xref: /dflybsd-src/sys/vfs/hammer2/hammer2_vfsops.c (revision 97157a3876f6bf4907a92b7c75f2260a85c5acdf)
1 /*-
2  * Copyright (c) 2011-2013 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 
62 #include "hammer2.h"
63 #include "hammer2_lz4.h"
64 
65 #include "zlib/hammer2_zlib.h"
66 
67 #define REPORT_REFS_ERRORS 1	/* XXX remove me */
68 
69 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
70 
71 struct hammer2_sync_info {
72 	hammer2_trans_t trans;
73 	int error;
74 	int waitfor;
75 };
76 
77 TAILQ_HEAD(hammer2_mntlist, hammer2_mount);
78 static struct hammer2_mntlist hammer2_mntlist;
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 long hammer2_limit_dirty_chains;
86 long hammer2_iod_file_read;
87 long hammer2_iod_meta_read;
88 long hammer2_iod_indr_read;
89 long hammer2_iod_fmap_read;
90 long hammer2_iod_volu_read;
91 long hammer2_iod_file_write;
92 long hammer2_iod_meta_write;
93 long hammer2_iod_indr_write;
94 long hammer2_iod_fmap_write;
95 long hammer2_iod_volu_write;
96 long hammer2_ioa_file_read;
97 long hammer2_ioa_meta_read;
98 long hammer2_ioa_indr_read;
99 long hammer2_ioa_fmap_read;
100 long hammer2_ioa_volu_read;
101 long hammer2_ioa_fmap_write;
102 long hammer2_ioa_file_write;
103 long hammer2_ioa_meta_write;
104 long hammer2_ioa_indr_write;
105 long hammer2_ioa_volu_write;
106 
107 MALLOC_DECLARE(C_BUFFER);
108 MALLOC_DEFINE(C_BUFFER, "compbuffer", "Buffer used for compression.");
109 
110 MALLOC_DECLARE(D_BUFFER);
111 MALLOC_DEFINE(D_BUFFER, "decompbuffer", "Buffer used for decompression.");
112 
113 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
114 
115 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
116 	   &hammer2_debug, 0, "");
117 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW,
118 	   &hammer2_cluster_enable, 0, "");
119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
120 	   &hammer2_hardlink_enable, 0, "");
121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
122 	   &hammer2_flush_pipe, 0, "");
123 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
124 	   &hammer2_limit_dirty_chains, 0, "");
125 
126 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
127 	   &hammer2_iod_file_read, 0, "");
128 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
129 	   &hammer2_iod_meta_read, 0, "");
130 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
131 	   &hammer2_iod_indr_read, 0, "");
132 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
133 	   &hammer2_iod_fmap_read, 0, "");
134 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
135 	   &hammer2_iod_volu_read, 0, "");
136 
137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
138 	   &hammer2_iod_file_write, 0, "");
139 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
140 	   &hammer2_iod_meta_write, 0, "");
141 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
142 	   &hammer2_iod_indr_write, 0, "");
143 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
144 	   &hammer2_iod_fmap_write, 0, "");
145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
146 	   &hammer2_iod_volu_write, 0, "");
147 
148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
149 	   &hammer2_ioa_file_read, 0, "");
150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
151 	   &hammer2_ioa_meta_read, 0, "");
152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
153 	   &hammer2_ioa_indr_read, 0, "");
154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
155 	   &hammer2_ioa_fmap_read, 0, "");
156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
157 	   &hammer2_ioa_volu_read, 0, "");
158 
159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
160 	   &hammer2_ioa_file_write, 0, "");
161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
162 	   &hammer2_ioa_meta_write, 0, "");
163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
164 	   &hammer2_ioa_indr_write, 0, "");
165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
166 	   &hammer2_ioa_fmap_write, 0, "");
167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
168 	   &hammer2_ioa_volu_write, 0, "");
169 
170 static int hammer2_vfs_init(struct vfsconf *conf);
171 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
172 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
173 				struct ucred *cred);
174 static int hammer2_remount(hammer2_mount_t *, struct mount *, char *,
175 				struct vnode *, struct ucred *);
176 static int hammer2_recovery(hammer2_mount_t *hmp);
177 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
178 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
179 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
180 				struct ucred *cred);
181 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
182 				struct ucred *cred);
183 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
184 				ino_t ino, struct vnode **vpp);
185 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
186 				struct fid *fhp, struct vnode **vpp);
187 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
188 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
189 				int *exflagsp, struct ucred **credanonp);
190 
191 static int hammer2_install_volume_header(hammer2_mount_t *hmp);
192 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
193 
194 static void hammer2_write_thread(void *arg);
195 
196 static void hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp);
197 static void hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp);
198 
199 /*
200  * Functions for compression in threads,
201  * from hammer2_vnops.c
202  */
203 static void hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
204 				hammer2_inode_t *ip,
205 				hammer2_inode_data_t *ipdata,
206 				hammer2_chain_t **parentp,
207 				hammer2_key_t lbase, int ioflag, int pblksize,
208 				int *errorp);
209 static void hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
210 				hammer2_inode_t *ip,
211 				hammer2_inode_data_t *ipdata,
212 				hammer2_chain_t **parentp,
213 				hammer2_key_t lbase, int ioflag,
214 				int pblksize, int *errorp, int comp_algo);
215 static void hammer2_zero_check_and_write(struct buf *bp,
216 				hammer2_trans_t *trans, hammer2_inode_t *ip,
217 				hammer2_inode_data_t *ipdata,
218 				hammer2_chain_t **parentp,
219 				hammer2_key_t lbase,
220 				int ioflag, int pblksize, int *errorp);
221 static int test_block_zeros(const char *buf, size_t bytes);
222 static void zero_write(struct buf *bp, hammer2_trans_t *trans,
223 				hammer2_inode_t *ip,
224 				hammer2_inode_data_t *ipdata,
225 				hammer2_chain_t **parentp,
226 				hammer2_key_t lbase,
227 				int *errorp);
228 static void hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp,
229 				int ioflag, int pblksize, int *errorp);
230 
231 static int hammer2_rcvdmsg(kdmsg_msg_t *msg);
232 static void hammer2_autodmsg(kdmsg_msg_t *msg);
233 
234 
235 /*
236  * HAMMER2 vfs operations.
237  */
238 static struct vfsops hammer2_vfsops = {
239 	.vfs_init	= hammer2_vfs_init,
240 	.vfs_uninit = hammer2_vfs_uninit,
241 	.vfs_sync	= hammer2_vfs_sync,
242 	.vfs_mount	= hammer2_vfs_mount,
243 	.vfs_unmount	= hammer2_vfs_unmount,
244 	.vfs_root 	= hammer2_vfs_root,
245 	.vfs_statfs	= hammer2_vfs_statfs,
246 	.vfs_statvfs	= hammer2_vfs_statvfs,
247 	.vfs_vget	= hammer2_vfs_vget,
248 	.vfs_vptofh	= hammer2_vfs_vptofh,
249 	.vfs_fhtovp	= hammer2_vfs_fhtovp,
250 	.vfs_checkexp	= hammer2_vfs_checkexp
251 };
252 
253 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
254 
255 VFS_SET(hammer2_vfsops, hammer2, 0);
256 MODULE_VERSION(hammer2, 1);
257 
258 static
259 int
260 hammer2_vfs_init(struct vfsconf *conf)
261 {
262 	static struct objcache_malloc_args margs_read;
263 	static struct objcache_malloc_args margs_write;
264 
265 	int error;
266 
267 	error = 0;
268 
269 	if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
270 		error = EINVAL;
271 	if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
272 		error = EINVAL;
273 	if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
274 		error = EINVAL;
275 
276 	if (error)
277 		kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
278 
279 	margs_read.objsize = 65536;
280 	margs_read.mtype = D_BUFFER;
281 
282 	margs_write.objsize = 32768;
283 	margs_write.mtype = C_BUFFER;
284 
285 	cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
286 				0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
287 				objcache_malloc_free, &margs_read);
288 	cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
289 				0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
290 				objcache_malloc_free, &margs_write);
291 
292 	lockinit(&hammer2_mntlk, "mntlk", 0, 0);
293 	TAILQ_INIT(&hammer2_mntlist);
294 
295 	hammer2_limit_dirty_chains = desiredvnodes / 10;
296 
297 	return (error);
298 }
299 
300 static
301 int
302 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
303 {
304 	objcache_destroy(cache_buffer_read);
305 	objcache_destroy(cache_buffer_write);
306 	return 0;
307 }
308 
309 /*
310  * Mount or remount HAMMER2 fileystem from physical media
311  *
312  *	mountroot
313  *		mp		mount point structure
314  *		path		NULL
315  *		data		<unused>
316  *		cred		<unused>
317  *
318  *	mount
319  *		mp		mount point structure
320  *		path		path to mount point
321  *		data		pointer to argument structure in user space
322  *			volume	volume path (device@LABEL form)
323  *			hflags	user mount flags
324  *		cred		user credentials
325  *
326  * RETURNS:	0	Success
327  *		!0	error number
328  */
329 static
330 int
331 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
332 		  struct ucred *cred)
333 {
334 	struct hammer2_mount_info info;
335 	hammer2_pfsmount_t *pmp;
336 	hammer2_mount_t *hmp;
337 	hammer2_key_t key_next;
338 	hammer2_key_t key_dummy;
339 	hammer2_key_t lhc;
340 	struct vnode *devvp;
341 	struct nlookupdata nd;
342 	hammer2_chain_t *parent;
343 	hammer2_chain_t *schain;
344 	hammer2_chain_t *rchain;
345 	struct file *fp;
346 	char devstr[MNAMELEN];
347 	size_t size;
348 	size_t done;
349 	char *dev;
350 	char *label;
351 	int ronly = 1;
352 	int error;
353 	int cache_index;
354 	int i;
355 
356 	hmp = NULL;
357 	pmp = NULL;
358 	dev = NULL;
359 	label = NULL;
360 	devvp = NULL;
361 	cache_index = -1;
362 
363 	kprintf("hammer2_mount\n");
364 
365 	if (path == NULL) {
366 		/*
367 		 * Root mount
368 		 */
369 		bzero(&info, sizeof(info));
370 		info.cluster_fd = -1;
371 		return (EOPNOTSUPP);
372 	} else {
373 		/*
374 		 * Non-root mount or updating a mount
375 		 */
376 		error = copyin(data, &info, sizeof(info));
377 		if (error)
378 			return (error);
379 
380 		error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
381 		if (error)
382 			return (error);
383 
384 		/* Extract device and label */
385 		dev = devstr;
386 		label = strchr(devstr, '@');
387 		if (label == NULL ||
388 		    ((label + 1) - dev) > done) {
389 			return (EINVAL);
390 		}
391 		*label = '\0';
392 		label++;
393 		if (*label == '\0')
394 			return (EINVAL);
395 
396 		if (mp->mnt_flag & MNT_UPDATE) {
397 			/* Update mount */
398 			/* HAMMER2 implements NFS export via mountctl */
399 			pmp = MPTOPMP(mp);
400 			for (i = 0; i < pmp->cluster.nchains; ++i) {
401 				hmp = pmp->cluster.chains[i]->hmp;
402 				devvp = hmp->devvp;
403 				error = hammer2_remount(hmp, mp, path,
404 							devvp, cred);
405 				if (error)
406 					break;
407 			}
408 			hammer2_inode_install_hidden(pmp);
409 
410 			return error;
411 		}
412 	}
413 
414 	/*
415 	 * PFS mount
416 	 *
417 	 * Lookup name and verify it refers to a block device.
