xref: /dflybsd-src/sys/vfs/hammer2/hammer2_vfsops.c (revision a413fe45675e6d132bfaa7a0a089b5a6e670bb07)
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_mount);
76 TAILQ_HEAD(hammer2_pfslist, hammer2_pfsmount);
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(C_BUFFER);
110 MALLOC_DEFINE(C_BUFFER, "compbuffer", "Buffer used for compression.");
111 
112 MALLOC_DECLARE(D_BUFFER);
113 MALLOC_DEFINE(D_BUFFER, "decompbuffer", "Buffer used for decompression.");
114 
115 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
116 
117 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
118 	   &hammer2_debug, 0, "");
119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW,
120 	   &hammer2_cluster_enable, 0, "");
121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
122 	   &hammer2_hardlink_enable, 0, "");
123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
124 	   &hammer2_flush_pipe, 0, "");
125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
126 	   &hammer2_synchronous_flush, 0, "");
127 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
128 	   &hammer2_limit_dirty_chains, 0, "");
129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
130 	   &hammer2_dio_count, 0, "");
131 
132 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
133 	   &hammer2_iod_file_read, 0, "");
134 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
135 	   &hammer2_iod_meta_read, 0, "");
136 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
137 	   &hammer2_iod_indr_read, 0, "");
138 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
139 	   &hammer2_iod_fmap_read, 0, "");
140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
141 	   &hammer2_iod_volu_read, 0, "");
142 
143 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
144 	   &hammer2_iod_file_write, 0, "");
145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
146 	   &hammer2_iod_meta_write, 0, "");
147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
148 	   &hammer2_iod_indr_write, 0, "");
149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
150 	   &hammer2_iod_fmap_write, 0, "");
151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
152 	   &hammer2_iod_volu_write, 0, "");
153 
154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
155 	   &hammer2_ioa_file_read, 0, "");
156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
157 	   &hammer2_ioa_meta_read, 0, "");
158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
159 	   &hammer2_ioa_indr_read, 0, "");
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
161 	   &hammer2_ioa_fmap_read, 0, "");
162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
163 	   &hammer2_ioa_volu_read, 0, "");
164 
165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
166 	   &hammer2_ioa_file_write, 0, "");
167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
168 	   &hammer2_ioa_meta_write, 0, "");
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
170 	   &hammer2_ioa_indr_write, 0, "");
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
172 	   &hammer2_ioa_fmap_write, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
174 	   &hammer2_ioa_volu_write, 0, "");
175 
176 static int hammer2_vfs_init(struct vfsconf *conf);
177 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
178 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
179 				struct ucred *cred);
180 static int hammer2_remount(hammer2_mount_t *, struct mount *, char *,
181 				struct vnode *, struct ucred *);
182 static int hammer2_recovery(hammer2_mount_t *hmp);
183 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
184 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
185 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
186 				struct ucred *cred);
187 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
188 				struct ucred *cred);
189 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
190 				ino_t ino, struct vnode **vpp);
191 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
192 				struct fid *fhp, struct vnode **vpp);
193 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
194 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
195 				int *exflagsp, struct ucred **credanonp);
196 
197 static int hammer2_install_volume_header(hammer2_mount_t *hmp);
198 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
199 
200 static void hammer2_write_thread(void *arg);
201 
202 static void hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp);
203 static void hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp);
204 
205 /*
206  * Functions for compression in threads,
207  * from hammer2_vnops.c
208  */
209 static void hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
210 				hammer2_inode_t *ip,
211 				const hammer2_inode_data_t *ripdata,
212 				hammer2_cluster_t *cparent,
213 				hammer2_key_t lbase, int ioflag, int pblksize,
214 				int *errorp);
215 static void hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
216 				hammer2_inode_t *ip,
217 				const hammer2_inode_data_t *ripdata,
218 				hammer2_cluster_t *cparent,
219 				hammer2_key_t lbase, int ioflag,
220 				int pblksize, int *errorp,
221 				int comp_algo, int check_algo);
222 static void hammer2_zero_check_and_write(struct buf *bp,
223 				hammer2_trans_t *trans, hammer2_inode_t *ip,
224 				const hammer2_inode_data_t *ripdata,
225 				hammer2_cluster_t *cparent,
226 				hammer2_key_t lbase,
227 				int ioflag, int pblksize, int *errorp,
228 				int check_algo);
229 static int test_block_zeros(const char *buf, size_t bytes);
230 static void zero_write(struct buf *bp, hammer2_trans_t *trans,
231 				hammer2_inode_t *ip,
232 				const hammer2_inode_data_t *ripdata,
233 				hammer2_cluster_t *cparent,
234 				hammer2_key_t lbase,
235 				int *errorp);
236 static void hammer2_write_bp(hammer2_cluster_t *cluster, struct buf *bp,
237 				int ioflag, int pblksize, int *errorp,
238 				int check_algo);
239 
240 /*
241  * HAMMER2 vfs operations.
