xref: /dflybsd-src/sys/vfs/hammer2/hammer2_vfsops.c (revision 25ca8c794df06033ca37cb3fdf9b2992a7b5a9e6)
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 	hammer2_trans_t trans;
71 	int error;
72 	int waitfor;
73 };
74 
75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
76 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
77 static struct hammer2_mntlist hammer2_mntlist;
78 static struct hammer2_pfslist hammer2_pfslist;
79 static struct lock hammer2_mntlk;
80 
81 int hammer2_debug;
82 int hammer2_cluster_enable = 1;
83 int hammer2_hardlink_enable = 1;
84 int hammer2_flush_pipe = 100;
85 int hammer2_synchronous_flush = 1;
86 int hammer2_dio_count;
87 long hammer2_limit_dirty_chains;
88 long hammer2_iod_file_read;
89 long hammer2_iod_meta_read;
90 long hammer2_iod_indr_read;
91 long hammer2_iod_fmap_read;
92 long hammer2_iod_volu_read;
93 long hammer2_iod_file_write;
94 long hammer2_iod_meta_write;
95 long hammer2_iod_indr_write;
96 long hammer2_iod_fmap_write;
97 long hammer2_iod_volu_write;
98 long hammer2_ioa_file_read;
99 long hammer2_ioa_meta_read;
100 long hammer2_ioa_indr_read;
101 long hammer2_ioa_fmap_read;
102 long hammer2_ioa_volu_read;
103 long hammer2_ioa_fmap_write;
104 long hammer2_ioa_file_write;
105 long hammer2_ioa_meta_write;
106 long hammer2_ioa_indr_write;
107 long hammer2_ioa_volu_write;
108 
109 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
110 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
111 		"Buffer used for compression.");
112 
113 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
114 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
115 		"Buffer used for decompression.");
116 
117 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
118 
119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
120 	   &hammer2_debug, 0, "");
121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW,
122 	   &hammer2_cluster_enable, 0, "");
123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
124 	   &hammer2_hardlink_enable, 0, "");
125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
126 	   &hammer2_flush_pipe, 0, "");
127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
128 	   &hammer2_synchronous_flush, 0, "");
129 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
130 	   &hammer2_limit_dirty_chains, 0, "");
131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
132 	   &hammer2_dio_count, 0, "");
133 
134 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
135 	   &hammer2_iod_file_read, 0, "");
136 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
137 	   &hammer2_iod_meta_read, 0, "");
138 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
139 	   &hammer2_iod_indr_read, 0, "");
140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
141 	   &hammer2_iod_fmap_read, 0, "");
142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
143 	   &hammer2_iod_volu_read, 0, "");
144 
145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
146 	   &hammer2_iod_file_write, 0, "");
147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
148 	   &hammer2_iod_meta_write, 0, "");
149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
150 	   &hammer2_iod_indr_write, 0, "");
151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
152 	   &hammer2_iod_fmap_write, 0, "");
153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
154 	   &hammer2_iod_volu_write, 0, "");
155 
156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
157 	   &hammer2_ioa_file_read, 0, "");
158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
159 	   &hammer2_ioa_meta_read, 0, "");
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
161 	   &hammer2_ioa_indr_read, 0, "");
162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
163 	   &hammer2_ioa_fmap_read, 0, "");
164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
165 	   &hammer2_ioa_volu_read, 0, "");
166 
167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
168 	   &hammer2_ioa_file_write, 0, "");
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
170 	   &hammer2_ioa_meta_write, 0, "");
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
172 	   &hammer2_ioa_indr_write, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
174 	   &hammer2_ioa_fmap_write, 0, "");
175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
176 	   &hammer2_ioa_volu_write, 0, "");
177 
178 static int hammer2_vfs_init(struct vfsconf *conf);
179 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
180 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
181 				struct ucred *cred);
182 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
183 				struct vnode *, struct ucred *);
184 static int hammer2_recovery(hammer2_dev_t *hmp);
185 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
186 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
187 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
188 				struct ucred *cred);
189 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
190 				struct ucred *cred);
191 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
192 				ino_t ino, struct vnode **vpp);
193 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
194 				struct fid *fhp, struct vnode **vpp);
195 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
196 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
197 				int *exflagsp, struct ucred **credanonp);
198 
199 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
200 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
201 
202 static void hammer2_update_pmps(hammer2_dev_t *hmp);
203 
204 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
205 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
206 				hammer2_dev_t *hmp);
207 
208 /*
209  * HAMMER2 vfs operations.
210  */
211 static struct vfsops hammer2_vfsops = {
212 	.vfs_init	= hammer2_vfs_init,
213 	.vfs_uninit	= hammer2_vfs_uninit,
214 	.vfs_sync	= hammer2_vfs_sync,
215 	.vfs_mount	= hammer2_vfs_mount,
216 	.vfs_unmount	= hammer2_vfs_unmount,
217 	.vfs_root 	= hammer2_vfs_root,
218 	.vfs_statfs	= hammer2_vfs_statfs,
219 	.vfs_statvfs	= hammer2_vfs_statvfs,
220 	.vfs_vget	= hammer2_vfs_vget,
221 	.vfs_vptofh	= hammer2_vfs_vptofh,
222 	.vfs_fhtovp	= hammer2_vfs_fhtovp,
223 	.vfs_checkexp	= hammer2_vfs_checkexp
224 };
225 
226 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
227 
228 VFS_SET(hammer2_vfsops, hammer2, 0);
229 MODULE_VERSION(hammer2, 1);
230 
231 static
232 int
233 hammer2_vfs_init(struct vfsconf *conf)
234 {
235 	static struct objcache_malloc_args margs_read;
236 	static struct objcache_malloc_args margs_write;
237 	static struct objcache_malloc_args margs_vop;
238 
239 	int error;
240 
241 	error = 0;
242 
243 	if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
244 		error = EINVAL;
245 	if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
246 		error = EINVAL;
247 	if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
248 		error = EINVAL;
249 
250 	if (error)
251 		kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
252 
253 	margs_read.objsize = 65536;
254 	margs_read.mtype = M_HAMMER2_DEBUFFER;
255 
256 	margs_write.objsize = 32768;
257 	margs_write.mtype = M_HAMMER2_CBUFFER;
258 
259 	margs_vop.objsize = sizeof(hammer2_xop_t);
260 	margs_vop.mtype = M_HAMMER2;
261 
262 	cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
263 				0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
264 				objcache_malloc_free, &margs_read);
265 	cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
266 				0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
267 				objcache_malloc_free, &margs_write);
268 	cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
269 				0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
270 				objcache_malloc_free, &margs_vop);
271 
272 
273 	lockinit(&hammer2_mntlk, "mntlk", 0, 0);
274 	TAILQ_INIT(&hammer2_mntlist);
275 	TAILQ_INIT(&hammer2_pfslist);
276 
277 	hammer2_limit_dirty_chains = desiredvnodes / 10;
278 
279 	return (error);
280 }
281 
282 static
283 int
284 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
285 {
286 	objcache_destroy(cache_buffer_read);
287 	objcache_destroy(cache_buffer_write);
288 	objcache_destroy(cache_xops);
289 	return 0;
290 }
291 
292 /*
293  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
294  * mounts and the spmp structure for media (hmp) structures.
295  *
296  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
297  * transactions.  Note that synchronization does not use this field.
298  * (typically frontend operations and synchronization cannot run on the
299  * same PFS node at the same time).
300  *
301  * XXX check locking
302  */
303 hammer2_pfs_t *
304 hammer2_pfsalloc(hammer2_cluster_t *cluster,
305 		 const hammer2_inode_data_t *ripdata,
306 		 hammer2_tid_t modify_tid)
307 {
308 	hammer2_chain_t *rchain;
309 	hammer2_inode_t *iroot;
310 	hammer2_pfs_t *pmp;
311 	int count;
312 	int i;
313 	int j;
314 
315 	/*
316 	 * Locate or create the PFS based on the cluster id.  If ripdata
317 	 * is NULL this is a spmp which is unique and is always allocated.
318 	 */
319 	if (ripdata) {
320 		TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
321 			if (bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
322 				 sizeof(pmp->pfs_clid)) == 0) {
323 					break;
324 			}
325 		}
326 	} else {
327 		pmp = NULL;
328 	}
329 
330 	if (pmp == NULL) {
331 		pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
332 		hammer2_trans_manage_init(&pmp->tmanage);
333 		kmalloc_create(&pmp->minode, "HAMMER2-inodes");
334 		kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
335 		lockinit(&pmp->lock, "pfslk", 0, 0);
336 		spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
337 		RB_INIT(&pmp->inum_tree);
338 		TAILQ_INIT(&pmp->unlinkq);
339 		spin_init(&pmp->list_spin, "hm2pfsalloc_list");
340 
341 		for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
342 			hammer2_xop_group_init(pmp, &pmp->xop_groups[j]);
343 
344 		/*
345 		 * Save the last media transaction id for the flusher.  Set
346 		 * initial
347 		 */
348 		if (ripdata)
349 			pmp->pfs_clid = ripdata->meta.pfs_clid;
350 		hammer2_mtx_init(&pmp->wthread_mtx, "h2wthr");
351 		bioq_init(&pmp->wthread_bioq);
352 		TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
353 
354 		/*
355 		 * The synchronization thread may start too early, make
356 		 * sure it stays frozen until we are ready to let it go.
357 		 * XXX
358 		 */
359 		/*
360 		pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
361 					 HAMMER2_THREAD_REMASTER;
362 		*/
363 	}
364 
365 	/*
366 	 * Create the PFS's root inode.
367 	 */
368 	if ((iroot = pmp->iroot) == NULL) {
369 		iroot = hammer2_inode_get(pmp, NULL, NULL);
370 		pmp->iroot = iroot;
371 		hammer2_inode_ref(iroot);
372 		hammer2_inode_unlock(iroot, NULL);
373 	}
374 
375 	/*
376 	 * Stop here if no cluster is passed in.