418 	 */
419 	error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
420 	if (error == 0)
421 		error = nlookup(&nd);
422 	if (error == 0)
423 		error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
424 	nlookup_done(&nd);
425 
426 	if (error == 0) {
427 		if (vn_isdisk(devvp, &error))
428 			error = vfs_mountedon(devvp);
429 	}
430 
431 	/*
432 	 * Determine if the device has already been mounted.  After this
433 	 * check hmp will be non-NULL if we are doing the second or more
434 	 * hammer2 mounts from the same device.
435 	 */
436 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
437 	TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
438 		if (hmp->devvp == devvp)
439 			break;
440 	}
441 
442 	/*
443 	 * Open the device if this isn't a secondary mount and construct
444 	 * the H2 device mount (hmp).
445 	 */
446 	if (hmp == NULL) {
447 		if (error == 0 && vcount(devvp) > 0)
448 			error = EBUSY;
449 
450 		/*
451 		 * Now open the device
452 		 */
453 		if (error == 0) {
454 			ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
455 			vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
456 			error = vinvalbuf(devvp, V_SAVE, 0, 0);
457 			if (error == 0) {
458 				error = VOP_OPEN(devvp,
459 						 ronly ? FREAD : FREAD | FWRITE,
460 						 FSCRED, NULL);
461 			}
462 			vn_unlock(devvp);
463 		}
464 		if (error && devvp) {
465 			vrele(devvp);
466 			devvp = NULL;
467 		}
468 		if (error) {
469 			lockmgr(&hammer2_mntlk, LK_RELEASE);
470 			return error;
471 		}
472 		hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
473 		hmp->ronly = ronly;
474 		hmp->devvp = devvp;
475 		kmalloc_create(&hmp->mchain, "HAMMER2-chains");
476 		TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
477 		RB_INIT(&hmp->iotree);
478 
479 		lockinit(&hmp->alloclk, "h2alloc", 0, 0);
480 		lockinit(&hmp->voldatalk, "voldata", 0, LK_CANRECURSE);
481 		TAILQ_INIT(&hmp->transq);
482 
483 		/*
484 		 * vchain setup. vchain.data is embedded.
485 		 * vchain.refs is initialized and will never drop to 0.
486 		 *
487 		 * NOTE! voldata is not yet loaded.
488 		 */
489 		hmp->vchain.hmp = hmp;
490 		hmp->vchain.refs = 1;
491 		hmp->vchain.data = (void *)&hmp->voldata;
492 		hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
493 		hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
494 		hmp->vchain.delete_tid = HAMMER2_MAX_TID;
495 
496 		hammer2_chain_core_alloc(NULL, &hmp->vchain, NULL);
497 		/* hmp->vchain.u.xxx is left NULL */
498 
499 		/*
500 		 * fchain setup.  fchain.data is embedded.
501 		 * fchain.refs is initialized and will never drop to 0.
502 		 *
503 		 * The data is not used but needs to be initialized to
504 		 * pass assertion muster.  We use this chain primarily
505 		 * as a placeholder for the freemap's top-level RBTREE
506 		 * so it does not interfere with the volume's topology
507 		 * RBTREE.
508 		 */
509 		hmp->fchain.hmp = hmp;
510 		hmp->fchain.refs = 1;
511 		hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
512 		hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
513 		hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
514 		hmp->fchain.bref.methods =
515 			HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
516 			HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
517 		hmp->fchain.delete_tid = HAMMER2_MAX_TID;
518 
519 		hammer2_chain_core_alloc(NULL, &hmp->fchain, NULL);
520 		/* hmp->fchain.u.xxx is left NULL */
521 
522 		/*
523 		 * Install the volume header and initialize fields from
524 		 * voldata.
525 		 */
526 		error = hammer2_install_volume_header(hmp);
527 		if (error) {
528 			++hmp->pmp_count;
529 			hammer2_vfs_unmount_hmp1(mp, hmp);
530 			hammer2_vfs_unmount_hmp2(mp, hmp);
531 			hammer2_vfs_unmount(mp, MNT_FORCE);
532 			return error;
533 		}
534 
535 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
536 		hmp->vchain.modify_tid = hmp->voldata.mirror_tid;
537 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
538 		hmp->fchain.modify_tid = hmp->voldata.freemap_tid;
539 
540 		/*
541 		 * First locate the super-root inode, which is key 0
542 		 * relative to the volume header's blockset.
543 		 *
544 		 * Then locate the root inode by scanning the directory keyspace
545 		 * represented by the label.
546 		 */
547 		parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
548 		schain = hammer2_chain_lookup(&parent, &key_dummy,
549 				      HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
550 				      &cache_index, 0);
551 		hammer2_chain_lookup_done(parent);
552 		if (schain == NULL) {
553 			kprintf("hammer2_mount: invalid super-root\n");
554 			++hmp->pmp_count;
555 			hammer2_vfs_unmount_hmp1(mp, hmp);
556 			hammer2_vfs_unmount_hmp2(mp, hmp);
557 			hammer2_vfs_unmount(mp, MNT_FORCE);
558 			return EINVAL;
559 		}
560 
561 		/*
562 		 * NOTE: inode_get sucks up schain's lock.
563 		 */
564 		atomic_set_int(&schain->flags, HAMMER2_CHAIN_PFSROOT);
565 		hmp->sroot = hammer2_inode_get(NULL, NULL, schain);
566 		hammer2_inode_ref(hmp->sroot);
567 		hammer2_inode_unlock_ex(hmp->sroot, schain);
568 		schain = NULL;
569 		/* leave hmp->sroot with one ref */
570 
571 		if ((mp->mnt_flag & MNT_RDONLY) == 0) {
572 			error = hammer2_recovery(hmp);
573 			/* XXX do something with error */
574 		}
575 	}
576 
577 	/*
578 	 * Block device opened successfully, finish initializing the
579 	 * mount structure.
580 	 *
581 	 * From this point on we have to call hammer2_unmount() on failure.
582 	 */
583 	pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
584 
585 	kmalloc_create(&pmp->minode, "HAMMER2-inodes");
586 	kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
587 	lockinit(&pmp->lock, "pfslk", 0, 0);
588 	spin_init(&pmp->inum_spin);
589 	RB_INIT(&pmp->inum_tree);
590 
591 	kdmsg_iocom_init(&pmp->iocom, pmp,
592 			 KDMSG_IOCOMF_AUTOCONN |
593 			 KDMSG_IOCOMF_AUTOSPAN |
594 			 KDMSG_IOCOMF_AUTOCIRC,
595 			 pmp->mmsg, hammer2_rcvdmsg);
596 
597 	ccms_domain_init(&pmp->ccms_dom);
598 	++hmp->pmp_count;
599 	lockmgr(&hammer2_mntlk, LK_RELEASE);
600 	kprintf("hammer2_mount hmp=%p pmp=%p pmpcnt=%d\n",
601 		hmp, pmp, hmp->pmp_count);
602 
603 	mp->mnt_flag = MNT_LOCAL;
604 	mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;	/* all entry pts are SMP */
605 	mp->mnt_kern_flag |= MNTK_THR_SYNC;	/* new vsyncscan semantics */
606 
607 	/*
608 	 * required mount structure initializations
609 	 */
610 	mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
611 	mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
612 
613 	mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
614 	mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
615 
616 	/*
617 	 * Optional fields
618 	 */
619 	mp->mnt_iosize_max = MAXPHYS;
620 	mp->mnt_data = (qaddr_t)pmp;
621 	pmp->mp = mp;
622 
623 	/*
624 	 * Lookup mount point under the media-localized super-root.
625 	 */
626 	parent = hammer2_inode_lock_ex(hmp->sroot);
627 	lhc = hammer2_dirhash(label, strlen(label));
628 	rchain = hammer2_chain_lookup(&parent, &key_next,
629 				      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
630 				      &cache_index, 0);
631 	while (rchain) {
632 		if (rchain->bref.type == HAMMER2_BREF_TYPE_INODE &&
633 		    strcmp(label, rchain->data->ipdata.filename) == 0) {
634 			break;
635 		}
636 		rchain = hammer2_chain_next(&parent, rchain, &key_next,
637 					    key_next,
638 					    lhc + HAMMER2_DIRHASH_LOMASK,
639 					    &cache_index, 0);
640 	}
641 	hammer2_inode_unlock_ex(hmp->sroot, parent);
642 
643 	if (rchain == NULL) {
644 		kprintf("hammer2_mount: PFS label not found\n");
645 		hammer2_vfs_unmount_hmp1(mp, hmp);
646 		hammer2_vfs_unmount_hmp2(mp, hmp);
647 		hammer2_vfs_unmount(mp, MNT_FORCE);
648 		return EINVAL;
649 	}
650 	if (rchain->flags & HAMMER2_CHAIN_MOUNTED) {
651 		hammer2_chain_unlock(rchain);
652 		kprintf("hammer2_mount: PFS label already mounted!\n");
653 		hammer2_vfs_unmount_hmp1(mp, hmp);
654 		hammer2_vfs_unmount_hmp2(mp, hmp);
655 		hammer2_vfs_unmount(mp, MNT_FORCE);
656 		return EBUSY;
657 	}
658 #if 0
659 	if (rchain->flags & HAMMER2_CHAIN_RECYCLE) {
660 		kprintf("hammer2_mount: PFS label currently recycling\n");
661 		hammer2_vfs_unmount_hmp1(mp, hmp);
662 		hammer2_vfs_unmount_hmp2(mp, hmp);
663 		hammer2_vfs_unmount(mp, MNT_FORCE);
664 		return EBUSY;
665 	}
666 #endif
667 	/*
668 	 * After this point hammer2_vfs_unmount() has visibility on hmp
669 	 * and manual hmp1/hmp2 calls are not needed on fatal errors.
670 	 */
671 
672 	atomic_set_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
673 
674 	/*
675 	 * NOTE: *_get() integrates chain's lock into the inode lock.
676 	 */
677 	hammer2_chain_ref(rchain);		/* for pmp->rchain */
678 	pmp->cluster.nchains = 1;
679 	pmp->cluster.chains[0] = rchain;
680 	pmp->iroot = hammer2_inode_get(pmp, NULL, rchain);
681 	hammer2_inode_ref(pmp->iroot);		/* ref for pmp->iroot */
682 
683 	KKASSERT(rchain->pmp == NULL);		/* tracking pmp for rchain */
684 	rchain->pmp = pmp;
685 
686 	hammer2_inode_unlock_ex(pmp->iroot, rchain);
687 
688 	kprintf("iroot %p\n", pmp->iroot);
689 
690 	/*
691 	 * The logical file buffer bio write thread handles things
692 	 * like physical block assignment and compression.
693 	 */
694 	mtx_init(&pmp->wthread_mtx);
695 	bioq_init(&pmp->wthread_bioq);
696 	pmp->wthread_destroy = 0;
697 	lwkt_create(hammer2_write_thread, pmp,
698 		    &pmp->wthread_td, NULL, 0, -1, "hwrite-%s", label);
699 
700 	/*
701 	 * Ref the cluster management messaging descriptor.  The mount
702 	 * program deals with the other end of the communications pipe.