242  */
243 static struct vfsops hammer2_vfsops = {
244 	.vfs_init	= hammer2_vfs_init,
245 	.vfs_uninit	= hammer2_vfs_uninit,
246 	.vfs_sync	= hammer2_vfs_sync,
247 	.vfs_mount	= hammer2_vfs_mount,
248 	.vfs_unmount	= hammer2_vfs_unmount,
249 	.vfs_root 	= hammer2_vfs_root,
250 	.vfs_statfs	= hammer2_vfs_statfs,
251 	.vfs_statvfs	= hammer2_vfs_statvfs,
252 	.vfs_vget	= hammer2_vfs_vget,
253 	.vfs_vptofh	= hammer2_vfs_vptofh,
254 	.vfs_fhtovp	= hammer2_vfs_fhtovp,
255 	.vfs_checkexp	= hammer2_vfs_checkexp
256 };
257 
258 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
259 
260 VFS_SET(hammer2_vfsops, hammer2, 0);
261 MODULE_VERSION(hammer2, 1);
262 
263 static
264 int
265 hammer2_vfs_init(struct vfsconf *conf)
266 {
267 	static struct objcache_malloc_args margs_read;
268 	static struct objcache_malloc_args margs_write;
269 
270 	int error;
271 
272 	error = 0;
273 
274 	if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
275 		error = EINVAL;
276 	if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
277 		error = EINVAL;
278 	if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
279 		error = EINVAL;
280 
281 	if (error)
282 		kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
283 
284 	margs_read.objsize = 65536;
285 	margs_read.mtype = D_BUFFER;
286 
287 	margs_write.objsize = 32768;
288 	margs_write.mtype = C_BUFFER;
289 
290 	cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
291 				0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
292 				objcache_malloc_free, &margs_read);
293 	cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
294 				0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
295 				objcache_malloc_free, &margs_write);
296 
297 	lockinit(&hammer2_mntlk, "mntlk", 0, 0);
298 	TAILQ_INIT(&hammer2_mntlist);
299 	TAILQ_INIT(&hammer2_pfslist);
300 
301 	hammer2_limit_dirty_chains = desiredvnodes / 10;
302 
303 	hammer2_trans_manage_init();
304 
305 	return (error);
306 }
307 
308 static
309 int
310 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
311 {
312 	objcache_destroy(cache_buffer_read);
313 	objcache_destroy(cache_buffer_write);
314 	return 0;
315 }
316 
317 /*
318  * Core PFS allocator.  Used to allocate the pmp structure for PFS cluster
319  * mounts and the spmp structure for media (hmp) structures.
320  */
321 static hammer2_pfsmount_t *
322 hammer2_pfsalloc(const hammer2_inode_data_t *ripdata, hammer2_tid_t alloc_tid)
323 {
324 	hammer2_pfsmount_t *pmp;
325 
326 	pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
327 	kmalloc_create(&pmp->minode, "HAMMER2-inodes");
328 	kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
329 	lockinit(&pmp->lock, "pfslk", 0, 0);
330 	spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
331 	RB_INIT(&pmp->inum_tree);
332 	TAILQ_INIT(&pmp->unlinkq);
333 	spin_init(&pmp->list_spin, "hm2pfsalloc_list");
334 
335 	pmp->alloc_tid = alloc_tid + 1;	  /* our first media transaction id */
336 	pmp->flush_tid = pmp->alloc_tid;
337 	if (ripdata) {
338 		pmp->inode_tid = ripdata->pfs_inum + 1;
339 		pmp->pfs_clid = ripdata->pfs_clid;
340 	}
341 	hammer2_mtx_init(&pmp->wthread_mtx, "h2wthr");
342 	bioq_init(&pmp->wthread_bioq);
343 
344 	return pmp;
345 }
346 
347 /*
348  * Mount or remount HAMMER2 fileystem from physical media
349  *
350  *	mountroot
351  *		mp		mount point structure
352  *		path		NULL
353  *		data		<unused>
354  *		cred		<unused>
355  *
356  *	mount
357  *		mp		mount point structure
358  *		path		path to mount point
359  *		data		pointer to argument structure in user space
360  *			volume	volume path (device@LABEL form)
361  *			hflags	user mount flags
362  *		cred		user credentials
363  *
364  * RETURNS:	0	Success
365  *		!0	error number
366  */
367 static
368 int
369 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
370 		  struct ucred *cred)
371 {
372 	struct hammer2_mount_info info;
373 	hammer2_pfsmount_t *pmp;
374 	hammer2_pfsmount_t *spmp;
375 	hammer2_mount_t *hmp;
376 	hammer2_key_t key_next;
377 	hammer2_key_t key_dummy;
378 	hammer2_key_t lhc;
379 	struct vnode *devvp;
380 	struct nlookupdata nd;
381 	hammer2_chain_t *parent;