377 	 */
378 	if (cluster == NULL)
379 		goto done;
380 
381 	/*
382 	 * When a cluster is passed in we must add the cluster's chains
383 	 * to the PFS's root inode, update pmp->pfs_types[], and update
384 	 * the syncronization threads.
385 	 *
386 	 * At the moment empty spots can develop due to removals or failures.
387 	 * Ultimately we want to re-fill these spots but doing so might
388 	 * confused running code. XXX
389 	 */
390 	hammer2_inode_ref(iroot);
391 	hammer2_mtx_ex(&iroot->lock);
392 	j = iroot->cluster.nchains;
393 
394 	kprintf("add PFS to pmp %p[%d]\n", pmp, j);
395 
396 	for (i = 0; i < cluster->nchains; ++i) {
397 		if (j == HAMMER2_MAXCLUSTER)
398 			break;
399 		rchain = cluster->array[i].chain;
400 		KKASSERT(rchain->pmp == NULL);
401 		rchain->pmp = pmp;
402 		hammer2_chain_ref(rchain);
403 		iroot->cluster.array[j].chain = rchain;
404 		pmp->pfs_types[j] = ripdata->meta.pfs_type;
405 		pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
406 
407 		/*
408 		 * If the PFS is already mounted we must account
409 		 * for the mount_count here.
410 		 */
411 		if (pmp->mp)
412 			++rchain->hmp->mount_count;
413 
414 		/*
415 		 * May have to fixup dirty chain tracking.  Previous
416 		 * pmp was NULL so nothing to undo.
417 		 */
418 		if (rchain->flags & HAMMER2_CHAIN_MODIFIED)
419 			hammer2_pfs_memory_inc(pmp);
420 		++j;
421 	}
422 	iroot->cluster.nchains = j;
423 
424 	if (i != cluster->nchains) {
425 		kprintf("hammer2_mount: cluster full!\n");
426 		/* XXX fatal error? */
427 	}
428 
429 	/*
430 	 * Update nmasters from any PFS inode which is part of the cluster.
431 	 * It is possible that this will result in a value which is too
432 	 * high.  MASTER PFSs are authoritative for pfs_nmasters and will
433 	 * override this value later on.
434 	 *
435 	 * (This informs us of masters that might not currently be
436 	 *  discoverable by this mount).
437 	 */
438 	if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
439 		pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
440 	}
441 
442 	/*
443 	 * Count visible masters.  Masters are usually added with
444 	 * ripdata->meta.pfs_nmasters set to 1.  This detects when there
445 	 * are more (XXX and must update the master inodes).
446 	 */
447 	count = 0;
448 	for (i = 0; i < iroot->cluster.nchains; ++i) {
449 		if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
450 			++count;
451 	}
452 	if (pmp->pfs_nmasters < count)
453 		pmp->pfs_nmasters = count;
454 
455 	/*
456 	 * Create missing synchronization and support threads.
457 	 *
458 	 * Single-node masters (including snapshots) have nothing to
459 	 * synchronize and do not require this thread.
460 	 *
461 	 * Multi-node masters or any number of soft masters, slaves, copy,
462 	 * or other PFS types need the thread.
463 	 *
464 	 * Each thread is responsible for its particular cluster index.
465 	 * We use independent threads so stalls or mismatches related to
466 	 * any given target do not affect other targets.
467 	 */
468 	for (i = 0; i < iroot->cluster.nchains; ++i) {
469 		/*
470 		 * Single-node masters (including snapshots) have nothing
471 		 * to synchronize and will make direct xops support calls,
472 		 * thus they do not require this thread.
473 		 *
474 		 * Note that there can be thousands of snapshots.  We do not
475 		 * want to create thousands of threads.
476 		 */
477 		if (pmp->pfs_nmasters <= 1 &&
478 		    pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
479 			continue;
480 		}
481 
482 		/*
483 		 * Sync support thread
484 		 */
485 		if (pmp->sync_thrs[i].td == NULL) {
486 			hammer2_thr_create(&pmp->sync_thrs[i], pmp,
487 					   "h2nod", i, 0,
488 					   hammer2_primary_sync_thread);
489 		}
490 
491 		/*
492 		 * Xops support threads
493 		 */
494 		for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
495 			if (pmp->xop_groups[j].thrs[i].td)
496 				continue;
497 			hammer2_thr_create(&pmp->xop_groups[j].thrs[i], pmp,
498 					   "h2xop", i, j,
499 					   hammer2_primary_xops_thread);
500 		}
501 	}
502 
503 	hammer2_mtx_unlock(&iroot->lock);
504 	hammer2_inode_drop(iroot);
505 done:
506 	return pmp;
507 }
508 
509 /*
510  * Destroy a PFS, typically only occurs after the last mount on a device
511  * has gone away.
512  */
513 static void
514 hammer2_pfsfree(hammer2_pfs_t *pmp)
515 {
516 	hammer2_inode_t *iroot;
517 	int i;
518 	int j;
519 
520 	/*
521 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
522 	 * by the flush code.
523 	 */
524 	TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
525 
526 	iroot = pmp->iroot;
527 	if (iroot) {
528 		for (i = 0; i < iroot->cluster.nchains; ++i) {
529 			hammer2_thr_delete(&pmp->sync_thrs[i]);
530 			for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
531 				hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
532 		}
533 #if REPORT_REFS_ERRORS
534 		if (pmp->iroot->refs != 1)
535 			kprintf("PMP->IROOT %p REFS WRONG %d\n",
536 				pmp->iroot, pmp->iroot->refs);
537 #else
538 		KKASSERT(pmp->iroot->refs == 1);
539 #endif
540 		/* ref for pmp->iroot */
541 		hammer2_inode_drop(pmp->iroot);
542 		pmp->iroot = NULL;
543 	}
544 
545 	kmalloc_destroy(&pmp->mmsg);
546 	kmalloc_destroy(&pmp->minode);
547 
548 	kfree(pmp, M_HAMMER2);
549 }
550 
551 /*
552  * Remove all references to hmp from the pfs list.  Any PFS which becomes
553  * empty is terminated and freed.
554  *
555  * XXX inefficient.
556  */
557 static void
558 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
559 {
560 	hammer2_pfs_t *pmp;
561 	hammer2_inode_t *iroot;
562 	hammer2_cluster_t *cluster;
563 	hammer2_chain_t *rchain;
564 	int didfreeze;
565 	int i;
566 	int j;
567 
568 again:
569 	TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
570 		if ((iroot = pmp->iroot) == NULL)
571 			continue;
572 		if (hmp->spmp == pmp) {
573 			kprintf("unmount hmp %p remove spmp %p\n",
574 				hmp, pmp);
575 			hmp->spmp = NULL;
576 		}
577 
578 		/*
579 		 * Determine if this PFS is affected.  If it is we must
580 		 * freeze all management threads and lock its iroot.
581 		 *
582 		 * Freezing a management thread forces it idle, operations
583 		 * in-progress will be aborted and it will have to start
584 		 * over again when unfrozen, or exit if told to exit.
585 		 */
586 		cluster = &iroot->cluster;
587 		for (i = 0; i < cluster->nchains; ++i) {
588 			rchain = cluster->array[i].chain;
589 			if (rchain == NULL || rchain->hmp != hmp)
590 				continue;
591 			break;
592 		}
593 		if (i != cluster->nchains) {
594 			/*
595 			 * Make sure all synchronization threads are locked
596 			 * down.
597 			 */
598 			for (i = 0; i < iroot->cluster.nchains; ++i) {
599 				hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
600 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
601 					hammer2_thr_freeze_async(
602 						&pmp->xop_groups[j].thrs[i]);
603 				}
604 			}
605 			for (i = 0; i < iroot->cluster.nchains; ++i) {
606 				hammer2_thr_freeze(&pmp->sync_thrs[i]);
607 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
608 					hammer2_thr_freeze(
609 						&pmp->xop_groups[j].thrs[i]);
610 				}
611 			}
612 
613 			/*
614 			 * Lock the inode and clean out matching chains.
615 			 * Note that we cannot use hammer2_inode_lock_*()
616 			 * here because that would attempt to validate the
617 			 * cluster that we are in the middle of ripping
618 			 * apart.
619 			 *
620 			 * WARNING! We are working directly on the inodes
621 			 *	    embedded cluster.
622 			 */
623 			hammer2_mtx_ex(&iroot->lock);
624 
625 			/*
626 			 * Remove the chain from matching elements of the PFS.
627 			 */
628 			for (i = 0; i < cluster->nchains; ++i) {
629 				rchain = cluster->array[i].chain;
630 				if (rchain == NULL || rchain->hmp != hmp)
631 					continue;
632 				hammer2_thr_delete(&pmp->sync_thrs[i]);
633 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
634 					hammer2_thr_delete(
635 						&pmp->xop_groups[j].thrs[i]);
636 				}
637 				rchain = cluster->array[i].chain;
638 				cluster->array[i].chain = NULL;
639 				pmp->pfs_types[i] = 0;
640 				if (pmp->pfs_names[i]) {
641 					kfree(pmp->pfs_names[i], M_HAMMER2);
642 					pmp->pfs_names[i] = NULL;
643 				}
644 				hammer2_chain_drop(rchain);
645 
646 				/* focus hint */
647 				if (cluster->focus == rchain)
648 					cluster->focus = NULL;
649 			}
650 			hammer2_mtx_unlock(&iroot->lock);
651 			didfreeze = 1;	/* remaster, unfreeze down below */
652 		} else {
653 			didfreeze = 0;
654 		}
655 
656 		/*
657 		 * Cleanup trailing chains.  Do not reorder chains (for now).
658 		 * XXX might remove more than we intended.
659 		 */
660 		while (i > 0) {
661 			if (cluster->array[i - 1].chain)
662 				break;
663 			--i;
664 		}
665 		cluster->nchains = i;
666 
667 		/*
668 		 * If the PMP has no elements remaining we can destroy it.
669 		 * (this will transition management threads from frozen->exit).