703 	 */
704 	fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
705 	if (fp == NULL) {
706 		kprintf("hammer2_mount: bad cluster_fd!\n");
707 		hammer2_vfs_unmount(mp, MNT_FORCE);
708 		return EBADF;
709 	}
710 	hammer2_cluster_reconnect(pmp, fp);
711 
712 	/*
713 	 * With the cluster operational install ihidden.
714 	 */
715 	hammer2_inode_install_hidden(pmp);
716 
717 	/*
718 	 * Finish setup
719 	 */
720 	vfs_getnewfsid(mp);
721 	vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
722 	vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
723 	vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
724 
725 	copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
726 	bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
727 	bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
728 	copyinstr(path, mp->mnt_stat.f_mntonname,
729 		  sizeof(mp->mnt_stat.f_mntonname) - 1,
730 		  &size);
731 
732 	/*
733 	 * Initial statfs to prime mnt_stat.
734 	 */
735 	hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
736 
737 	return 0;
738 }
739 
740 /*
741  * Handle bioq for strategy write
742  */
743 static
744 void
745 hammer2_write_thread(void *arg)
746 {
747 	hammer2_pfsmount_t *pmp;
748 	struct bio *bio;
749 	struct buf *bp;
750 	hammer2_trans_t trans;
751 	struct vnode *vp;
752 	hammer2_inode_t *ip;
753 	hammer2_chain_t *parent;
754 	hammer2_chain_t **parentp;
755 	hammer2_inode_data_t *ipdata;
756 	hammer2_key_t lbase;
757 	int lblksize;
758 	int pblksize;
759 	int error;
760 
761 	pmp = arg;
762 
763 	mtx_lock(&pmp->wthread_mtx);
764 	while (pmp->wthread_destroy == 0) {
765 		if (bioq_first(&pmp->wthread_bioq) == NULL) {
766 			mtxsleep(&pmp->wthread_bioq, &pmp->wthread_mtx,
767 				 0, "h2bioqw", 0);
768 		}
769 		parent = NULL;
770 		parentp = &parent;
771 
772 		hammer2_trans_init(&trans, pmp, NULL, HAMMER2_TRANS_BUFCACHE);
773 
774 		while ((bio = bioq_takefirst(&pmp->wthread_bioq)) != NULL) {
775 			/*
776 			 * dummy bio for synchronization.  The transaction
777 			 * must be reinitialized.
778 			 */
779 			if (bio->bio_buf == NULL) {
780 				bio->bio_flags |= BIO_DONE;
781 				wakeup(bio);
782 				hammer2_trans_done(&trans);
783 				hammer2_trans_init(&trans, pmp, NULL,
784 						   HAMMER2_TRANS_BUFCACHE);
785 				continue;
786 			}
787 
788 			/*
789 			 * else normal bio processing
790 			 */
791 			mtx_unlock(&pmp->wthread_mtx);
792 
793 			hammer2_lwinprog_drop(pmp);
794 
795 			error = 0;
796 			bp = bio->bio_buf;
797 			vp = bp->b_vp;
798 			ip = VTOI(vp);
799 
800 			/*
801 			 * Inode is modified, flush size and mtime changes
802 			 * to ensure that the file size remains consistent
803 			 * with the buffers being flushed.
804 			 */
805 			parent = hammer2_inode_lock_ex(ip);
806 			if (ip->flags & (HAMMER2_INODE_RESIZED |
807 					 HAMMER2_INODE_MTIME)) {
808 				hammer2_inode_fsync(&trans, ip, parentp);
809 			}
810 			ipdata = hammer2_chain_modify_ip(&trans, ip,
811 							 parentp, 0);
812 			lblksize = hammer2_calc_logical(ip, bio->bio_offset,
813 							&lbase, NULL);
814 			pblksize = hammer2_calc_physical(ip, lbase);
815 			hammer2_write_file_core(bp, &trans, ip, ipdata,
816 						parentp,
817 						lbase, IO_ASYNC,
818 						pblksize, &error);
819 			hammer2_inode_unlock_ex(ip, parent);
820 			if (error) {
821 				kprintf("hammer2: error in buffer write\n");
822 				bp->b_flags |= B_ERROR;
823 				bp->b_error = EIO;
824 			}
825 			biodone(bio);
826 			mtx_lock(&pmp->wthread_mtx);
827 		}
828 		hammer2_trans_done(&trans);
829 	}
830 	pmp->wthread_destroy = -1;
831 	wakeup(&pmp->wthread_destroy);
832 
833 	mtx_unlock(&pmp->wthread_mtx);
834 }
835 
836 void
837 hammer2_bioq_sync(hammer2_pfsmount_t *pmp)
838 {
839 	struct bio sync_bio;
840 
841 	bzero(&sync_bio, sizeof(sync_bio));	/* dummy with no bio_buf */
842 	mtx_lock(&pmp->wthread_mtx);
843 	if (pmp->wthread_destroy == 0) {
844 		if (TAILQ_EMPTY(&pmp->wthread_bioq.queue)) {
845 		       bioq_insert_tail(&pmp->wthread_bioq, &sync_bio);
846 		       wakeup(&pmp->wthread_bioq);
847 		} else {
848 		       bioq_insert_tail(&pmp->wthread_bioq, &sync_bio);
849 		}
850 		while ((sync_bio.bio_flags & BIO_DONE) == 0)
851 			mtxsleep(&sync_bio, &pmp->wthread_mtx, 0, "h2bioq", 0);
852 	}
853 	mtx_unlock(&pmp->wthread_mtx);
854 }
855 
856 /*
857  * Return a chain suitable for I/O, creating the chain if necessary
858  * and assigning its physical block.
859  */
860 static
861 hammer2_chain_t *
862 hammer2_assign_physical(hammer2_trans_t *trans,
863 			hammer2_inode_t *ip, hammer2_chain_t **parentp,
864 			hammer2_key_t lbase, int pblksize, int *errorp)
865 {
866 	hammer2_chain_t *parent;
867 	hammer2_chain_t *chain;
868 	hammer2_off_t pbase;
869 	hammer2_key_t key_dummy;
870 	int pradix = hammer2_getradix(pblksize);
871 	int cache_index = -1;
872 
873 	/*
874 	 * Locate the chain associated with lbase, return a locked chain.
875 	 * However, do not instantiate any data reference (which utilizes a
876 	 * device buffer) because we will be using direct IO via the
877 	 * logical buffer cache buffer.
878 	 */
879 	*errorp = 0;
880 	KKASSERT(pblksize >= HAMMER2_MIN_ALLOC);
881 retry:
882 	parent = *parentp;
883 	hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); /* extra lock */
884 	chain = hammer2_chain_lookup(&parent, &key_dummy,
885 				     lbase, lbase,
886 				     &cache_index, HAMMER2_LOOKUP_NODATA);
887 
888 	if (chain == NULL) {
889 		/*
890 		 * We found a hole, create a new chain entry.
891 		 *
892 		 * NOTE: DATA chains are created without device backing
893 		 *	 store (nor do we want any).
894 		 */
895 		*errorp = hammer2_chain_create(trans, &parent, &chain,
896 					       lbase, HAMMER2_PBUFRADIX,
897 					       HAMMER2_BREF_TYPE_DATA,
898 					       pblksize);
899 		if (chain == NULL) {
900 			hammer2_chain_lookup_done(parent);
901 			panic("hammer2_chain_create: par=%p error=%d\n",
902 				parent, *errorp);
903 			goto retry;
904 		}
905 
906 		pbase = chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX;
907 		/*ip->delta_dcount += pblksize;*/
908 	} else {
909 		switch (chain->bref.type) {
910 		case HAMMER2_BREF_TYPE_INODE:
911 			/*
912 			 * The data is embedded in the inode.  The
913 			 * caller is responsible for marking the inode
914 			 * modified and copying the data to the embedded
915 			 * area.
916 			 */
917 			pbase = NOOFFSET;
918 			break;
919 		case HAMMER2_BREF_TYPE_DATA:
920 			if (chain->bytes != pblksize) {
921 				hammer2_chain_resize(trans, ip,
922 						     parent, &chain,
923 						     pradix,
924 						     HAMMER2_MODIFY_OPTDATA);
925 			}
926 			hammer2_chain_modify(trans, &chain,
927 					     HAMMER2_MODIFY_OPTDATA);
928 			pbase = chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX;
929 			break;
930 		default:
931 			panic("hammer2_assign_physical: bad type");
932 			/* NOT REACHED */
933 			pbase = NOOFFSET;
934 			break;
935 		}
936 	}
937 
938 	/*
939 	 * Cleanup.  If chain wound up being the inode (i.e. DIRECTDATA),
940 	 * we might have to replace *parentp.
941 	 */
942 	hammer2_chain_lookup_done(parent);
943 	if (chain) {
944 		if (*parentp != chain &&
945 		    (*parentp)->core == chain->core) {
946 			parent = *parentp;
947 			*parentp = chain;		/* eats lock */
948 			hammer2_chain_unlock(parent);
949 			hammer2_chain_lock(chain, 0);	/* need another */
950 		}
951 		/* else chain already locked for return */
952 	}
953 	return (chain);
954 }
955 
956 /*
957  * From hammer2_vnops.c.
958  * The core write function which determines which path to take
959  * depending on compression settings.
960  */
961 static
962 void
963 hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
964 			hammer2_inode_t *ip, hammer2_inode_data_t *ipdata,
965 			hammer2_chain_t **parentp,
966 			hammer2_key_t lbase, int ioflag, int pblksize,
967 			int *errorp)
968 {
969 	hammer2_chain_t *chain;
970 
971 	switch(HAMMER2_DEC_COMP(ipdata->comp_algo)) {
972 	case HAMMER2_COMP_NONE:
973 		/*
974 		 * We have to assign physical storage to the buffer
975 		 * we intend to dirty or write now to avoid deadlocks
976 		 * in the strategy code later.
977 		 *
978 		 * This can return NOOFFSET for inode-embedded data.
979 		 * The strategy code will take care of it in that case.
980 		 */
981 		chain = hammer2_assign_physical(trans, ip, parentp,
982 						lbase, pblksize,
983 						errorp);
984 		hammer2_write_bp(chain, bp, ioflag, pblksize, errorp);
985 		if (chain)
986 			hammer2_chain_unlock(chain);
987 		break;
988 	case HAMMER2_COMP_AUTOZERO:
989 		/*
990 		 * Check for zero-fill only
991 		 */
992 		hammer2_zero_check_and_write(bp, trans, ip,
993 				    ipdata, parentp, lbase,
994 				    ioflag, pblksize, errorp);
995 		break;
996 	case HAMMER2_COMP_LZ4:
997 	case HAMMER2_COMP_ZLIB:
998 	default:
999 		/*
1000 		 * Check for zero-fill and attempt compression.
1001 		 */
1002 		hammer2_compress_and_write(bp, trans, ip,
1003 					   ipdata, parentp,
1004 					   lbase, ioflag,
1005 					   pblksize, errorp,
1006 					   ipdata->comp_algo);
1007 		break;
1008 	}
1009 	ipdata = &ip->chain->data->ipdata;	/* reload */
1010 }
1011 
1012 /*
1013  * From hammer2_vnops.c
1014  * Generic function that will perform the compression in compression
1015  * write path. The compression algorithm is determined by the settings
1016  * obtained from inode.