382 	hammer2_chain_t *rchain;
383 	hammer2_cluster_t *cluster;
384 	hammer2_cluster_t *cparent;
385 	const hammer2_inode_data_t *ripdata;
386 	hammer2_blockref_t bref;
387 	struct file *fp;
388 	char devstr[MNAMELEN];
389 	size_t size;
390 	size_t done;
391 	char *dev;
392 	char *label;
393 	int ronly = 1;
394 	int error;
395 	int cache_index;
396 	int ddflag;
397 	int i;
398 
399 	hmp = NULL;
400 	pmp = NULL;
401 	dev = NULL;
402 	label = NULL;
403 	devvp = NULL;
404 	cache_index = -1;
405 
406 	kprintf("hammer2_mount\n");
407 
408 	if (path == NULL) {
409 		/*
410 		 * Root mount
411 		 */
412 		bzero(&info, sizeof(info));
413 		info.cluster_fd = -1;
414 		return (EOPNOTSUPP);
415 	} else {
416 		/*
417 		 * Non-root mount or updating a mount
418 		 */
419 		error = copyin(data, &info, sizeof(info));
420 		if (error)
421 			return (error);
422 
423 		error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
424 		if (error)
425 			return (error);
426 
427 		/* Extract device and label */
428 		dev = devstr;
429 		label = strchr(devstr, '@');
430 		if (label == NULL ||
431 		    ((label + 1) - dev) > done) {
432 			return (EINVAL);
433 		}
434 		*label = '\0';
435 		label++;
436 		if (*label == '\0')
437 			return (EINVAL);
438 
439 		if (mp->mnt_flag & MNT_UPDATE) {
440 			/*
441 			 * Update mount.  Note that pmp->iroot->cluster is
442 			 * an inode-embedded cluster and thus cannot be
443 			 * directly locked.
444 			 *
445 			 * XXX HAMMER2 needs to implement NFS export via
446 			 *     mountctl.
447 			 */
448 			pmp = MPTOPMP(mp);
449 			cluster = &pmp->iroot->cluster;
450 			for (i = 0; i < cluster->nchains; ++i) {
451 				hmp = cluster->array[i].chain->hmp;
452 				devvp = hmp->devvp;
453 				error = hammer2_remount(hmp, mp, path,
454 							devvp, cred);
455 				if (error)
456 					break;
457 			}
458 			/*hammer2_inode_install_hidden(pmp);*/
459 
460 			return error;
461 		}
462 	}
463 
464 	/*
465 	 * HMP device mount
466 	 *
467 	 * Lookup name and verify it refers to a block device.
468 	 */
469 	error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
470 	if (error == 0)
471 		error = nlookup(&nd);
472 	if (error == 0)
473 		error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
474 	nlookup_done(&nd);
475 
476 	if (error == 0) {
477 		if (vn_isdisk(devvp, &error))
478 			error = vfs_mountedon(devvp);
479 	}
480 
481 	/*
482 	 * Determine if the device has already been mounted.  After this
483 	 * check hmp will be non-NULL if we are doing the second or more
484 	 * hammer2 mounts from the same device.
485 	 */
486 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
487 	TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
488 		if (hmp->devvp == devvp)
489 			break;
490 	}
491 
492 	/*
493 	 * Open the device if this isn't a secondary mount and construct
494 	 * the H2 device mount (hmp).
495 	 */
496 	if (hmp == NULL) {
497 		hammer2_chain_t *schain;
498 		hammer2_xid_t xid;
499 
500 		if (error == 0 && vcount(devvp) > 0)
501 			error = EBUSY;
502 
503 		/*
504 		 * Now open the device
505 		 */
506 		if (error == 0) {
507 			ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
508 			vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
509 			error = vinvalbuf(devvp, V_SAVE, 0, 0);
510 			if (error == 0) {
511 				error = VOP_OPEN(devvp,
512 						 ronly ? FREAD : FREAD | FWRITE,
513 						 FSCRED, NULL);
514 			}
515 			vn_unlock(devvp);
516 		}
517 		if (error && devvp) {
518 			vrele(devvp);
519 			devvp = NULL;
520 		}
521 		if (error) {
522 			lockmgr(&hammer2_mntlk, LK_RELEASE);
523 			return error;
524 		}
525 		hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
526 		hmp->ronly = ronly;
527 		hmp->devvp = devvp;
528 		kmalloc_create(&hmp->mchain, "HAMMER2-chains");
529 		TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
530 		RB_INIT(&hmp->iotree);
531 		spin_init(&hmp->io_spin, "hm2mount_io");
532 		spin_init(&hmp->list_spin, "hm2mount_list");
533 		TAILQ_INIT(&hmp->flushq);
534 
535 		lockinit(&hmp->vollk, "h2vol", 0, 0);
536 
537 		/*
538 		 * vchain setup. vchain.data is embedded.
