xref: /dflybsd-src/sys/vfs/hammer2/hammer2_vfsops.c (revision 8d378610e3b5687c707bc8aad4e11a3a96bea2fc)
1 /*
2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  *
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in
16  *    the documentation and/or other materials provided with the
17  *    distribution.
18  * 3. Neither the name of The DragonFly Project nor the names of its
19  *    contributors may be used to endorse or promote products derived
20  *    from this software without specific, prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/nlookup.h>
39 #include <sys/vnode.h>
40 #include <sys/mount.h>
41 #include <sys/fcntl.h>
42 #include <sys/buf.h>
43 #include <sys/uuid.h>
44 #include <sys/vfsops.h>
45 #include <sys/sysctl.h>
46 #include <sys/socket.h>
47 #include <sys/objcache.h>
48 
49 #include <sys/proc.h>
50 #include <sys/namei.h>
51 #include <sys/mountctl.h>
52 #include <sys/dirent.h>
53 #include <sys/uio.h>
54 
55 #include <sys/mutex.h>
56 #include <sys/mutex2.h>
57 
58 #include "hammer2.h"
59 #include "hammer2_disk.h"
60 #include "hammer2_mount.h"
61 #include "hammer2_lz4.h"
62 
63 #include "zlib/hammer2_zlib.h"
64 
65 #define REPORT_REFS_ERRORS 1	/* XXX remove me */
66 
67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
68 
69 struct hammer2_sync_info {
70 	int error;
71 	int waitfor;
72 };
73 
74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
75 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
76 static struct hammer2_mntlist hammer2_mntlist;
77 static struct hammer2_pfslist hammer2_pfslist;
78 static struct lock hammer2_mntlk;
79 
80 int hammer2_debug;
81 int hammer2_cluster_read = 4;		/* physical read-ahead */
82 int hammer2_cluster_write = 0;		/* bdwrite() so later inval works */
83 int hammer2_dedup_enable = 1;
84 int hammer2_inval_enable = 0;
85 int hammer2_flush_pipe = 100;
86 int hammer2_synchronous_flush = 1;
87 int hammer2_dio_count;
88 long hammer2_chain_allocs;
89 long hammer2_chain_frees;
90 long hammer2_limit_dirty_chains;
91 long hammer2_count_modified_chains;
92 long hammer2_iod_invals;
93 long hammer2_iod_file_read;
94 long hammer2_iod_meta_read;
95 long hammer2_iod_indr_read;
96 long hammer2_iod_fmap_read;
97 long hammer2_iod_volu_read;
98 long hammer2_iod_file_write;
99 long hammer2_iod_file_wembed;
100 long hammer2_iod_file_wzero;
101 long hammer2_iod_file_wdedup;
102 long hammer2_iod_meta_write;
103 long hammer2_iod_indr_write;
104 long hammer2_iod_fmap_write;
105 long hammer2_iod_volu_write;
106 
107 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
108 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
109 		"Buffer used for compression.");
110 
111 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
112 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
113 		"Buffer used for decompression.");
114 
115 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
116 
117 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
118 	   &hammer2_debug, 0, "");
119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW,
120 	   &hammer2_cluster_read, 0, "");
121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
122 	   &hammer2_cluster_write, 0, "");
123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
124 	   &hammer2_dedup_enable, 0, "");
125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
126 	   &hammer2_inval_enable, 0, "");
127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
128 	   &hammer2_flush_pipe, 0, "");
129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
130 	   &hammer2_synchronous_flush, 0, "");
131 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
132 	   &hammer2_chain_allocs, 0, "");
133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
134 	   &hammer2_chain_frees, 0, "");
135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
136 	   &hammer2_limit_dirty_chains, 0, "");
137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
138 	   &hammer2_count_modified_chains, 0, "");
139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
140 	   &hammer2_dio_count, 0, "");
141 
142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
143 	   &hammer2_iod_invals, 0, "");
144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
145 	   &hammer2_iod_file_read, 0, "");
146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
147 	   &hammer2_iod_meta_read, 0, "");
148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
149 	   &hammer2_iod_indr_read, 0, "");
150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
151 	   &hammer2_iod_fmap_read, 0, "");
152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
153 	   &hammer2_iod_volu_read, 0, "");
154 
155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
156 	   &hammer2_iod_file_write, 0, "");
157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
158 	   &hammer2_iod_file_wembed, 0, "");
159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
160 	   &hammer2_iod_file_wzero, 0, "");
161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
162 	   &hammer2_iod_file_wdedup, 0, "");
163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
164 	   &hammer2_iod_meta_write, 0, "");
165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
166 	   &hammer2_iod_indr_write, 0, "");
167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
168 	   &hammer2_iod_fmap_write, 0, "");
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
170 	   &hammer2_iod_volu_write, 0, "");
171 
172 long hammer2_check_icrc32;
173 long hammer2_check_xxhash64;
174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
175 	   &hammer2_check_icrc32, 0, "");
176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
177 	   &hammer2_check_xxhash64, 0, "");
178 
179 static int hammer2_vfs_init(struct vfsconf *conf);
180 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
181 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
182 				struct ucred *cred);
183 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
184 				struct vnode *, struct ucred *);
185 static int hammer2_recovery(hammer2_dev_t *hmp);
186 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
187 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
188 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
189 				struct ucred *cred);
190 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
191 				struct ucred *cred);
192 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
193 				struct fid *fhp, struct vnode **vpp);
194 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
195 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
196 				int *exflagsp, struct ucred **credanonp);
197 
198 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
199 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
200 
201 static void hammer2_update_pmps(hammer2_dev_t *hmp);
202 
203 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
204 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
205 				hammer2_dev_t *hmp);
206 
207 /*
208  * HAMMER2 vfs operations.
209  */
210 static struct vfsops hammer2_vfsops = {
211 	.vfs_init	= hammer2_vfs_init,
212 	.vfs_uninit	= hammer2_vfs_uninit,
213 	.vfs_sync	= hammer2_vfs_sync,
214 	.vfs_mount	= hammer2_vfs_mount,
215 	.vfs_unmount	= hammer2_vfs_unmount,
216 	.vfs_root 	= hammer2_vfs_root,
217 	.vfs_statfs	= hammer2_vfs_statfs,
218 	.vfs_statvfs	= hammer2_vfs_statvfs,
219 	.vfs_vget	= hammer2_vfs_vget,
220 	.vfs_vptofh	= hammer2_vfs_vptofh,
221 	.vfs_fhtovp	= hammer2_vfs_fhtovp,
222 	.vfs_checkexp	= hammer2_vfs_checkexp
223 };
224 
225 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
226 
227 VFS_SET(hammer2_vfsops, hammer2, 0);
228 MODULE_VERSION(hammer2, 1);
229 
230 static
231 int
232 hammer2_vfs_init(struct vfsconf *conf)
233 {
234 	static struct objcache_malloc_args margs_read;
235 	static struct objcache_malloc_args margs_write;
236 	static struct objcache_malloc_args margs_vop;
237 
238 	int error;
239 
240 	error = 0;
241 
242 	if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
243 		error = EINVAL;
244 	if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
245 		error = EINVAL;
246 	if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
247 		error = EINVAL;
248 
249 	if (error)
250 		kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
251 
252 	margs_read.objsize = 65536;
253 	margs_read.mtype = M_HAMMER2_DEBUFFER;
254 
255 	margs_write.objsize = 32768;
256 	margs_write.mtype = M_HAMMER2_CBUFFER;
257 
258 	margs_vop.objsize = sizeof(hammer2_xop_t);
259 	margs_vop.mtype = M_HAMMER2;
260 
261 	/*
262 	 * Note thaht for the XOPS cache we want backing store allocations
263 	 * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
264 	 * confusion), so use the backing store function that does it.  This
265 	 * means that initial XOPS objects are zerod but REUSED objects are
266 	 * not.  So we are responsible for cleaning the object up sufficiently
267 	 * for our needs before objcache_put()ing it back (typically just the
268 	 * FIFO indices).
269 	 */
270 	cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
271 				0, 1, NULL, NULL, NULL,
272 				objcache_malloc_alloc,
273 				objcache_malloc_free,
274 				&margs_read);
275 	cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
276 				0, 1, NULL, NULL, NULL,
277 				objcache_malloc_alloc,
278 				objcache_malloc_free,
279 				&margs_write);
280 	cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
281 				0, 1, NULL, NULL, NULL,
282 				objcache_malloc_alloc_zero,
283 				objcache_malloc_free,
284 				&margs_vop);
285 
286 
287 	lockinit(&hammer2_mntlk, "mntlk", 0, 0);
288 	TAILQ_INIT(&hammer2_mntlist);
289 	TAILQ_INIT(&hammer2_pfslist);
290 
291 	hammer2_limit_dirty_chains = maxvnodes / 10;
292 
293 	return (error);
294 }
295 
296 static
297 int
298 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
299 {
300 	objcache_destroy(cache_buffer_read);
301 	objcache_destroy(cache_buffer_write);
302 	objcache_destroy(cache_xops);
303 	return 0;
304 }
305 
306 /*
307  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
308  * mounts and the spmp structure for media (hmp) structures.
309  *
310  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
311  * transactions.  Note that synchronization does not use this field.
312  * (typically frontend operations and synchronization cannot run on the
313  * same PFS node at the same time).
314  *
315  * XXX check locking
316  */
317 hammer2_pfs_t *
318 hammer2_pfsalloc(hammer2_chain_t *chain, const hammer2_inode_data_t *ripdata,
319 		 hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
320 {
321 	hammer2_inode_t *iroot;
322 	hammer2_pfs_t *pmp;
323 	int count;
324 	int i;
325 	int j;
326 
327 	/*
328 	 * Locate or create the PFS based on the cluster id.  If ripdata
329 	 * is NULL this is a spmp which is unique and is always allocated.
330 	 *
331 	 * If the device is mounted in local mode all PFSs are considered
332 	 * independent and not part of any cluster (for debugging only).
