1476 
1477 	if (pmp->mp) {
1478 		kprintf("hammer2_mount: PFS already mounted!\n");
1479 		hammer2_unmount_helper(mp, NULL, hmp);
1480 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1481 		hammer2_vfs_unmount(mp, MNT_FORCE);
1482 
1483 		return EBUSY;
1484 	}
1485 
1486 	pmp->hflags = info.hflags;
1487         mp->mnt_flag |= MNT_LOCAL;
1488         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1489         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1490 
1491         /*
1492          * required mount structure initializations
1493          */
1494         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1495         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1496 
1497         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1498         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1499 
1500         /*
1501          * Optional fields
1502          */
1503         mp->mnt_iosize_max = MAXPHYS;
1504 
1505 	/*
1506 	 * Connect up mount pointers.
1507 	 */
1508 	hammer2_mount_helper(mp, pmp);
1509         lockmgr(&hammer2_mntlk, LK_RELEASE);
1510 
1511 	/*
1512 	 * Finish setup
1513 	 */
1514 	vfs_getnewfsid(mp);
1515 	vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1516 	vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1517 	vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1518 
1519 	if (path) {
1520 		copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1521 			  MNAMELEN - 1, &size);
1522 		bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1523 	} /* else root mount, already in there */
1524 
1525 	bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1526 	if (path) {
1527 		copyinstr(path, mp->mnt_stat.f_mntonname,
1528 			  sizeof(mp->mnt_stat.f_mntonname) - 1,
1529 			  &size);
1530 	} else {
1531 		/* root mount */
1532 		mp->mnt_stat.f_mntonname[0] = '/';
1533 	}
1534 
1535 	/*
1536 	 * Initial statfs to prime mnt_stat.
1537 	 */
1538 	hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1539 
1540 	return 0;
1541 }
1542 
1543 /*
1544  * Scan PFSs under the super-root and create hammer2_pfs structures.
1545  */
1546 static
1547 void
1548 hammer2_update_pmps(hammer2_dev_t *hmp)
1549 {
1550 	const hammer2_inode_data_t *ripdata;
1551 	hammer2_chain_t *parent;
1552 	hammer2_chain_t *chain;
1553 	hammer2_blockref_t bref;
1554 	hammer2_dev_t *force_local;
1555 	hammer2_pfs_t *spmp;
1556 	hammer2_pfs_t *pmp;
1557 	hammer2_key_t key_next;
1558 	int error;
1559 
1560 	/*
1561 	 * Force local mount (disassociate all PFSs from their clusters).
1562 	 * Used primarily for debugging.
1563 	 */
1564 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1565 
1566 	/*
1567 	 * Lookup mount point under the media-localized super-root.
1568 	 *
1569 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1570 	 * up later on.
1571 	 */
1572 	spmp = hmp->spmp;
1573 	hammer2_inode_lock(spmp->iroot, 0);
1574 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1575 	chain = hammer2_chain_lookup(&parent, &key_next,
1576 					 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1577 					 &error, 0);
1578 	while (chain) {
1579 		if (chain->error) {
1580 			kprintf("I/O error scanning PFS labels\n");
1581 		} else if (chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
1582 			kprintf("Non inode chain type %d under super-root\n",
1583 				chain->bref.type);
1584 		} else {
1585 			ripdata = &chain->data->ipdata;
1586 			bref = chain->bref;
1587 			pmp = hammer2_pfsalloc(chain, ripdata,
1588 					       bref.modify_tid, force_local);
1589 		}
1590 		chain = hammer2_chain_next(&parent, chain, &key_next,
1591 					   key_next, HAMMER2_KEY_MAX,
1592 					   &error, 0);
1593 	}
1594 	if (parent) {
1595 		hammer2_chain_unlock(parent);
1596 		hammer2_chain_drop(parent);
1597 	}
1598 	hammer2_inode_unlock(spmp->iroot);
1599 }
1600 
1601 static
1602 int
1603 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1604 		struct ucred *cred)
1605 {
1606 	hammer2_volume_t *vol;
1607 	struct vnode *devvp;
1608 	int i, error, result = 0;
1609 
1610 	if (!(hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)))
1611 		return 0;
1612 
1613 	for (i = 0; i < hmp->nvolumes; ++i) {
1614 		vol = &hmp->volumes[i];
1615 		devvp = vol->dev->devvp;
1616 		KKASSERT(devvp);
1617 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1618 		VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1619 		vn_unlock(devvp);
1620 		error = 0;
1621 		if (vol->id == HAMMER2_ROOT_VOLUME) {
1622 			error = hammer2_recovery(hmp);
1623 			if (error == 0)
1624 				error |= hammer2_fixup_pfses(hmp);
1625 		}
1626 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1627 		if (error == 0) {
1628 			VOP_CLOSE(devvp, FREAD, NULL);
1629 		} else {
1630 			VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1631 		}
1632 		vn_unlock(devvp);
1633 		result |= error;
1634 	}
1635 	if (result == 0) {
1636 		kprintf("hammer2: enable read/write\n");
1637 		hmp->ronly = 0;
1638 	}
1639 
1640 	return result;
1641 }
1642 
1643 static
1644 int
1645 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1646 {
1647 	hammer2_pfs_t *pmp;
1648 	int flags;
1649 	int error = 0;
1650 
1651 	pmp = MPTOPMP(mp);
1652 
1653 	if (pmp == NULL)
1654 		return(0);
1655 
1656 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1657 
1658 	/*
1659 	 * If mount initialization proceeded far enough we must flush
1660 	 * its vnodes and sync the underlying mount points.  Three syncs
1661 	 * are required to fully flush the filesystem (freemap updates lag
1662 	 * by one flush, and one extra for safety).
1663 	 */
1664 	if (mntflags & MNT_FORCE)
1665 		flags = FORCECLOSE;
1666 	else
1667 		flags = 0;
1668 	if (pmp->iroot) {
1669 		error = vflush(mp, 0, flags);
1670 		if (error)
1671 			goto failed;
1672 		hammer2_vfs_sync(mp, MNT_WAIT);
1673 		hammer2_vfs_sync(mp, MNT_WAIT);
1674 		hammer2_vfs_sync(mp, MNT_WAIT);
1675 	}
1676 
1677 	/*
1678 	 * Cleanup the frontend support XOPS threads
1679 	 */
1680 	hammer2_xop_helper_cleanup(pmp);
1681 
1682 	if (pmp->mp)
1683 		hammer2_unmount_helper(mp, pmp, NULL);
1684 
1685 	error = 0;
1686 failed:
1687 	lockmgr(&hammer2_mntlk, LK_RELEASE);
1688 
1689 	return (error);
1690 }
1691 
1692 /*
1693  * Mount helper, hook the system mount into our PFS.
1694  * The mount lock is held.
1695  *
1696  * We must bump the mount_count on related devices for any
1697  * mounted PFSs.
1698  */
1699 static
1700 void
1701 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1702 {
1703 	hammer2_cluster_t *cluster;
1704 	hammer2_chain_t *rchain;
1705 	int i;
1706 
1707         mp->mnt_data = (qaddr_t)pmp;
1708 	pmp->mp = mp;
1709 
1710 	/*
1711 	 * After pmp->mp is set we have to adjust hmp->mount_count.
1712 	 */
1713 	cluster = &pmp->iroot->cluster;
1714 	for (i = 0; i < cluster->nchains; ++i) {
1715 		rchain = cluster->array[i].chain;
1716 		if (rchain == NULL)
1717 			continue;
1718 		++rchain->hmp->mount_count;
1719 	}
1720 
1721 	/*
1722 	 * Create missing Xop threads
1723 	 */
1724 	hammer2_xop_helper_create(pmp);
1725 }
1726 
1727 /*
1728  * Mount helper, unhook the system mount from our PFS.