1017  */
1018 static
1019 void
1020 hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
1021 	hammer2_inode_t *ip, hammer2_inode_data_t *ipdata,
1022 	hammer2_chain_t **parentp,
1023 	hammer2_key_t lbase, int ioflag, int pblksize,
1024 	int *errorp, int comp_algo)
1025 {
1026 	hammer2_chain_t *chain;
1027 	int comp_size;
1028 	int comp_block_size;
1029 	char *comp_buffer;
1030 
1031 	if (test_block_zeros(bp->b_data, pblksize)) {
1032 		zero_write(bp, trans, ip, ipdata, parentp, lbase, errorp);
1033 		return;
1034 	}
1035 
1036 	comp_size = 0;
1037 	comp_buffer = NULL;
1038 
1039 	KKASSERT(pblksize / 2 <= 32768);
1040 
1041 	if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
1042 		z_stream strm_compress;
1043 		int comp_level;
1044 		int ret;
1045 
1046 		switch(HAMMER2_DEC_COMP(comp_algo)) {
1047 		case HAMMER2_COMP_LZ4:
1048 			comp_buffer = objcache_get(cache_buffer_write,
1049 						   M_INTWAIT);
1050 			comp_size = LZ4_compress_limitedOutput(
1051 					bp->b_data,
1052 					&comp_buffer[sizeof(int)],
1053 					pblksize,
1054 					pblksize / 2 - sizeof(int));
1055 			/*
1056 			 * We need to prefix with the size, LZ4
1057 			 * doesn't do it for us.  Add the related
1058 			 * overhead.
1059 			 */
1060 			*(int *)comp_buffer = comp_size;
1061 			if (comp_size)
1062 				comp_size += sizeof(int);
1063 			break;
1064 		case HAMMER2_COMP_ZLIB:
1065 			comp_level = HAMMER2_DEC_LEVEL(comp_algo);
1066 			if (comp_level == 0)
1067 				comp_level = 6;	/* default zlib compression */
1068 			else if (comp_level < 6)
1069 				comp_level = 6;
1070 			else if (comp_level > 9)
1071 				comp_level = 9;
1072 			ret = deflateInit(&strm_compress, comp_level);
1073 			if (ret != Z_OK) {
1074 				kprintf("HAMMER2 ZLIB: fatal error "
1075 					"on deflateInit.\n");
1076 			}
1077 
1078 			comp_buffer = objcache_get(cache_buffer_write,
1079 						   M_INTWAIT);
1080 			strm_compress.next_in = bp->b_data;
1081 			strm_compress.avail_in = pblksize;
1082 			strm_compress.next_out = comp_buffer;
1083 			strm_compress.avail_out = pblksize / 2;
1084 			ret = deflate(&strm_compress, Z_FINISH);
1085 			if (ret == Z_STREAM_END) {
1086 				comp_size = pblksize / 2 -
1087 					    strm_compress.avail_out;
1088 			} else {
1089 				comp_size = 0;
1090 			}
1091 			ret = deflateEnd(&strm_compress);
1092 			break;
1093 		default:
1094 			kprintf("Error: Unknown compression method.\n");
1095 			kprintf("Comp_method = %d.\n", comp_algo);
1096 			break;
1097 		}
1098 	}
1099 
1100 	if (comp_size == 0) {
1101 		/*
1102 		 * compression failed or turned off
1103 		 */
1104 		comp_block_size = pblksize;	/* safety */
1105 		if (++ip->comp_heuristic > 128)
1106 			ip->comp_heuristic = 8;
1107 	} else {
1108 		/*
1109 		 * compression succeeded
1110 		 */
1111 		ip->comp_heuristic = 0;
1112 		if (comp_size <= 1024) {
1113 			comp_block_size = 1024;
1114 		} else if (comp_size <= 2048) {
1115 			comp_block_size = 2048;
1116 		} else if (comp_size <= 4096) {
1117 			comp_block_size = 4096;
1118 		} else if (comp_size <= 8192) {
1119 			comp_block_size = 8192;
1120 		} else if (comp_size <= 16384) {
1121 			comp_block_size = 16384;
1122 		} else if (comp_size <= 32768) {
1123 			comp_block_size = 32768;
1124 		} else {
1125 			panic("hammer2: WRITE PATH: "
1126 			      "Weird comp_size value.");
1127 			/* NOT REACHED */
1128 			comp_block_size = pblksize;
1129 		}
1130 	}
1131 
1132 	chain = hammer2_assign_physical(trans, ip, parentp,
1133 					lbase, comp_block_size,
1134 					errorp);
1135 	ipdata = &ip->chain->data->ipdata;	/* RELOAD */
1136 
1137 	if (*errorp) {
1138 		kprintf("WRITE PATH: An error occurred while "
1139 			"assigning physical space.\n");
1140 		KKASSERT(chain == NULL);
1141 	} else {
1142 		/* Get device offset */
1143 		hammer2_io_t *dio;
1144 		char *bdata;
1145 		int temp_check;
1146 
1147 		KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1148 
1149 		switch(chain->bref.type) {
1150 		case HAMMER2_BREF_TYPE_INODE:
1151 			KKASSERT(chain->data->ipdata.op_flags &
1152 				 HAMMER2_OPFLAG_DIRECTDATA);
1153 			KKASSERT(bp->b_loffset == 0);
1154 			bcopy(bp->b_data, chain->data->ipdata.u.data,
1155 			      HAMMER2_EMBEDDED_BYTES);
1156 			break;
1157 		case HAMMER2_BREF_TYPE_DATA:
1158 			temp_check = HAMMER2_DEC_CHECK(chain->bref.methods);
1159 
1160 			/*
1161 			 * Optimize out the read-before-write
1162 			 * if possible.
1163 			 */
1164 			*errorp = hammer2_io_newnz(chain->hmp,
1165 						   chain->bref.data_off,
1166 						   chain->bytes,
1167 						   &dio);
1168 			if (*errorp) {
1169 				hammer2_io_brelse(&dio);
1170 				kprintf("hammer2: WRITE PATH: "
1171 					"dbp bread error\n");
1172 				break;
1173 			}
1174 			bdata = hammer2_io_data(dio, chain->bref.data_off);
1175 
1176 			/*
1177 			 * When loading the block make sure we don't
1178 			 * leave garbage after the compressed data.
1179 			 */
1180 			if (comp_size) {
1181 				chain->bref.methods =
1182 					HAMMER2_ENC_COMP(comp_algo) +
1183 					HAMMER2_ENC_CHECK(temp_check);
1184 				bcopy(comp_buffer, bdata, comp_size);
1185 				if (comp_size != comp_block_size) {
1186 					bzero(bdata + comp_size,
1187 					      comp_block_size - comp_size);
1188 				}
1189 			} else {
1190 				chain->bref.methods =
1191 					HAMMER2_ENC_COMP(
1192 						HAMMER2_COMP_NONE) +
1193 					HAMMER2_ENC_CHECK(temp_check);
1194 				bcopy(bp->b_data, bdata, pblksize);
1195 			}
1196 
1197 			/*
1198 			 * Device buffer is now valid, chain is no
1199 			 * longer in the initial state.
1200 			 */
1201 			atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1202 
1203 			/* Now write the related bdp. */
1204 			if (ioflag & IO_SYNC) {
1205 				/*
1206 				 * Synchronous I/O requested.
1207 				 */
1208 				hammer2_io_bwrite(&dio);
1209 			/*
1210 			} else if ((ioflag & IO_DIRECT) &&
1211 				   loff + n == pblksize) {
1212 				hammer2_io_bdwrite(&dio);
1213 			*/
1214 			} else if (ioflag & IO_ASYNC) {
1215 				hammer2_io_bawrite(&dio);
1216 			} else {
1217 				hammer2_io_bdwrite(&dio);
1218 			}
1219 			break;
1220 		default:
1221 			panic("hammer2_write_bp: bad chain type %d\n",
1222 				chain->bref.type);
1223 			/* NOT REACHED */
1224 			break;
1225 		}
1226 
1227 		hammer2_chain_unlock(chain);
1228 	}
1229 	if (comp_buffer)
1230 		objcache_put(cache_buffer_write, comp_buffer);
1231 }
1232 
1233 /*
1234  * Function that performs zero-checking and writing without compression,
1235  * it corresponds to default zero-checking path.
1236  */
1237 static
1238 void
1239 hammer2_zero_check_and_write(struct buf *bp, hammer2_trans_t *trans,
1240 	hammer2_inode_t *ip, hammer2_inode_data_t *ipdata,
1241 	hammer2_chain_t **parentp,
1242 	hammer2_key_t lbase, int ioflag, int pblksize, int *errorp)
1243 {
1244 	hammer2_chain_t *chain;
1245 
1246 	if (test_block_zeros(bp->b_data, pblksize)) {
1247 		zero_write(bp, trans, ip, ipdata, parentp, lbase, errorp);
1248 	} else {
1249 		chain = hammer2_assign_physical(trans, ip, parentp,
1250 						lbase, pblksize, errorp);
1251 		hammer2_write_bp(chain, bp, ioflag, pblksize, errorp);
1252 		if (chain)
1253 			hammer2_chain_unlock(chain);
1254 	}
1255 }
1256 
1257 /*
1258  * A function to test whether a block of data contains only zeros,
1259  * returns TRUE (non-zero) if the block is all zeros.
1260  */
1261 static
1262 int
1263 test_block_zeros(const char *buf, size_t bytes)
1264 {
1265 	size_t i;
1266 
1267 	for (i = 0; i < bytes; i += sizeof(long)) {
1268 		if (*(const long *)(buf + i) != 0)
1269 			return (0);
1270 	}
1271 	return (1);
1272 }
1273 
1274 /*
1275  * Function to "write" a block that contains only zeros.
1276  */
1277 static
1278 void
1279 zero_write(struct buf *bp, hammer2_trans_t *trans, hammer2_inode_t *ip,
1280 	hammer2_inode_data_t *ipdata, hammer2_chain_t **parentp,
1281 	hammer2_key_t lbase, int *errorp __unused)
1282 {
1283 	hammer2_chain_t *parent;
1284 	hammer2_chain_t *chain;
1285 	hammer2_key_t key_dummy;
1286 	int cache_index = -1;
1287 
1288 	parent = hammer2_chain_lookup_init(*parentp, 0);
1289 
1290 	chain = hammer2_chain_lookup(&parent, &key_dummy, lbase, lbase,
1291 				     &cache_index, HAMMER2_LOOKUP_NODATA);
1292 	if (chain) {
1293 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1294 			bzero(chain->data->ipdata.u.data,
1295 			      HAMMER2_EMBEDDED_BYTES);
1296 		} else {
1297 			hammer2_chain_delete(trans, chain, 0);
1298 		}
1299 		hammer2_chain_unlock(chain);
1300 	}
1301 	hammer2_chain_lookup_done(parent);
1302 }
1303 
1304 /*
1305  * Function to write the data as it is, without performing any sort of
1306  * compression. This function is used in path without compression and
1307  * default zero-checking path.