333 	 */
334 	if (ripdata) {
335 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
336 			if (force_local != pmp->force_local)
337 				continue;
338 			if (force_local == NULL &&
339 			    bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
340 				 sizeof(pmp->pfs_clid)) == 0) {
341 					break;
342 			} else if (force_local && pmp->pfs_names[0] &&
343 			    strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
344 					break;
345 			}
346 		}
347 	} else {
348 		pmp = NULL;
349 	}
350 
351 	if (pmp == NULL) {
352 		pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
353 		pmp->force_local = force_local;
354 		hammer2_trans_manage_init(pmp);
355 		kmalloc_create(&pmp->minode, "HAMMER2-inodes");
356 		kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
357 		lockinit(&pmp->lock, "pfslk", 0, 0);
358 		lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
359 		spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
360 		spin_init(&pmp->xop_spin, "h2xop");
361 		spin_init(&pmp->lru_spin, "h2lru");
362 		RB_INIT(&pmp->inum_tree);
363 		TAILQ_INIT(&pmp->sideq);
364 		TAILQ_INIT(&pmp->lru_list);
365 		spin_init(&pmp->list_spin, "hm2pfsalloc_list");
366 
367 		/*
368 		 * Distribute backend operations to threads
369 		 */
370 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
371 			for (j = 0; j < HAMMER2_XOPGROUPS +
372 					HAMMER2_SPECTHREADS; ++j) {
373 				TAILQ_INIT(&pmp->xopq[i][j]);
374 			}
375 		}
376 		for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
377 			hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
378 
379 		/*
380 		 * Save the last media transaction id for the flusher.  Set
381 		 * initial
382 		 */
383 		if (ripdata)
384 			pmp->pfs_clid = ripdata->meta.pfs_clid;
385 		TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
386 
387 		/*
388 		 * The synchronization thread may start too early, make
389 		 * sure it stays frozen until we are ready to let it go.
390 		 * XXX
391 		 */
392 		/*
393 		pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
394 					 HAMMER2_THREAD_REMASTER;
395 		*/
396 	}
397 
398 	/*
399 	 * Create the PFS's root inode.
400 	 */
401 	if ((iroot = pmp->iroot) == NULL) {
402 		iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
403 		pmp->iroot = iroot;
404 		hammer2_inode_ref(iroot);
405 		hammer2_inode_unlock(iroot);
406 	}
407 
408 	/*
409 	 * Stop here if no chain is passed in.
410 	 */
411 	if (chain == NULL)
412 		goto done;
413 
414 	/*
415 	 * When a chain is passed in we must add it to the PFS's root
416 	 * inode, update pmp->pfs_types[], and update the syncronization
417 	 * threads.
418 	 *
419 	 * When forcing local mode, mark the PFS as a MASTER regardless.
420 	 *
421 	 * At the moment empty spots can develop due to removals or failures.
422 	 * Ultimately we want to re-fill these spots but doing so might
423 	 * confused running code. XXX
424 	 */
425 	hammer2_inode_ref(iroot);
426 	hammer2_mtx_ex(&iroot->lock);
427 	j = iroot->cluster.nchains;
428 
429 	kprintf("add PFS to pmp %p[%d]\n", pmp, j);
430 
431 	if (j == HAMMER2_MAXCLUSTER) {
432 		kprintf("hammer2_mount: cluster full!\n");
433 		/* XXX fatal error? */
434 	} else {
435 		KKASSERT(chain->pmp == NULL);
436 		chain->pmp = pmp;
437 		hammer2_chain_ref(chain);
438 		iroot->cluster.array[j].chain = chain;
439 		if (force_local)
440 			pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
441 		else
442 			pmp->pfs_types[j] = ripdata->meta.pfs_type;
443 		pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
444 		pmp->pfs_hmps[j] = chain->hmp;
445 
446 		/*
447 		 * If the PFS is already mounted we must account
448 		 * for the mount_count here.
449 		 */
450 		if (pmp->mp)
451 			++chain->hmp->mount_count;
452 
453 		/*
454 		 * May have to fixup dirty chain tracking.  Previous
455 		 * pmp was NULL so nothing to undo.
456 		 */
457 		if (chain->flags & HAMMER2_CHAIN_MODIFIED)
458 			hammer2_pfs_memory_inc(pmp);
459 		++j;
460 	}
461 	iroot->cluster.nchains = j;
462 
463 	/*
464 	 * Update nmasters from any PFS inode which is part of the cluster.
465 	 * It is possible that this will result in a value which is too
466 	 * high.  MASTER PFSs are authoritative for pfs_nmasters and will
467 	 * override this value later on.
468 	 *
469 	 * (This informs us of masters that might not currently be
470 	 *  discoverable by this mount).
471 	 */
472 	if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
473 		pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
474 	}
475 
476 	/*
477 	 * Count visible masters.  Masters are usually added with
478 	 * ripdata->meta.pfs_nmasters set to 1.  This detects when there
479 	 * are more (XXX and must update the master inodes).
480 	 */
481 	count = 0;
482 	for (i = 0; i < iroot->cluster.nchains; ++i) {
483 		if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
484 			++count;
485 	}
486 	if (pmp->pfs_nmasters < count)
487 		pmp->pfs_nmasters = count;
488 
489 	/*
490 	 * Create missing synchronization and support threads.
491 	 *
492 	 * Single-node masters (including snapshots) have nothing to
493 	 * synchronize and do not require this thread.
494 	 *
495 	 * Multi-node masters or any number of soft masters, slaves, copy,
496 	 * or other PFS types need the thread.
497 	 *
498 	 * Each thread is responsible for its particular cluster index.
499 	 * We use independent threads so stalls or mismatches related to
500 	 * any given target do not affect other targets.
501 	 */
502 	for (i = 0; i < iroot->cluster.nchains; ++i) {
503 		/*
504 		 * Single-node masters (including snapshots) have nothing
505 		 * to synchronize and will make direct xops support calls,
506 		 * thus they do not require this thread.
507 		 *
508 		 * Note that there can be thousands of snapshots.  We do not
509 		 * want to create thousands of threads.
510 		 */
511 		if (pmp->pfs_nmasters <= 1 &&
512 		    pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
513 			continue;
514 		}
515 
516 		/*
517 		 * Sync support thread
518 		 */
519 		if (pmp->sync_thrs[i].td == NULL) {
520 			hammer2_thr_create(&pmp->sync_thrs[i], pmp,
521 					   "h2nod", i, -1,
522 					   hammer2_primary_sync_thread);
523 		}
524 	}
525 
526 	/*
527 	 * Create missing Xop threads
528 	 */
529 	if (pmp->mp)
530 		hammer2_xop_helper_create(pmp);
531 
532 	hammer2_mtx_unlock(&iroot->lock);
533 	hammer2_inode_drop(iroot);
534 done:
535 	return pmp;
536 }
537 
538 /*
539  * Destroy a PFS, typically only occurs after the last mount on a device
540  * has gone away.
541  */
542 static void
543 hammer2_pfsfree(hammer2_pfs_t *pmp)
544 {
545 	hammer2_inode_t *iroot;
546 	hammer2_chain_t *chain;
547 	int i;
548 	int j;
549 
550 	/*
551 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
552 	 * by the flush code.
553 	 */
554 	TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
555 
556 	iroot = pmp->iroot;
557 	if (iroot) {
558 		for (i = 0; i < iroot->cluster.nchains; ++i) {
559 			hammer2_thr_delete(&pmp->sync_thrs[i]);
560 			for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
561 				hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
562 		}
563 #if REPORT_REFS_ERRORS
564 		if (pmp->iroot->refs != 1)
565 			kprintf("PMP->IROOT %p REFS WRONG %d\n",
566 				pmp->iroot, pmp->iroot->refs);
567 #else
568 		KKASSERT(pmp->iroot->refs == 1);
569 #endif
570 		/* ref for pmp->iroot */
571 		hammer2_inode_drop(pmp->iroot);
572 		pmp->iroot = NULL;
573 	}
574 
575 	/*
576 	 * Cleanup chains remaining on LRU list.
577 	 */
578 	kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
579 	while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
580 		hammer2_chain_ref(chain);
581 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
582 		hammer2_chain_drop(chain);
583 	}
584 
585 	/*
586 	 * Free remaining pmp resources
587 	 */
588 	kmalloc_destroy(&pmp->mmsg);
589 	kmalloc_destroy(&pmp->minode);
590 
591 	kfree(pmp, M_HAMMER2);
592 }
593 
594 /*
595  * Remove all references to hmp from the pfs list.  Any PFS which becomes
596  * empty is terminated and freed.
597  *
598  * XXX inefficient.
599  */
600 static void
601 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
602 {
603 	hammer2_pfs_t *pmp;
604 	hammer2_inode_t *iroot;
605 	hammer2_chain_t *rchain;
606 	int didfreeze;
607 	int i;
608 	int j;
609 
610 again:
611 	TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
612 		if ((iroot = pmp->iroot) == NULL)
613 			continue;
614 		if (hmp->spmp == pmp) {
615 			kprintf("unmount hmp %p remove spmp %p\n",
616 				hmp, pmp);
617 			hmp->spmp = NULL;
618 		}
619 
620 		/*
621 		 * Determine if this PFS is affected.  If it is we must
622 		 * freeze all management threads and lock its iroot.
623 		 *
624 		 * Freezing a management thread forces it idle, operations
625 		 * in-progress will be aborted and it will have to start
626 		 * over again when unfrozen, or exit if told to exit.
627 		 */
628 		for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
629 			if (pmp->pfs_hmps[i] == hmp)
630 				break;
631 		}
632 		if (i != HAMMER2_MAXCLUSTER) {
633 			/*
634 			 * Make sure all synchronization threads are locked
635 			 * down.
636 			 */
637 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
638 				if (pmp->pfs_hmps[i] == NULL)
639 					continue;
640 				hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
641 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
642 					hammer2_thr_freeze_async(
643 						&pmp->xop_groups[j].thrs[i]);
644 				}
645 			}
646 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
647 				if (pmp->pfs_hmps[i] == NULL)
648 					continue;
649 				hammer2_thr_freeze(&pmp->sync_thrs[i]);
650 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
651 					hammer2_thr_freeze(
652 						&pmp->xop_groups[j].thrs[i]);
653 				}
654 			}
655 
656 			/*
657 			 * Lock the inode and clean out matching chains.
658 			 * Note that we cannot use hammer2_inode_lock_*()
659 			 * here because that would attempt to validate the
660 			 * cluster that we are in the middle of ripping
661 			 * apart.