1729  * The mount lock is held.
1730  *
1731  * If hmp is supplied a mount responsible for being the first to open
1732  * the block device failed and the block device and all PFSs using the
1733  * block device must be cleaned up.
1734  *
1735  * If pmp is supplied multiple devices might be backing the PFS and each
1736  * must be disconnected.  This might not be the last PFS using some of the
1737  * underlying devices.  Also, we have to adjust our hmp->mount_count
1738  * accounting for the devices backing the pmp which is now undergoing an
1739  * unmount.
1740  */
1741 static
1742 void
1743 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1744 {
1745 	hammer2_cluster_t *cluster;
1746 	hammer2_chain_t *rchain;
1747 	int dumpcnt;
1748 	int i;
1749 
1750 	/*
1751 	 * If no device supplied this is a high-level unmount and we have to
1752 	 * to disconnect the mount, adjust mount_count, and locate devices
1753 	 * that might now have no mounts.
1754 	 */
1755 	if (pmp) {
1756 		KKASSERT(hmp == NULL);
1757 		KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1758 		pmp->mp = NULL;
1759 		mp->mnt_data = NULL;
1760 
1761 		/*
1762 		 * After pmp->mp is cleared we have to account for
1763 		 * mount_count.
1764 		 */
1765 		cluster = &pmp->iroot->cluster;
1766 		for (i = 0; i < cluster->nchains; ++i) {
1767 			rchain = cluster->array[i].chain;
1768 			if (rchain == NULL)
1769 				continue;
1770 			--rchain->hmp->mount_count;
1771 			/* scrapping hmp now may invalidate the pmp */
1772 		}
1773 again:
1774 		TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1775 			if (hmp->mount_count == 0) {
1776 				hammer2_unmount_helper(NULL, NULL, hmp);
1777 				goto again;
1778 			}
1779 		}
1780 		return;
1781 	}
1782 
1783 	/*
1784 	 * Try to terminate the block device.  We can't terminate it if
1785 	 * there are still PFSs referencing it.
1786 	 */
1787 	if (hmp->mount_count)
1788 		return;
1789 
1790 	/*
1791 	 * Decomission the network before we start messing with the
1792 	 * device and PFS.
1793 	 */
1794 	hammer2_iocom_uninit(hmp);
1795 
1796 	hammer2_bulkfree_uninit(hmp);
1797 	hammer2_pfsfree_scan(hmp, 0);
1798 
1799 	/*
1800 	 * Cycle the volume data lock as a safety (probably not needed any
1801 	 * more).  To ensure everything is out we need to flush at least
1802 	 * three times.  (1) The running of the sideq can dirty the
1803 	 * filesystem, (2) A normal flush can dirty the freemap, and
1804 	 * (3) ensure that the freemap is fully synchronized.
1805 	 *
1806 	 * The next mount's recovery scan can clean everything up but we want
1807 	 * to leave the filesystem in a 100% clean state on a normal unmount.
1808 	 */
1809 #if 0
1810 	hammer2_voldata_lock(hmp);
1811 	hammer2_voldata_unlock(hmp);
1812 #endif
1813 
1814 	/*
1815 	 * Flush whatever is left.  Unmounted but modified PFS's might still
1816 	 * have some dirty chains on them.
1817 	 */
1818 	hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1819 	hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1820 
1821 	if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1822 		hammer2_voldata_modify(hmp);
1823 		hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1824 					    HAMMER2_FLUSH_ALL);
1825 	}
1826 	hammer2_chain_unlock(&hmp->fchain);
1827 
1828 	if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1829 		hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1830 					    HAMMER2_FLUSH_ALL);
1831 	}
1832 	hammer2_chain_unlock(&hmp->vchain);
1833 
1834 	if ((hmp->vchain.flags | hmp->fchain.flags) &
1835 	    HAMMER2_CHAIN_FLUSH_MASK) {
1836 		kprintf("hammer2_unmount: chains left over after final sync\n");
1837 		kprintf("    vchain %08x\n", hmp->vchain.flags);
1838 		kprintf("    fchain %08x\n", hmp->fchain.flags);
1839 
1840 		if (hammer2_debug & 0x0010)
1841 			Debugger("entered debugger");
1842 	}
1843 
1844 	hammer2_pfsfree_scan(hmp, 1);
1845 
1846 	KKASSERT(hmp->spmp == NULL);
1847 
1848 	/*
1849 	 * Finish up with the device vnode
1850 	 */
1851 	if (!TAILQ_EMPTY(&hmp->devvpl)) {
1852 		hammer2_close_devvp(&hmp->devvpl, hmp->ronly);
1853 		hammer2_cleanup_devvp(&hmp->devvpl);
1854 	}
1855 	KKASSERT(TAILQ_EMPTY(&hmp->devvpl));
1856 
1857 	/*
1858 	 * Clear vchain/fchain flags that might prevent final cleanup
1859 	 * of these chains.
1860 	 */
1861 	if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1862 		atomic_add_long(&hammer2_count_modified_chains, -1);
1863 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1864 		hammer2_pfs_memory_wakeup(hmp->vchain.pmp, -1);
1865 	}
1866 	if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1867 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1868 	}
1869 
1870 	if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1871 		atomic_add_long(&hammer2_count_modified_chains, -1);
1872 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1873 		hammer2_pfs_memory_wakeup(hmp->fchain.pmp, -1);
1874 	}
1875 	if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1876 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1877 	}
1878 
1879 	/*
1880 	 * Final drop of embedded freemap root chain to
1881 	 * clean up fchain.core (fchain structure is not
1882 	 * flagged ALLOCATED so it is cleaned out and then
1883 	 * left to rot).
1884 	 */
1885 	hammer2_chain_drop(&hmp->fchain);
1886 
1887 	/*
1888 	 * Final drop of embedded volume root chain to clean
1889 	 * up vchain.core (vchain structure is not flagged
1890 	 * ALLOCATED so it is cleaned out and then left to
1891 	 * rot).