539 		 * vchain.refs is initialized and will never drop to 0.
540 		 *
541 		 * NOTE! voldata is not yet loaded.
542 		 */
543 		hmp->vchain.hmp = hmp;
544 		hmp->vchain.refs = 1;
545 		hmp->vchain.data = (void *)&hmp->voldata;
546 		hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
547 		hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
548 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
549 
550 		hammer2_chain_core_alloc(NULL, &hmp->vchain);
551 		/* hmp->vchain.u.xxx is left NULL */
552 
553 		/*
554 		 * fchain setup.  fchain.data is embedded.
555 		 * fchain.refs is initialized and will never drop to 0.
556 		 *
557 		 * The data is not used but needs to be initialized to
558 		 * pass assertion muster.  We use this chain primarily
559 		 * as a placeholder for the freemap's top-level RBTREE
560 		 * so it does not interfere with the volume's topology
561 		 * RBTREE.
562 		 */
563 		hmp->fchain.hmp = hmp;
564 		hmp->fchain.refs = 1;
565 		hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
566 		hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
567 		hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
568 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
569 		hmp->fchain.bref.methods =
570 			HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
571 			HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
572 
573 		hammer2_chain_core_alloc(NULL, &hmp->fchain);
574 		/* hmp->fchain.u.xxx is left NULL */
575 
576 		/*
577 		 * Install the volume header and initialize fields from
578 		 * voldata.
579 		 */
580 		error = hammer2_install_volume_header(hmp);
581 		if (error) {
582 			++hmp->pmp_count;
583 			hammer2_vfs_unmount_hmp1(mp, hmp);
584 			hammer2_vfs_unmount_hmp2(mp, hmp);
585 			lockmgr(&hammer2_mntlk, LK_RELEASE);
586 			hammer2_vfs_unmount(mp, MNT_FORCE);
587 			return error;
588 		}
589 
590 		/*
591 		 * Really important to get these right or flush will get
592 		 * confused.
593 		 */
594 		hmp->spmp = hammer2_pfsalloc(NULL, hmp->voldata.mirror_tid);
595 		kprintf("alloc spmp %p tid %016jx\n",
596 			hmp->spmp, hmp->voldata.mirror_tid);
597 		spmp = hmp->spmp;
598 		spmp->inode_tid = 1;
599 
600 		xid = 0;
601 		hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
602 		hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
603 		hmp->vchain.pmp = spmp;
604 		hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
605 		hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
606 		hmp->fchain.pmp = spmp;
607 
608 		/*
609 		 * First locate the super-root inode, which is key 0
610 		 * relative to the volume header's blockset.
611 		 *
612 		 * Then locate the root inode by scanning the directory keyspace
613 		 * represented by the label.
614 		 */
615 		parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
616 		schain = hammer2_chain_lookup(&parent, &key_dummy,
617 				      HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
618 				      &cache_index, 0, &ddflag);
619 		hammer2_chain_lookup_done(parent);
620 		if (schain == NULL) {
621 			kprintf("hammer2_mount: invalid super-root\n");
622 			++hmp->pmp_count;
623 			hammer2_vfs_unmount_hmp1(mp, hmp);
624 			hammer2_vfs_unmount_hmp2(mp, hmp);
625 			lockmgr(&hammer2_mntlk, LK_RELEASE);
626 			hammer2_vfs_unmount(mp, MNT_FORCE);
627 			return EINVAL;
628 		}
629 
630 		/*
631 		 * Sanity-check schain's pmp, finish initializing spmp.
632 		 */
633 		ripdata = &hammer2_chain_rdata(schain)->ipdata;
634 		KKASSERT(schain->pmp == spmp);
635 		spmp->pfs_clid = ripdata->pfs_clid;
636 
637 		/*
638 		 * NOTE: inode_get sucks up schain's lock.
639 		 */
640 		cluster = hammer2_cluster_from_chain(schain);
641 		spmp->iroot = hammer2_inode_get(spmp, NULL, cluster);
642 		spmp->spmp_hmp = hmp;
643 		hammer2_inode_ref(spmp->iroot);
644 		hammer2_inode_unlock_ex(spmp->iroot, cluster);
645 		schain = NULL;
646 		/* leave spmp->iroot with one ref */
647 
648 		if ((mp->mnt_flag & MNT_RDONLY) == 0) {
649 			error = hammer2_recovery(hmp);
650 			/* XXX do something with error */
651 		}
652 		++hmp->pmp_count;
653 
654 		hammer2_iocom_init(hmp);
655 
656 		/*
657 		 * Ref the cluster management messaging descriptor.  The mount
658 		 * program deals with the other end of the communications pipe.