670 		 */
671 		if (cluster->nchains == 0) {
672 			kprintf("unmount hmp %p last ref to PMP=%p\n",
673 				hmp, pmp);
674 			hammer2_pfsfree(pmp);
675 			goto again;
676 		}
677 
678 		/*
679 		 * If elements still remain we need to set the REMASTER
680 		 * flag and unfreeze it.
681 		 */
682 		if (didfreeze) {
683 			for (i = 0; i < iroot->cluster.nchains; ++i) {
684 				hammer2_thr_remaster(&pmp->sync_thrs[i]);
685 				hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
686 				for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
687 					hammer2_thr_remaster(
688 						&pmp->xop_groups[j].thrs[i]);
689 					hammer2_thr_unfreeze(
690 						&pmp->xop_groups[j].thrs[i]);
691 				}
692 			}
693 		}
694 	}
695 }
696 
697 /*
698  * Mount or remount HAMMER2 fileystem from physical media
699  *
700  *	mountroot
701  *		mp		mount point structure
702  *		path		NULL
703  *		data		<unused>
704  *		cred		<unused>
705  *
706  *	mount
707  *		mp		mount point structure
708  *		path		path to mount point
709  *		data		pointer to argument structure in user space
710  *			volume	volume path (device@LABEL form)
711  *			hflags	user mount flags
712  *		cred		user credentials
713  *
714  * RETURNS:	0	Success
715  *		!0	error number
716  */
717 static
718 int
719 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
720 		  struct ucred *cred)
721 {
722 	struct hammer2_mount_info info;
723 	hammer2_pfs_t *pmp;
724 	hammer2_pfs_t *spmp;
725 	hammer2_dev_t *hmp;
726 	hammer2_key_t key_next;
727 	hammer2_key_t key_dummy;
728 	hammer2_key_t lhc;
729 	struct vnode *devvp;
730 	struct nlookupdata nd;
731 	hammer2_chain_t *parent;
732 	hammer2_cluster_t *cluster;
733 	hammer2_cluster_t *cparent;
734 	const hammer2_inode_data_t *ripdata;
735 	hammer2_blockref_t bref;
736 	struct file *fp;
737 	char devstr[MNAMELEN];
738 	size_t size;
739 	size_t done;
740 	char *dev;
741 	char *label;
742 	int ronly = 1;
743 	int error;
744 	int cache_index;
745 	int i;
746 
747 	hmp = NULL;
748 	pmp = NULL;
749 	dev = NULL;
750 	label = NULL;
751 	devvp = NULL;
752 	cache_index = -1;
753 
754 	kprintf("hammer2_mount\n");
755 
756 	if (path == NULL) {
757 		/*
758 		 * Root mount
759 		 */
760 		bzero(&info, sizeof(info));
761 		info.cluster_fd = -1;
762 		return (EOPNOTSUPP);
763 	} else {
764 		/*
765 		 * Non-root mount or updating a mount
766 		 */
767 		error = copyin(data, &info, sizeof(info));
768 		if (error)
769 			return (error);
770 
771 		error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
772 		if (error)
773 			return (error);
774 
775 		/* Extract device and label */
776 		dev = devstr;
777 		label = strchr(devstr, '@');
778 		if (label == NULL ||
779 		    ((label + 1) - dev) > done) {
780 			return (EINVAL);
781 		}
782 		*label = '\0';
783 		label++;
784 		if (*label == '\0')
785 			return (EINVAL);
786 
787 		if (mp->mnt_flag & MNT_UPDATE) {
788 			/*
789 			 * Update mount.  Note that pmp->iroot->cluster is
790 			 * an inode-embedded cluster and thus cannot be
791 			 * directly locked.
792 			 *
793 			 * XXX HAMMER2 needs to implement NFS export via
794 			 *     mountctl.
795 			 */
796 			pmp = MPTOPMP(mp);
797 			cluster = &pmp->iroot->cluster;
798 			for (i = 0; i < cluster->nchains; ++i) {
799 				if (cluster->array[i].chain == NULL)
800 					continue;
801 				hmp = cluster->array[i].chain->hmp;
802 				devvp = hmp->devvp;
803 				error = hammer2_remount(hmp, mp, path,
804 							devvp, cred);
805 				if (error)
806 					break;
807 			}
808 			/*hammer2_inode_install_hidden(pmp);*/
809 
810 			return error;
811 		}
812 	}
813 
814 	/*
815 	 * HMP device mount
816 	 *
817 	 * Lookup name and verify it refers to a block device.
818 	 */
819 	error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
820 	if (error == 0)
821 		error = nlookup(&nd);
822 	if (error == 0)
823 		error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
824 	nlookup_done(&nd);
825 
826 	if (error == 0) {
827 		if (vn_isdisk(devvp, &error))
828 			error = vfs_mountedon(devvp);
829 	}
830 
831 	/*
832 	 * Determine if the device has already been mounted.  After this
833 	 * check hmp will be non-NULL if we are doing the second or more
834 	 * hammer2 mounts from the same device.
835 	 */
836 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
837 	TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
838 		if (hmp->devvp == devvp)
839 			break;
840 	}
841 
842 	/*
843 	 * Open the device if this isn't a secondary mount and construct
844 	 * the H2 device mount (hmp).
845 	 */
846 	if (hmp == NULL) {
847 		hammer2_chain_t *schain;
848 		hammer2_xid_t xid;
849 
850 		if (error == 0 && vcount(devvp) > 0)
851 			error = EBUSY;
852 
853 		/*
854 		 * Now open the device
855 		 */
856 		if (error == 0) {
857 			ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
858 			vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
859 			error = vinvalbuf(devvp, V_SAVE, 0, 0);
860 			if (error == 0) {
861 				error = VOP_OPEN(devvp,
862 						 ronly ? FREAD : FREAD | FWRITE,
863 						 FSCRED, NULL);
864 			}
865 			vn_unlock(devvp);
866 		}
867 		if (error && devvp) {
868 			vrele(devvp);
869 			devvp = NULL;
870 		}
871 		if (error) {
872 			lockmgr(&hammer2_mntlk, LK_RELEASE);
873 			return error;
874 		}
875 		hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
876 		ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
877 		hmp->ronly = ronly;
878 		hmp->devvp = devvp;
879 		kmalloc_create(&hmp->mchain, "HAMMER2-chains");
880 		TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
881 		RB_INIT(&hmp->iotree);
882 		spin_init(&hmp->io_spin, "hm2mount_io");
883 		spin_init(&hmp->list_spin, "hm2mount_list");
884 		TAILQ_INIT(&hmp->flushq);
885 
886 		lockinit(&hmp->vollk, "h2vol", 0, 0);
887 
888 		/*
889 		 * vchain setup. vchain.data is embedded.
890 		 * vchain.refs is initialized and will never drop to 0.
891 		 *
892 		 * NOTE! voldata is not yet loaded.
893 		 */
894 		hmp->vchain.hmp = hmp;
895 		hmp->vchain.refs = 1;
896 		hmp->vchain.data = (void *)&hmp->voldata;
897 		hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
898 		hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
899 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
900 
901 		hammer2_chain_core_init(&hmp->vchain);
902 		/* hmp->vchain.u.xxx is left NULL */
903 
904 		/*
905 		 * fchain setup.  fchain.data is embedded.
906 		 * fchain.refs is initialized and will never drop to 0.
907 		 *
908 		 * The data is not used but needs to be initialized to
909 		 * pass assertion muster.  We use this chain primarily
910 		 * as a placeholder for the freemap's top-level RBTREE
911 		 * so it does not interfere with the volume's topology
912 		 * RBTREE.
913 		 */
914 		hmp->fchain.hmp = hmp;
915 		hmp->fchain.refs = 1;
916 		hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
917 		hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
918 		hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
919 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
920 		hmp->fchain.bref.methods =
921 			HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
922 			HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
923 
924 		hammer2_chain_core_init(&hmp->fchain);
925 		/* hmp->fchain.u.xxx is left NULL */
926 
927 		/*
928 		 * Install the volume header and initialize fields from
929 		 * voldata.
930 		 */
931 		error = hammer2_install_volume_header(hmp);
932 		if (error) {
933 			hammer2_unmount_helper(mp, NULL, hmp);
934 			lockmgr(&hammer2_mntlk, LK_RELEASE);
935 			hammer2_vfs_unmount(mp, MNT_FORCE);
936 			return error;
937 		}
938 
939 		/*
940 		 * Really important to get these right or flush will get
941 		 * confused.
942 		 */
943 		hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0);
944 		kprintf("alloc spmp %p tid %016jx\n",
945 			hmp->spmp, hmp->voldata.mirror_tid);
946 		spmp = hmp->spmp;
947 
948 		/*
949 		 * Dummy-up vchain and fchain's modify_tid.  mirror_tid
950 		 * is inherited from the volume header.
951 		 */
952 		xid = 0;
953 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
954 		hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
955 		hmp->vchain.pmp = spmp;
956 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
957 		hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
958 		hmp->fchain.pmp = spmp;
959 
960 		/*
961 		 * First locate the super-root inode, which is key 0
962 		 * relative to the volume header's blockset.
963 		 *
964 		 * Then locate the root inode by scanning the directory keyspace
965 		 * represented by the label.
966 		 */
967 		parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
968 		schain = hammer2_chain_lookup(&parent, &key_dummy,
969 				      HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
970 				      &cache_index, 0);
971 		hammer2_chain_lookup_done(parent);
972 		if (schain == NULL) {
973 			kprintf("hammer2_mount: invalid super-root\n");
974 			hammer2_unmount_helper(mp, NULL, hmp);
975 			lockmgr(&hammer2_mntlk, LK_RELEASE);
976 			hammer2_vfs_unmount(mp, MNT_FORCE);
977 			return EINVAL;
978 		}
979 		if (schain->error) {
980 			kprintf("hammer2_mount: error %s reading super-root\n",
981 				hammer2_error_str(schain->error));
982 			hammer2_chain_unlock(schain);
983 			hammer2_chain_drop(schain);
984 			schain = NULL;
985 			hammer2_unmount_helper(mp, NULL, hmp);
986 			lockmgr(&hammer2_mntlk, LK_RELEASE);
987 			hammer2_vfs_unmount(mp, MNT_FORCE);
988 			return EINVAL;
989 		}
990 
991 		/*
992 		 * The super-root always uses an inode_tid of 1 when
993 		 * creating PFSs.