1892 	 */
1893 	dumpcnt = 50;
1894 	hammer2_dump_chain(&hmp->vchain, 0, 0, &dumpcnt, 'v', (u_int)-1);
1895 	dumpcnt = 50;
1896 	hammer2_dump_chain(&hmp->fchain, 0, 0, &dumpcnt, 'f', (u_int)-1);
1897 
1898 	hammer2_chain_drop(&hmp->vchain);
1899 
1900 	hammer2_io_cleanup(hmp, &hmp->iotree);
1901 	if (hmp->iofree_count) {
1902 		kprintf("io_cleanup: %d I/O's left hanging\n",
1903 			hmp->iofree_count);
1904 	}
1905 
1906 	TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1907 	kmalloc_destroy_obj(&hmp->mchain);
1908 	kmalloc_destroy(&hmp->mmsg);
1909 	kfree(hmp, M_HAMMER2);
1910 }
1911 
1912 int
1913 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1914 		 ino_t ino, struct vnode **vpp)
1915 {
1916 	hammer2_xop_lookup_t *xop;
1917 	hammer2_pfs_t *pmp;
1918 	hammer2_inode_t *ip;
1919 	hammer2_tid_t inum;
1920 	int error;
1921 
1922 	inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1923 
1924 	error = 0;
1925 	pmp = MPTOPMP(mp);
1926 
1927 	/*
1928 	 * Easy if we already have it cached
1929 	 */
1930 	ip = hammer2_inode_lookup(pmp, inum);
1931 	if (ip) {
1932 		hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1933 		*vpp = hammer2_igetv(ip, &error);
1934 		hammer2_inode_unlock(ip);
1935 		hammer2_inode_drop(ip);		/* from lookup */
1936 
1937 		return error;
1938 	}
1939 
1940 	/*
1941 	 * Otherwise we have to find the inode
1942 	 */
1943 	xop = hammer2_xop_alloc(pmp->iroot, 0);
1944 	xop->lhc = inum;
1945 	hammer2_xop_start(&xop->head, &hammer2_lookup_desc);
1946 	error = hammer2_xop_collect(&xop->head, 0);
1947 
1948 	if (error == 0)
1949 		ip = hammer2_inode_get(pmp, &xop->head, -1, -1);
1950 	hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1951 
1952 	if (ip) {
1953 		*vpp = hammer2_igetv(ip, &error);
1954 		hammer2_inode_unlock(ip);
1955 	} else {
1956 		*vpp = NULL;
1957 		error = ENOENT;
1958 	}
1959 	return (error);
1960 }
1961 
1962 static
1963 int
1964 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1965 {
1966 	hammer2_pfs_t *pmp;
1967 	struct vnode *vp;
1968 	int error;
1969 
1970 	pmp = MPTOPMP(mp);
1971 	if (pmp->iroot == NULL) {
1972 		kprintf("hammer2 (%s): no root inode\n",
1973 			mp->mnt_stat.f_mntfromname);
1974 		*vpp = NULL;
1975 		return EINVAL;
1976 	}
1977 
1978 	error = 0;
1979 	hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1980 
1981 	while (pmp->inode_tid == 0) {
1982 		hammer2_xop_ipcluster_t *xop;
1983 		const hammer2_inode_meta_t *meta;
1984 
1985 		xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1986 		hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc);
1987 		error = hammer2_xop_collect(&xop->head, 0);
1988 
1989 		if (error == 0) {
1990 			meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1991 			pmp->iroot->meta = *meta;
1992 			pmp->inode_tid = meta->pfs_inum + 1;
1993 			hammer2_xop_pdata(&xop->head);
1994 			/* meta invalid */
1995 
1996 			if (pmp->inode_tid < HAMMER2_INODE_START)
1997 				pmp->inode_tid = HAMMER2_INODE_START;
1998 			pmp->modify_tid =
1999 				xop->head.cluster.focus->bref.modify_tid + 1;
2000 #if 0
2001 			kprintf("PFS: Starting inode %jd\n",
2002 				(intmax_t)pmp->inode_tid);
2003 			kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
2004 				pmp->inode_tid, pmp->modify_tid);
2005 #endif
2006 			wakeup(&pmp->iroot);
2007 
2008 			hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2009 
2010 			/*
2011 			 * Prime the mount info.
2012 			 */
2013 			hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
2014 			break;
2015 		}
2016 
2017 		/*
2018 		 * Loop, try again
2019 		 */
2020 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2021 		hammer2_inode_unlock(pmp->iroot);
2022 		error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
2023 		hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
2024 		if (error == EINTR)
2025 			break;
2026 	}
2027 
2028 	if (error) {
2029 		hammer2_inode_unlock(pmp->iroot);
2030 		*vpp = NULL;
2031 	} else {
2032 		vp = hammer2_igetv(pmp->iroot, &error);
2033 		hammer2_inode_unlock(pmp->iroot);
2034 		*vpp = vp;
2035 	}
2036 
2037 	return (error);
2038 }
2039 
2040 /*
2041  * Filesystem status
2042  *
2043  * XXX incorporate ipdata->meta.inode_quota and data_quota
2044  */
2045 static
2046 int
2047 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
2048 {
2049 	hammer2_pfs_t *pmp;
2050 	hammer2_dev_t *hmp;
2051 	hammer2_blockref_t bref;
2052 	struct statfs tmp;
2053 	int i;
2054 
2055 	/*
2056 	 * NOTE: iroot might not have validated the cluster yet.
2057 	 */
2058 	pmp = MPTOPMP(mp);
2059 
2060 	bzero(&tmp, sizeof(tmp));
2061 
2062 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2063 		hmp = pmp->pfs_hmps[i];
2064 		if (hmp == NULL)
2065 			continue;
2066 		if (pmp->iroot->cluster.array[i].chain)
2067 			bref = pmp->iroot->cluster.array[i].chain->bref;
2068 		else
2069 			bzero(&bref, sizeof(bref));
2070 
2071 		tmp.f_files = bref.embed.stats.inode_count;
2072 		tmp.f_ffree = 0;
2073 		tmp.f_blocks = hmp->voldata.allocator_size /
2074 			       mp->mnt_vstat.f_bsize;
2075 		tmp.f_bfree = hmp->voldata.allocator_free /
2076 			      mp->mnt_vstat.f_bsize;
2077 		tmp.f_bavail = tmp.f_bfree;
2078 
2079 		if (cred && cred->cr_uid != 0) {
2080 			uint64_t adj;
2081 
2082 			/* 5% */
2083 			adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2084 			tmp.f_blocks -= adj;
2085 			tmp.f_bfree -= adj;
2086 			tmp.f_bavail -= adj;
2087 		}
2088 
2089 		mp->mnt_stat.f_blocks = tmp.f_blocks;
2090 		mp->mnt_stat.f_bfree = tmp.f_bfree;
2091 		mp->mnt_stat.f_bavail = tmp.f_bavail;
2092 		mp->mnt_stat.f_files = tmp.f_files;
2093 		mp->mnt_stat.f_ffree = tmp.f_ffree;
2094 
2095 		*sbp = mp->mnt_stat;
2096 	}
2097 	return (0);
2098 }
2099 
2100 static
2101 int
2102 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2103 {
2104 	hammer2_pfs_t *pmp;
2105 	hammer2_dev_t *hmp;
2106 	hammer2_blockref_t bref;
2107 	struct statvfs tmp;
2108 	int i;
2109 
2110 	/*
2111 	 * NOTE: iroot might not have validated the cluster yet.
2112 	 */
2113 	pmp = MPTOPMP(mp);
2114 	bzero(&tmp, sizeof(tmp));
2115 
2116 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2117 		hmp = pmp->pfs_hmps[i];
2118 		if (hmp == NULL)
2119 			continue;
2120 		if (pmp->iroot->cluster.array[i].chain)
2121 			bref = pmp->iroot->cluster.array[i].chain->bref;
2122 		else
2123 			bzero(&bref, sizeof(bref));
2124 
2125 		tmp.f_files = bref.embed.stats.inode_count;
2126 		tmp.f_ffree = 0;
2127 		tmp.f_blocks = hmp->voldata.allocator_size /
2128 			       mp->mnt_vstat.f_bsize;
2129 		tmp.f_bfree = hmp->voldata.allocator_free /
2130 			      mp->mnt_vstat.f_bsize;
2131 		tmp.f_bavail = tmp.f_bfree;
2132 
2133 		if (cred && cred->cr_uid != 0) {
2134 			uint64_t adj;
2135 
2136 			/* 5% */
2137 			adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2138 			tmp.f_blocks -= adj;
2139 			tmp.f_bfree -= adj;
2140 			tmp.f_bavail -= adj;
2141 		}
2142 
2143 		mp->mnt_vstat.f_blocks = tmp.f_blocks;
2144 		mp->mnt_vstat.f_bfree = tmp.f_bfree;
2145 		mp->mnt_vstat.f_bavail = tmp.f_bavail;
2146 		mp->mnt_vstat.f_files = tmp.f_files;
2147 		mp->mnt_vstat.f_ffree = tmp.f_ffree;
2148 
2149 		*sbp = mp->mnt_vstat;
2150 	}
2151 	return (0);
2152 }
2153 
2154 /*
2155  * Mount-time recovery (RW mounts)
2156  *
2157  * Updates to the free block table are allowed to lag flushes by one
2158  * transaction.  In case of a crash, then on a fresh mount we must do an
2159  * incremental scan of the last committed transaction id and make sure that
2160  * all related blocks have been marked allocated.