659 		 */
660 		fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
661 		if (fp) {
662 			hammer2_cluster_reconnect(hmp, fp);
663 		} else {
664 			kprintf("hammer2_mount: bad cluster_fd!\n");
665 		}
666 	} else {
667 		spmp = hmp->spmp;
668 		++hmp->pmp_count;
669 	}
670 
671 	/*
672 	 * Lookup mount point under the media-localized super-root.
673 	 *
674 	 * cluster->pmp will incorrectly point to spmp and must be fixed
675 	 * up later on.
676 	 */
677 	cparent = hammer2_inode_lock_ex(spmp->iroot);
678 	lhc = hammer2_dirhash(label, strlen(label));
679 	cluster = hammer2_cluster_lookup(cparent, &key_next,
680 				      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
681 				      0, &ddflag);
682 	while (cluster) {
683 		if (hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE &&
684 		    strcmp(label,
685 		       hammer2_cluster_rdata(cluster)->ipdata.filename) == 0) {
686 			break;
687 		}
688 		cluster = hammer2_cluster_next(cparent, cluster, &key_next,
689 					    key_next,
690 					    lhc + HAMMER2_DIRHASH_LOMASK, 0);
691 	}
692 	hammer2_inode_unlock_ex(spmp->iroot, cparent);
693 
694 	if (cluster == NULL) {
695 		kprintf("hammer2_mount: PFS label not found\n");
696 		hammer2_vfs_unmount_hmp1(mp, hmp);
697 		hammer2_vfs_unmount_hmp2(mp, hmp);
698 		lockmgr(&hammer2_mntlk, LK_RELEASE);
699 		hammer2_vfs_unmount(mp, MNT_FORCE);
700 		return EINVAL;
701 	}
702 
703 	for (i = 0; i < cluster->nchains; ++i) {
704 		rchain = cluster->array[i].chain;
705 		if (rchain->flags & HAMMER2_CHAIN_MOUNTED) {
706 			kprintf("hammer2_mount: PFS label already mounted!\n");
707 			hammer2_cluster_unlock(cluster);
708 			hammer2_vfs_unmount_hmp1(mp, hmp);
709 			hammer2_vfs_unmount_hmp2(mp, hmp);
710 			lockmgr(&hammer2_mntlk, LK_RELEASE);
711 			hammer2_vfs_unmount(mp, MNT_FORCE);
712 			return EBUSY;
713 		}
714 		KKASSERT(rchain->pmp == NULL);
715 #if 0
716 		if (rchain->flags & HAMMER2_CHAIN_RECYCLE) {
717 			kprintf("hammer2_mount: PFS label is recycling\n");
718 			hammer2_cluster_unlock(cluster);
719 			hammer2_vfs_unmount_hmp1(mp, hmp);
720 			hammer2_vfs_unmount_hmp2(mp, hmp);
721 			lockmgr(&hammer2_mntlk, LK_RELEASE);
722 			hammer2_vfs_unmount(mp, MNT_FORCE);
723 			return EBUSY;
724 		}
725 #endif
726 	}
727 
728 	/*
729 	 * Check to see if the cluster id is already mounted at the mount
730 	 * point.  If it is, add us to the cluster.
731 	 */
732 	ripdata = &hammer2_cluster_rdata(cluster)->ipdata;
733 	hammer2_cluster_bref(cluster, &bref);
734 	TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
735 		if (pmp->spmp_hmp == NULL &&
736 		    bcmp(&pmp->pfs_clid, &ripdata->pfs_clid,
737 			 sizeof(pmp->pfs_clid)) == 0) {
738 			break;
739 		}
740 	}
741 
742 	if (pmp) {
743 		int i;
744 		int j;
745 
746 		/*
747 		 * Directly lock the inode->lock, do not run through
748 		 * hammer2_inode_lock*().