994 		 */
995 		spmp->inode_tid = 1;
996 		spmp->modify_tid = schain->bref.modify_tid;
997 
998 		/*
999 		 * Sanity-check schain's pmp and finish initialization.
1000 		 * Any chain belonging to the super-root topology should
1001 		 * have a NULL pmp (not even set to spmp).
1002 		 */
1003 		ripdata = &hammer2_chain_rdata(schain)->ipdata;
1004 		KKASSERT(schain->pmp == NULL);
1005 		spmp->pfs_clid = ripdata->meta.pfs_clid;
1006 
1007 		/*
1008 		 * Replace the dummy spmp->iroot with a real one.  It's
1009 		 * easier to just do a wholesale replacement than to try
1010 		 * to update the chain and fixup the iroot fields.
1011 		 *
1012 		 * The returned inode is locked with the supplied cluster.
1013 		 */
1014 		cluster = hammer2_cluster_from_chain(schain);
1015 		hammer2_inode_drop(spmp->iroot);
1016 		spmp->iroot = NULL;
1017 		spmp->iroot = hammer2_inode_get(spmp, NULL, cluster);
1018 		spmp->spmp_hmp = hmp;
1019 		spmp->pfs_types[0] = ripdata->meta.pfs_type;
1020 		hammer2_inode_ref(spmp->iroot);
1021 		hammer2_inode_unlock(spmp->iroot, cluster);
1022 		schain = NULL;
1023 		/* leave spmp->iroot with one ref */
1024 
1025 		if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1026 			error = hammer2_recovery(hmp);
1027 			/* XXX do something with error */
1028 		}
1029 		hammer2_update_pmps(hmp);
1030 		hammer2_iocom_init(hmp);
1031 
1032 		/*
1033 		 * Ref the cluster management messaging descriptor.  The mount
1034 		 * program deals with the other end of the communications pipe.
1035 		 */
1036 		fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1037 		if (fp) {
1038 			hammer2_cluster_reconnect(hmp, fp);
1039 		} else {
1040 			kprintf("hammer2_mount: bad cluster_fd!\n");
1041 		}
1042 	} else {
1043 		spmp = hmp->spmp;
1044 	}
1045 
1046 	/*
1047 	 * Lookup the mount point under the media-localized super-root.
1048 	 * Scanning hammer2_pfslist doesn't help us because it represents
1049 	 * PFS cluster ids which can aggregate several named PFSs together.
1050 	 *
1051 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1052 	 * up later on.
1053 	 */
1054 	hammer2_inode_lock(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1055 	cparent = hammer2_inode_cluster(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1056 	lhc = hammer2_dirhash(label, strlen(label));
1057 	cluster = hammer2_cluster_lookup(cparent, &key_next,
1058 				      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1059 				      0);
1060 	while (cluster) {
1061 		if (hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE &&
1062 		    strcmp(label,
1063 		       hammer2_cluster_rdata(cluster)->ipdata.filename) == 0) {
1064 			break;
1065 		}
1066 		cluster = hammer2_cluster_next(cparent, cluster, &key_next,
1067 					    key_next,
1068 					    lhc + HAMMER2_DIRHASH_LOMASK, 0);
1069 	}
1070 	hammer2_inode_unlock(spmp->iroot, cparent);
1071 
1072 	/*
1073 	 * PFS could not be found?
1074 	 */
1075 	if (cluster == NULL) {
1076 		kprintf("hammer2_mount: PFS label not found\n");
1077 		hammer2_unmount_helper(mp, NULL, hmp);
1078 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1079 		hammer2_vfs_unmount(mp, MNT_FORCE);
1080 
1081 		return EINVAL;
1082 	}
1083 
1084 	/*
1085 	 * Acquire the pmp structure (it should have already been allocated
1086 	 * via hammer2_update_pmps() so do not pass cluster in to add to
1087 	 * available chains).
1088 	 *
1089 	 * Check if the cluster has already been mounted.  A cluster can
1090 	 * only be mounted once, use null mounts to mount additional copies.
1091 	 */
1092 	ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1093 	hammer2_cluster_bref(cluster, &bref);
1094 	pmp = hammer2_pfsalloc(NULL, ripdata, bref.modify_tid);
1095 	hammer2_cluster_unlock(cluster);
1096 	hammer2_cluster_drop(cluster);
1097 
1098 	if (pmp->mp) {
1099 		kprintf("hammer2_mount: PFS already mounted!\n");
1100 		hammer2_unmount_helper(mp, NULL, hmp);
1101 		lockmgr(&hammer2_mntlk, LK_RELEASE);
1102 		hammer2_vfs_unmount(mp, MNT_FORCE);
1103 
1104 		return EBUSY;
1105 	}
1106 
1107 	/*
1108 	 * Finish the mount
1109 	 */
1110         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1111 
1112         mp->mnt_flag = MNT_LOCAL;
1113         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1114         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1115 
1116         /*
1117          * required mount structure initializations
1118          */
1119         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1120         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1121 
1122         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1123         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1124 
1125         /*
1126          * Optional fields
1127          */
1128         mp->mnt_iosize_max = MAXPHYS;
1129 
1130 	/*
1131 	 * Connect up mount pointers.
1132 	 */
1133 	hammer2_mount_helper(mp, pmp);
1134 
1135         lockmgr(&hammer2_mntlk, LK_RELEASE);
1136 
1137 	/*
1138 	 * A mounted PFS needs a write thread for logical buffers and
1139 	 * a hidden directory for deletions of open files.  These features
1140 	 * are not used by unmounted PFSs.
1141 	 *
1142 	 * The logical file buffer bio write thread handles things like
1143 	 * physical block assignment and compression.
1144 	 */
1145 	pmp->wthread_destroy = 0;
1146 	lwkt_create(hammer2_write_thread, pmp,
1147 		    &pmp->wthread_td, NULL, 0, -1, "h2pfs-%s", label);
1148 
1149 	/*
1150 	 * With the cluster operational install ihidden.
1151 	 * (only applicable to pfs mounts, not applicable to spmp)
1152 	 */
1153 	hammer2_inode_install_hidden(pmp);
1154 
1155 	/*
1156 	 * Finish setup
1157 	 */
1158 	vfs_getnewfsid(mp);
1159 	vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1160 	vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1161 	vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1162 
1163 	copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
1164 	bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1165 	bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1166 	copyinstr(path, mp->mnt_stat.f_mntonname,
1167 		  sizeof(mp->mnt_stat.f_mntonname) - 1,
1168 		  &size);
1169 
1170 	/*
1171 	 * Initial statfs to prime mnt_stat.
1172 	 */
1173 	hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1174 
1175 	return 0;
1176 }
1177 
1178 /*
1179  * Scan PFSs under the super-root and create hammer2_pfs structures.
1180  */
1181 static
1182 void
1183 hammer2_update_pmps(hammer2_dev_t *hmp)
1184 {
1185 	const hammer2_inode_data_t *ripdata;
1186 	hammer2_cluster_t *cparent;
1187 	hammer2_cluster_t *cluster;
1188 	hammer2_blockref_t bref;
1189 	hammer2_pfs_t *spmp;
1190 	hammer2_pfs_t *pmp;
1191 	hammer2_key_t key_next;
1192 
1193 	/*
1194 	 * Lookup mount point under the media-localized super-root.
1195 	 *
1196 	 * cluster->pmp will incorrectly point to spmp and must be fixed
1197 	 * up later on.
1198 	 */
1199 	spmp = hmp->spmp;
1200 	hammer2_inode_lock(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1201 	cparent = hammer2_inode_cluster(spmp->iroot, HAMMER2_RESOLVE_ALWAYS);
1202 	cluster = hammer2_cluster_lookup(cparent, &key_next,
1203 					 HAMMER2_KEY_MIN,
1204 					 HAMMER2_KEY_MAX,
1205 					 0);
1206 	while (cluster) {
1207 		if (hammer2_cluster_type(cluster) != HAMMER2_BREF_TYPE_INODE)
1208 			continue;
1209 		ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
1210 		hammer2_cluster_bref(cluster, &bref);
1211 		kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1212 
1213 		pmp = hammer2_pfsalloc(cluster, ripdata, bref.modify_tid);
1214 		cluster = hammer2_cluster_next(cparent, cluster,
1215 					       &key_next,
1216 					       key_next,
1217 					       HAMMER2_KEY_MAX,
1218 					       0);
1219 	}
1220 	hammer2_inode_unlock(spmp->iroot, cparent);
1221 }
1222 
1223 static
1224 int
1225 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path,
1226 		struct vnode *devvp, struct ucred *cred)
1227 {
1228 	int error;
1229 
1230 	if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1231 		error = hammer2_recovery(hmp);
1232 	} else {
1233 		error = 0;
1234 	}
1235 	return error;
1236 }
1237 
1238 static
1239 int
1240 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1241 {
1242 	hammer2_pfs_t *pmp;
1243 	int flags;
1244 	int error = 0;
1245 
1246 	pmp = MPTOPMP(mp);
1247 
1248 	if (pmp == NULL)
1249 		return(0);
1250 
1251 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1252 
1253 	/*
1254 	 * If mount initialization proceeded far enough we must flush
1255 	 * its vnodes and sync the underlying mount points.  Three syncs
1256 	 * are required to fully flush the filesystem (freemap updates lag
1257 	 * by one flush, and one extra for safety).