2161  */
2162 struct hammer2_recovery_elm {
2163 	TAILQ_ENTRY(hammer2_recovery_elm) entry;
2164 	hammer2_chain_t *chain;
2165 	hammer2_tid_t sync_tid;
2166 };
2167 
2168 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2169 
2170 struct hammer2_recovery_info {
2171 	struct hammer2_recovery_list list;
2172 	hammer2_tid_t	mtid;
2173 	int	depth;
2174 };
2175 
2176 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2177 			hammer2_chain_t *parent,
2178 			struct hammer2_recovery_info *info,
2179 			hammer2_tid_t sync_tid);
2180 
2181 #define HAMMER2_RECOVERY_MAXDEPTH	10
2182 
2183 static
2184 int
2185 hammer2_recovery(hammer2_dev_t *hmp)
2186 {
2187 	struct hammer2_recovery_info info;
2188 	struct hammer2_recovery_elm *elm;
2189 	hammer2_chain_t *parent;
2190 	hammer2_tid_t sync_tid;
2191 	hammer2_tid_t mirror_tid;
2192 	int error;
2193 
2194 	hammer2_trans_init(hmp->spmp, 0);
2195 
2196 	sync_tid = hmp->voldata.freemap_tid;
2197 	mirror_tid = hmp->voldata.mirror_tid;
2198 
2199 	kprintf("hammer2_mount: \"%s\": ", hmp->devrepname);
2200 	if (sync_tid >= mirror_tid) {
2201 		kprintf("no recovery needed\n");
2202 	} else {
2203 		kprintf("freemap recovery %016jx-%016jx\n",
2204 			sync_tid + 1, mirror_tid);
2205 	}
2206 
2207 	TAILQ_INIT(&info.list);
2208 	info.depth = 0;
2209 	parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2210 	error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2211 	hammer2_chain_lookup_done(parent);
2212 
2213 	while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2214 		TAILQ_REMOVE(&info.list, elm, entry);
2215 		parent = elm->chain;
2216 		sync_tid = elm->sync_tid;
2217 		kfree(elm, M_HAMMER2);
2218 
2219 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2220 		error |= hammer2_recovery_scan(hmp, parent, &info,
2221 					      hmp->voldata.freemap_tid);
2222 		hammer2_chain_unlock(parent);
2223 		hammer2_chain_drop(parent);	/* drop elm->chain ref */
2224 	}
2225 
2226 	hammer2_trans_done(hmp->spmp, 0);
2227 
2228 	return error;
2229 }
2230 
2231 static
2232 int
2233 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2234 		      struct hammer2_recovery_info *info,
2235 		      hammer2_tid_t sync_tid)
2236 {
2237 	const hammer2_inode_data_t *ripdata;
2238 	hammer2_chain_t *chain;
2239 	hammer2_blockref_t bref;
2240 	int tmp_error;
2241 	int rup_error;
2242 	int error;
2243 	int first;
2244 
2245 	/*
2246 	 * Adjust freemap to ensure that the block(s) are marked allocated.
2247 	 */
2248 	if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2249 		hammer2_freemap_adjust(hmp, &parent->bref,
2250 				       HAMMER2_FREEMAP_DORECOVER);
2251 	}
2252 
2253 	/*
2254 	 * Check type for recursive scan
2255 	 */
2256 	switch(parent->bref.type) {
2257 	case HAMMER2_BREF_TYPE_VOLUME:
2258 		/* data already instantiated */
2259 		break;
2260 	case HAMMER2_BREF_TYPE_INODE:
2261 		/*
2262 		 * Must instantiate data for DIRECTDATA test and also
2263 		 * for recursion.
2264 		 */
2265 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2266 		ripdata = &parent->data->ipdata;
2267 		if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2268 			/* not applicable to recovery scan */
2269 			hammer2_chain_unlock(parent);
2270 			return 0;
2271 		}
2272 		hammer2_chain_unlock(parent);
2273 		break;
2274 	case HAMMER2_BREF_TYPE_INDIRECT:
2275 		/*
2276 		 * Must instantiate data for recursion
2277 		 */
2278 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2279 		hammer2_chain_unlock(parent);
2280 		break;
2281 	case HAMMER2_BREF_TYPE_DIRENT:
2282 	case HAMMER2_BREF_TYPE_DATA:
2283 	case HAMMER2_BREF_TYPE_FREEMAP:
2284 	case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2285 	case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2286 		/* not applicable to recovery scan */
2287 		return 0;
2288 		break;
2289 	default:
2290 		return HAMMER2_ERROR_BADBREF;
2291 	}
2292 
2293 	/*
2294 	 * Defer operation if depth limit reached.
2295 	 */
2296 	if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2297 		struct hammer2_recovery_elm *elm;
2298 
2299 		elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2300 		elm->chain = parent;
2301 		elm->sync_tid = sync_tid;
2302 		hammer2_chain_ref(parent);
2303 		TAILQ_INSERT_TAIL(&info->list, elm, entry);
2304 		/* unlocked by caller */
2305 
2306 		return(0);
2307 	}
2308 
2309 
2310 	/*
2311 	 * Recursive scan of the last flushed transaction only.  We are
2312 	 * doing this without pmp assignments so don't leave the chains
2313 	 * hanging around after we are done with them.
2314 	 *
2315 	 * error	Cumulative error this level only
2316 	 * rup_error	Cumulative error for recursion
2317 	 * tmp_error	Specific non-cumulative recursion error
2318 	 */
2319 	chain = NULL;
2320 	first = 1;
2321 	rup_error = 0;
2322 	error = 0;
2323 
2324 	for (;;) {
2325 		error |= hammer2_chain_scan(parent, &chain, &bref,
2326 					    &first,
2327 					    HAMMER2_LOOKUP_NODATA);
2328 
2329 		/*
2330 		 * Problem during scan or EOF
2331 		 */
2332 		if (error)
2333 			break;
2334 
2335 		/*
2336 		 * If this is a leaf
2337 		 */
2338 		if (chain == NULL) {
2339 			if (bref.mirror_tid > sync_tid) {
2340 				hammer2_freemap_adjust(hmp, &bref,
2341 						     HAMMER2_FREEMAP_DORECOVER);
2342 			}
2343 			continue;
2344 		}
2345 
2346 		/*
2347 		 * This may or may not be a recursive node.
2348 		 */
2349 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2350 		if (bref.mirror_tid > sync_tid) {
2351 			++info->depth;
2352 			tmp_error = hammer2_recovery_scan(hmp, chain,
2353 							   info, sync_tid);
2354 			--info->depth;
2355 		} else {
2356 			tmp_error = 0;
2357 		}
2358 
2359 		/*
2360 		 * Flush the recovery at the PFS boundary to stage it for
2361 		 * the final flush of the super-root topology.
2362 		 */
2363 		if (tmp_error == 0 &&
2364 		    (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2365 		    (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2366 			hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2367 					     HAMMER2_FLUSH_ALL);
2368 		}
2369 		rup_error |= tmp_error;
2370 	}
2371 	return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2372 }
2373 
2374 /*
2375  * This fixes up an error introduced in earlier H2 implementations where
2376  * moving a PFS inode into an indirect block wound up causing the
2377  * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2378  */
2379 static
2380 int
2381 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2382 {
2383 	const hammer2_inode_data_t *ripdata;
2384 	hammer2_chain_t *parent;
2385 	hammer2_chain_t *chain;
2386 	hammer2_key_t key_next;
2387 	hammer2_pfs_t *spmp;
2388 	int error;
2389 
2390 	error = 0;
2391 
2392 	/*
2393 	 * Lookup mount point under the media-localized super-root.