749 		 */
750 		hammer2_inode_ref(pmp->iroot);
751 		hammer2_mtx_ex(&pmp->iroot->lock);
752 
753 		if (pmp->iroot->cluster.nchains + cluster->nchains >
754 		    HAMMER2_MAXCLUSTER) {
755 			kprintf("hammer2_mount: cluster full!\n");
756 
757 			hammer2_mtx_unlock(&pmp->iroot->lock);
758 			hammer2_inode_drop(pmp->iroot);
759 
760 			hammer2_cluster_unlock(cluster);
761 			hammer2_vfs_unmount_hmp1(mp, hmp);
762 			hammer2_vfs_unmount_hmp2(mp, hmp);
763 			lockmgr(&hammer2_mntlk, LK_RELEASE);
764 			hammer2_vfs_unmount(mp, MNT_FORCE);
765 			return EBUSY;
766 		}
767 		kprintf("hammer2_vfs_mount: Adding pfs to existing cluster\n");
768 		j = pmp->iroot->cluster.nchains;
769 		for (i = 0; i < cluster->nchains; ++i) {
770 			rchain = cluster->array[i].chain;
771 			KKASSERT(rchain->pmp == NULL);
772 			rchain->pmp = pmp;
773 			hammer2_chain_ref(rchain);
774 			pmp->iroot->cluster.array[j].chain = rchain;
775 			++j;
776 		}
777 		pmp->iroot->cluster.nchains = j;
778 		hammer2_mtx_unlock(&pmp->iroot->lock);
779 		hammer2_inode_drop(pmp->iroot);
780 		hammer2_cluster_unlock(cluster);
781 		lockmgr(&hammer2_mntlk, LK_RELEASE);
782 
783 		kprintf("ok\n");
784 		hammer2_inode_install_hidden(pmp);
785 
786 		return ERANGE;
787 	}
788 
789 	/*
790 	 * Block device opened successfully, finish initializing the
791 	 * mount structure.
792 	 *
793 	 * From this point on we have to call hammer2_unmount() on failure.
794 	 */
795 	pmp = hammer2_pfsalloc(ripdata, bref.mirror_tid);
796 	kprintf("PMP mirror_tid is %016jx\n", bref.mirror_tid);
797 	for (i = 0; i < cluster->nchains; ++i) {
798 		rchain = cluster->array[i].chain;
799 		KKASSERT(rchain->pmp == NULL);
800 		rchain->pmp = pmp;
801 		atomic_set_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
802 	}
803 	cluster->pmp = pmp;
804 
805 	TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
806 	lockmgr(&hammer2_mntlk, LK_RELEASE);
807 
808 	kprintf("hammer2_mount hmp=%p pmp=%p pmpcnt=%d\n",
809 		hmp, pmp, hmp->pmp_count);
810 
811 	mp->mnt_flag = MNT_LOCAL;
812 	mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;	/* all entry pts are SMP */
813 	mp->mnt_kern_flag |= MNTK_THR_SYNC;	/* new vsyncscan semantics */
814 
815 	/*
816 	 * required mount structure initializations
817 	 */
818 	mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
819 	mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
820 
821 	mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
822 	mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
823 
824 	/*
825 	 * Optional fields
826 	 */
827 	mp->mnt_iosize_max = MAXPHYS;
828 	mp->mnt_data = (qaddr_t)pmp;
829 	pmp->mp = mp;
830 
831 	/*
832 	 * After this point hammer2_vfs_unmount() has visibility on hmp
833 	 * and manual hmp1/hmp2 calls are not needed on fatal errors.
834 	 */
835 	pmp->iroot = hammer2_inode_get(pmp, NULL, cluster);
836 	hammer2_inode_ref(pmp->iroot);		/* ref for pmp->iroot */
837 	hammer2_inode_unlock_ex(pmp->iroot, cluster);
838 
839 	/*
840 	 * The logical file buffer bio write thread handles things
841 	 * like physical block assignment and compression.
842 	 *
843 	 * (only applicable to pfs mounts, not applicable to spmp)
844 	 */
845 	pmp->wthread_destroy = 0;
846 	lwkt_create(hammer2_write_thread, pmp,
847 		    &pmp->wthread_td, NULL, 0, -1, "hwrite-%s", label);
848 
849 	/*
850 	 * With the cluster operational install ihidden.
851 	 * (only applicable to pfs mounts, not applicable to spmp)
852 	 */
853 	hammer2_inode_install_hidden(pmp);
854 
855 	/*
856 	 * Finish setup
857 	 */
858 	vfs_getnewfsid(mp);
859 	vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
860 	vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
861 	vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
862 
863 	copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
864 	bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
865 	bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
866 	copyinstr(path, mp->mnt_stat.f_mntonname,
867 		  sizeof(mp->mnt_stat.f_mntonname) - 1,
868 		  &size);
869 
870 	/*
871 	 * Initial statfs to prime mnt_stat.