662 			 *
663 			 * WARNING! We are working directly on the inodes
664 			 *	    embedded cluster.
665 			 */
666 			hammer2_mtx_ex(&iroot->lock);
667 
668 			/*
669 			 * Remove the chain from matching elements of the PFS.
670 			 */
671 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
672 				if (pmp->pfs_hmps[i] != hmp)
673 					continue;
674 				hammer2_thr_delete(&pmp->sync_thrs[i]);
675 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
676 					hammer2_thr_delete(
677 						&pmp->xop_groups[j].thrs[i]);
678 				}
679 				rchain = iroot->cluster.array[i].chain;
680 				iroot->cluster.array[i].chain = NULL;
681 				pmp->pfs_types[i] = 0;
682 				if (pmp->pfs_names[i]) {
683 					kfree(pmp->pfs_names[i], M_HAMMER2);
684 					pmp->pfs_names[i] = NULL;
685 				}
686 				if (rchain) {
687 					hammer2_chain_drop(rchain);
688 					/* focus hint */
689 					if (iroot->cluster.focus == rchain)
690 						iroot->cluster.focus = NULL;
691 				}
692 				pmp->pfs_hmps[i] = NULL;
693 			}
694 			hammer2_mtx_unlock(&iroot->lock);
695 			didfreeze = 1;	/* remaster, unfreeze down below */
696 		} else {
697 			didfreeze = 0;
698 		}
699 
700 		/*
701 		 * Cleanup trailing chains.  Gaps may remain.
702 		 */
703 		for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
704 			if (pmp->pfs_hmps[i])
705 				break;
706 		}
707 		iroot->cluster.nchains = i + 1;
708 
709 		/*
710 		 * If the PMP has no elements remaining we can destroy it.
711 		 * (this will transition management threads from frozen->exit).
712 		 */
713 		if (iroot->cluster.nchains == 0) {
714 			kprintf("unmount hmp %p last ref to PMP=%p\n",
715 				hmp, pmp);
716 			hammer2_pfsfree(pmp);
717 			goto again;
718 		}
719 
720 		/*
721 		 * If elements still remain we need to set the REMASTER
722 		 * flag and unfreeze it.
723 		 */
724 		if (didfreeze) {
725 			for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
726 				if (pmp->pfs_hmps[i] == NULL)
727 					continue;
728 				hammer2_thr_remaster(&pmp->sync_thrs[i]);
729 				hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
730 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
731 					hammer2_thr_remaster(
732 						&pmp->xop_groups[j].thrs[i]);
733 					hammer2_thr_unfreeze(
734 						&pmp->xop_groups[j].thrs[i]);
735 				}
736 			}
737 		}
738 	}
739 }
740 
741 /*
742  * Mount or remount HAMMER2 fileystem from physical media
743  *
744  *	mountroot
745  *		mp		mount point structure
746  *		path		NULL
747  *		data		<unused>
748  *		cred		<unused>
749  *
750  *	mount
751  *		mp		mount point structure
752  *		path		path to mount point
753  *		data		pointer to argument structure in user space
754  *			volume	volume path (device@LABEL form)
755  *			hflags	user mount flags
756  *		cred		user credentials
757  *
758  * RETURNS:	0	Success
759  *		!0	error number
760  */
761 static
762 int
763 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
764 		  struct ucred *cred)
765 {
766 	struct hammer2_mount_info info;
767 	hammer2_pfs_t *pmp;
768 	hammer2_pfs_t *spmp;
769 	hammer2_dev_t *hmp;
770 	hammer2_dev_t *force_local;
771 	hammer2_key_t key_next;
772 	hammer2_key_t key_dummy;
773 	hammer2_key_t lhc;
774 	struct vnode *devvp;
775 	struct nlookupdata nd;
776 	hammer2_chain_t *parent;
777 	hammer2_chain_t *chain;
778 	hammer2_cluster_t *cluster;
779 	const hammer2_inode_data_t *ripdata;
780 	hammer2_blockref_t bref;
781 	struct file *fp;
782 	char devstr[MNAMELEN];
783 	size_t size;
784 	size_t done;
785 	char *dev;
786 	char *label;
787 	int ronly = 1;
788 	int error;
789 	int cache_index;
790 	int i;
791 
792 	hmp = NULL;
793 	pmp = NULL;
794 	dev = NULL;
795 	label = NULL;
796 	devvp = NULL;
797 	cache_index = -1;
798 
799 	kprintf("hammer2_mount\n");
800 
801 	if (path == NULL) {
802 		/*
803 		 * Root mount
804 		 */
805 		bzero(&info, sizeof(info));
806 		info.cluster_fd = -1;
807 		ksnprintf(devstr, sizeof(devstr), "%s",
808 			  mp->mnt_stat.f_mntfromname);
809 		kprintf("hammer2_mount: root '%s'\n", devstr);
810 	} else {
811 		/*
812 		 * Non-root mount or updating a mount
813 		 */
814 		error = copyin(data, &info, sizeof(info));
815 		if (error)
816 			return (error);
817 
818 		error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
819 		if (error)
820 			return (error);
821 	}
822 
823 	/* Extract device and label */
824 	dev = devstr;
825 	label = strchr(devstr, '@');
826 	if (label == NULL ||
827 	    ((label + 1) - dev) > done) {
828 		return (EINVAL);
829 	}
830 	*label = '\0';
831 	label++;
832 	if (*label == '\0')
833 		return (EINVAL);
834 
835 	if (mp->mnt_flag & MNT_UPDATE) {
836 		/*
837 		 * Update mount.  Note that pmp->iroot->cluster is
838 		 * an inode-embedded cluster and thus cannot be
839 		 * directly locked.
840 		 *
841 		 * XXX HAMMER2 needs to implement NFS export via
842 		 *     mountctl.
843 		 */
844 		pmp = MPTOPMP(mp);
845 		pmp->hflags = info.hflags;
846 		cluster = &pmp->iroot->cluster;
847 		for (i = 0; i < cluster->nchains; ++i) {
848 			if (cluster->array[i].chain == NULL)
849 				continue;
850 			hmp = cluster->array[i].chain->hmp;
851 			devvp = hmp->devvp;
852 			error = hammer2_remount(hmp, mp, path,
853 						devvp, cred);
854 			if (error)
855 				break;
856 		}
857 
858 		return error;
859 	}
860 
861 	/*
862 	 * HMP device mount
863 	 *
864 	 * Lookup name and verify it refers to a block device.
865 	 */
866 	if (path) {
867 		error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
868 		if (error == 0)
869 			error = nlookup(&nd);
870 		if (error == 0)
871 			error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
872 		nlookup_done(&nd);
873 	} else {
874 		/* root mount */
875 		cdev_t cdev = kgetdiskbyname(dev);
876 		error = bdevvp(cdev, &devvp);
877 		if (error)
878 			kprintf("hammer2: cannot find '%s'\n", dev);
879 	}
880 
881 	if (error == 0) {
882 		if (vn_isdisk(devvp, &error))
883 			error = vfs_mountedon(devvp);
884 	}
885 
886 	/*
887 	 * Determine if the device has already been mounted.  After this
888 	 * check hmp will be non-NULL if we are doing the second or more
889 	 * hammer2 mounts from the same device.
890 	 */
891 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
892 	TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
893 		if (hmp->devvp == devvp)
894 			break;
895 	}
896 
897 	/*
898 	 * Open the device if this isn't a secondary mount and construct
899 	 * the H2 device mount (hmp).
900 	 */
901 	if (hmp == NULL) {
902 		hammer2_chain_t *schain;
903 		hammer2_xid_t xid;
904 
905 		if (error == 0 && vcount(devvp) > 0)
906 			error = EBUSY;
907 
908 		/*
909 		 * Now open the device
910 		 */
911 		if (error == 0) {
912 			ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
913 			vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
914 			error = vinvalbuf(devvp, V_SAVE, 0, 0);
915 			if (error == 0) {
916 				error = VOP_OPEN(devvp,
917 						 ronly ? FREAD : FREAD | FWRITE,
918 						 FSCRED, NULL);
919 			}
920 			vn_unlock(devvp);
921 		}
922 		if (error && devvp) {
923 			vrele(devvp);
924 			devvp = NULL;
925 		}
926 		if (error) {
927 			lockmgr(&hammer2_mntlk, LK_RELEASE);
928 			return error;
929 		}
930 		hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
931 		ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
932 		hmp->ronly = ronly;
933 		hmp->devvp = devvp;
934 		hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
935 		kmalloc_create(&hmp->mchain, "HAMMER2-chains");
936 		TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
937 		RB_INIT(&hmp->iotree);
938 		spin_init(&hmp->io_spin, "hm2mount_io");
939 		spin_init(&hmp->list_spin, "hm2mount_list");
940 		TAILQ_INIT(&hmp->flushq);
941 
942 		lockinit(&hmp->vollk, "h2vol", 0, 0);
943 		lockinit(&hmp->bulklk, "h2bulk", 0, 0);
944 
945 		/*
946 		 * vchain setup. vchain.data is embedded.
947 		 * vchain.refs is initialized and will never drop to 0.
948 		 *
949 		 * NOTE! voldata is not yet loaded.
950 		 */
951 		hmp->vchain.hmp = hmp;
952 		hmp->vchain.refs = 1;
953 		hmp->vchain.data = (void *)&hmp->voldata;
954 		hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
955 		hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
956 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
957 
958 		hammer2_chain_core_init(&hmp->vchain);
959 		/* hmp->vchain.u.xxx is left NULL */
960 
961 		/*
962 		 * fchain setup.  fchain.data is embedded.
963 		 * fchain.refs is initialized and will never drop to 0.
964 		 *
965 		 * The data is not used but needs to be initialized to
966 		 * pass assertion muster.  We use this chain primarily
967 		 * as a placeholder for the freemap's top-level RBTREE
968 		 * so it does not interfere with the volume's topology
969 		 * RBTREE.
970 		 */
971 		hmp->fchain.hmp = hmp;
972 		hmp->fchain.refs = 1;
973 		hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
974 		hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
975 		hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
976 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
977 		hmp->fchain.bref.methods =
978 			HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
979 			HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
980 
981 		hammer2_chain_core_init(&hmp->fchain);
982 		/* hmp->fchain.u.xxx is left NULL */
983 
984 		/*
985 		 * Install the volume header and initialize fields from
986 		 * voldata.