2394 	 *
2395 	 * cluster->pmp will incorrectly point to spmp and must be fixed
2396 	 * up later on.
2397 	 */
2398 	spmp = hmp->spmp;
2399 	hammer2_inode_lock(spmp->iroot, 0);
2400 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2401 	chain = hammer2_chain_lookup(&parent, &key_next,
2402 					 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2403 					 &error, 0);
2404 	while (chain) {
2405 		if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2406 			continue;
2407 		if (chain->error) {
2408 			kprintf("I/O error scanning PFS labels\n");
2409 			error |= chain->error;
2410 		} else if ((chain->bref.flags &
2411 			    HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2412 			int error2;
2413 
2414 			ripdata = &chain->data->ipdata;
2415 			hammer2_trans_init(hmp->spmp, 0);
2416 			error2 = hammer2_chain_modify(chain,
2417 						      chain->bref.modify_tid,
2418 						      0, 0);
2419 			if (error2 == 0) {
2420 				kprintf("hammer2: Correct mis-flagged PFS %s\n",
2421 					ripdata->filename);
2422 				chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2423 			} else {
2424 				error |= error2;
2425 			}
2426 			hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2427 					     HAMMER2_FLUSH_ALL);
2428 			hammer2_trans_done(hmp->spmp, 0);
2429 		}
2430 		chain = hammer2_chain_next(&parent, chain, &key_next,
2431 					   key_next, HAMMER2_KEY_MAX,
2432 					   &error, 0);
2433 	}
2434 	if (parent) {
2435 		hammer2_chain_unlock(parent);
2436 		hammer2_chain_drop(parent);
2437 	}
2438 	hammer2_inode_unlock(spmp->iroot);
2439 
2440 	return error;
2441 }
2442 
2443 /*
2444  * Sync a mount point; this is called periodically on a per-mount basis from
2445  * the filesystem syncer, and whenever a user issues a sync.
2446  */
2447 int
2448 hammer2_vfs_sync(struct mount *mp, int waitfor)
2449 {
2450 	int error;
2451 
2452 	error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor);
2453 
2454 	return error;
2455 }
2456 
2457 /*
2458  * Because frontend operations lock vnodes before we get a chance to
2459  * lock the related inode, we can't just acquire a vnode lock without
2460  * risking a deadlock.  The frontend may be holding a vnode lock while
2461  * also blocked on our SYNCQ flag while trying to get the inode lock.
2462  *
2463  * To deal with this situation we can check the vnode lock situation
2464  * after locking the inode and perform a work-around.
2465  */
2466 int
2467 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor)
2468 {
2469 	struct mount *mp;
2470 	/*hammer2_xop_flush_t *xop;*/
2471 	/*struct hammer2_sync_info info;*/
2472 	hammer2_inode_t *ip;
2473 	hammer2_depend_t *depend;
2474 	hammer2_depend_t *depend_next;
2475 	struct vnode *vp;
2476 	uint32_t pass2;
2477 	int error;
2478 	int wakecount;
2479 	int dorestart;
2480 
2481 	mp = pmp->mp;
2482 
2483 	/*
2484 	 * Move all inodes on sideq to syncq.  This will clear sideq.
2485 	 * This should represent all flushable inodes.  These inodes
2486 	 * will already have refs due to being on syncq or sideq.  We
2487 	 * must do this all at once with the spinlock held to ensure that
2488 	 * all inode dependencies are part of the same flush.
2489 	 *
2490 	 * We should be able to do this asynchronously from frontend
2491 	 * operations because we will be locking the inodes later on
2492 	 * to actually flush them, and that will partition any frontend
2493 	 * op using the same inode.  Either it has already locked the
2494 	 * inode and we will block, or it has not yet locked the inode
2495 	 * and it will block until we are finished flushing that inode.
2496 	 *
2497 	 * When restarting, only move the inodes flagged as PASS2 from
2498 	 * SIDEQ to SYNCQ.  PASS2 propagation by inode_lock4() and
2499 	 * inode_depend() are atomic with the spin-lock.
2500 	 */
2501 	hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2502 #ifdef HAMMER2_DEBUG_SYNC
2503 	kprintf("FILESYSTEM SYNC BOUNDARY\n");
2504 #endif
2505 	dorestart = 0;
2506 
2507 	/*
2508 	 * Move inodes from depq to syncq, releasing the related
2509 	 * depend structures.
2510 	 */
2511 restart:
2512 #ifdef HAMMER2_DEBUG_SYNC
2513 	kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart);
2514 #endif
2515 	hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/);
2516 	hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN);
2517 
2518 	/*
2519 	 * Move inodes from depq to syncq.  When restarting, only depq's
2520 	 * marked pass2 are moved.
2521 	 */
2522 	hammer2_spin_ex(&pmp->list_spin);
2523 	depend_next = TAILQ_FIRST(&pmp->depq);
2524 	wakecount = 0;
2525 
2526 	while ((depend = depend_next) != NULL) {
2527 		depend_next = TAILQ_NEXT(depend, entry);
2528 		if (dorestart && depend->pass2 == 0)
2529 			continue;
2530 		TAILQ_FOREACH(ip, &depend->sideq, entry) {
2531 			KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ);
2532 			atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ);
2533 			atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ);
2534 			ip->depend = NULL;
2535 		}
2536 
2537 		/*
2538 		 * NOTE: pmp->sideq_count includes both sideq and syncq
2539 		 */
2540 		TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry);
2541 
2542 		depend->count = 0;
2543 		depend->pass2 = 0;
2544 		TAILQ_REMOVE(&pmp->depq, depend, entry);
2545 	}
2546 
2547 	hammer2_spin_unex(&pmp->list_spin);
2548 	hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/
2549 				      HAMMER2_TRANS_WAITING);
2550 	dorestart = 0;
2551 
2552 	/*
2553 	 * sideq_count may have dropped enough to allow us to unstall
2554 	 * the frontend.
2555 	 */
2556 	hammer2_pfs_memory_wakeup(pmp, 0);
2557 
2558 	/*
2559 	 * Now run through all inodes on syncq.
2560 	 *
2561 	 * Flush transactions only interlock with other flush transactions.
2562 	 * Any conflicting frontend operations will block on the inode, but
2563 	 * may hold a vnode lock while doing so.
2564 	 */
2565 	hammer2_spin_ex(&pmp->list_spin);
2566 	while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) {
2567 		/*
2568 		 * Remove the inode from the SYNCQ, transfer the syncq ref
2569 		 * to us.  We must clear SYNCQ to allow any potential
2570 		 * front-end deadlock to proceed.  We must set PASS2 so
2571 		 * the dependency code knows what to do.
2572 		 */
2573 		pass2 = ip->flags;
2574 		cpu_ccfence();
2575 		if (atomic_cmpset_int(&ip->flags,
2576 			      pass2,
2577 			      (pass2 & ~(HAMMER2_INODE_SYNCQ |
2578 					 HAMMER2_INODE_SYNCQ_WAKEUP)) |
2579 			      HAMMER2_INODE_SYNCQ_PASS2) == 0) {
2580 			continue;
2581 		}
2582 		TAILQ_REMOVE(&pmp->syncq, ip, entry);
2583 		--pmp->sideq_count;
2584 		hammer2_spin_unex(&pmp->list_spin);
2585 
2586 		/*
2587 		 * Tickle anyone waiting on ip->flags or the hysteresis
2588 		 * on the dirty inode count.