1308  */
1309 static
1310 void
1311 hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp, int ioflag,
1312 				int pblksize, int *errorp)
1313 {
1314 	hammer2_io_t *dio;
1315 	char *bdata;
1316 	int error;
1317 	int temp_check = HAMMER2_DEC_CHECK(chain->bref.methods);
1318 
1319 	KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1320 
1321 	switch(chain->bref.type) {
1322 	case HAMMER2_BREF_TYPE_INODE:
1323 		KKASSERT(chain->data->ipdata.op_flags &
1324 			 HAMMER2_OPFLAG_DIRECTDATA);
1325 		KKASSERT(bp->b_loffset == 0);
1326 		bcopy(bp->b_data, chain->data->ipdata.u.data,
1327 		      HAMMER2_EMBEDDED_BYTES);
1328 		error = 0;
1329 		break;
1330 	case HAMMER2_BREF_TYPE_DATA:
1331 		error = hammer2_io_newnz(chain->hmp, chain->bref.data_off,
1332 					 chain->bytes, &dio);
1333 		if (error) {
1334 			hammer2_io_bqrelse(&dio);
1335 			kprintf("hammer2: WRITE PATH: dbp bread error\n");
1336 			break;
1337 		}
1338 		bdata = hammer2_io_data(dio, chain->bref.data_off);
1339 
1340 		chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1341 				      HAMMER2_ENC_CHECK(temp_check);
1342 		bcopy(bp->b_data, bdata, chain->bytes);
1343 
1344 		/*
1345 		 * Device buffer is now valid, chain is no
1346 		 * longer in the initial state.
1347 		 */
1348 		atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1349 
1350 		if (ioflag & IO_SYNC) {
1351 			/*
1352 			 * Synchronous I/O requested.
1353 			 */
1354 			hammer2_io_bwrite(&dio);
1355 		/*
1356 		} else if ((ioflag & IO_DIRECT) && loff + n == pblksize) {
1357 			hammer2_io_bdwrite(&dio);
1358 		*/
1359 		} else if (ioflag & IO_ASYNC) {
1360 			hammer2_io_bawrite(&dio);
1361 		} else {
1362 			hammer2_io_bdwrite(&dio);
1363 		}
1364 		break;
1365 	default:
1366 		panic("hammer2_write_bp: bad chain type %d\n",
1367 		      chain->bref.type);
1368 		/* NOT REACHED */
1369 		error = 0;
1370 		break;
1371 	}
1372 	*errorp = error;
1373 }
1374 
1375 static
1376 int
1377 hammer2_remount(hammer2_mount_t *hmp, struct mount *mp, char *path,
1378 		struct vnode *devvp, struct ucred *cred)
1379 {
1380 	int error;
1381 
1382 	if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1383 		error = hammer2_recovery(hmp);
1384 	} else {
1385 		error = 0;
1386 	}
1387 	return error;
1388 }
1389 
1390 static
1391 int
1392 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1393 {
1394 	hammer2_pfsmount_t *pmp;
1395 	hammer2_mount_t *hmp;
1396 	hammer2_chain_t *rchain;
1397 	int flags;
1398 	int error = 0;
1399 	int i;
1400 
1401 	pmp = MPTOPMP(mp);
1402 
1403 	if (pmp == NULL)
1404 		return(0);
1405 
1406 	ccms_domain_uninit(&pmp->ccms_dom);
1407 	kdmsg_iocom_uninit(&pmp->iocom);	/* XXX chain dependency */
1408 
1409 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1410 
1411 	/*
1412 	 * If mount initialization proceeded far enough we must flush
1413 	 * its vnodes.
1414 	 */
1415 	if (mntflags & MNT_FORCE)
1416 		flags = FORCECLOSE;
1417 	else
1418 		flags = 0;
1419 	if (pmp->iroot) {
1420 		error = vflush(mp, 0, flags);
1421 		if (error)
1422 			goto failed;
1423 	}
1424 
1425 	if (pmp->wthread_td) {
1426 		mtx_lock(&pmp->wthread_mtx);
1427 		pmp->wthread_destroy = 1;
1428 		wakeup(&pmp->wthread_bioq);
1429 		while (pmp->wthread_destroy != -1) {
1430 			mtxsleep(&pmp->wthread_destroy,
1431 				&pmp->wthread_mtx, 0,
1432 				"umount-sleep",	0);
1433 		}
1434 		mtx_unlock(&pmp->wthread_mtx);
1435 		pmp->wthread_td = NULL;
1436 	}
1437 
1438 	/*
1439 	 * Cleanup our reference on ihidden.
1440 	 */
1441 	if (pmp->ihidden) {
1442 		hammer2_inode_drop(pmp->ihidden);
1443 		pmp->ihidden = NULL;
1444 	}
1445 
1446 	/*
1447 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
1448 	 * by the flush code.
1449 	 */
1450 	if (pmp->iroot) {
1451 #if REPORT_REFS_ERRORS
1452 		if (pmp->iroot->refs != 1)
1453 			kprintf("PMP->IROOT %p REFS WRONG %d\n",
1454 				pmp->iroot, pmp->iroot->refs);
1455 #else
1456 		KKASSERT(pmp->iroot->refs == 1);
1457 #endif
1458 		/* ref for pmp->iroot */
1459 		hammer2_inode_drop(pmp->iroot);
1460 		pmp->iroot = NULL;
1461 	}
1462 
1463 	for (i = 0; i < pmp->cluster.nchains; ++i) {
1464 		hmp = pmp->cluster.chains[i]->hmp;
1465 
1466 		hammer2_vfs_unmount_hmp1(mp, hmp);
1467 
1468 		rchain = pmp->cluster.chains[i];
1469 		if (rchain) {
1470 			atomic_clear_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
1471 #if REPORT_REFS_ERRORS
1472 			if (rchain->refs != 1)
1473 				kprintf("PMP->RCHAIN %p REFS WRONG %d\n",
1474 					rchain, rchain->refs);
1475 #else
1476 			KKASSERT(rchain->refs == 1);
1477 #endif
1478 			hammer2_chain_drop(rchain);
1479 			pmp->cluster.chains[i] = NULL;
1480 		}
1481 
1482 		hammer2_vfs_unmount_hmp2(mp, hmp);
1483 	}
1484 
1485 	pmp->mp = NULL;
1486 	mp->mnt_data = NULL;
1487 
1488 	kmalloc_destroy(&pmp->mmsg);
1489 	kmalloc_destroy(&pmp->minode);
1490 
1491 	kfree(pmp, M_HAMMER2);
1492 	error = 0;
1493 
1494 failed:
1495 	lockmgr(&hammer2_mntlk, LK_RELEASE);
1496 
1497 	return (error);
1498 }
1499 
1500 static
1501 void
1502 hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp)
1503 {
1504 	hammer2_mount_exlock(hmp);
1505 	--hmp->pmp_count;
1506 
1507 	kprintf("hammer2_unmount hmp=%p pmpcnt=%d\n", hmp, hmp->pmp_count);
1508 
1509 	/*
1510 	 * Flush any left over chains.  The voldata lock is only used
1511 	 * to synchronize against HAMMER2_CHAIN_MODIFIED_AUX.
1512 	 *
1513 	 * Flush twice to ensure that the freemap is completely
1514 	 * synchronized.  If we only do it once the next mount's
1515 	 * recovery scan will have to do some fixups (which isn't
1516 	 * bad, but we don't want it to have to do it except when
1517 	 * recovering from a crash).
1518 	 */
1519 	hammer2_voldata_lock(hmp);
1520 	if (((hmp->vchain.flags | hmp->fchain.flags) &
1521 	     HAMMER2_CHAIN_MODIFIED) ||
1522 	    hmp->vchain.core->update_hi > hmp->voldata.mirror_tid ||
1523 	    hmp->fchain.core->update_hi > hmp->voldata.freemap_tid) {
1524 		hammer2_voldata_unlock(hmp, 0);
1525 		hammer2_vfs_sync(mp, MNT_WAIT);
1526 		/*hammer2_vfs_sync(mp, MNT_WAIT);*/
1527 	} else {
1528 		hammer2_voldata_unlock(hmp, 0);
1529 	}
1530 	if (hmp->pmp_count == 0) {
1531 		if (((hmp->vchain.flags | hmp->fchain.flags) &
1532 		     HAMMER2_CHAIN_MODIFIED) ||
1533 		    (hmp->vchain.core->update_hi >
1534 		     hmp->voldata.mirror_tid) ||
1535 		    (hmp->fchain.core->update_hi >
1536 		     hmp->voldata.freemap_tid)) {
1537 			kprintf("hammer2_unmount: chains left over "
1538 				"after final sync\n");
1539 			kprintf("    vchain %08x update_hi %jx/%jx\n",
1540 				hmp->vchain.flags,
1541 				hmp->voldata.mirror_tid,
1542 				hmp->vchain.core->update_hi);
1543 			kprintf("    fchain %08x update_hi %jx/%jx\n",
1544 				hmp->fchain.flags,
1545 				hmp->voldata.freemap_tid,
1546 				hmp->fchain.core->update_hi);
1547 
1548 			if (hammer2_debug & 0x0010)
1549 				Debugger("entered debugger");
1550 		}
1551 	}
1552 }
1553 
1554 static
1555 void
1556 hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp)
1557 {
1558 	struct vnode *devvp;
1559 	int dumpcnt;
1560 	int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1561 
1562 	/*
1563 	 * If no PFS's left drop the master hammer2_mount for the
1564 	 * device.
1565 	 */
1566 	if (hmp->pmp_count == 0) {
1567 		if (hmp->sroot) {
1568 			hammer2_inode_drop(hmp->sroot);
1569 			hmp->sroot = NULL;
1570 		}
1571 
1572 		/*
1573 		 * Finish up with the device vnode
1574 		 */
1575 		if ((devvp = hmp->devvp) != NULL) {
1576 			vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1577 			vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1578 			hmp->devvp = NULL;
1579 			VOP_CLOSE(devvp,
1580 				  (ronly ? FREAD : FREAD|FWRITE));
1581 			vn_unlock(devvp);
1582 			vrele(devvp);
1583 			devvp = NULL;
1584 		}
1585 
1586 		/*
1587 		 * Final drop of embedded freemap root chain to
1588 		 * clean up fchain.core (fchain structure is not
1589 		 * flagged ALLOCATED so it is cleaned out and then
1590 		 * left to rot).
1591 		 */
1592 		hammer2_chain_drop(&hmp->fchain);
1593 
1594 		/*
1595 		 * Final drop of embedded volume root chain to clean
1596 		 * up vchain.core (vchain structure is not flagged
1597 		 * ALLOCATED so it is cleaned out and then left to
1598 		 * rot).