1258 	 */
1259 	if (mntflags & MNT_FORCE)
1260 		flags = FORCECLOSE;
1261 	else
1262 		flags = 0;
1263 	if (pmp->iroot) {
1264 		error = vflush(mp, 0, flags);
1265 		if (error)
1266 			goto failed;
1267 		hammer2_vfs_sync(mp, MNT_WAIT);
1268 		hammer2_vfs_sync(mp, MNT_WAIT);
1269 		hammer2_vfs_sync(mp, MNT_WAIT);
1270 	}
1271 
1272 	if (pmp->wthread_td) {
1273 		hammer2_mtx_ex(&pmp->wthread_mtx);
1274 		pmp->wthread_destroy = 1;
1275 		wakeup(&pmp->wthread_bioq);
1276 		while (pmp->wthread_destroy != -1) {
1277 			mtxsleep(&pmp->wthread_destroy,
1278 				&pmp->wthread_mtx, 0,
1279 				"umount-sleep",	0);
1280 		}
1281 		hammer2_mtx_unlock(&pmp->wthread_mtx);
1282 		pmp->wthread_td = NULL;
1283 	}
1284 
1285 	/*
1286 	 * Cleanup our reference on ihidden.
1287 	 */
1288 	if (pmp->ihidden) {
1289 		hammer2_inode_drop(pmp->ihidden);
1290 		pmp->ihidden = NULL;
1291 	}
1292 	if (pmp->mp)
1293 		hammer2_unmount_helper(mp, pmp, NULL);
1294 
1295 	error = 0;
1296 failed:
1297 	lockmgr(&hammer2_mntlk, LK_RELEASE);
1298 
1299 	return (error);
1300 }
1301 
1302 /*
1303  * Mount helper, hook the system mount into our PFS.
1304  * The mount lock is held.
1305  *
1306  * We must bump the mount_count on related devices for any
1307  * mounted PFSs.
1308  */
1309 static
1310 void
1311 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1312 {
1313 	hammer2_cluster_t *cluster;
1314 	hammer2_chain_t *rchain;
1315 	int i;
1316 
1317         mp->mnt_data = (qaddr_t)pmp;
1318 	pmp->mp = mp;
1319 
1320 	/*
1321 	 * After pmp->mp is set we have to adjust hmp->mount_count.
1322 	 */
1323 	cluster = &pmp->iroot->cluster;
1324 	for (i = 0; i < cluster->nchains; ++i) {
1325 		rchain = cluster->array[i].chain;
1326 		if (rchain == NULL)
1327 			continue;
1328 		++rchain->hmp->mount_count;
1329 		kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1330 			rchain->hmp, rchain->hmp->mount_count);
1331 	}
1332 }
1333 
1334 /*
1335  * Mount helper, unhook the system mount from our PFS.
1336  * The mount lock is held.
1337  *
1338  * If hmp is supplied a mount responsible for being the first to open
1339  * the block device failed and the block device and all PFSs using the
1340  * block device must be cleaned up.
1341  *
1342  * If pmp is supplied multiple devices might be backing the PFS and each
1343  * must be disconnect.  This might not be the last PFS using some of the
1344  * underlying devices.  Also, we have to adjust our hmp->mount_count
1345  * accounting for the devices backing the pmp which is now undergoing an
1346  * unmount.
1347  */
1348 static
1349 void
1350 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1351 {
1352 	hammer2_cluster_t *cluster;
1353 	hammer2_chain_t *rchain;
1354 	struct vnode *devvp;
1355 	int dumpcnt;
1356 	int ronly = 0;
1357 	int i;
1358 
1359 	/*
1360 	 * If no device supplied this is a high-level unmount and we have to
1361 	 * to disconnect the mount, adjust mount_count, and locate devices
1362 	 * that might now have no mounts.
1363 	 */
1364 	if (pmp) {
1365 		KKASSERT(hmp == NULL);
1366 		KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1367 		pmp->mp = NULL;
1368 		mp->mnt_data = NULL;
1369 
1370 		/*
1371 		 * After pmp->mp is cleared we have to account for
1372 		 * mount_count.
1373 		 */
1374 		cluster = &pmp->iroot->cluster;
1375 		for (i = 0; i < cluster->nchains; ++i) {
1376 			rchain = cluster->array[i].chain;
1377 			if (rchain == NULL)
1378 				continue;
1379 			--rchain->hmp->mount_count;
1380 			kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1381 				rchain->hmp, rchain->hmp->mount_count);
1382 			/* scrapping hmp now may invalidate the pmp */
1383 		}
1384 again:
1385 		TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1386 			if (hmp->mount_count == 0) {
1387 				hammer2_unmount_helper(NULL, NULL, hmp);
1388 				goto again;
1389 			}
1390 		}
1391 		return;
1392 	}
1393 
1394 	/*
1395 	 * Try to terminate the block device.  We can't terminate it if
1396 	 * there are still PFSs referencing it.
1397 	 */
1398 	kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1399 		hmp, hmp->mount_count);
1400 	if (hmp->mount_count)
1401 		return;
1402 
1403 	hammer2_pfsfree_scan(hmp);
1404 	hammer2_dev_exlock(hmp);	/* XXX order */
1405 
1406 	/*
1407 	 * Cycle the volume data lock as a safety (probably not needed any
1408 	 * more).  To ensure everything is out we need to flush at least
1409 	 * three times.  (1) The running of the unlinkq can dirty the
1410 	 * filesystem, (2) A normal flush can dirty the freemap, and
1411 	 * (3) ensure that the freemap is fully synchronized.
1412 	 *
1413 	 * The next mount's recovery scan can clean everything up but we want
1414 	 * to leave the filesystem in a 100% clean state on a normal unmount.
1415 	 */
1416 #if 0
1417 	hammer2_voldata_lock(hmp);
1418 	hammer2_voldata_unlock(hmp);
1419 #endif
1420 	hammer2_iocom_uninit(hmp);
1421 
1422 	if ((hmp->vchain.flags | hmp->fchain.flags) &
1423 	    HAMMER2_CHAIN_FLUSH_MASK) {
1424 		kprintf("hammer2_unmount: chains left over "
1425 			"after final sync\n");
1426 		kprintf("    vchain %08x\n", hmp->vchain.flags);
1427 		kprintf("    fchain %08x\n", hmp->fchain.flags);
1428 
1429 		if (hammer2_debug & 0x0010)
1430 			Debugger("entered debugger");
1431 	}
1432 
1433 	KKASSERT(hmp->spmp == NULL);
1434 
1435 	/*
1436 	 * Finish up with the device vnode
1437 	 */
1438 	if ((devvp = hmp->devvp) != NULL) {
1439 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1440 		vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1441 		hmp->devvp = NULL;
1442 		VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1443 		vn_unlock(devvp);
1444 		vrele(devvp);
1445 		devvp = NULL;
1446 	}
1447 
1448 	/*
1449 	 * Clear vchain/fchain flags that might prevent final cleanup
1450 	 * of these chains.
1451 	 */
1452 	if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1453 		atomic_clear_int(&hmp->vchain.flags,
1454 				 HAMMER2_CHAIN_MODIFIED);
1455 		hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1456 		hammer2_chain_drop(&hmp->vchain);
1457 	}
1458 	if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1459 		atomic_clear_int(&hmp->vchain.flags,
1460 				 HAMMER2_CHAIN_UPDATE);
1461 		hammer2_chain_drop(&hmp->vchain);
1462 	}
1463 
1464 	if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1465 		atomic_clear_int(&hmp->fchain.flags,
1466 				 HAMMER2_CHAIN_MODIFIED);
1467 		hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1468 		hammer2_chain_drop(&hmp->fchain);
1469 	}
1470 	if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1471 		atomic_clear_int(&hmp->fchain.flags,
1472 				 HAMMER2_CHAIN_UPDATE);
1473 		hammer2_chain_drop(&hmp->fchain);
1474 	}
1475 
1476 	/*
1477 	 * Final drop of embedded freemap root chain to
1478 	 * clean up fchain.core (fchain structure is not
1479 	 * flagged ALLOCATED so it is cleaned out and then
1480 	 * left to rot).
1481 	 */
1482 	hammer2_chain_drop(&hmp->fchain);
1483 
1484 	/*
1485 	 * Final drop of embedded volume root chain to clean
1486 	 * up vchain.core (vchain structure is not flagged
1487 	 * ALLOCATED so it is cleaned out and then left to
1488 	 * rot).
1489 	 */
1490 	dumpcnt = 50;
1491 	hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1492 	dumpcnt = 50;
1493 	hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1494 	hammer2_dev_unlock(hmp);
1495 	hammer2_chain_drop(&hmp->vchain);
1496 
1497 	hammer2_io_cleanup(hmp, &hmp->iotree);
1498 	if (hmp->iofree_count) {
1499 		kprintf("io_cleanup: %d I/O's left hanging\n",
1500 			hmp->iofree_count);
1501 	}
1502 
1503 	TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1504 	kmalloc_destroy(&hmp->mchain);
1505 	kfree(hmp, M_HAMMER2);
1506 }
1507 
1508 static
1509 int
1510 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1511 	     ino_t ino, struct vnode **vpp)
1512 {
1513 	kprintf("hammer2_vget\n");
1514 	return (EOPNOTSUPP);
1515 }
1516 
1517 static
1518 int
1519 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1520 {
1521 	hammer2_pfs_t *pmp;
1522 	hammer2_cluster_t *cparent;
1523 	int error;
1524 	struct vnode *vp;
1525 
1526 	pmp = MPTOPMP(mp);
1527 	if (pmp->iroot == NULL) {
1528 		*vpp = NULL;
1529 		error = EINVAL;
1530 	} else {
1531 		hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_ALWAYS |
1532 					       HAMMER2_RESOLVE_SHARED);
1533 		cparent = hammer2_inode_cluster(pmp->iroot,
1534 						HAMMER2_RESOLVE_ALWAYS |
1535 					        HAMMER2_RESOLVE_SHARED);
1536 
1537 		/*
1538 		 * Initialize pmp->inode_tid and pmp->modify_tid on first access
1539 		 * to the root of mount that resolves good.
1540 		 * XXX probably not the best place for this.