872 	 */
873 	hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
874 
875 	return 0;
876 }
877 
878 /*
879  * Handle bioq for strategy write
880  */
881 static
882 void
883 hammer2_write_thread(void *arg)
884 {
885 	hammer2_pfsmount_t *pmp;
886 	struct bio *bio;
887 	struct buf *bp;
888 	hammer2_trans_t trans;
889 	struct vnode *vp;
890 	hammer2_inode_t *ip;
891 	hammer2_cluster_t *cparent;
892 	hammer2_inode_data_t *wipdata;
893 	hammer2_key_t lbase;
894 	int lblksize;
895 	int pblksize;
896 	int error;
897 
898 	pmp = arg;
899 
900 	hammer2_mtx_ex(&pmp->wthread_mtx);
901 	while (pmp->wthread_destroy == 0) {
902 		if (bioq_first(&pmp->wthread_bioq) == NULL) {
903 			mtxsleep(&pmp->wthread_bioq, &pmp->wthread_mtx,
904 				 0, "h2bioqw", 0);
905 		}
906 		cparent = NULL;
907 
908 		hammer2_trans_init(&trans, pmp, HAMMER2_TRANS_BUFCACHE);
909 
910 		while ((bio = bioq_takefirst(&pmp->wthread_bioq)) != NULL) {
911 			/*
912 			 * dummy bio for synchronization.  The transaction
913 			 * must be reinitialized.
914 			 */
915 			if (bio->bio_buf == NULL) {
916 				bio->bio_flags |= BIO_DONE;
917 				wakeup(bio);
918 				hammer2_trans_done(&trans);
919 				hammer2_trans_init(&trans, pmp,
920 						   HAMMER2_TRANS_BUFCACHE);
921 				continue;
922 			}
923 
924 			/*
925 			 * else normal bio processing
926 			 */
927 			hammer2_mtx_unlock(&pmp->wthread_mtx);
928 
929 			hammer2_lwinprog_drop(pmp);
930 
931 			error = 0;
932 			bp = bio->bio_buf;
933 			vp = bp->b_vp;
934 			ip = VTOI(vp);
935 
936 			/*
937 			 * Inode is modified, flush size and mtime changes
938 			 * to ensure that the file size remains consistent
939 			 * with the buffers being flushed.
940 			 *
941 			 * NOTE: The inode_fsync() call only flushes the
942 			 *	 inode's meta-data state, it doesn't try
943 			 *	 to flush underlying buffers or chains.
944 			 */
945 			cparent = hammer2_inode_lock_ex(ip);
946 			if (ip->flags & (HAMMER2_INODE_RESIZED |
947 					 HAMMER2_INODE_MTIME)) {
948 				hammer2_inode_fsync(&trans, ip, cparent);
949 			}
950 			wipdata = hammer2_cluster_modify_ip(&trans, ip,
951 							 cparent, 0);
952 			lblksize = hammer2_calc_logical(ip, bio->bio_offset,
953 							&lbase, NULL);
954 			pblksize = hammer2_calc_physical(ip, wipdata, lbase);
955 			hammer2_write_file_core(bp, &trans, ip, wipdata,
956 						cparent,
957 						lbase, IO_ASYNC,
958 						pblksize, &error);
959 			hammer2_cluster_modsync(cparent);
960 			hammer2_inode_unlock_ex(ip, cparent);
961 			if (error) {
962 				kprintf("hammer2: error in buffer write\n");
963 				bp->b_flags |= B_ERROR;
964 				bp->b_error = EIO;
965 			}
966 			biodone(bio);
967 			hammer2_mtx_ex(&pmp->wthread_mtx);
968 		}
969 		hammer2_trans_done(&trans);
970 	}
971 	pmp->wthread_destroy = -1;
972 	wakeup(&pmp->wthread_destroy);
973 
974 	hammer2_mtx_unlock(&pmp->wthread_mtx);
975 }
976 
977 void
978 hammer2_bioq_sync(hammer2_pfsmount_t *pmp)
979 {
980 	struct bio sync_bio;
981 
982 	bzero(&sync_bio, sizeof(sync_bio));	/* dummy with no bio_buf */
983 	hammer2_mtx_ex(&pmp->wthread_mtx);
984 	if (pmp->wthread_destroy == 0 &&
985 	    TAILQ_FIRST(&pmp->wthread_bioq.queue)) {
986 		bioq_insert_tail(&pmp->wthread_bioq, &sync_bio);
987 		while ((sync_bio.bio_flags & BIO_DONE) == 0)
988 			mtxsleep(&sync_bio, &pmp->wthread_mtx, 0, "h2bioq", 0);
989 	}
990 	hammer2_mtx_unlock(&pmp->wthread_mtx);
991 }
992 
993 /*
994  * Return a chain suitable for I/O, creating the chain if necessary
995  * and assigning its physical block.
996  */
997 static
998 hammer2_cluster_t *
999 hammer2_assign_physical(hammer2_trans_t *trans,
1000 			hammer2_inode_t *ip, hammer2_cluster_t *cparent,
1001 			hammer2_key_t lbase, int pblksize, int *errorp)
1002 {
1003 	hammer2_cluster_t *cluster;
1004 	hammer2_cluster_t *dparent;
1005 	hammer2_key_t key_dummy;
1006 	int pradix = hammer2_getradix(pblksize);
1007 	int ddflag;
1008 
1009 	/*
1010 	 * Locate the chain associated with lbase, return a locked chain.