1599 		 */
1600 		dumpcnt = 50;
1601 		hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt);
1602 		dumpcnt = 50;
1603 		hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt);
1604 		hammer2_mount_unlock(hmp);
1605 		hammer2_chain_drop(&hmp->vchain);
1606 
1607 		hammer2_io_cleanup(hmp, &hmp->iotree);
1608 		if (hmp->iofree_count) {
1609 			kprintf("io_cleanup: %d I/O's left hanging\n",
1610 				hmp->iofree_count);
1611 		}
1612 
1613 		TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1614 		kmalloc_destroy(&hmp->mchain);
1615 		kfree(hmp, M_HAMMER2);
1616 	} else {
1617 		hammer2_mount_unlock(hmp);
1618 	}
1619 }
1620 
1621 static
1622 int
1623 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1624 	     ino_t ino, struct vnode **vpp)
1625 {
1626 	kprintf("hammer2_vget\n");
1627 	return (EOPNOTSUPP);
1628 }
1629 
1630 static
1631 int
1632 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1633 {
1634 	hammer2_pfsmount_t *pmp;
1635 	hammer2_chain_t *parent;
1636 	int error;
1637 	struct vnode *vp;
1638 
1639 	pmp = MPTOPMP(mp);
1640 	if (pmp->iroot == NULL) {
1641 		*vpp = NULL;
1642 		error = EINVAL;
1643 	} else {
1644 		parent = hammer2_inode_lock_sh(pmp->iroot);
1645 		vp = hammer2_igetv(pmp->iroot, &error);
1646 		hammer2_inode_unlock_sh(pmp->iroot, parent);
1647 		*vpp = vp;
1648 		if (vp == NULL)
1649 			kprintf("vnodefail\n");
1650 	}
1651 
1652 	return (error);
1653 }
1654 
1655 /*
1656  * Filesystem status
1657  *
1658  * XXX incorporate ipdata->inode_quota and data_quota
1659  */
1660 static
1661 int
1662 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1663 {
1664 	hammer2_pfsmount_t *pmp;
1665 	hammer2_mount_t *hmp;
1666 
1667 	pmp = MPTOPMP(mp);
1668 	KKASSERT(pmp->cluster.nchains >= 1);
1669 	hmp = pmp->cluster.chains[0]->hmp;	/* XXX */
1670 
1671 	mp->mnt_stat.f_files = pmp->inode_count;
1672 	mp->mnt_stat.f_ffree = 0;
1673 	mp->mnt_stat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1674 	mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1675 	mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1676 
1677 	*sbp = mp->mnt_stat;
1678 	return (0);
1679 }
1680 
1681 static
1682 int
1683 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1684 {
1685 	hammer2_pfsmount_t *pmp;
1686 	hammer2_mount_t *hmp;
1687 
1688 	pmp = MPTOPMP(mp);
1689 	KKASSERT(pmp->cluster.nchains >= 1);
1690 	hmp = pmp->cluster.chains[0]->hmp;	/* XXX */
1691 
1692 	mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1693 	mp->mnt_vstat.f_files = pmp->inode_count;
1694 	mp->mnt_vstat.f_ffree = 0;
1695 	mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1696 	mp->mnt_vstat.f_bfree =  hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1697 	mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1698 
1699 	*sbp = mp->mnt_vstat;
1700 	return (0);
1701 }
1702 
1703 /*
1704  * Mount-time recovery (RW mounts)
1705  *
1706  * Updates to the free block table are allowed to lag flushes by one
1707  * transaction.  In case of a crash, then on a fresh mount we must do an
1708  * incremental scan of transaction id voldata.mirror_tid and make sure the
1709  * related blocks have been marked allocated.
1710  *
1711  */
1712 struct hammer2_recovery_elm {
1713 	TAILQ_ENTRY(hammer2_recovery_elm) entry;
1714 	hammer2_chain_t *chain;
1715 };
1716 
1717 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1718 
1719 static int hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1720 			hammer2_chain_t *parent,
1721 			struct hammer2_recovery_list *list, int depth);
1722 
1723 #define HAMMER2_RECOVERY_MAXDEPTH	10
1724 
1725 static
1726 int
1727 hammer2_recovery(hammer2_mount_t *hmp)
1728 {
1729 	hammer2_trans_t trans;
1730 	struct hammer2_recovery_list list;
1731 	struct hammer2_recovery_elm *elm;
1732 	hammer2_chain_t *parent;
1733 	int error;
1734 	int cumulative_error = 0;
1735 
1736 	hammer2_trans_init(&trans, NULL, hmp, 0);
1737 
1738 	TAILQ_INIT(&list);
1739 	parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1740 	cumulative_error = hammer2_recovery_scan(&trans, hmp, parent, &list, 0);
1741 	hammer2_chain_lookup_done(parent);
1742 
1743 	while ((elm = TAILQ_FIRST(&list)) != NULL) {
1744 		TAILQ_REMOVE(&list, elm, entry);
1745 		parent = elm->chain;
1746 		kfree(elm, M_HAMMER2);
1747 
1748 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
1749 					   HAMMER2_RESOLVE_NOREF);
1750 		error = hammer2_recovery_scan(&trans, hmp, parent, &list, 0);
1751 		hammer2_chain_unlock(parent);
1752 		if (error)
1753 			cumulative_error = error;
1754 	}
1755 	hammer2_trans_done(&trans);
1756 
1757 	return cumulative_error;
1758 }
1759 
1760 static
1761 int
1762 hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1763 		      hammer2_chain_t *parent,
1764 		      struct hammer2_recovery_list *list, int depth)
1765 {
1766 	hammer2_chain_t *chain;
1767 	int cache_index;
1768 	int cumulative_error = 0;
1769 	int error;
1770 
1771 	/*
1772 	 * Defer operation if depth limit reached.
1773 	 */
1774 	if (depth >= HAMMER2_RECOVERY_MAXDEPTH) {
1775 		struct hammer2_recovery_elm *elm;
1776 
1777 		elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
1778 		elm->chain = parent;
1779 		hammer2_chain_ref(parent);
1780 		TAILQ_INSERT_TAIL(list, elm, entry);
1781 		/* unlocked by caller */
1782 
1783 		return(0);
1784 	}
1785 
1786 	/*
1787 	 * Adjust freemap to ensure that the block(s) are marked allocated.
1788 	 */
1789 	if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1790 		hammer2_freemap_adjust(trans, hmp, &parent->bref,
1791 				       HAMMER2_FREEMAP_DORECOVER);
1792 	}
1793 
1794 	/*
1795 	 * Check type for recursive scan
1796 	 */
1797 	switch(parent->bref.type) {
1798 	case HAMMER2_BREF_TYPE_VOLUME:
1799 		/* data already instantiated */
1800 		break;
1801 	case HAMMER2_BREF_TYPE_INODE:
1802 		/*
1803 		 * Must instantiate data for DIRECTDATA test and also
1804 		 * for recursion.
1805 		 */
1806 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1807 		hammer2_chain_unlock(parent);
1808 		if (parent->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
1809 			/* not applicable to recovery scan */
1810 			return 0;
1811 		}
1812 		break;
1813 	case HAMMER2_BREF_TYPE_INDIRECT:
1814 		/*
1815 		 * Must instantiate data for recursion
1816 		 */
1817 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1818 		hammer2_chain_unlock(parent);
1819 		break;
1820 	case HAMMER2_BREF_TYPE_DATA:
1821 	case HAMMER2_BREF_TYPE_FREEMAP:
1822 	case HAMMER2_BREF_TYPE_FREEMAP_NODE:
1823 	case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
1824 		/* not applicable to recovery scan */
1825 		return 0;
1826 		break;
1827 	default:
1828 		return EDOM;
1829 	}
1830 
1831 	/*
1832 	 * Recursive scan of the last flushed transaction only.  We are
1833 	 * doing this without pmp assignments so don't leave the chains
1834 	 * hanging around after we are done with them.
1835 	 */
1836 	cache_index = 0;
1837 	chain = hammer2_chain_scan(parent, NULL, &cache_index,
1838 				   HAMMER2_LOOKUP_NODATA);
1839 	while (chain) {
1840 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
1841 		if (chain->bref.mirror_tid >= hmp->voldata.mirror_tid) {
1842 			error = hammer2_recovery_scan(trans, hmp, chain,
1843 						      list, depth + 1);
1844 			if (error)
1845 				cumulative_error = error;
1846 		}
1847 		chain = hammer2_chain_scan(parent, chain, &cache_index,
1848 					   HAMMER2_LOOKUP_NODATA);
1849 	}
1850 
1851 	return cumulative_error;
1852 }
1853 
1854 /*
1855  * Sync the entire filesystem; this is called from the filesystem syncer
1856  * process periodically and whenever a user calls sync(1) on the hammer
1857  * mountpoint.
1858  *
1859  * Currently is actually called from the syncer! \o/
1860  *
1861  * This task will have to snapshot the state of the dirty inode chain.
1862  * From that, it will have to make sure all of the inodes on the dirty
1863  * chain have IO initiated. We make sure that io is initiated for the root
1864  * block.
1865  *
1866  * If waitfor is set, we wait for media to acknowledge the new rootblock.
1867  *
1868  * THINKS: side A vs side B, to have sync not stall all I/O?
1869  */
1870 int
1871 hammer2_vfs_sync(struct mount *mp, int waitfor)
1872 {
1873 	struct hammer2_sync_info info;
1874 	hammer2_chain_t *chain;
1875 	hammer2_pfsmount_t *pmp;
1876 	hammer2_mount_t *hmp;
1877 	int flags;
1878 	int error;
1879 	int total_error;
1880 	int force_fchain;
1881 	int i;
1882 
1883 	pmp = MPTOPMP(mp);
1884 
1885 	/*
1886 	 * We can't acquire locks on existing vnodes while in a transaction
1887 	 * without risking a deadlock.  This assumes that vfsync() can be
1888 	 * called without the vnode locked (which it can in DragonFly).
1889 	 * Otherwise we'd have to implement a multi-pass or flag the lock
1890 	 * failures and retry.
1891 	 *
1892 	 * The reclamation code interlocks with the sync list's token
1893 	 * (by removing the vnode from the scan list) before unlocking
1894 	 * the inode, giving us time to ref the inode.
1895 	 */
1896 	/*flags = VMSC_GETVP;*/
1897 	flags = 0;
1898 	if (waitfor & MNT_LAZY)
1899 		flags |= VMSC_ONEPASS;
1900 
1901 	/*
1902 	 * Initialize a normal transaction and sync everything out, then
1903 	 * wait for pending I/O to finish (so it gets a transaction id
1904 	 * that the meta-data flush will catch).
1905 	 */
1906 	hammer2_trans_init(&info.trans, pmp, NULL, 0);
1907 	info.error = 0;
1908 	info.waitfor = MNT_NOWAIT;
1909 	vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
1910 
1911 	if (info.error == 0 && (waitfor & MNT_WAIT)) {
1912 		info.waitfor = waitfor;
1913 		    vsyncscan(mp, flags, hammer2_sync_scan2, &info);
1914 
1915 	}
1916 	hammer2_trans_done(&info.trans);
1917 	hammer2_bioq_sync(info.trans.pmp);
1918 
1919 	/*
1920 	 * Start the flush transaction and flush all meta-data.
1921 	 */
1922 	hammer2_trans_init(&info.trans, pmp, NULL, HAMMER2_TRANS_ISFLUSH);
1923 
1924 	total_error = 0;
1925 	for (i = 0; i < pmp->cluster.nchains; ++i) {
1926 		hmp = pmp->cluster.chains[i]->hmp;
1927 
1928 		/*
1929 		 * Media mounts have two 'roots', vchain for the topology
1930 		 * and fchain for the free block table.  Flush both.
1931 		 *
1932 		 * Note that the topology and free block table are handled
1933 		 * independently, so the free block table can wind up being
1934 		 * ahead of the topology.  We depend on the bulk free scan
1935 		 * code to deal with any loose ends.