987 		 */
988 		error = hammer2_install_volume_header(hmp);
989 		if (error) {
990 			hammer2_unmount_helper(mp, NULL, hmp);
991 			lockmgr(&hammer2_mntlk, LK_RELEASE);
992 			hammer2_vfs_unmount(mp, MNT_FORCE);
993 			return error;
994 		}
995 
996 		/*
997 		 * Really important to get these right or flush will get
998 		 * confused.
999 		 */
1000 		hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1001 		kprintf("alloc spmp %p tid %016jx\n",
1002 			hmp->spmp, hmp->voldata.mirror_tid);
1003 		spmp = hmp->spmp;
1004 
1005 		/*
1006 		 * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1007 		 * is inherited from the volume header.
1008 		 */
1009 		xid = 0;
1010 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1011 		hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1012 		hmp->vchain.pmp = spmp;
1013 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1014 		hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1015 		hmp->fchain.pmp = spmp;
1016 
1017 		/*
1018 		 * First locate the super-root inode, which is key 0
1019 		 * relative to the volume header's blockset.
1020 		 *
1021 		 * Then locate the root inode by scanning the directory keyspace
1022 		 * represented by the label.
1023 		 */
1024 		parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1025 		schain = hammer2_chain_lookup(&parent, &key_dummy,
1026 				      HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1027 				      &cache_index, 0);
1028 		hammer2_chain_lookup_done(parent);
1029 		if (schain == NULL) {
1030 			kprintf("hammer2_mount: invalid super-root\n");
1031 			hammer2_unmount_helper(mp, NULL, hmp);
1032 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1033 			hammer2_vfs_unmount(mp, MNT_FORCE);
1034 			return EINVAL;
1035 		}
1036 		if (schain->error) {
1037 			kprintf("hammer2_mount: error %s reading super-root\n",
1038 				hammer2_error_str(schain->error));
1039 			hammer2_chain_unlock(schain);
1040 			hammer2_chain_drop(schain);
1041 			schain = NULL;
1042 			hammer2_unmount_helper(mp, NULL, hmp);
1043 			lockmgr(&hammer2_mntlk, LK_RELEASE);
1044 			hammer2_vfs_unmount(mp, MNT_FORCE);
1045 			return EINVAL;
1046 		}
1047 
1048 		/*
1049 		 * The super-root always uses an inode_tid of 1 when
1050 		 * creating PFSs.
1051 		 */
1052 		spmp->inode_tid = 1;
1053 		spmp->modify_tid = schain->bref.modify_tid + 1;
1054 
1055 		/*
1056 		 * Sanity-check schain's pmp and finish initialization.
1057 		 * Any chain belonging to the super-root topology should
1058 		 * have a NULL pmp (not even set to spmp).
1059 		 */
1060 		ripdata = &hammer2_chain_rdata(schain)->ipdata;
1061 		KKASSERT(schain->pmp == NULL);
1062 		spmp->pfs_clid = ripdata->meta.pfs_clid;
1063 
1064 		/*
1065 		 * Replace the dummy spmp->iroot with a real one.  It's
1066 		 * easier to just do a wholesale replacement than to try
1067 		 * to update the chain and fixup the iroot fields.
1068 		 *
1069 		 * The returned inode is locked with the supplied cluster.
1070 		 */
1071 		cluster = hammer2_cluster_from_chain(schain);
1072 		hammer2_inode_drop(spmp->iroot);
1073 		spmp->iroot = NULL;
1074 		spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1075 		spmp->spmp_hmp = hmp;
1076 		spmp->pfs_types[0] = ripdata->meta.pfs_type;
1077 		spmp->pfs_hmps[0] = hmp;
1078 		hammer2_inode_ref(spmp->iroot);
1079 		hammer2_inode_unlock(spmp->iroot);
1080 		hammer2_cluster_unlock(cluster);
1081 		hammer2_cluster_drop(cluster);
1082 		schain = NULL;
1083 		/* leave spmp->iroot with one ref */
1084 
1085 		if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1086 			error = hammer2_recovery(hmp);
1087 			/* XXX do something with error */
1088 		}
1089 		hammer2_update_pmps(hmp);
1090 		hammer2_iocom_init(hmp);
1091 
1092 		/*
1093 		 * Ref the cluster management messaging descriptor.  The mount
1094 		 * program deals with the other end of the communications pipe.
1095 		 *
1096 		 * Root mounts typically do not supply one.
1097 		 */
1098 		if (info.cluster_fd >= 0) {
1099 			fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1100 			if (fp) {
1101 				hammer2_cluster_reconnect(hmp, fp);
1102 			} else {
1103 				kprintf("hammer2_mount: bad cluster_fd!\n");
1104 			}
1105 		}
1106 	} else {
1107 		spmp = hmp->spmp;
1108 		if (info.hflags & HMNT2_DEVFLAGS) {
1109 			kprintf("hammer2: Warning: mount flags pertaining "
1110 				"to the whole device may only be specified "
1111 				"on the first mount of the device: %08x\n",
1112 				info.hflags & HMNT2_DEVFLAGS);
1113 		}
1114 	}
1115 
1116 	/*
1117 	 * Force local mount (disassociate all PFSs from their clusters).
1118 	 * Used primarily for debugging.
1119 	 */
1120 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1121 
1122 	/*
1123 	 * Lookup the mount point under the media-localized super-root.
1124 	 * Scanning hammer2_pfslist doesn't help us because it represents
1125 	 * PFS cluster ids which can aggregate several named PFSs together.
1126 	 *
1127 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1128 	 * up later on.
1129 	 */
1130 	hammer2_inode_lock(spmp->iroot, 0);
1131 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1132 	lhc = hammer2_dirhash(label, strlen(label));
1133 	chain = hammer2_chain_lookup(&parent, &key_next,
1134 				     lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1135 				     &cache_index, 0);
1136 	while (chain) {
1137 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1138 		    strcmp(label, chain->data->ipdata.filename) == 0) {
1139 			break;
1140 		}
1141 		chain = hammer2_chain_next(&parent, chain, &key_next,
1142 					    key_next,
1143 					    lhc + HAMMER2_DIRHASH_LOMASK,
1144 					    &cache_index, 0);
1145 	}
1146 	if (parent) {
1147 		hammer2_chain_unlock(parent);
1148 		hammer2_chain_drop(parent);
1149 	}
1150 	hammer2_inode_unlock(spmp->iroot);
1151 
1152 	/*
1153 	 * PFS could not be found?
1154 	 */
1155 	if (chain == NULL) {
1156 		kprintf("hammer2_mount: PFS label not found\n");
1157 		hammer2_unmount_helper(mp, NULL, hmp);
1158 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1159 		hammer2_vfs_unmount(mp, MNT_FORCE);
1160 
1161 		return EINVAL;
1162 	}
1163 
1164 	/*
1165 	 * Acquire the pmp structure (it should have already been allocated
1166 	 * via hammer2_update_pmps() so do not pass cluster in to add to
1167 	 * available chains).
1168 	 *
1169 	 * Check if the cluster has already been mounted.  A cluster can
1170 	 * only be mounted once, use null mounts to mount additional copies.
1171 	 */
1172 	ripdata = &chain->data->ipdata;
1173 	bref = chain->bref;
1174 	pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid, force_local);
1175 	hammer2_chain_unlock(chain);
1176 	hammer2_chain_drop(chain);
1177 
1178 	if (pmp->mp) {
1179 		kprintf("hammer2_mount: PFS already mounted!\n");
1180 		hammer2_unmount_helper(mp, NULL, hmp);
1181 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1182 		hammer2_vfs_unmount(mp, MNT_FORCE);
1183 
1184 		return EBUSY;
1185 	}
1186 
1187 	/*
1188 	 * Finish the mount
1189 	 */
1190         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1191 
1192 	pmp->hflags = info.hflags;
1193         mp->mnt_flag = MNT_LOCAL;
1194         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1195         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1196 
1197         /*
1198          * required mount structure initializations
1199          */
1200         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1201         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1202 
1203         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1204         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1205 
1206         /*
1207          * Optional fields
1208          */
1209         mp->mnt_iosize_max = MAXPHYS;
1210 
1211 	/*
1212 	 * Connect up mount pointers.
1213 	 */
1214 	hammer2_mount_helper(mp, pmp);
1215 
1216         lockmgr(&hammer2_mntlk, LK_RELEASE);
1217 
1218 	/*
1219 	 * Finish setup
1220 	 */
1221 	vfs_getnewfsid(mp);
1222 	vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1223 	vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1224 	vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1225 
1226 	if (path) {
1227 		copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1228 			  MNAMELEN - 1, &size);
1229 		bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1230 	} /* else root mount, already in there */
1231 
1232 	bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1233 	if (path) {
1234 		copyinstr(path, mp->mnt_stat.f_mntonname,
1235 			  sizeof(mp->mnt_stat.f_mntonname) - 1,
1236 			  &size);
1237 	} else {
1238 		/* root mount */
1239 		mp->mnt_stat.f_mntonname[0] = '/';
1240 	}
1241 
1242 	/*
1243 	 * Initial statfs to prime mnt_stat.
1244 	 */
1245 	hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1246 
1247 	return 0;
1248 }
1249 
1250 /*
1251  * Scan PFSs under the super-root and create hammer2_pfs structures.
1252  */
1253 static
1254 void
1255 hammer2_update_pmps(hammer2_dev_t *hmp)
1256 {
1257 	const hammer2_inode_data_t *ripdata;
1258 	hammer2_chain_t *parent;
1259 	hammer2_chain_t *chain;
1260 	hammer2_blockref_t bref;
1261 	hammer2_dev_t *force_local;
1262 	hammer2_pfs_t *spmp;
1263 	hammer2_pfs_t *pmp;
1264 	hammer2_key_t key_next;
1265 	int cache_index = -1;
1266 
1267 	/*
1268 	 * Force local mount (disassociate all PFSs from their clusters).
1269 	 * Used primarily for debugging.
1270 	 */
1271 	force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1272 
1273 	/*
1274 	 * Lookup mount point under the media-localized super-root.