1541 		 */
1542 		if (pmp->inode_tid == 0 &&
1543 		    cparent->error == 0 && cparent->focus) {
1544 			const hammer2_inode_data_t *ripdata;
1545 			hammer2_blockref_t bref;
1546 
1547 			ripdata = &hammer2_cluster_rdata(cparent)->ipdata;
1548 			hammer2_cluster_bref(cparent, &bref);
1549 			pmp->inode_tid = ripdata->meta.pfs_inum + 1;
1550 			pmp->modify_tid = bref.modify_tid;
1551 			pmp->iroot->meta = ripdata->meta;
1552 			hammer2_cluster_bref(cparent, &pmp->iroot->bref);
1553 			kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1554 				pmp->inode_tid, pmp->modify_tid);
1555 		}
1556 
1557 		vp = hammer2_igetv(pmp->iroot, cparent, &error);
1558 		hammer2_inode_unlock(pmp->iroot, cparent);
1559 		*vpp = vp;
1560 		if (vp == NULL)
1561 			kprintf("vnodefail\n");
1562 	}
1563 
1564 	return (error);
1565 }
1566 
1567 /*
1568  * Filesystem status
1569  *
1570  * XXX incorporate ipdata->meta.inode_quota and data_quota
1571  */
1572 static
1573 int
1574 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1575 {
1576 	hammer2_pfs_t *pmp;
1577 	hammer2_dev_t *hmp;
1578 	hammer2_blockref_t bref;
1579 
1580 	pmp = MPTOPMP(mp);
1581 	KKASSERT(pmp->iroot->cluster.nchains >= 1);
1582 	hmp = pmp->iroot->cluster.focus->hmp;	/* iroot retains focus */
1583 	bref = pmp->iroot->cluster.focus->bref;	/* no lock */
1584 
1585 	mp->mnt_stat.f_files = bref.inode_count;
1586 	mp->mnt_stat.f_ffree = 0;
1587 	mp->mnt_stat.f_blocks = (bref.data_count +
1588 				 hmp->voldata.allocator_free) /
1589 				mp->mnt_vstat.f_bsize;
1590 	mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free /
1591 				mp->mnt_vstat.f_bsize;
1592 	mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1593 
1594 	*sbp = mp->mnt_stat;
1595 	return (0);
1596 }
1597 
1598 static
1599 int
1600 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1601 {
1602 	hammer2_pfs_t *pmp;
1603 	hammer2_dev_t *hmp;
1604 	hammer2_blockref_t bref;
1605 
1606 	pmp = MPTOPMP(mp);
1607 	KKASSERT(pmp->iroot->cluster.nchains >= 1);
1608 	hmp = pmp->iroot->cluster.focus->hmp;	/* iroot retains focus */
1609 	bref = pmp->iroot->cluster.focus->bref;	/* no lock */
1610 
1611 	mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1612 	mp->mnt_vstat.f_files = bref.inode_count;
1613 	mp->mnt_vstat.f_ffree = 0;
1614 	mp->mnt_vstat.f_blocks = (bref.data_count +
1615 				 hmp->voldata.allocator_free) /
1616 				mp->mnt_vstat.f_bsize;
1617 	mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1618 				mp->mnt_vstat.f_bsize;
1619 	mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1620 
1621 	*sbp = mp->mnt_vstat;
1622 	return (0);
1623 }
1624 
1625 /*
1626  * Mount-time recovery (RW mounts)
1627  *
1628  * Updates to the free block table are allowed to lag flushes by one
1629  * transaction.  In case of a crash, then on a fresh mount we must do an
1630  * incremental scan of the last committed transaction id and make sure that
1631  * all related blocks have been marked allocated.
1632  *
1633  * The super-root topology and each PFS has its own transaction id domain,
1634  * so we must track PFS boundary transitions.
1635  */
1636 struct hammer2_recovery_elm {
1637 	TAILQ_ENTRY(hammer2_recovery_elm) entry;
1638 	hammer2_chain_t *chain;
1639 	hammer2_tid_t sync_tid;
1640 };
1641 
1642 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1643 
1644 struct hammer2_recovery_info {
1645 	struct hammer2_recovery_list list;
1646 	int	depth;
1647 };
1648 
1649 static int hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_dev_t *hmp,
1650 			hammer2_chain_t *parent,
1651 			struct hammer2_recovery_info *info,
1652 			hammer2_tid_t sync_tid);
1653 
1654 #define HAMMER2_RECOVERY_MAXDEPTH	10
1655 
1656 static
1657 int
1658 hammer2_recovery(hammer2_dev_t *hmp)
1659 {
1660 	hammer2_trans_t trans;
1661 	struct hammer2_recovery_info info;
1662 	struct hammer2_recovery_elm *elm;
1663 	hammer2_chain_t *parent;
1664 	hammer2_tid_t sync_tid;
1665 	hammer2_tid_t mirror_tid;
1666 	int error;
1667 	int cumulative_error = 0;
1668 
1669 	hammer2_trans_init(&trans, hmp->spmp, 0);
1670 
1671 	sync_tid = hmp->voldata.freemap_tid;
1672 	mirror_tid = hmp->voldata.mirror_tid;
1673 
1674 	kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1675 	if (sync_tid >= mirror_tid) {
1676 		kprintf(" no recovery needed\n");
1677 	} else {
1678 		kprintf(" freemap recovery %016jx-%016jx\n",
1679 			sync_tid + 1, mirror_tid);
1680 	}
1681 
1682 	TAILQ_INIT(&info.list);
1683 	info.depth = 0;
1684 	parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1685 	cumulative_error = hammer2_recovery_scan(&trans, hmp, parent,
1686 						 &info, sync_tid);
1687 	hammer2_chain_lookup_done(parent);
1688 
1689 	while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1690 		TAILQ_REMOVE(&info.list, elm, entry);
1691 		parent = elm->chain;
1692 		sync_tid = elm->sync_tid;
1693 		kfree(elm, M_HAMMER2);
1694 
1695 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1696 		error = hammer2_recovery_scan(&trans, hmp, parent,
1697 					      &info,
1698 					      hmp->voldata.freemap_tid);
1699 		hammer2_chain_unlock(parent);
1700 		hammer2_chain_drop(parent);	/* drop elm->chain ref */
1701 		if (error)
1702 			cumulative_error = error;
1703 	}
1704 	hammer2_trans_done(&trans);
1705 
1706 	return cumulative_error;
1707 }
1708 
1709 static
1710 int
1711 hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_dev_t *hmp,
1712 		      hammer2_chain_t *parent,
1713 		      struct hammer2_recovery_info *info,
1714 		      hammer2_tid_t sync_tid)
1715 {
1716 	const hammer2_inode_data_t *ripdata;
1717 	hammer2_chain_t *chain;
1718 	int cache_index;
1719 	int cumulative_error = 0;
1720 	int error;
1721 
1722 	/*
1723 	 * Adjust freemap to ensure that the block(s) are marked allocated.
1724 	 */
1725 	if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1726 		hammer2_freemap_adjust(trans, hmp, &parent->bref,
1727 				       HAMMER2_FREEMAP_DORECOVER);
1728 	}
1729 
1730 	/*
1731 	 * Check type for recursive scan
1732 	 */
1733 	switch(parent->bref.type) {
1734 	case HAMMER2_BREF_TYPE_VOLUME:
1735 		/* data already instantiated */
1736 		break;
1737 	case HAMMER2_BREF_TYPE_INODE:
1738 		/*
1739 		 * Must instantiate data for DIRECTDATA test and also
1740 		 * for recursion.
1741 		 */
1742 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1743 		ripdata = &hammer2_chain_rdata(parent)->ipdata;
1744 		if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1745 			/* not applicable to recovery scan */
1746 			hammer2_chain_unlock(parent);
1747 			return 0;
1748 		}
1749 		hammer2_chain_unlock(parent);
1750 		break;
1751 	case HAMMER2_BREF_TYPE_INDIRECT:
1752 		/*
1753 		 * Must instantiate data for recursion
1754 		 */
1755 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1756 		hammer2_chain_unlock(parent);
1757 		break;
1758 	case HAMMER2_BREF_TYPE_DATA:
1759 	case HAMMER2_BREF_TYPE_FREEMAP:
1760 	case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1761 	case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1762 		/* not applicable to recovery scan */
1763 		return 0;
1764 		break;
1765 	default:
1766 		return EDOM;
1767 	}
1768 
1769 	/*
1770 	 * Defer operation if depth limit reached or if we are crossing a
1771 	 * PFS boundary.
1772 	 */
1773 	if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1774 		struct hammer2_recovery_elm *elm;
1775 
1776 		elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1777 		elm->chain = parent;
1778 		elm->sync_tid = sync_tid;
1779 		hammer2_chain_ref(parent);
1780 		TAILQ_INSERT_TAIL(&info->list, elm, entry);
1781 		/* unlocked by caller */
1782 
1783 		return(0);
1784 	}
1785 
1786 
1787 	/*
1788 	 * Recursive scan of the last flushed transaction only.  We are
1789 	 * doing this without pmp assignments so don't leave the chains
1790 	 * hanging around after we are done with them.
1791 	 */
1792 	cache_index = 0;
1793 	chain = hammer2_chain_scan(parent, NULL, &cache_index,
1794 				   HAMMER2_LOOKUP_NODATA);
1795 	while (chain) {
1796 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
1797 		if (chain->bref.mirror_tid > sync_tid) {
1798 			++info->depth;
1799 			error = hammer2_recovery_scan(trans, hmp, chain,
1800 						      info, sync_tid);
1801 			--info->depth;
1802 			if (error)
1803 				cumulative_error = error;
1804 		}
1805 
1806 		/*
1807 		 * Flush the recovery at the PFS boundary to stage it for
1808 		 * the final flush of the super-root topology.
1809 		 */
1810 		if ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
1811 		    (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
1812 			hammer2_flush(trans, chain, 1);
1813 		}
1814 		chain = hammer2_chain_scan(parent, chain, &cache_index,
1815 					   HAMMER2_LOOKUP_NODATA);
1816 	}
1817 
1818 	return cumulative_error;
1819 }
1820 
1821 /*
1822  * Sync a mount point; this is called on a per-mount basis from the
1823  * filesystem syncer process periodically and whenever a user issues
1824  * a sync.