1936 		 */
1937 #if 1
1938 		hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1939 		if ((hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) ||
1940 		    hmp->fchain.core->update_hi > hmp->voldata.freemap_tid) {
1941 			/* this will also modify vchain as a side effect */
1942 			chain = &hmp->fchain;
1943 			hammer2_chain_flush(&info.trans, &chain);
1944 			KKASSERT(chain == &hmp->fchain);
1945 		}
1946 		hammer2_chain_unlock(&hmp->fchain);
1947 #endif
1948 
1949 		hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1950 		if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) ||
1951 		    hmp->vchain.core->update_hi > hmp->voldata.mirror_tid) {
1952 			chain = &hmp->vchain;
1953 			hammer2_chain_flush(&info.trans, &chain);
1954 			KKASSERT(chain == &hmp->vchain);
1955 			force_fchain = 1;
1956 		} else {
1957 			force_fchain = 0;
1958 		}
1959 		hammer2_chain_unlock(&hmp->vchain);
1960 
1961 #if 0
1962 		hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1963 		if ((hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) ||
1964 		    hmp->fchain.core->update_hi > hmp->voldata.freemap_tid ||
1965 		    force_fchain) {
1966 			/* this will also modify vchain as a side effect */
1967 			chain = &hmp->fchain;
1968 			hammer2_chain_flush(&info.trans, &chain);
1969 			KKASSERT(chain == &hmp->fchain);
1970 		}
1971 		hammer2_chain_unlock(&hmp->fchain);
1972 #endif
1973 
1974 		error = 0;
1975 
1976 		/*
1977 		 * We can't safely flush the volume header until we have
1978 		 * flushed any device buffers which have built up.
1979 		 *
1980 		 * XXX this isn't being incremental
1981 		 */
1982 		vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
1983 		error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
1984 		vn_unlock(hmp->devvp);
1985 
1986 		/*
1987 		 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
1988 		 * volume header needs synchronization via hmp->volsync.
1989 		 *
1990 		 * XXX synchronize the flag & data with only this flush XXX
1991 		 */
1992 		if (error == 0 &&
1993 		    (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
1994 			struct buf *bp;
1995 
1996 			/*
1997 			 * Synchronize the disk before flushing the volume
1998 			 * header.
1999 			 */
2000 			bp = getpbuf(NULL);
2001 			bp->b_bio1.bio_offset = 0;
2002 			bp->b_bufsize = 0;
2003 			bp->b_bcount = 0;
2004 			bp->b_cmd = BUF_CMD_FLUSH;
2005 			bp->b_bio1.bio_done = biodone_sync;
2006 			bp->b_bio1.bio_flags |= BIO_SYNC;
2007 			vn_strategy(hmp->devvp, &bp->b_bio1);
2008 			biowait(&bp->b_bio1, "h2vol");
2009 			relpbuf(bp, NULL);
2010 
2011 			/*
2012 			 * Then we can safely flush the version of the
2013 			 * volume header synchronized by the flush code.
2014 			 */
2015 			i = hmp->volhdrno + 1;
2016 			if (i >= HAMMER2_NUM_VOLHDRS)
2017 				i = 0;
2018 			if (i * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
2019 			    hmp->volsync.volu_size) {
2020 				i = 0;
2021 			}
2022 			kprintf("sync volhdr %d %jd\n",
2023 				i, (intmax_t)hmp->volsync.volu_size);
2024 			bp = getblk(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2025 				    HAMMER2_PBUFSIZE, 0, 0);
2026 			atomic_clear_int(&hmp->vchain.flags,
2027 					 HAMMER2_CHAIN_VOLUMESYNC);
2028 			bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
2029 			bawrite(bp);
2030 			hmp->volhdrno = i;
2031 		}
2032 		if (error)
2033 			total_error = error;
2034 	}
2035 	hammer2_trans_done(&info.trans);
2036 
2037 	return (total_error);
2038 }
2039 
2040 /*
2041  * Sync passes.
2042  *
2043  * NOTE: We don't test update_lo/update_hi or MOVED here because the fsync
2044  *	 code won't flush on those flags.  The syncer code above will do a
2045  *	 general meta-data flush globally that will catch these flags.
2046  */
2047 
2048 static int
2049 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2050 {
2051 	struct hammer2_sync_info *info = data;
2052 	hammer2_inode_t *ip;
2053 	int error;
2054 
2055 	/*
2056 	 *
2057 	 */
2058 	ip = VTOI(vp);
2059 	if (ip == NULL)
2060 		return(0);
2061 	if (vp->v_type == VNON || vp->v_type == VBAD) {
2062 		vclrisdirty(vp);
2063 		return(0);
2064 	}
2065 	if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2066 	    RB_EMPTY(&vp->v_rbdirty_tree)) {
2067 		vclrisdirty(vp);
2068 		return(0);
2069 	}
2070 
2071 	/*
2072 	 * VOP_FSYNC will start a new transaction so replicate some code
2073 	 * here to do it inline (see hammer2_vop_fsync()).
2074 	 *
2075 	 * WARNING: The vfsync interacts with the buffer cache and might
2076 	 *          block, we can't hold the inode lock at that time.
2077 	 *	    However, we MUST ref ip before blocking to ensure that
2078 	 *	    it isn't ripped out from under us (since we do not
2079 	 *	    hold a lock on the vnode).
2080 	 */
2081 	hammer2_inode_ref(ip);
2082 	atomic_clear_int(&ip->flags, HAMMER2_INODE_MODIFIED);
2083 	if (vp)
2084 		vfsync(vp, MNT_NOWAIT, 1, NULL, NULL);
2085 
2086 #if 0
2087 	/*
2088 	 * XXX this interferes with flush operations mainly because the
2089 	 *     same transaction id is being used by asynchronous buffer
2090 	 *     operations above and can be reordered after the flush
2091 	 *     below.
2092 	 */
2093 	parent = hammer2_inode_lock_ex(ip);
2094 	hammer2_chain_flush(&info->trans, &parent);
2095 	hammer2_inode_unlock_ex(ip, parent);
2096 #endif
2097 	hammer2_inode_drop(ip);
2098 	error = 0;
2099 #if 0
2100 	error = VOP_FSYNC(vp, MNT_NOWAIT, 0);
2101 #endif
2102 	if (error)
2103 		info->error = error;
2104 	return(0);
2105 }
2106 
2107 static
2108 int
2109 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2110 {
2111 	return (0);
2112 }
2113 
2114 static
2115 int
2116 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2117 	       struct fid *fhp, struct vnode **vpp)
2118 {
2119 	return (0);
2120 }
2121 
2122 static
2123 int
2124 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2125 		 int *exflagsp, struct ucred **credanonp)
2126 {
2127 	return (0);
2128 }
2129 
2130 /*
2131  * Support code for hammer2_mount().  Read, verify, and install the volume
2132  * header into the HMP
2133  *
2134  * XXX read four volhdrs and use the one with the highest TID whos CRC
2135  *     matches.
2136  *
2137  * XXX check iCRCs.
2138  *
2139  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2140  *     nonexistant locations.
2141  *
2142  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2143  */
2144 static
2145 int
2146 hammer2_install_volume_header(hammer2_mount_t *hmp)
2147 {
2148 	hammer2_volume_data_t *vd;
2149 	struct buf *bp;
2150 	hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2151 	int error_reported;
2152 	int error;
2153 	int valid;
2154 	int i;
2155 
2156 	error_reported = 0;
2157 	error = 0;
2158 	valid = 0;
2159 	bp = NULL;
2160 
2161 	/*
2162 	 * There are up to 4 copies of the volume header (syncs iterate
2163 	 * between them so there is no single master).  We don't trust the
2164 	 * volu_size field so we don't know precisely how large the filesystem
2165 	 * is, so depend on the OS to return an error if we go beyond the
2166 	 * block device's EOF.
2167 	 */
2168 	for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2169 		error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2170 			      HAMMER2_VOLUME_BYTES, &bp);
2171 		if (error) {
2172 			brelse(bp);
2173 			bp = NULL;
2174 			continue;
2175 		}
2176 
2177 		vd = (struct hammer2_volume_data *) bp->b_data;
2178 		if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2179 		    (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2180 			brelse(bp);
2181 			bp = NULL;
2182 			continue;
2183 		}
2184 
2185 		if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2186 			/* XXX: Reversed-endianness filesystem */
2187 			kprintf("hammer2: reverse-endian filesystem detected");
2188 			brelse(bp);
2189 			bp = NULL;
2190 			continue;
2191 		}
2192 
2193 		crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2194 		crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2195 				      HAMMER2_VOLUME_ICRC0_SIZE);
2196 		bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2197 		bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2198 				       HAMMER2_VOLUME_ICRC1_SIZE);
2199 		if ((crc0 != crc) || (bcrc0 != bcrc)) {
2200 			kprintf("hammer2 volume header crc "
2201 				"mismatch copy #%d %08x/%08x\n",
2202 				i, crc0, crc);
2203 			error_reported = 1;
2204 			brelse(bp);
2205 			bp = NULL;
2206 			continue;
2207 		}
2208 		if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2209 			valid = 1;
2210 			hmp->voldata = *vd;
2211 			hmp->volhdrno = i;
2212 		}
2213 		brelse(bp);
2214 		bp = NULL;
2215 	}
2216 	if (valid) {
2217 		hmp->volsync = hmp->voldata;
2218 		error = 0;
2219 		if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2220 			kprintf("hammer2: using volume header #%d\n",
2221 				hmp->volhdrno);
2222 		}
2223 	} else {
2224 		error = EINVAL;
2225 		kprintf("hammer2: no valid volume headers found!\n");
2226 	}
2227 	return (error);
2228 }
2229 
2230 /*
2231  * Reconnect using the passed file pointer.  The caller must ref the
2232  * fp for us.
2233  */
2234 void
2235 hammer2_cluster_reconnect(hammer2_pfsmount_t *pmp, struct file *fp)
2236 {
2237 	hammer2_inode_data_t *ipdata;
2238 	hammer2_chain_t *parent;
2239 	hammer2_mount_t *hmp;
2240 	size_t name_len;
2241 
2242 	hmp = pmp->cluster.chains[0]->hmp;	/* XXX */
2243 
2244 	/*
2245 	 * Closes old comm descriptor, kills threads, cleans up
2246 	 * states, then installs the new descriptor and creates
2247 	 * new threads.
2248 	 */
2249 	kdmsg_iocom_reconnect(&pmp->iocom, fp, "hammer2");
2250 
2251 	/*
2252 	 * Setup LNK_CONN fields for autoinitiated state machine
2253 	 */
2254 	parent = hammer2_inode_lock_ex(pmp->iroot);
2255 	ipdata = &parent->data->ipdata;
2256 	pmp->iocom.auto_lnk_conn.pfs_clid = ipdata->pfs_clid;
2257 	pmp->iocom.auto_lnk_conn.pfs_fsid = ipdata->pfs_fsid;
2258 	pmp->iocom.auto_lnk_conn.pfs_type = ipdata->pfs_type;
2259 	pmp->iocom.auto_lnk_conn.proto_version = DMSG_SPAN_PROTO_1;
2260 	pmp->iocom.auto_lnk_conn.peer_type = hmp->voldata.peer_type;
2261 
2262 	/*
2263 	 * Filter adjustment.  Clients do not need visibility into other
2264 	 * clients (otherwise millions of clients would present a serious
2265 	 * problem).  The fs_label also serves to restrict the namespace.