1275 	 *
1276 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1277 	 * up later on.
1278 	 */
1279 	spmp = hmp->spmp;
1280 	hammer2_inode_lock(spmp->iroot, 0);
1281 	parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1282 	chain = hammer2_chain_lookup(&parent, &key_next,
1283 					 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1284 					 &cache_index, 0);
1285 	while (chain) {
1286 		if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1287 			continue;
1288 		ripdata = &chain->data->ipdata;
1289 		bref = chain->bref;
1290 		kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1291 
1292 		pmp = hammer2_pfsalloc(chain, ripdata,
1293 				       bref.modify_tid, force_local);
1294 		chain = hammer2_chain_next(&parent, chain, &key_next,
1295 					   key_next, HAMMER2_KEY_MAX,
1296 					   &cache_index, 0);
1297 	}
1298 	if (parent) {
1299 		hammer2_chain_unlock(parent);
1300 		hammer2_chain_drop(parent);
1301 	}
1302 	hammer2_inode_unlock(spmp->iroot);
1303 }
1304 
1305 static
1306 int
1307 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1308 		struct vnode *devvp, struct ucred *cred)
1309 {
1310 	int error;
1311 
1312 	if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1313 		error = hammer2_recovery(hmp);
1314 	} else {
1315 		error = 0;
1316 	}
1317 	return error;
1318 }
1319 
1320 static
1321 int
1322 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1323 {
1324 	hammer2_pfs_t *pmp;
1325 	int flags;
1326 	int error = 0;
1327 
1328 	pmp = MPTOPMP(mp);
1329 
1330 	if (pmp == NULL)
1331 		return(0);
1332 
1333 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1334 
1335 	/*
1336 	 * If mount initialization proceeded far enough we must flush
1337 	 * its vnodes and sync the underlying mount points.  Three syncs
1338 	 * are required to fully flush the filesystem (freemap updates lag
1339 	 * by one flush, and one extra for safety).
1340 	 */
1341 	if (mntflags & MNT_FORCE)
1342 		flags = FORCECLOSE;
1343 	else
1344 		flags = 0;
1345 	if (pmp->iroot) {
1346 		error = vflush(mp, 0, flags);
1347 		if (error)
1348 			goto failed;
1349 		hammer2_vfs_sync(mp, MNT_WAIT);
1350 		hammer2_vfs_sync(mp, MNT_WAIT);
1351 		hammer2_vfs_sync(mp, MNT_WAIT);
1352 	}
1353 
1354 	/*
1355 	 * Cleanup the frontend support XOPS threads
1356 	 */
1357 	hammer2_xop_helper_cleanup(pmp);
1358 
1359 	if (pmp->mp)
1360 		hammer2_unmount_helper(mp, pmp, NULL);
1361 
1362 	error = 0;
1363 failed:
1364 	lockmgr(&hammer2_mntlk, LK_RELEASE);
1365 
1366 	return (error);
1367 }
1368 
1369 /*
1370  * Mount helper, hook the system mount into our PFS.
1371  * The mount lock is held.
1372  *
1373  * We must bump the mount_count on related devices for any
1374  * mounted PFSs.
1375  */
1376 static
1377 void
1378 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1379 {
1380 	hammer2_cluster_t *cluster;
1381 	hammer2_chain_t *rchain;
1382 	int i;
1383 
1384         mp->mnt_data = (qaddr_t)pmp;
1385 	pmp->mp = mp;
1386 
1387 	/*
1388 	 * After pmp->mp is set we have to adjust hmp->mount_count.
1389 	 */
1390 	cluster = &pmp->iroot->cluster;
1391 	for (i = 0; i < cluster->nchains; ++i) {
1392 		rchain = cluster->array[i].chain;
1393 		if (rchain == NULL)
1394 			continue;
1395 		++rchain->hmp->mount_count;
1396 		kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1397 			rchain->hmp, rchain->hmp->mount_count);
1398 	}
1399 
1400 	/*
1401 	 * Create missing Xop threads
1402 	 */
1403 	hammer2_xop_helper_create(pmp);
1404 }
1405 
1406 /*
1407  * Mount helper, unhook the system mount from our PFS.
1408  * The mount lock is held.
1409  *
1410  * If hmp is supplied a mount responsible for being the first to open
1411  * the block device failed and the block device and all PFSs using the
1412  * block device must be cleaned up.
1413  *
1414  * If pmp is supplied multiple devices might be backing the PFS and each
1415  * must be disconnected.  This might not be the last PFS using some of the
1416  * underlying devices.  Also, we have to adjust our hmp->mount_count
1417  * accounting for the devices backing the pmp which is now undergoing an
1418  * unmount.
1419  */
1420 static
1421 void
1422 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1423 {
1424 	hammer2_cluster_t *cluster;
1425 	hammer2_chain_t *rchain;
1426 	struct vnode *devvp;
1427 	int dumpcnt;
1428 	int ronly = 0;
1429 	int i;
1430 
1431 	/*
1432 	 * If no device supplied this is a high-level unmount and we have to
1433 	 * to disconnect the mount, adjust mount_count, and locate devices
1434 	 * that might now have no mounts.
1435 	 */
1436 	if (pmp) {
1437 		KKASSERT(hmp == NULL);
1438 		KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1439 		pmp->mp = NULL;
1440 		mp->mnt_data = NULL;
1441 
1442 		/*
1443 		 * After pmp->mp is cleared we have to account for
1444 		 * mount_count.
1445 		 */
1446 		cluster = &pmp->iroot->cluster;
1447 		for (i = 0; i < cluster->nchains; ++i) {
1448 			rchain = cluster->array[i].chain;
1449 			if (rchain == NULL)
1450 				continue;
1451 			--rchain->hmp->mount_count;
1452 			kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1453 				rchain->hmp, rchain->hmp->mount_count);
1454 			/* scrapping hmp now may invalidate the pmp */
1455 		}
1456 again:
1457 		TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1458 			if (hmp->mount_count == 0) {
1459 				hammer2_unmount_helper(NULL, NULL, hmp);
1460 				goto again;
1461 			}
1462 		}
1463 		return;
1464 	}
1465 
1466 	/*
1467 	 * Try to terminate the block device.  We can't terminate it if
1468 	 * there are still PFSs referencing it.
1469 	 */
1470 	kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1471 		hmp, hmp->mount_count);
1472 	if (hmp->mount_count)
1473 		return;
1474 
1475 	hammer2_pfsfree_scan(hmp);
1476 	hammer2_dev_exlock(hmp);	/* XXX order */
1477 
1478 	/*
1479 	 * Cycle the volume data lock as a safety (probably not needed any
1480 	 * more).  To ensure everything is out we need to flush at least
1481 	 * three times.  (1) The running of the sideq can dirty the
1482 	 * filesystem, (2) A normal flush can dirty the freemap, and
1483 	 * (3) ensure that the freemap is fully synchronized.
1484 	 *
1485 	 * The next mount's recovery scan can clean everything up but we want
1486 	 * to leave the filesystem in a 100% clean state on a normal unmount.
1487 	 */
1488 #if 0
1489 	hammer2_voldata_lock(hmp);
1490 	hammer2_voldata_unlock(hmp);
1491 #endif
1492 	hammer2_iocom_uninit(hmp);
1493 
1494 	if ((hmp->vchain.flags | hmp->fchain.flags) &
1495 	    HAMMER2_CHAIN_FLUSH_MASK) {
1496 		kprintf("hammer2_unmount: chains left over "
1497 			"after final sync\n");
1498 		kprintf("    vchain %08x\n", hmp->vchain.flags);
1499 		kprintf("    fchain %08x\n", hmp->fchain.flags);
1500 
1501 		if (hammer2_debug & 0x0010)
1502 			Debugger("entered debugger");
1503 	}
1504 
1505 	KKASSERT(hmp->spmp == NULL);
1506 
1507 	/*
1508 	 * Finish up with the device vnode
1509 	 */
1510 	if ((devvp = hmp->devvp) != NULL) {
1511 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1512 		vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1513 		hmp->devvp = NULL;
1514 		VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1515 		vn_unlock(devvp);
1516 		vrele(devvp);
1517 		devvp = NULL;
1518 	}
1519 
1520 	/*
1521 	 * Clear vchain/fchain flags that might prevent final cleanup
1522 	 * of these chains.
1523 	 */
1524 	if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1525 		atomic_add_long(&hammer2_count_modified_chains, -1);
1526 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1527 		hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1528 	}
1529 	if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1530 		atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1531 	}
1532 
1533 	if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1534 		atomic_add_long(&hammer2_count_modified_chains, -1);
1535 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1536 		hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1537 	}
1538 	if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1539 		atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1540 	}
1541 
1542 	/*
1543 	 * Final drop of embedded freemap root chain to
1544 	 * clean up fchain.core (fchain structure is not
1545 	 * flagged ALLOCATED so it is cleaned out and then
1546 	 * left to rot).
1547 	 */
1548 	hammer2_chain_drop(&hmp->fchain);
1549 
1550 	/*
1551 	 * Final drop of embedded volume root chain to clean
1552 	 * up vchain.core (vchain structure is not flagged
1553 	 * ALLOCATED so it is cleaned out and then left to
1554 	 * rot).