1825  */
1826 int
1827 hammer2_vfs_sync(struct mount *mp, int waitfor)
1828 {
1829 	struct hammer2_sync_info info;
1830 	hammer2_inode_t *iroot;
1831 	hammer2_chain_t *chain;
1832 	hammer2_chain_t *parent;
1833 	hammer2_pfs_t *pmp;
1834 	hammer2_dev_t *hmp;
1835 	int flags;
1836 	int error;
1837 	int total_error;
1838 	int i;
1839 	int j;
1840 
1841 	pmp = MPTOPMP(mp);
1842 	iroot = pmp->iroot;
1843 	KKASSERT(iroot);
1844 	KKASSERT(iroot->pmp == pmp);
1845 
1846 	/*
1847 	 * We can't acquire locks on existing vnodes while in a transaction
1848 	 * without risking a deadlock.  This assumes that vfsync() can be
1849 	 * called without the vnode locked (which it can in DragonFly).
1850 	 * Otherwise we'd have to implement a multi-pass or flag the lock
1851 	 * failures and retry.
1852 	 *
1853 	 * The reclamation code interlocks with the sync list's token
1854 	 * (by removing the vnode from the scan list) before unlocking
1855 	 * the inode, giving us time to ref the inode.
1856 	 */
1857 	/*flags = VMSC_GETVP;*/
1858 	flags = 0;
1859 	if (waitfor & MNT_LAZY)
1860 		flags |= VMSC_ONEPASS;
1861 
1862 #if 0
1863 	/*
1864 	 * Preflush the vnodes using a normal transaction before interlocking
1865 	 * with a flush transaction.
1866 	 */
1867 	hammer2_trans_init(&info.trans, pmp, 0);
1868 	info.error = 0;
1869 	info.waitfor = MNT_NOWAIT;
1870 	vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1871 	hammer2_trans_done(&info.trans);
1872 #endif
1873 
1874 	/*
1875 	 * Start our flush transaction.  This does not return until all
1876 	 * concurrent transactions have completed and will prevent any
1877 	 * new transactions from running concurrently, except for the
1878 	 * buffer cache transactions.
1879 	 *
1880 	 * For efficiency do an async pass before making sure with a
1881 	 * synchronous pass on all related buffer cache buffers.  It
1882 	 * should theoretically not be possible for any new file buffers
1883 	 * to be instantiated during this sequence.
1884 	 */
1885 	hammer2_trans_init(&info.trans, pmp, HAMMER2_TRANS_ISFLUSH |
1886 					     HAMMER2_TRANS_PREFLUSH);
1887 	hammer2_run_unlinkq(&info.trans, pmp);
1888 
1889 	info.error = 0;
1890 	info.waitfor = MNT_NOWAIT;
1891 	vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1892 	info.waitfor = MNT_WAIT;
1893 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
1894 
1895 	/*
1896 	 * Clear PREFLUSH.  This prevents (or asserts on) any new logical
1897 	 * buffer cache flushes which occur during the flush.  Device buffers
1898 	 * are not affected.
1899 	 */
1900 	hammer2_bioq_sync(info.trans.pmp);
1901 	atomic_clear_int(&info.trans.flags, HAMMER2_TRANS_PREFLUSH);
1902 
1903 	total_error = 0;
1904 
1905 	/*
1906 	 * Flush all nodes to synchronize the PFSROOT subtopology to the media.
1907 	 *
1908 	 * Note that this flush will not be visible on crash recovery until
1909 	 * we flush the super-root topology in the next loop.
1910 	 */
1911 	for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
1912 		chain = iroot->cluster.array[i].chain;
1913 		if (chain == NULL)
1914 			continue;
1915 
1916 		hammer2_chain_ref(chain);
1917 		hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
1918 		if (chain->flags & HAMMER2_CHAIN_FLUSH_MASK) {
1919 			hammer2_flush(&info.trans, chain, 1);
1920 			parent = chain->parent;
1921 			KKASSERT(chain->pmp != parent->pmp);
1922 			hammer2_chain_setflush(&info.trans, parent);
1923 		}
1924 		hammer2_chain_unlock(chain);
1925 		hammer2_chain_drop(chain);
1926 	}
1927 	hammer2_trans_done(&info.trans);
1928 
1929 	/*
1930 	 * Flush all volume roots to synchronize PFS flushes with the
1931 	 * storage media volume header.  This will flush the freemap and
1932 	 * the superroot topology but stops when it reaches a PFSROOT
1933 	 * (which we already flushed above).
1934 	 *
1935 	 * This is the last step which connects the volume root to the
1936 	 * PFSROOT dirs flushed above.
1937 	 *
1938 	 * Each spmp (representing the hmp's super-root) requires its own
1939 	 * transaction.
1940 	 */
1941 	for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
1942 		hammer2_chain_t *tmp;
1943 
1944 		chain = iroot->cluster.array[i].chain;
1945 		if (chain == NULL)
1946 			continue;
1947 
1948 		hmp = chain->hmp;
1949 
1950 		/*
1951 		 * We only have to flush each hmp once
1952 		 */
1953 		for (j = i - 1; j >= 0; --j) {
1954 			if ((tmp = iroot->cluster.array[j].chain) != NULL) {
1955 				if (tmp->hmp == hmp)
1956 					break;
1957 			}
1958 		}
1959 		if (j >= 0)
1960 			continue;
1961 
1962 		/*
1963 		 * spmp transaction.  The super-root is never directly
1964 		 * mounted so there shouldn't be any vnodes, let alone any
1965 		 * dirty vnodes associated with it.
1966 		 */
1967 		hammer2_trans_init(&info.trans, hmp->spmp,
1968 				   HAMMER2_TRANS_ISFLUSH);
1969 
1970 		/*
1971 		 * Media mounts have two 'roots', vchain for the topology
1972 		 * and fchain for the free block table.  Flush both.
1973 		 *
1974 		 * Note that the topology and free block table are handled
1975 		 * independently, so the free block table can wind up being
1976 		 * ahead of the topology.  We depend on the bulk free scan
1977 		 * code to deal with any loose ends.
1978 		 */
1979 		hammer2_chain_ref(&hmp->vchain);
1980 		hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1981 		hammer2_chain_ref(&hmp->fchain);
1982 		hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1983 		if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1984 			/*
1985 			 * This will also modify vchain as a side effect,
1986 			 * mark vchain as modified now.
1987 			 */
1988 			hammer2_voldata_modify(hmp);
1989 			chain = &hmp->fchain;
1990 			hammer2_flush(&info.trans, chain, 1);
1991 			KKASSERT(chain == &hmp->fchain);
1992 		}
1993 		hammer2_chain_unlock(&hmp->fchain);
1994 		hammer2_chain_unlock(&hmp->vchain);
1995 		hammer2_chain_drop(&hmp->fchain);
1996 		/* vchain dropped down below */
1997 
1998 		hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1999 		if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2000 			chain = &hmp->vchain;
2001 			hammer2_flush(&info.trans, chain, 1);
2002 			KKASSERT(chain == &hmp->vchain);
2003 		}
2004 		hammer2_chain_unlock(&hmp->vchain);
2005 		hammer2_chain_drop(&hmp->vchain);
2006 
2007 		error = 0;
2008 
2009 		/*
2010 		 * We can't safely flush the volume header until we have
2011 		 * flushed any device buffers which have built up.
2012 		 *
2013 		 * XXX this isn't being incremental
2014 		 */
2015 		vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
2016 		error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
2017 		vn_unlock(hmp->devvp);
2018 
2019 		/*
2020 		 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
2021 		 * volume header needs synchronization via hmp->volsync.
2022 		 *
2023 		 * XXX synchronize the flag & data with only this flush XXX
2024 		 */
2025 		if (error == 0 &&
2026 		    (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
2027 			struct buf *bp;
2028 
2029 			/*
2030 			 * Synchronize the disk before flushing the volume
2031 			 * header.
2032 			 */
2033 			bp = getpbuf(NULL);
2034 			bp->b_bio1.bio_offset = 0;
2035 			bp->b_bufsize = 0;
2036 			bp->b_bcount = 0;
2037 			bp->b_cmd = BUF_CMD_FLUSH;
2038 			bp->b_bio1.bio_done = biodone_sync;
2039 			bp->b_bio1.bio_flags |= BIO_SYNC;
2040 			vn_strategy(hmp->devvp, &bp->b_bio1);
2041 			biowait(&bp->b_bio1, "h2vol");
2042 			relpbuf(bp, NULL);
2043 
2044 			/*
2045 			 * Then we can safely flush the version of the
2046 			 * volume header synchronized by the flush code.
2047 			 */
2048 			i = hmp->volhdrno + 1;
2049 			if (i >= HAMMER2_NUM_VOLHDRS)
2050 				i = 0;
2051 			if (i * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
2052 			    hmp->volsync.volu_size) {
2053 				i = 0;
2054 			}
2055 			kprintf("sync volhdr %d %jd\n",
2056 				i, (intmax_t)hmp->volsync.volu_size);
2057 			bp = getblk(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2058 				    HAMMER2_PBUFSIZE, 0, 0);
2059 			atomic_clear_int(&hmp->vchain.flags,
2060 					 HAMMER2_CHAIN_VOLUMESYNC);
2061 			bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
2062 			bawrite(bp);
2063 			hmp->volhdrno = i;
2064 		}
2065 		if (error)
2066 			total_error = error;
2067 
2068 		hammer2_trans_done(&info.trans);	/* spmp trans */
2069 	}
2070 	return (total_error);
2071 }
2072 
2073 /*
2074  * Sync passes.
2075  */
2076 static int
2077 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2078 {
2079 	struct hammer2_sync_info *info = data;
2080 	hammer2_inode_t *ip;
2081 	int error;
2082 
2083 	/*
2084 	 * Degenerate cases.  Note that ip == NULL typically means the
2085 	 * syncer vnode itself and we don't want to vclrisdirty() in that
2086 	 * situation.
2087 	 */
2088 	ip = VTOI(vp);
2089 	if (ip == NULL) {
2090 		return(0);
2091 	}
2092 	if (vp->v_type == VNON || vp->v_type == VBAD) {
2093 		vclrisdirty(vp);
2094 		return(0);
2095 	}
2096 
2097 	/*
2098 	 * VOP_FSYNC will start a new transaction so replicate some code
2099 	 * here to do it inline (see hammer2_vop_fsync()).