2266 	 */
2267 	pmp->iocom.auto_lnk_conn.peer_mask = 1LLU << HAMMER2_PEER_HAMMER2;
2268 	pmp->iocom.auto_lnk_conn.pfs_mask = (uint64_t)-1;
2269 	switch (ipdata->pfs_type) {
2270 	case DMSG_PFSTYPE_CLIENT:
2271 		pmp->iocom.auto_lnk_conn.peer_mask &=
2272 				~(1LLU << DMSG_PFSTYPE_CLIENT);
2273 		break;
2274 	default:
2275 		break;
2276 	}
2277 
2278 	name_len = ipdata->name_len;
2279 	if (name_len >= sizeof(pmp->iocom.auto_lnk_conn.fs_label))
2280 		name_len = sizeof(pmp->iocom.auto_lnk_conn.fs_label) - 1;
2281 	bcopy(ipdata->filename,
2282 	      pmp->iocom.auto_lnk_conn.fs_label,
2283 	      name_len);
2284 	pmp->iocom.auto_lnk_conn.fs_label[name_len] = 0;
2285 
2286 	/*
2287 	 * Setup LNK_SPAN fields for autoinitiated state machine
2288 	 */
2289 	pmp->iocom.auto_lnk_span.pfs_clid = ipdata->pfs_clid;
2290 	pmp->iocom.auto_lnk_span.pfs_fsid = ipdata->pfs_fsid;
2291 	pmp->iocom.auto_lnk_span.pfs_type = ipdata->pfs_type;
2292 	pmp->iocom.auto_lnk_span.peer_type = hmp->voldata.peer_type;
2293 	pmp->iocom.auto_lnk_span.proto_version = DMSG_SPAN_PROTO_1;
2294 	name_len = ipdata->name_len;
2295 	if (name_len >= sizeof(pmp->iocom.auto_lnk_span.fs_label))
2296 		name_len = sizeof(pmp->iocom.auto_lnk_span.fs_label) - 1;
2297 	bcopy(ipdata->filename,
2298 	      pmp->iocom.auto_lnk_span.fs_label,
2299 	      name_len);
2300 	pmp->iocom.auto_lnk_span.fs_label[name_len] = 0;
2301 	hammer2_inode_unlock_ex(pmp->iroot, parent);
2302 
2303 	kdmsg_iocom_autoinitiate(&pmp->iocom, hammer2_autodmsg);
2304 }
2305 
2306 static int
2307 hammer2_rcvdmsg(kdmsg_msg_t *msg)
2308 {
2309 	switch(msg->any.head.cmd & DMSGF_TRANSMASK) {
2310 	case DMSG_DBG_SHELL:
2311 		/*
2312 		 * (non-transaction)
2313 		 * Execute shell command (not supported atm)
2314 		 */
2315 		kdmsg_msg_reply(msg, DMSG_ERR_NOSUPP);
2316 		break;
2317 	case DMSG_DBG_SHELL | DMSGF_REPLY:
2318 		/*
2319 		 * (non-transaction)
2320 		 */
2321 		if (msg->aux_data) {
2322 			msg->aux_data[msg->aux_size - 1] = 0;
2323 			kprintf("HAMMER2 DBG: %s\n", msg->aux_data);
2324 		}
2325 		break;
2326 	default:
2327 		/*
2328 		 * Unsupported message received.  We only need to
2329 		 * reply if it's a transaction in order to close our end.
2330 		 * Ignore any one-way messages are any further messages
2331 		 * associated with the transaction.
2332 		 *
2333 		 * NOTE: This case also includes DMSG_LNK_ERROR messages
2334 		 *	 which might be one-way, replying to those would
2335 		 *	 cause an infinite ping-pong.
2336 		 */
2337 		if (msg->any.head.cmd & DMSGF_CREATE)
2338 			kdmsg_msg_reply(msg, DMSG_ERR_NOSUPP);
2339 		break;
2340 	}
2341 	return(0);
2342 }
2343 
2344 /*
2345  * This function is called after KDMSG has automatically handled processing
2346  * of a LNK layer message (typically CONN, SPAN, or CIRC).
2347  *
2348  * We tag off the LNK_CONN to trigger our LNK_VOLCONF messages which
2349  * advertises all available hammer2 super-root volumes.
2350  */
2351 static void
2352 hammer2_autodmsg(kdmsg_msg_t *msg)
2353 {
2354 	hammer2_pfsmount_t *pmp = msg->iocom->handle;
2355 	hammer2_mount_t *hmp = pmp->cluster.chains[0]->hmp; /* XXX */
2356 	int copyid;
2357 
2358 	/*
2359 	 * We only care about replies to our LNK_CONN auto-request.  kdmsg
2360 	 * has already processed the reply, we use this calback as a shim
2361 	 * to know when we can advertise available super-root volumes.
2362 	 */
2363 	if ((msg->any.head.cmd & DMSGF_TRANSMASK) !=
2364 	    (DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_REPLY) ||
2365 	    msg->state == NULL) {
2366 		return;
2367 	}
2368 
2369 	kprintf("LNK_CONN REPLY RECEIVED CMD %08x\n", msg->any.head.cmd);
2370 
2371 	if (msg->any.head.cmd & DMSGF_CREATE) {
2372 		kprintf("HAMMER2: VOLDATA DUMP\n");
2373 
2374 		/*
2375 		 * Dump the configuration stored in the volume header
2376 		 */
2377 		hammer2_voldata_lock(hmp);
2378 		for (copyid = 0; copyid < HAMMER2_COPYID_COUNT; ++copyid) {
2379 			if (hmp->voldata.copyinfo[copyid].copyid == 0)
2380 				continue;
2381 			hammer2_volconf_update(pmp, copyid);
2382 		}
2383 		hammer2_voldata_unlock(hmp, 0);
2384 	}
2385 	if ((msg->any.head.cmd & DMSGF_DELETE) &&
2386 	    msg->state && (msg->state->txcmd & DMSGF_DELETE) == 0) {
2387 		kprintf("HAMMER2: CONN WAS TERMINATED\n");
2388 	}
2389 }
2390 
2391 /*
2392  * Volume configuration updates are passed onto the userland service
2393  * daemon via the open LNK_CONN transaction.
2394  */
2395 void
2396 hammer2_volconf_update(hammer2_pfsmount_t *pmp, int index)
2397 {
2398 	hammer2_mount_t *hmp = pmp->cluster.chains[0]->hmp;	/* XXX */
2399 	kdmsg_msg_t *msg;
2400 
2401 	/* XXX interlock against connection state termination */
2402 	kprintf("volconf update %p\n", pmp->iocom.conn_state);
2403 	if (pmp->iocom.conn_state) {
2404 		kprintf("TRANSMIT VOLCONF VIA OPEN CONN TRANSACTION\n");
2405 		msg = kdmsg_msg_alloc_state(pmp->iocom.conn_state,
2406 					    DMSG_LNK_VOLCONF, NULL, NULL);
2407 		msg->any.lnk_volconf.copy = hmp->voldata.copyinfo[index];
2408 		msg->any.lnk_volconf.mediaid = hmp->voldata.fsid;
2409 		msg->any.lnk_volconf.index = index;
2410 		kdmsg_msg_write(msg);
2411 	}
2412 }
2413 
2414 /*
2415  * This handles hysteresis on regular file flushes.  Because the BIOs are
2416  * routed to a thread it is possible for an excessive number to build up
2417  * and cause long front-end stalls long before the runningbuffspace limit
2418  * is hit, so we implement hammer2_flush_pipe to control the
2419  * hysteresis.
2420  *
2421  * This is a particular problem when compression is used.
2422  */
2423 void
2424 hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp)
2425 {
2426 	atomic_add_int(&pmp->count_lwinprog, 1);
2427 }
2428 
2429 void
2430 hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp)
2431 {
2432 	int lwinprog;
2433 
2434 	lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2435 	if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2436 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2437 		atomic_clear_int(&pmp->count_lwinprog,
2438 				 HAMMER2_LWINPROG_WAITING);
2439 		wakeup(&pmp->count_lwinprog);
2440 	}
2441 }
2442 
2443 void
2444 hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp)
2445 {
2446 	int lwinprog;
2447 
2448 	for (;;) {
2449 		lwinprog = pmp->count_lwinprog;
2450 		cpu_ccfence();
2451 		if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2452 			break;
2453 		tsleep_interlock(&pmp->count_lwinprog, 0);
2454 		atomic_set_int(&pmp->count_lwinprog, HAMMER2_LWINPROG_WAITING);
2455 		lwinprog = pmp->count_lwinprog;
2456 		if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2457 			break;
2458 		tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2459 	}
2460 }
2461 
2462 void
2463 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp)
2464 {
2465 	hammer2_chain_layer_t *layer;
2466 	hammer2_chain_t *scan;
2467 	hammer2_chain_t *first_parent;
2468 
2469 	--*countp;
2470 	if (*countp == 0) {
2471 		kprintf("%*.*s...\n", tab, tab, "");
2472 		return;
2473 	}
2474 	if (*countp < 0)
2475 		return;
2476 	first_parent = chain->core ? TAILQ_FIRST(&chain->core->ownerq) : NULL;
2477 	kprintf("%*.*schain %p.%d %016jx/%d mir=%016jx\n",
2478 		tab, tab, "",
2479 		chain, chain->bref.type,
2480 		chain->bref.key, chain->bref.keybits,
2481 		chain->bref.mirror_tid);
2482 
2483 	kprintf("%*.*s      [%08x] (%s) dt=%016jx refs=%d\n",
2484 		tab, tab, "",
2485 		chain->flags,
2486 		((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2487 		chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2488 		chain->delete_tid,
2489 		chain->refs);
2490 
2491 	kprintf("%*.*s      core %p [%08x] lo=%08jx hi=%08jx fp=%p np=%p",
2492 		tab, tab, "",
2493 		chain->core, (chain->core ? chain->core->flags : 0),
2494 		(chain->core ? chain->core->update_lo : -1),
2495 		(chain->core ? chain->core->update_hi : -1),
2496 		first_parent,
2497 		(first_parent ? TAILQ_NEXT(chain, core_entry) : NULL));
2498 
2499 	if (first_parent)
2500 		kprintf(" [fpflags %08x fprefs %d\n",
2501 			first_parent->flags,
2502 			first_parent->refs);
2503 	if (chain->core == NULL || TAILQ_EMPTY(&chain->core->layerq))
2504 		kprintf("\n");
2505 	else
2506 		kprintf(" {\n");
2507 	if (chain->core) {
2508 		TAILQ_FOREACH(layer, &chain->core->layerq, entry) {
2509 			RB_FOREACH(scan, hammer2_chain_tree, &layer->rbtree) {
2510 				hammer2_dump_chain(scan, tab + 4, countp);
2511 			}
2512 		}
2513 	}
2514 	if (chain->core && !TAILQ_EMPTY(&chain->core->layerq)) {
2515 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2516 			kprintf("%*.*s}(%s)\n", tab, tab, "",
2517 				chain->data->ipdata.filename);
2518 		else
2519 			kprintf("%*.*s}\n", tab, tab, "");
2520 	}
2521 }
2522