1555 	 */
1556 	dumpcnt = 50;
1557 	hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1558 	dumpcnt = 50;
1559 	hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1560 	hammer2_dev_unlock(hmp);
1561 	hammer2_chain_drop(&hmp->vchain);
1562 
1563 	hammer2_io_cleanup(hmp, &hmp->iotree);
1564 	if (hmp->iofree_count) {
1565 		kprintf("io_cleanup: %d I/O's left hanging\n",
1566 			hmp->iofree_count);
1567 	}
1568 
1569 	TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1570 	kmalloc_destroy(&hmp->mchain);
1571 	kfree(hmp, M_HAMMER2);
1572 }
1573 
1574 int
1575 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1576 		 ino_t ino, struct vnode **vpp)
1577 {
1578 	hammer2_xop_lookup_t *xop;
1579 	hammer2_pfs_t *pmp;
1580 	hammer2_inode_t *ip;
1581 	hammer2_tid_t inum;
1582 	int error;
1583 
1584 	inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1585 
1586 	error = 0;
1587 	pmp = MPTOPMP(mp);
1588 
1589 	/*
1590 	 * Easy if we already have it cached
1591 	 */
1592 	ip = hammer2_inode_lookup(pmp, inum);
1593 	if (ip) {
1594 		hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1595 		*vpp = hammer2_igetv(ip, &error);
1596 		hammer2_inode_unlock(ip);
1597 		hammer2_inode_drop(ip);		/* from lookup */
1598 
1599 		return error;
1600 	}
1601 
1602 	/*
1603 	 * Otherwise we have to find the inode
1604 	 */
1605 	xop = hammer2_xop_alloc(pmp->iroot, 0);
1606 	xop->lhc = inum;
1607 	hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1608 	error = hammer2_xop_collect(&xop->head, 0);
1609 
1610 	if (error == 0) {
1611 		if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1612 			kprintf("vget: no collect error but also no rdata\n");
1613 			kprintf("xop %p\n", xop);
1614 			while ((hammer2_debug & 0x80000) == 0) {
1615 				tsleep(xop, PCATCH, "wait", hz * 10);
1616 			}
1617 			ip = NULL;
1618 		} else {
1619 			ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1620 		}
1621 	}
1622 	hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1623 
1624 	if (ip) {
1625 		*vpp = hammer2_igetv(ip, &error);
1626 		hammer2_inode_unlock(ip);
1627 	} else {
1628 		*vpp = NULL;
1629 		error = ENOENT;
1630 	}
1631 	return (error);
1632 }
1633 
1634 static
1635 int
1636 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1637 {
1638 	hammer2_pfs_t *pmp;
1639 	struct vnode *vp;
1640 	int error;
1641 
1642 	pmp = MPTOPMP(mp);
1643 	if (pmp->iroot == NULL) {
1644 		*vpp = NULL;
1645 		return EINVAL;
1646 	}
1647 
1648 	error = 0;
1649 	hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1650 
1651 	while (pmp->inode_tid == 0) {
1652 		hammer2_xop_ipcluster_t *xop;
1653 		hammer2_inode_meta_t *meta;
1654 
1655 		xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1656 		hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1657 		error = hammer2_xop_collect(&xop->head, 0);
1658 
1659 		if (error == 0) {
1660 			meta = &xop->head.cluster.focus->data->ipdata.meta;
1661 			pmp->iroot->meta = *meta;
1662 			pmp->inode_tid = meta->pfs_inum + 1;
1663 			if (pmp->inode_tid < HAMMER2_INODE_START)
1664 				pmp->inode_tid = HAMMER2_INODE_START;
1665 			pmp->modify_tid =
1666 				xop->head.cluster.focus->bref.modify_tid + 1;
1667 			kprintf("PFS: Starting inode %jd\n",
1668 				(intmax_t)pmp->inode_tid);
1669 			kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1670 				pmp->inode_tid, pmp->modify_tid);
1671 			wakeup(&pmp->iroot);
1672 
1673 			hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1674 
1675 			/*
1676 			 * Prime the mount info.
1677 			 */
1678 			hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1679 			break;
1680 		}
1681 
1682 		/*
1683 		 * Loop, try again
1684 		 */
1685 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1686 		hammer2_inode_unlock(pmp->iroot);
1687 		error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1688 		hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1689 		if (error == EINTR)
1690 			break;
1691 	}
1692 
1693 	if (error) {
1694 		hammer2_inode_unlock(pmp->iroot);
1695 		*vpp = NULL;
1696 	} else {
1697 		vp = hammer2_igetv(pmp->iroot, &error);
1698 		hammer2_inode_unlock(pmp->iroot);
1699 		*vpp = vp;
1700 	}
1701 
1702 	return (error);
1703 }
1704 
1705 /*
1706  * Filesystem status
1707  *
1708  * XXX incorporate ipdata->meta.inode_quota and data_quota
1709  */
1710 static
1711 int
1712 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1713 {
1714 	hammer2_pfs_t *pmp;
1715 	hammer2_dev_t *hmp;
1716 	hammer2_blockref_t bref;
1717 	int i;
1718 
1719 	/*
1720 	 * NOTE: iroot might not have validated the cluster yet.
1721 	 */
1722 	pmp = MPTOPMP(mp);
1723 
1724 	mp->mnt_stat.f_files = 0;
1725 	mp->mnt_stat.f_ffree = 0;
1726 	mp->mnt_stat.f_blocks = 0;
1727 	mp->mnt_stat.f_bfree = 0;
1728 	mp->mnt_stat.f_bavail = 0;
1729 
1730 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1731 		hmp = pmp->pfs_hmps[i];
1732 		if (hmp == NULL)
1733 			continue;
1734 		if (pmp->iroot->cluster.array[i].chain)
1735 			bref = pmp->iroot->cluster.array[i].chain->bref;
1736 		else
1737 			bzero(&bref, sizeof(bref));
1738 
1739 		mp->mnt_stat.f_files = bref.inode_count;
1740 		mp->mnt_stat.f_ffree = 0;
1741 #if 0
1742 		mp->mnt_stat.f_blocks = (bref.data_count +
1743 					 hmp->voldata.allocator_free) /
1744 					mp->mnt_vstat.f_bsize;
1745 		mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free /
1746 					mp->mnt_vstat.f_bsize;
1747 #endif
1748 		mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1749 					mp->mnt_vstat.f_bsize;
1750 		mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1751 					mp->mnt_vstat.f_bsize;
1752 		mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1753 
1754 		*sbp = mp->mnt_stat;
1755 	}
1756 	return (0);
1757 }
1758 
1759 static
1760 int
1761 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1762 {
1763 	hammer2_pfs_t *pmp;
1764 	hammer2_dev_t *hmp;
1765 	hammer2_blockref_t bref;
1766 	int i;
1767 
1768 	/*
1769 	 * NOTE: iroot might not have validated the cluster yet.
1770 	 */
1771 	pmp = MPTOPMP(mp);
1772 
1773 	mp->mnt_vstat.f_bsize = 0;
1774 	mp->mnt_vstat.f_files = 0;
1775 	mp->mnt_vstat.f_ffree = 0;
1776 	mp->mnt_vstat.f_blocks = 0;
1777 	mp->mnt_vstat.f_bfree = 0;
1778 	mp->mnt_vstat.f_bavail = 0;
1779 
1780 	for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1781 		hmp = pmp->pfs_hmps[i];
1782 		if (hmp == NULL)
1783 			continue;
1784 		if (pmp->iroot->cluster.array[i].chain)
1785 			bref = pmp->iroot->cluster.array[i].chain->bref;
1786 		else
1787 			bzero(&bref, sizeof(bref));
1788 
1789 		mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1790 		mp->mnt_vstat.f_files = bref.inode_count;
1791 		mp->mnt_vstat.f_ffree = 0;
1792 #if 0
1793 		mp->mnt_vstat.f_blocks = (bref.data_count +
1794 					 hmp->voldata.allocator_free) /
1795 					mp->mnt_vstat.f_bsize;
1796 		mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1797 					mp->mnt_vstat.f_bsize;
1798 #endif
1799 		mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1800 					mp->mnt_vstat.f_bsize;
1801 		mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1802 					mp->mnt_vstat.f_bsize;
1803 		mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1804 
1805 		*sbp = mp->mnt_vstat;
1806 	}
1807 	return (0);
1808 }
1809 
1810 /*
1811  * Mount-time recovery (RW mounts)
1812  *
1813  * Updates to the free block table are allowed to lag flushes by one
1814  * transaction.  In case of a crash, then on a fresh mount we must do an
1815  * incremental scan of the last committed transaction id and make sure that
1816  * all related blocks have been marked allocated.
1817  *
1818  * The super-root topology and each PFS has its own transaction id domain,
1819  * so we must track PFS boundary transitions.
1820  */
1821 struct hammer2_recovery_elm {
1822 	TAILQ_ENTRY(hammer2_recovery_elm) entry;
1823 	hammer2_chain_t *chain;
1824 	hammer2_tid_t sync_tid;
1825 };
1826 
1827 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1828 
1829 struct hammer2_recovery_info {
1830 	struct hammer2_recovery_list list;
1831 	hammer2_tid_t	mtid;
1832 	int	depth;
1833 };
1834 
1835 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1836 			hammer2_chain_t *parent,
1837 			struct hammer2_recovery_info *info,
1838 			hammer2_tid_t sync_tid);
1839 
1840 #define HAMMER2_RECOVERY_MAXDEPTH	10
1841 
1842 static
1843 int
1844 hammer2_recovery(hammer2_dev_t *hmp)
1845 {
1846 	struct hammer2_recovery_info info;
1847 	struct hammer2_recovery_elm *elm;
1848 	hammer2_chain_t *parent;
1849 	hammer2_tid_t sync_tid;
1850 	hammer2_tid_t mirror_tid;
1851 	int error;
1852 	int cumulative_error = 0;
1853 
1854 	hammer2_trans_init(hmp->spmp, 0);
1855 
1856 	sync_tid = hmp->voldata.freemap_tid;
1857 	mirror_tid = hmp->voldata.mirror_tid;
1858 
1859 	kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1860 	if (sync_tid >= mirror_tid) {
1861 		kprintf(" no recovery needed\n");
1862 	} else {
1863 		kprintf(" freemap recovery %016jx-%016jx\n",
1864 			sync_tid + 1, mirror_tid);
1865 	}
1866 
1867 	TAILQ_INIT(&info.list);
1868 	info.depth = 0;
1869 	parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1870 	cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
1871 	hammer2_chain_lookup_done(parent);
1872 
1873 	while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1874 		TAILQ_REMOVE(&info.list, elm, entry);
1875 		parent = elm->chain;
1876 		sync_tid = elm->sync_tid;
1877 		kfree(elm, M_HAMMER2);
1878 
1879 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1880 		error = hammer2_recovery_scan(hmp, parent, &info,
1881 					      hmp->voldata.freemap_tid);
1882 		hammer2_chain_unlock(parent);
1883 		hammer2_chain_drop(parent);	/* drop elm->chain ref */
1884 		if (error)
1885 			cumulative_error = error;
1886 	}
1887 	hammer2_trans_done(hmp->spmp);
1888 
1889 	return cumulative_error;
1890 }
1891 
1892 static
1893 int
1894 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1895 		      struct hammer2_recovery_info *info,
1896 		      hammer2_tid_t sync_tid)
1897 {
1898 	const hammer2_inode_data_t *ripdata;
1899 	hammer2_chain_t *chain;
1900 	hammer2_blockref_t bref;
1901 	int cache_index;
1902 	int cumulative_error = 0;
1903 	int error;
1904 	int first;
1905 
1906 	/*
1907 	 * Adjust freemap to ensure that the block(s) are marked allocated.