2589 		 */
2590 		if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP)
2591 			wakeup(&ip->flags);
2592 		if (++wakecount >= hammer2_limit_dirty_inodes / 20 + 1) {
2593 			wakecount = 0;
2594 			hammer2_pfs_memory_wakeup(pmp, 0);
2595 		}
2596 
2597 		/*
2598 		 * Relock the inode, and we inherit a ref from the above.
2599 		 * We will check for a race after we acquire the vnode.
2600 		 */
2601 		hammer2_mtx_ex(&ip->lock);
2602 
2603 		/*
2604 		 * We need the vp in order to vfsync() dirty buffers, so if
2605 		 * one isn't attached we can skip it.
2606 		 *
2607 		 * Ordering the inode lock and then the vnode lock has the
2608 		 * potential to deadlock.  If we had left SYNCQ set that could
2609 		 * also deadlock us against the frontend even if we don't hold
2610 		 * any locks, but the latter is not a problem now since we
2611 		 * cleared it.  igetv will temporarily release the inode lock
2612 		 * in a safe manner to work-around the deadlock.
2613 		 *
2614 		 * Unfortunately it is still possible to deadlock when the
2615 		 * frontend obtains multiple inode locks, because all the
2616 		 * related vnodes are already locked (nor can the vnode locks
2617 		 * be released and reacquired without messing up RECLAIM and
2618 		 * INACTIVE sequencing).
2619 		 *
2620 		 * The solution for now is to move the vp back onto SIDEQ
2621 		 * and set dorestart, which will restart the flush after we
2622 		 * exhaust the current SYNCQ.  Note that additional
2623 		 * dependencies may build up, so we definitely need to move
2624 		 * the whole SIDEQ back to SYNCQ when we restart.
2625 		 */
2626 		vp = ip->vp;
2627 		if (vp) {
2628 			if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) {
2629 				/*
2630 				 * Failed to get the vnode, requeue the inode
2631 				 * (PASS2 is already set so it will be found
2632 				 * again on the restart).
2633 				 *
2634 				 * Then unlock, possibly sleep, and retry
2635 				 * later.  We sleep if PASS2 was *previously*
2636 				 * set, before we set it again above.
2637 				 */
2638 				vp = NULL;
2639 				dorestart = 1;
2640 #ifdef HAMMER2_DEBUG_SYNC
2641 				kprintf("inum %ld (sync delayed by vnode)\n",
2642 					(long)ip->meta.inum);
2643 #endif
2644 				hammer2_inode_delayed_sideq(ip);
2645 
2646 				hammer2_mtx_unlock(&ip->lock);
2647 				hammer2_inode_drop(ip);
2648 
2649 				if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) {
2650 					tsleep(&dorestart, 0, "h2syndel", 2);
2651 				}
2652 				hammer2_spin_ex(&pmp->list_spin);
2653 				continue;
2654 			}
2655 		} else {
2656 			vp = NULL;
2657 		}
2658 
2659 		/*
2660 		 * If the inode wound up on a SIDEQ again it will already be
2661 		 * prepped for another PASS2.  In this situation if we flush
2662 		 * it now we will just wind up flushing it again in the same
2663 		 * syncer run, so we might as well not flush it now.
2664 		 */
2665 		if (ip->flags & HAMMER2_INODE_SIDEQ) {
2666 			hammer2_mtx_unlock(&ip->lock);
2667 			hammer2_inode_drop(ip);
2668 			if (vp)
2669 				vput(vp);
2670 			dorestart = 1;
2671 			hammer2_spin_ex(&pmp->list_spin);
2672 			continue;
2673 		}
2674 
2675 		/*
2676 		 * Ok we have the inode exclusively locked and if vp is
2677 		 * not NULL that will also be exclusively locked.  Do the
2678 		 * meat of the flush.
2679 		 *
2680 		 * vp token needed for v_rbdirty_tree check / vclrisdirty
2681 		 * sequencing.  Though we hold the vnode exclusively so
2682 		 * we shouldn't need to hold the token also in this case.
2683 		 */
2684 		if (vp) {
2685 			vfsync(vp, MNT_WAIT, 1, NULL, NULL);
2686 			bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */
2687 		}
2688 
2689 		/*
2690 		 * If the inode has not yet been inserted into the tree
2691 		 * we must do so.  Then sync and flush it.  The flush should
2692 		 * update the parent.
2693 		 */
2694 		if (ip->flags & HAMMER2_INODE_DELETING) {
2695 #ifdef HAMMER2_DEBUG_SYNC
2696 			kprintf("inum %ld destroy\n", (long)ip->meta.inum);
2697 #endif
2698 			hammer2_inode_chain_des(ip);
2699 			atomic_add_long(&hammer2_iod_inode_deletes, 1);
2700 		} else if (ip->flags & HAMMER2_INODE_CREATING) {
2701 #ifdef HAMMER2_DEBUG_SYNC
2702 			kprintf("inum %ld insert\n", (long)ip->meta.inum);
2703 #endif
2704 			hammer2_inode_chain_ins(ip);
2705 			atomic_add_long(&hammer2_iod_inode_creates, 1);
2706 		}
2707 #ifdef HAMMER2_DEBUG_SYNC
2708 		kprintf("inum %ld chain-sync\n", (long)ip->meta.inum);
2709 #endif
2710 
2711 		/*
2712 		 * Because I kinda messed up the design and index the inodes
2713 		 * under the root inode, along side the directory entries,
2714 		 * we can't flush the inode index under the iroot until the
2715 		 * end.  If we do it now we might miss effects created by
2716 		 * other inodes on the SYNCQ.
2717 		 *
2718 		 * Do a normal (non-FSSYNC) flush instead, which allows the
2719 		 * vnode code to work the same.  We don't want to force iroot
2720 		 * back onto the SIDEQ, and we also don't want the flush code
2721 		 * to update pfs_iroot_blocksets until the final flush later.
2722 		 *
2723 		 * XXX at the moment this will likely result in a double-flush
2724 		 * of the iroot chain.
2725 		 */
2726 		hammer2_inode_chain_sync(ip);
2727 		if (ip == pmp->iroot) {
2728 			hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP);
2729 		} else {
2730 			hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2731 						      HAMMER2_XOP_FSSYNC);
2732 		}
2733 		if (vp) {
2734 			lwkt_gettoken(&vp->v_token);
2735 			if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2736 					  HAMMER2_INODE_RESIZED |
2737 					  HAMMER2_INODE_DIRTYDATA)) == 0 &&
2738 			    RB_EMPTY(&vp->v_rbdirty_tree) &&
2739 			    !bio_track_active(&vp->v_track_write)) {
2740 				vclrisdirty(vp);
2741 			} else {
2742 				hammer2_inode_delayed_sideq(ip);
2743 			}
2744 			lwkt_reltoken(&vp->v_token);
2745 			vput(vp);
2746 			vp = NULL;	/* safety */
2747 		}
2748 		atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2);
2749 		hammer2_inode_unlock(ip);	/* unlock+drop */
2750 		/* ip pointer invalid */
2751 
2752 		/*
2753 		 * If the inode got dirted after we dropped our locks,
2754 		 * it will have already been moved back to the SIDEQ.
2755 		 */
2756 		hammer2_spin_ex(&pmp->list_spin);
2757 	}
2758 	hammer2_spin_unex(&pmp->list_spin);
2759 	hammer2_pfs_memory_wakeup(pmp, 0);
2760 
2761 	if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) {
2762 #ifdef HAMMER2_DEBUG_SYNC
2763 		kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n");
2764 		/*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/
2765 #endif
2766 		dorestart = 1;
2767 		goto restart;
2768 	}
2769 #ifdef HAMMER2_DEBUG_SYNC
2770 	kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n");
2771 	/*tsleep(&dorestart, 0, "h2STG2", hz*20);*/
2772 #endif
2773 
2774 	/*
2775 	 * We have to flush the PFS root last, even if it does not appear to
2776 	 * be dirty, because all the inodes in the PFS are indexed under it.