2100 	 *
2101 	 * WARNING: The vfsync interacts with the buffer cache and might
2102 	 *          block, we can't hold the inode lock at that time.
2103 	 *	    However, we MUST ref ip before blocking to ensure that
2104 	 *	    it isn't ripped out from under us (since we do not
2105 	 *	    hold a lock on the vnode).
2106 	 */
2107 	hammer2_inode_ref(ip);
2108 	if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2109 	    !RB_EMPTY(&vp->v_rbdirty_tree)) {
2110 		vfsync(vp, info->waitfor, 1, NULL, NULL);
2111 		hammer2_inode_fsync(&info->trans, ip, NULL);
2112 	}
2113 	if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2114 	    RB_EMPTY(&vp->v_rbdirty_tree)) {
2115 		vclrisdirty(vp);
2116 	}
2117 
2118 	hammer2_inode_drop(ip);
2119 #if 1
2120 	error = 0;
2121 	if (error)
2122 		info->error = error;
2123 #endif
2124 	return(0);
2125 }
2126 
2127 static
2128 int
2129 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2130 {
2131 	return (0);
2132 }
2133 
2134 static
2135 int
2136 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2137 	       struct fid *fhp, struct vnode **vpp)
2138 {
2139 	return (0);
2140 }
2141 
2142 static
2143 int
2144 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2145 		 int *exflagsp, struct ucred **credanonp)
2146 {
2147 	return (0);
2148 }
2149 
2150 /*
2151  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2152  * header into the HMP
2153  *
2154  * XXX read four volhdrs and use the one with the highest TID whos CRC
2155  *     matches.
2156  *
2157  * XXX check iCRCs.
2158  *
2159  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2160  *     nonexistant locations.
2161  *
2162  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2163  */
2164 static
2165 int
2166 hammer2_install_volume_header(hammer2_dev_t *hmp)
2167 {
2168 	hammer2_volume_data_t *vd;
2169 	struct buf *bp;
2170 	hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2171 	int error_reported;
2172 	int error;
2173 	int valid;
2174 	int i;
2175 
2176 	error_reported = 0;
2177 	error = 0;
2178 	valid = 0;
2179 	bp = NULL;
2180 
2181 	/*
2182 	 * There are up to 4 copies of the volume header (syncs iterate
2183 	 * between them so there is no single master).  We don't trust the
2184 	 * volu_size field so we don't know precisely how large the filesystem
2185 	 * is, so depend on the OS to return an error if we go beyond the
2186 	 * block device's EOF.
2187 	 */
2188 	for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2189 		error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2190 			      HAMMER2_VOLUME_BYTES, &bp);
2191 		if (error) {
2192 			brelse(bp);
2193 			bp = NULL;
2194 			continue;
2195 		}
2196 
2197 		vd = (struct hammer2_volume_data *) bp->b_data;
2198 		if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2199 		    (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2200 			brelse(bp);
2201 			bp = NULL;
2202 			continue;
2203 		}
2204 
2205 		if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2206 			/* XXX: Reversed-endianness filesystem */
2207 			kprintf("hammer2: reverse-endian filesystem detected");
2208 			brelse(bp);
2209 			bp = NULL;
2210 			continue;
2211 		}
2212 
2213 		crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2214 		crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2215 				      HAMMER2_VOLUME_ICRC0_SIZE);
2216 		bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2217 		bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2218 				       HAMMER2_VOLUME_ICRC1_SIZE);
2219 		if ((crc0 != crc) || (bcrc0 != bcrc)) {
2220 			kprintf("hammer2 volume header crc "
2221 				"mismatch copy #%d %08x/%08x\n",
2222 				i, crc0, crc);
2223 			error_reported = 1;
2224 			brelse(bp);
2225 			bp = NULL;
2226 			continue;
2227 		}
2228 		if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2229 			valid = 1;
2230 			hmp->voldata = *vd;
2231 			hmp->volhdrno = i;
2232 		}
2233 		brelse(bp);
2234 		bp = NULL;
2235 	}
2236 	if (valid) {
2237 		hmp->volsync = hmp->voldata;
2238 		error = 0;
2239 		if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2240 			kprintf("hammer2: using volume header #%d\n",
2241 				hmp->volhdrno);
2242 		}
2243 	} else {
2244 		error = EINVAL;
2245 		kprintf("hammer2: no valid volume headers found!\n");
2246 	}
2247 	return (error);
2248 }
2249 
2250 /*
2251  * This handles hysteresis on regular file flushes.  Because the BIOs are
2252  * routed to a thread it is possible for an excessive number to build up
2253  * and cause long front-end stalls long before the runningbuffspace limit
2254  * is hit, so we implement hammer2_flush_pipe to control the
2255  * hysteresis.
2256  *
2257  * This is a particular problem when compression is used.
2258  */
2259 void
2260 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2261 {
2262 	atomic_add_int(&pmp->count_lwinprog, 1);
2263 }
2264 
2265 void
2266 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2267 {
2268 	int lwinprog;
2269 
2270 	lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2271 	if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2272 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2273 		atomic_clear_int(&pmp->count_lwinprog,
2274 				 HAMMER2_LWINPROG_WAITING);
2275 		wakeup(&pmp->count_lwinprog);
2276 	}
2277 }
2278 
2279 void
2280 hammer2_lwinprog_wait(hammer2_pfs_t *pmp)
2281 {
2282 	int lwinprog;
2283 
2284 	for (;;) {
2285 		lwinprog = pmp->count_lwinprog;
2286 		cpu_ccfence();
2287 		if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2288 			break;
2289 		tsleep_interlock(&pmp->count_lwinprog, 0);
2290 		atomic_set_int(&pmp->count_lwinprog, HAMMER2_LWINPROG_WAITING);
2291 		lwinprog = pmp->count_lwinprog;
2292 		if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2293 			break;
2294 		tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2295 	}
2296 }
2297 
2298 /*
2299  * Manage excessive memory resource use for chain and related
2300  * structures.
2301  */
2302 void
2303 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2304 {
2305 	uint32_t waiting;
2306 	uint32_t count;
2307 	uint32_t limit;
2308 #if 0
2309 	static int zzticks;
2310 #endif
2311 
2312 	/*
2313 	 * Atomic check condition and wait.  Also do an early speedup of
2314 	 * the syncer to try to avoid hitting the wait.
2315 	 */
2316 	for (;;) {
2317 		waiting = pmp->inmem_dirty_chains;
2318 		cpu_ccfence();
2319 		count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2320 
2321 		limit = pmp->mp->mnt_nvnodelistsize / 10;
2322 		if (limit < hammer2_limit_dirty_chains)
2323 			limit = hammer2_limit_dirty_chains;
2324 		if (limit < 1000)
2325 			limit = 1000;
2326 
2327 #if 0
2328 		if ((int)(ticks - zzticks) > hz) {
2329 			zzticks = ticks;
2330 			kprintf("count %ld %ld\n", count, limit);
2331 		}
2332 #endif
2333 
2334 		/*
2335 		 * Block if there are too many dirty chains present, wait
2336 		 * for the flush to clean some out.
2337 		 */
2338 		if (count > limit) {
2339 			tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2340 			if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2341 					       waiting,
2342 				       waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2343 				speedup_syncer(pmp->mp);
2344 				tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2345 				       "chnmem", hz);
2346 			}
2347 			continue;	/* loop on success or fail */
2348 		}
2349 
2350 		/*
2351 		 * Try to start an early flush before we are forced to block.
2352 		 */
2353 		if (count > limit * 7 / 10)
2354 			speedup_syncer(pmp->mp);
2355 		break;
2356 	}
2357 }
2358 
2359 void
2360 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2361 {
2362 	if (pmp) {
2363 		atomic_add_int(&pmp->inmem_dirty_chains, 1);
2364 	}
2365 }
2366 
2367 void
2368 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2369 {
2370 	uint32_t waiting;
2371 
2372 	if (pmp == NULL)
2373 		return;
2374 
2375 	for (;;) {
2376 		waiting = pmp->inmem_dirty_chains;
2377 		cpu_ccfence();
2378 		if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2379 				       waiting,
2380 				       (waiting - 1) &
2381 					~HAMMER2_DIRTYCHAIN_WAITING)) {
2382 			break;
2383 		}
2384 	}
2385 
2386 	if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2387 		wakeup(&pmp->inmem_dirty_chains);
2388 }
2389 
2390 /*
2391  * Debugging
2392  */
2393 void
2394 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2395 {
2396 	hammer2_chain_t *scan;
2397 	hammer2_chain_t *parent;
2398 
2399 	--*countp;
2400 	if (*countp == 0) {
2401 		kprintf("%*.*s...\n", tab, tab, "");
2402 		return;
2403 	}
2404 	if (*countp < 0)
2405 		return;
2406 	kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2407 		tab, tab, "", pfx,
2408 		chain, chain->bref.type,
2409 		chain->bref.key, chain->bref.keybits,
2410 		chain->bref.mirror_tid);
2411 
2412 	kprintf("%*.*s      [%08x] (%s) refs=%d",
2413 		tab, tab, "",
2414 		chain->flags,
2415 		((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2416 		chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2417 		chain->refs);
2418 
2419 	parent = chain->parent;
2420 	if (parent)
2421 		kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2422 			tab, tab, "",
2423 			parent, parent->flags, parent->refs);
2424 	if (RB_EMPTY(&chain->core.rbtree)) {
2425 		kprintf("\n");
2426 	} else {
2427 		kprintf(" {\n");
2428 		RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2429 			hammer2_dump_chain(scan, tab + 4, countp, 'a');
2430 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2431 			kprintf("%*.*s}(%s)\n", tab, tab, "",
2432 				chain->data->ipdata.filename);
2433 		else
2434 			kprintf("%*.*s}\n", tab, tab, "");
2435 	}
2436 }
2437