1908 	 */
1909 	if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1910 		hammer2_freemap_adjust(hmp, &parent->bref,
1911 				       HAMMER2_FREEMAP_DORECOVER);
1912 	}
1913 
1914 	/*
1915 	 * Check type for recursive scan
1916 	 */
1917 	switch(parent->bref.type) {
1918 	case HAMMER2_BREF_TYPE_VOLUME:
1919 		/* data already instantiated */
1920 		break;
1921 	case HAMMER2_BREF_TYPE_INODE:
1922 		/*
1923 		 * Must instantiate data for DIRECTDATA test and also
1924 		 * for recursion.
1925 		 */
1926 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1927 		ripdata = &hammer2_chain_rdata(parent)->ipdata;
1928 		if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1929 			/* not applicable to recovery scan */
1930 			hammer2_chain_unlock(parent);
1931 			return 0;
1932 		}
1933 		hammer2_chain_unlock(parent);
1934 		break;
1935 	case HAMMER2_BREF_TYPE_INDIRECT:
1936 		/*
1937 		 * Must instantiate data for recursion
1938 		 */
1939 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1940 		hammer2_chain_unlock(parent);
1941 		break;
1942 	case HAMMER2_BREF_TYPE_DATA:
1943 	case HAMMER2_BREF_TYPE_FREEMAP:
1944 	case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1945 	case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1946 		/* not applicable to recovery scan */
1947 		return 0;
1948 		break;
1949 	default:
1950 		return EDOM;
1951 	}
1952 
1953 	/*
1954 	 * Defer operation if depth limit reached or if we are crossing a
1955 	 * PFS boundary.
1956 	 */
1957 	if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1958 		struct hammer2_recovery_elm *elm;
1959 
1960 		elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1961 		elm->chain = parent;
1962 		elm->sync_tid = sync_tid;
1963 		hammer2_chain_ref(parent);
1964 		TAILQ_INSERT_TAIL(&info->list, elm, entry);
1965 		/* unlocked by caller */
1966 
1967 		return(0);
1968 	}
1969 
1970 
1971 	/*
1972 	 * Recursive scan of the last flushed transaction only.  We are
1973 	 * doing this without pmp assignments so don't leave the chains
1974 	 * hanging around after we are done with them.
1975 	 */
1976 	cache_index = 0;
1977 	chain = NULL;
1978 	first = 1;
1979 
1980 	while (hammer2_chain_scan(parent, &chain, &bref,
1981 				  &first, &cache_index,
1982 				  HAMMER2_LOOKUP_NODATA) != NULL) {
1983 		/*
1984 		 * If this is a leaf
1985 		 */
1986 		if (chain == NULL) {
1987 			if (bref.mirror_tid > sync_tid) {
1988 				hammer2_freemap_adjust(hmp, &bref,
1989 						     HAMMER2_FREEMAP_DORECOVER);
1990 			}
1991 			continue;
1992 		}
1993 
1994 		/*
1995 		 * This may or may not be a recursive node.
1996 		 */
1997 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
1998 		if (bref.mirror_tid > sync_tid) {
1999 			++info->depth;
2000 			error = hammer2_recovery_scan(hmp, chain,
2001 						      info, sync_tid);
2002 			--info->depth;
2003 			if (error)
2004 				cumulative_error = error;
2005 		}
2006 
2007 		/*
2008 		 * Flush the recovery at the PFS boundary to stage it for
2009 		 * the final flush of the super-root topology.
2010 		 */
2011 		if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2012 		    (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2013 			hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2014 		}
2015 	}
2016 
2017 	return cumulative_error;
2018 }
2019 
2020 /*
2021  * Sync a mount point; this is called on a per-mount basis from the
2022  * filesystem syncer process periodically and whenever a user issues
2023  * a sync.
2024  */
2025 int
2026 hammer2_vfs_sync(struct mount *mp, int waitfor)
2027 {
2028 	hammer2_xop_flush_t *xop;
2029 	struct hammer2_sync_info info;
2030 	hammer2_inode_t *iroot;
2031 	hammer2_pfs_t *pmp;
2032 	int flags;
2033 	int error;
2034 
2035 	pmp = MPTOPMP(mp);
2036 	iroot = pmp->iroot;
2037 	KKASSERT(iroot);
2038 	KKASSERT(iroot->pmp == pmp);
2039 
2040 	/*
2041 	 * We can't acquire locks on existing vnodes while in a transaction
2042 	 * without risking a deadlock.  This assumes that vfsync() can be
2043 	 * called without the vnode locked (which it can in DragonFly).
2044 	 * Otherwise we'd have to implement a multi-pass or flag the lock
2045 	 * failures and retry.
2046 	 *
2047 	 * The reclamation code interlocks with the sync list's token
2048 	 * (by removing the vnode from the scan list) before unlocking
2049 	 * the inode, giving us time to ref the inode.
2050 	 */
2051 	/*flags = VMSC_GETVP;*/
2052 	flags = 0;
2053 	if (waitfor & MNT_LAZY)
2054 		flags |= VMSC_ONEPASS;
2055 
2056 #if 1
2057 	/*
2058 	 * Preflush the vnodes using a normal transaction before interlocking
2059 	 * with a flush transaction.  We do this to try to run as much of
2060 	 * the compression as possible outside the flush transaction.
2061 	 */
2062 	hammer2_trans_init(pmp, 0);
2063 	info.error = 0;
2064 	info.waitfor = MNT_NOWAIT;
2065 	vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2066 	hammer2_trans_done(pmp);
2067 #endif
2068 
2069 	/*
2070 	 * Start our flush transaction.  This does not return until all
2071 	 * concurrent transactions have completed and will prevent any
2072 	 * new transactions from running concurrently, except for the
2073 	 * buffer cache transactions.
2074 	 *
2075 	 * For efficiency do an async pass before making sure with a
2076 	 * synchronous pass on all related buffer cache buffers.  It
2077 	 * should theoretically not be possible for any new file buffers
2078 	 * to be instantiated during this sequence.
2079 	 */
2080 	hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH |
2081 			        HAMMER2_TRANS_PREFLUSH);
2082 	hammer2_inode_run_sideq(pmp);
2083 
2084 	info.error = 0;
2085 	info.waitfor = MNT_NOWAIT;
2086 	vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2087 	info.waitfor = MNT_WAIT;
2088 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2089 
2090 	/*
2091 	 * Clear PREFLUSH.  This prevents (or asserts on) any new logical
2092 	 * buffer cache flushes which occur during the flush.  Device buffers
2093 	 * are not affected.
2094 	 */
2095 	hammer2_bioq_sync(pmp);
2096 	hammer2_trans_clear_preflush(pmp);
2097 
2098 	/*
2099 	 * Use the XOP interface to concurrently flush all nodes to
2100 	 * synchronize the PFSROOT subtopology to the media.  A standard
2101 	 * end-of-scan ENOENT error indicates cluster sufficiency.
2102 	 *
2103 	 * Note that this flush will not be visible on crash recovery until
2104 	 * we flush the super-root topology in the next loop.
2105 	 *
2106 	 * XXX For now wait for all flushes to complete.
2107 	 */
2108 	if (iroot) {
2109 		xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2110 		hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2111 		error = hammer2_xop_collect(&xop->head,
2112 					    HAMMER2_XOP_COLLECT_WAITALL);
2113 		hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2114 		if (error == ENOENT)
2115 			error = 0;
2116 	} else {
2117 		error = 0;
2118 	}
2119 	hammer2_trans_done(pmp);
2120 
2121 	return (error);
2122 }
2123 
2124 /*
2125  * Sync passes.
2126  *
2127  * Note that we ignore the tranasction mtid we got above.  Instead,
2128  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2129  * transactions.
2130  */
2131 static int
2132 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2133 {
2134 	struct hammer2_sync_info *info = data;
2135 	hammer2_inode_t *ip;
2136 	int error;
2137 
2138 	/*
2139 	 * Degenerate cases.  Note that ip == NULL typically means the
2140 	 * syncer vnode itself and we don't want to vclrisdirty() in that
2141 	 * situation.
2142 	 */
2143 	ip = VTOI(vp);
2144 	if (ip == NULL) {
2145 		return(0);
2146 	}
2147 	if (vp->v_type == VNON || vp->v_type == VBAD) {
2148 		vclrisdirty(vp);
2149 		return(0);
2150 	}
2151 
2152 	/*
2153 	 * VOP_FSYNC will start a new transaction so replicate some code
2154 	 * here to do it inline (see hammer2_vop_fsync()).
2155 	 *
2156 	 * WARNING: The vfsync interacts with the buffer cache and might
2157 	 *          block, we can't hold the inode lock at that time.
2158 	 *	    However, we MUST ref ip before blocking to ensure that
2159 	 *	    it isn't ripped out from under us (since we do not
2160 	 *	    hold a lock on the vnode).