2777 	 * The normal flushing of iroot above would only occur if directory
2778 	 * entries under the root were changed.
2779 	 *
2780 	 * Specifying VOLHDR will cause an additionl flush of hmp->spmp
2781 	 * for the media making up the cluster.
2782 	 */
2783 	if ((ip = pmp->iroot) != NULL) {
2784 		hammer2_inode_ref(ip);
2785 		hammer2_mtx_ex(&ip->lock);
2786 		hammer2_inode_chain_sync(ip);
2787 		hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2788 					      HAMMER2_XOP_FSSYNC |
2789 					      HAMMER2_XOP_VOLHDR);
2790 		hammer2_inode_unlock(ip);	/* unlock+drop */
2791 	}
2792 #ifdef HAMMER2_DEBUG_SYNC
2793 	kprintf("FILESYSTEM SYNC STAGE 2 DONE\n");
2794 #endif
2795 
2796 	/*
2797 	 * device bioq sync
2798 	 */
2799 	hammer2_bioq_sync(pmp);
2800 
2801 #if 0
2802 	/*
2803 	 * Generally speaking we now want to flush the media topology from
2804 	 * the iroot through to the inodes.  The flush stops at any inode
2805 	 * boundary, which allows the frontend to continue running concurrent
2806 	 * modifying operations on inodes (including kernel flushes of
2807 	 * buffers) without interfering with the main sync.
2808 	 *
2809 	 * Use the XOP interface to concurrently flush all nodes to
2810 	 * synchronize the PFSROOT subtopology to the media.  A standard
2811 	 * end-of-scan ENOENT error indicates cluster sufficiency.
2812 	 *
2813 	 * Note that this flush will not be visible on crash recovery until
2814 	 * we flush the super-root topology in the next loop.
2815 	 *
2816 	 * XXX For now wait for all flushes to complete.
2817 	 */
2818 	if (mp && (ip = pmp->iroot) != NULL) {
2819 		/*
2820 		 * If unmounting try to flush everything including any
2821 		 * sub-trees under inodes, just in case there is dangling
2822 		 * modified data, as a safety.  Otherwise just flush up to
2823 		 * the inodes in this stage.
2824 		 */
2825 		kprintf("MP & IROOT\n");
2826 #ifdef HAMMER2_DEBUG_SYNC
2827 		kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n");
2828 #endif
2829 		if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2830 			xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2831 						    HAMMER2_XOP_VOLHDR |
2832 						    HAMMER2_XOP_FSSYNC |
2833 						    HAMMER2_XOP_INODE_STOP);
2834 		} else {
2835 			xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2836 						    HAMMER2_XOP_INODE_STOP |
2837 						    HAMMER2_XOP_VOLHDR |
2838 						    HAMMER2_XOP_FSSYNC |
2839 						    HAMMER2_XOP_INODE_STOP);
2840 		}
2841 		hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc);
2842 		error = hammer2_xop_collect(&xop->head,
2843 					    HAMMER2_XOP_COLLECT_WAITALL);
2844 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2845 #ifdef HAMMER2_DEBUG_SYNC
2846 		kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n");
2847 #endif
2848 		if (error == HAMMER2_ERROR_ENOENT)
2849 			error = 0;
2850 		else
2851 			error = hammer2_error_to_errno(error);
2852 	} else {
2853 		error = 0;
2854 	}
2855 #endif
2856 	error = 0;	/* XXX */
2857 	hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH);
2858 
2859 	return (error);
2860 }
2861 
2862 static
2863 int
2864 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2865 {
2866 	hammer2_inode_t *ip;
2867 
2868 	KKASSERT(MAXFIDSZ >= 16);
2869 	ip = VTOI(vp);
2870 	fhp->fid_len = offsetof(struct fid, fid_data[16]);
2871 	fhp->fid_ext = 0;
2872 	((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2873 	((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2874 
2875 	return 0;
2876 }
2877 
2878 static
2879 int
2880 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2881 	       struct fid *fhp, struct vnode **vpp)
2882 {
2883 	hammer2_pfs_t *pmp;
2884 	hammer2_tid_t inum;
2885 	int error;
2886 
2887 	pmp = MPTOPMP(mp);
2888 	inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2889 	if (vpp) {
2890 		if (inum == 1)
2891 			error = hammer2_vfs_root(mp, vpp);
2892 		else
2893 			error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2894 	} else {
2895 		error = 0;
2896 	}
2897 	if (error)
2898 		kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2899 	return error;
2900 }
2901 
2902 static
2903 int
2904 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2905 		 int *exflagsp, struct ucred **credanonp)
2906 {
2907 	hammer2_pfs_t *pmp;
2908 	struct netcred *np;
2909 	int error;
2910 
2911 	pmp = MPTOPMP(mp);
2912 	np = vfs_export_lookup(mp, &pmp->export, nam);
2913 	if (np) {
2914 		*exflagsp = np->netc_exflags;
2915 		*credanonp = &np->netc_anon;
2916 		error = 0;
2917 	} else {
2918 		error = EACCES;
2919 	}
2920 	return error;
2921 }
2922 
2923 /*
2924  * This handles hysteresis on regular file flushes.  Because the BIOs are
2925  * routed to a thread it is possible for an excessive number to build up
2926  * and cause long front-end stalls long before the runningbuffspace limit
2927  * is hit, so we implement hammer2_flush_pipe to control the
2928  * hysteresis.
2929  *
2930  * This is a particular problem when compression is used.
2931  */
2932 void
2933 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2934 {
2935 	atomic_add_int(&pmp->count_lwinprog, 1);
2936 }
2937 
2938 void
2939 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2940 {
2941 	int lwinprog;
2942 
2943 	lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2944 	if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2945 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2946 		atomic_clear_int(&pmp->count_lwinprog,
2947 				 HAMMER2_LWINPROG_WAITING);
2948 		wakeup(&pmp->count_lwinprog);
2949 	}
2950 	if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2951 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2952 		atomic_clear_int(&pmp->count_lwinprog,
2953 				 HAMMER2_LWINPROG_WAITING0);
2954 		wakeup(&pmp->count_lwinprog);
2955 	}
2956 }
2957 
2958 void
2959 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2960 {
2961 	int lwinprog;
2962 	int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2963 				    HAMMER2_LWINPROG_WAITING0;
2964 
2965 	for (;;) {
2966 		lwinprog = pmp->count_lwinprog;
2967 		cpu_ccfence();
2968 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2969 			break;
2970 		tsleep_interlock(&pmp->count_lwinprog, 0);
2971 		atomic_set_int(&pmp->count_lwinprog, lwflag);
2972 		lwinprog = pmp->count_lwinprog;
2973 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2974 			break;
2975 		tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2976 	}
2977 }
2978 
2979 /*
2980  * It is possible for an excessive number of dirty chains or dirty inodes
2981  * to build up.  When this occurs we start an asynchronous filesystem sync.
2982  * If the level continues to build up, we stall, waiting for it to drop,
2983  * with some hysteresis.
2984  *
2985  * This relies on the kernel calling hammer2_vfs_modifying() prior to
2986  * obtaining any vnode locks before making a modifying VOP call.
2987  */
2988 static int
2989 hammer2_vfs_modifying(struct mount *mp)
2990 {
2991 	if (mp->mnt_flag & MNT_RDONLY)
2992 		return EROFS;
2993 	hammer2_pfs_memory_wait(MPTOPMP(mp));
2994 
2995 	return 0;
2996 }
2997 
2998 /*
2999  * Initiate an asynchronous filesystem sync and, with hysteresis,
3000  * stall if the internal data structure count becomes too bloated.
3001  */
3002 void
3003 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
3004 {
3005 	uint32_t waiting;
3006 	int pcatch;
3007 	int error;
3008 
3009 	if (pmp == NULL || pmp->mp == NULL)
3010 		return;
3011 
3012 	for (;;) {
3013 		waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK;
3014 		cpu_ccfence();
3015 
3016 		/*
3017 		 * Start the syncer running at 1/2 the limit
3018 		 */
3019 		if (waiting > hammer2_limit_dirty_chains / 2 ||
3020 		    pmp->sideq_count > hammer2_limit_dirty_inodes / 2) {
3021 			trigger_syncer(pmp->mp);
3022 		}
3023 
3024 		/*
3025 		 * Stall at the limit waiting for the counts to drop.
3026 		 * This code will typically be woken up once the count
3027 		 * drops below 3/4 the limit, or in one second.
3028 		 */
3029 		if (waiting < hammer2_limit_dirty_chains &&
3030 		    pmp->sideq_count < hammer2_limit_dirty_inodes) {
3031 			break;
3032 		}
3033 
3034 		pcatch = curthread->td_proc ? PCATCH : 0;
3035 
3036 		tsleep_interlock(&pmp->inmem_dirty_chains, pcatch);
3037 		atomic_set_int(&pmp->inmem_dirty_chains,
3038 			       HAMMER2_DIRTYCHAIN_WAITING);
3039 		if (waiting < hammer2_limit_dirty_chains &&
3040 		    pmp->sideq_count < hammer2_limit_dirty_inodes) {
3041 			break;
3042 		}
3043 		trigger_syncer(pmp->mp);
3044 		error = tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED | pcatch,
3045 			       "h2memw", hz);
3046 		if (error == ERESTART)
3047 			break;
3048 	}
3049 }
3050 
3051 /*
3052  * Wake up any stalled frontend ops waiting, with hysteresis, using
3053  * 2/3 of the limit.
3054  */
3055 void
3056 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp, int count)
3057 {
3058 	uint32_t waiting;
3059 
3060 	if (pmp) {
3061 		waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, count);
3062 		/* don't need --waiting to test flag */
3063 
3064 		if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) &&
3065 		    (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <=
3066 		    hammer2_limit_dirty_chains * 2 / 3 &&
3067 		    pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) {
3068 			atomic_clear_int(&pmp->inmem_dirty_chains,
3069 					 HAMMER2_DIRTYCHAIN_WAITING);
3070 			wakeup(&pmp->inmem_dirty_chains);
3071 		}
3072 	}
3073 }
3074 
3075 void
3076 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
3077 {
3078 	if (pmp) {
3079 		atomic_add_int(&pmp->inmem_dirty_chains, 1);
3080 	}
3081 }
3082 
3083 /*
3084  * Volume header data locks
3085  */
3086 void
3087 hammer2_voldata_lock(hammer2_dev_t *hmp)
3088 {
3089 	lockmgr(&hmp->vollk, LK_EXCLUSIVE);
3090 }
3091 
3092 void
3093 hammer2_voldata_unlock(hammer2_dev_t *hmp)
3094 {
3095 	lockmgr(&hmp->vollk, LK_RELEASE);
3096 }
3097 
3098 void
3099 hammer2_voldata_modify(hammer2_dev_t *hmp)
3100 {
3101 	if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) == 0) {
3102 		atomic_add_long(&hammer2_count_modified_chains, 1);
3103 		atomic_set_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
3104 		hammer2_pfs_memory_inc(hmp->vchain.pmp);
3105 	}
3106 }
3107 
3108 /*
3109  * Returns 0 if the filesystem has tons of free space
3110  * Returns 1 if the filesystem has less than 10% remaining
3111  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
3112  */
3113 int
3114 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
3115 {
3116 	hammer2_pfs_t *pmp;
3117 	hammer2_dev_t *hmp;
3118 	hammer2_off_t free_reserved;
3119 	hammer2_off_t free_nominal;
3120 	int i;
3121 
3122 	pmp = ip->pmp;
3123 
3124 	if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
3125 		free_reserved = HAMMER2_SEGSIZE;
3126 		free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
3127 		for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
3128 			hmp = pmp->pfs_hmps[i];
3129 			if (hmp == NULL)
3130 				continue;
3131 			if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
3132 			    pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
3133 				continue;
3134 
3135 			if (free_nominal > hmp->voldata.allocator_free)
3136 				free_nominal = hmp->voldata.allocator_free;
3137 			if (free_reserved < hmp->free_reserved)
3138 				free_reserved = hmp->free_reserved;
3139 		}
3140 
3141 		/*
3142 		 * SMP races ok
3143 		 */
3144 		pmp->free_reserved = free_reserved;
3145 		pmp->free_nominal = free_nominal;
3146 		pmp->free_ticks = ticks;
3147 	} else {
3148 		free_reserved = pmp->free_reserved;
3149 		free_nominal = pmp->free_nominal;
3150 	}
3151 	if (cred && cred->cr_uid != 0) {
3152 		if ((int64_t)(free_nominal - bytes) <
3153 		    (int64_t)free_reserved) {
3154 			return 2;
3155 		}
3156 	} else {
3157 		if ((int64_t)(free_nominal - bytes) <
3158 		    (int64_t)free_reserved / 2) {
3159 			return 2;
3160 		}
3161 	}
3162 	if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
3163 		return 1;
3164 	return 0;
3165 }
3166 
3167 /*
3168  * Debugging
3169  */
3170 void
3171 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int bi, int *countp,
3172 		   char pfx, u_int flags)
3173 {
3174 	hammer2_chain_t *scan;
3175 	hammer2_chain_t *parent;
3176 
3177 	--*countp;
3178 	if (*countp == 0) {
3179 		kprintf("%*.*s...\n", tab, tab, "");
3180 		return;
3181 	}
3182 	if (*countp < 0)
3183 		return;
3184 	kprintf("%*.*s%c-chain %p %s.%-3d %016jx %016jx/%-2d mir=%016jx\n",
3185 		tab, tab, "", pfx, chain,
3186 		hammer2_bref_type_str(chain->bref.type), bi,
3187 		chain->bref.data_off, chain->bref.key, chain->bref.keybits,
3188 		chain->bref.mirror_tid);
3189 
3190 	kprintf("%*.*s      [%08x] (%s) refs=%d",
3191 		tab, tab, "",
3192 		chain->flags,
3193 		((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3194 		chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
3195 		chain->refs);
3196 
3197 	parent = chain->parent;
3198 	if (parent)
3199 		kprintf("\n%*.*s      p=%p [pflags %08x prefs %d]",
3200 			tab, tab, "",
3201 			parent, parent->flags, parent->refs);
3202 	if (RB_EMPTY(&chain->core.rbtree)) {
3203 		kprintf("\n");
3204 	} else {
3205 		int bi = 0;
3206 		kprintf(" {\n");
3207 		RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3208 			if ((scan->flags & flags) || flags == (u_int)-1) {
3209 				hammer2_dump_chain(scan, tab + 4, bi, countp,
3210 						   'a', flags);
3211 			}
3212 			bi++;
3213 		}
3214 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3215 			kprintf("%*.*s}(%s)\n", tab, tab, "",
3216 				chain->data->ipdata.filename);
3217 		else
3218 			kprintf("%*.*s}\n", tab, tab, "");
3219 	}
3220 }
3221