2161 	 */
2162 	hammer2_inode_ref(ip);
2163 	if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2164 	    !RB_EMPTY(&vp->v_rbdirty_tree)) {
2165 		vfsync(vp, info->waitfor, 1, NULL, NULL);
2166 		if (ip->flags & (HAMMER2_INODE_RESIZED |
2167 				 HAMMER2_INODE_MODIFIED)) {
2168 			hammer2_inode_lock(ip, 0);
2169 			hammer2_inode_chain_sync(ip);
2170 			hammer2_inode_unlock(ip);
2171 		}
2172 	}
2173 	if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2174 	    RB_EMPTY(&vp->v_rbdirty_tree)) {
2175 		vclrisdirty(vp);
2176 	}
2177 
2178 	hammer2_inode_drop(ip);
2179 #if 1
2180 	error = 0;
2181 	if (error)
2182 		info->error = error;
2183 #endif
2184 	return(0);
2185 }
2186 
2187 static
2188 int
2189 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2190 {
2191 	hammer2_inode_t *ip;
2192 
2193 	KKASSERT(MAXFIDSZ >= 16);
2194 	ip = VTOI(vp);
2195 	fhp->fid_len = offsetof(struct fid, fid_data[16]);
2196 	fhp->fid_ext = 0;
2197 	((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2198 	((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2199 
2200 	return 0;
2201 }
2202 
2203 static
2204 int
2205 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2206 	       struct fid *fhp, struct vnode **vpp)
2207 {
2208 	hammer2_pfs_t *pmp;
2209 	hammer2_tid_t inum;
2210 	int error;
2211 
2212 	pmp = MPTOPMP(mp);
2213 	inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2214 	if (vpp) {
2215 		if (inum == 1)
2216 			error = hammer2_vfs_root(mp, vpp);
2217 		else
2218 			error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2219 	} else {
2220 		error = 0;
2221 	}
2222 	if (error)
2223 		kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2224 	return error;
2225 }
2226 
2227 static
2228 int
2229 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2230 		 int *exflagsp, struct ucred **credanonp)
2231 {
2232 	hammer2_pfs_t *pmp;
2233 	struct netcred *np;
2234 	int error;
2235 
2236 	pmp = MPTOPMP(mp);
2237 	np = vfs_export_lookup(mp, &pmp->export, nam);
2238 	if (np) {
2239 		*exflagsp = np->netc_exflags;
2240 		*credanonp = &np->netc_anon;
2241 		error = 0;
2242 	} else {
2243 		error = EACCES;
2244 	}
2245 	return error;
2246 }
2247 
2248 /*
2249  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2250  * header into the HMP
2251  *
2252  * XXX read four volhdrs and use the one with the highest TID whos CRC
2253  *     matches.
2254  *
2255  * XXX check iCRCs.
2256  *
2257  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2258  *     nonexistant locations.
2259  *
2260  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2261  */
2262 static
2263 int
2264 hammer2_install_volume_header(hammer2_dev_t *hmp)
2265 {
2266 	hammer2_volume_data_t *vd;
2267 	struct buf *bp;
2268 	hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2269 	int error_reported;
2270 	int error;
2271 	int valid;
2272 	int i;
2273 
2274 	error_reported = 0;
2275 	error = 0;
2276 	valid = 0;
2277 	bp = NULL;
2278 
2279 	/*
2280 	 * There are up to 4 copies of the volume header (syncs iterate
2281 	 * between them so there is no single master).  We don't trust the
2282 	 * volu_size field so we don't know precisely how large the filesystem
2283 	 * is, so depend on the OS to return an error if we go beyond the
2284 	 * block device's EOF.
2285 	 */
2286 	for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2287 		error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2288 			      HAMMER2_VOLUME_BYTES, &bp);
2289 		if (error) {
2290 			brelse(bp);
2291 			bp = NULL;
2292 			continue;
2293 		}
2294 
2295 		vd = (struct hammer2_volume_data *) bp->b_data;
2296 		if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2297 		    (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2298 			brelse(bp);
2299 			bp = NULL;
2300 			continue;
2301 		}
2302 
2303 		if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2304 			/* XXX: Reversed-endianness filesystem */
2305 			kprintf("hammer2: reverse-endian filesystem detected");
2306 			brelse(bp);
2307 			bp = NULL;
2308 			continue;
2309 		}
2310 
2311 		crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2312 		crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2313 				      HAMMER2_VOLUME_ICRC0_SIZE);
2314 		bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2315 		bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2316 				       HAMMER2_VOLUME_ICRC1_SIZE);
2317 		if ((crc0 != crc) || (bcrc0 != bcrc)) {
2318 			kprintf("hammer2 volume header crc "
2319 				"mismatch copy #%d %08x/%08x\n",
2320 				i, crc0, crc);
2321 			error_reported = 1;
2322 			brelse(bp);
2323 			bp = NULL;
2324 			continue;
2325 		}
2326 		if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2327 			valid = 1;
2328 			hmp->voldata = *vd;
2329 			hmp->volhdrno = i;
2330 		}
2331 		brelse(bp);
2332 		bp = NULL;
2333 	}
2334 	if (valid) {
2335 		hmp->volsync = hmp->voldata;
2336 		error = 0;
2337 		if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2338 			kprintf("hammer2: using volume header #%d\n",
2339 				hmp->volhdrno);
2340 		}
2341 	} else {
2342 		error = EINVAL;
2343 		kprintf("hammer2: no valid volume headers found!\n");
2344 	}
2345 	return (error);
2346 }
2347 
2348 /*
2349  * This handles hysteresis on regular file flushes.  Because the BIOs are
2350  * routed to a thread it is possible for an excessive number to build up
2351  * and cause long front-end stalls long before the runningbuffspace limit
2352  * is hit, so we implement hammer2_flush_pipe to control the
2353  * hysteresis.
2354  *
2355  * This is a particular problem when compression is used.
2356  */
2357 void
2358 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2359 {
2360 	atomic_add_int(&pmp->count_lwinprog, 1);
2361 }
2362 
2363 void
2364 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2365 {
2366 	int lwinprog;
2367 
2368 	lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2369 	if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2370 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2371 		atomic_clear_int(&pmp->count_lwinprog,
2372 				 HAMMER2_LWINPROG_WAITING);
2373 		wakeup(&pmp->count_lwinprog);
2374 	}
2375 	if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2376 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2377 		atomic_clear_int(&pmp->count_lwinprog,
2378 				 HAMMER2_LWINPROG_WAITING0);
2379 		wakeup(&pmp->count_lwinprog);
2380 	}
2381 }
2382 
2383 void
2384 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2385 {
2386 	int lwinprog;
2387 	int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2388 				    HAMMER2_LWINPROG_WAITING0;
2389 
2390 	for (;;) {
2391 		lwinprog = pmp->count_lwinprog;
2392 		cpu_ccfence();
2393 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2394 			break;
2395 		tsleep_interlock(&pmp->count_lwinprog, 0);
2396 		atomic_set_int(&pmp->count_lwinprog, lwflag);
2397 		lwinprog = pmp->count_lwinprog;
2398 		if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2399 			break;
2400 		tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2401 	}
2402 }
2403 
2404 /*
2405  * Manage excessive memory resource use for chain and related
2406  * structures.
2407  */
2408 void
2409 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2410 {
2411 	uint32_t waiting;
2412 	uint32_t count;
2413 	uint32_t limit;
2414 #if 0
2415 	static int zzticks;
2416 #endif
2417 
2418 	/*
2419 	 * Atomic check condition and wait.  Also do an early speedup of
2420 	 * the syncer to try to avoid hitting the wait.
2421 	 */
2422 	for (;;) {
2423 		waiting = pmp->inmem_dirty_chains;
2424 		cpu_ccfence();
2425 		count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2426 
2427 		limit = pmp->mp->mnt_nvnodelistsize / 10;
2428 		if (limit < hammer2_limit_dirty_chains)
2429 			limit = hammer2_limit_dirty_chains;
2430 		if (limit < 1000)
2431 			limit = 1000;
2432 
2433 #if 0
2434 		if ((int)(ticks - zzticks) > hz) {
2435 			zzticks = ticks;
2436 			kprintf("count %ld %ld\n", count, limit);
2437 		}
2438 #endif
2439 
2440 		/*
2441 		 * Block if there are too many dirty chains present, wait
2442 		 * for the flush to clean some out.
2443 		 */
2444 		if (count > limit) {
2445 			tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2446 			if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2447 					       waiting,
2448 				       waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2449 				speedup_syncer(pmp->mp);
2450 				tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2451 				       "chnmem", hz);
2452 			}
2453 			continue;	/* loop on success or fail */
2454 		}
2455 
2456 		/*
2457 		 * Try to start an early flush before we are forced to block.
2458 		 */
2459 		if (count > limit * 7 / 10)
2460 			speedup_syncer(pmp->mp);
2461 		break;
2462 	}
2463 }
2464 
2465 void
2466 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2467 {
2468 	if (pmp) {
2469 		atomic_add_int(&pmp->inmem_dirty_chains, 1);
2470 	}
2471 }
2472 
2473 void
2474 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2475 {
2476 	uint32_t waiting;
2477 
2478 	if (pmp == NULL)
2479 		return;
2480 
2481 	for (;;) {
2482 		waiting = pmp->inmem_dirty_chains;
2483 		cpu_ccfence();
2484 		if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2485 				       waiting,
2486 				       (waiting - 1) &
2487 					~HAMMER2_DIRTYCHAIN_WAITING)) {
2488 			break;
2489 		}
2490 	}
2491 
2492 	if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2493 		wakeup(&pmp->inmem_dirty_chains);
2494 }
2495 
2496 /*
2497  * Debugging
2498  */
2499 void
2500 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2501 {
2502 	hammer2_chain_t *scan;
2503 	hammer2_chain_t *parent;
2504 
2505 	--*countp;
2506 	if (*countp == 0) {
2507 		kprintf("%*.*s...\n", tab, tab, "");
2508 		return;
2509 	}
2510 	if (*countp < 0)
2511 		return;
2512 	kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2513 		tab, tab, "", pfx,
2514 		chain, chain->bref.type,
2515 		chain->bref.key, chain->bref.keybits,
2516 		chain->bref.mirror_tid);
2517 
2518 	kprintf("%*.*s      [%08x] (%s) refs=%d",
2519 		tab, tab, "",
2520 		chain->flags,
2521 		((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2522 		chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2523 		chain->refs);
2524 
2525 	parent = chain->parent;
2526 	if (parent)
2527 		kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2528 			tab, tab, "",
2529 			parent, parent->flags, parent->refs);
2530 	if (RB_EMPTY(&chain->core.rbtree)) {
2531 		kprintf("\n");
2532 	} else {
2533 		kprintf(" {\n");
2534 		RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2535 			hammer2_dump_chain(scan, tab + 4, countp, 'a');
2536 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2537 			kprintf("%*.*s}(%s)\n", tab, tab, "",
2538 				chain->data->ipdata.filename);
2539 		else
2540 			kprintf("%*.*s}\n", tab, tab, "");
2541 	}
2542 }
2543