1011 	 * However, do not instantiate any data reference (which utilizes a
1012 	 * device buffer) because we will be using direct IO via the
1013 	 * logical buffer cache buffer.
1014 	 */
1015 	*errorp = 0;
1016 	KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
1017 retry:
1018 	dparent = hammer2_cluster_lookup_init(cparent, 0);
1019 	cluster = hammer2_cluster_lookup(dparent, &key_dummy,
1020 				     lbase, lbase,
1021 				     HAMMER2_LOOKUP_NODATA, &ddflag);
1022 
1023 	if (cluster == NULL) {
1024 		/*
1025 		 * We found a hole, create a new chain entry.
1026 		 *
1027 		 * NOTE: DATA chains are created without device backing
1028 		 *	 store (nor do we want any).
1029 		 */
1030 		*errorp = hammer2_cluster_create(trans, dparent, &cluster,
1031 					       lbase, HAMMER2_PBUFRADIX,
1032 					       HAMMER2_BREF_TYPE_DATA,
1033 					       pblksize, 0);
1034 		if (cluster == NULL) {
1035 			hammer2_cluster_lookup_done(dparent);
1036 			panic("hammer2_cluster_create: par=%p error=%d\n",
1037 				dparent->focus, *errorp);
1038 			goto retry;
1039 		}
1040 		/*ip->delta_dcount += pblksize;*/
1041 	} else {
1042 		switch (hammer2_cluster_type(cluster)) {
1043 		case HAMMER2_BREF_TYPE_INODE:
1044 			/*
1045 			 * The data is embedded in the inode.  The
1046 			 * caller is responsible for marking the inode
1047 			 * modified and copying the data to the embedded
1048 			 * area.
1049 			 */
1050 			break;
1051 		case HAMMER2_BREF_TYPE_DATA:
1052 			if (hammer2_cluster_need_resize(cluster, pblksize)) {
1053 				hammer2_cluster_resize(trans, ip,
1054 						     dparent, cluster,
1055 						     pradix,
1056 						     HAMMER2_MODIFY_OPTDATA);
1057 			}
1058 
1059 			/*
1060 			 * DATA buffers must be marked modified whether the
1061 			 * data is in a logical buffer or not.  We also have
1062 			 * to make this call to fixup the chain data pointers
1063 			 * after resizing in case this is an encrypted or
1064 			 * compressed buffer.
1065 			 */
1066 			hammer2_cluster_modify(trans, cluster,
1067 					       HAMMER2_MODIFY_OPTDATA);
1068 			break;
1069 		default:
1070 			panic("hammer2_assign_physical: bad type");
1071 			/* NOT REACHED */
1072 			break;
1073 		}
1074 	}
1075 
1076 	/*
1077 	 * Cleanup.  If cluster wound up being the inode itself, i.e.
1078 	 * the DIRECTDATA case for offset 0, then we need to update cparent.
1079 	 * The caller expects cparent to not become stale.
1080 	 */
1081 	hammer2_cluster_lookup_done(dparent);
1082 	/* dparent = NULL; safety */
1083 	if (cluster && ddflag)
1084 		hammer2_cluster_replace_locked(cparent, cluster);
1085 	return (cluster);
1086 }
1087 
1088 /*
1089  * bio queued from hammer2_vnops.c.
1090  *
1091  * The core write function which determines which path to take
1092  * depending on compression settings.  We also have to locate the
1093  * related clusters so we can calculate and set the check data for
1094  * the blockref.
1095  */
1096 static
1097 void
1098 hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
1099 			hammer2_inode_t *ip,
1100 			const hammer2_inode_data_t *ripdata,
1101 			hammer2_cluster_t *cparent,
1102 			hammer2_key_t lbase, int ioflag, int pblksize,
1103 			int *errorp)
1104 {
1105 	hammer2_cluster_t *cluster;
1106 
1107 	switch(HAMMER2_DEC_ALGO(ripdata->comp_algo)) {
1108 	case HAMMER2_COMP_NONE:
1109 		/*
1110 		 * We have to assign physical storage to the buffer
1111 		 * we intend to dirty or write now to avoid deadlocks
1112 		 * in the strategy code later.
1113 		 *
1114 		 * This can return NOOFFSET for inode-embedded data.
1115 		 * The strategy code will take care of it in that case.
1116 		 */
1117 		cluster = hammer2_assign_physical(trans, ip, cparent,
1118 						lbase, pblksize,
1119 						errorp);
1120 		hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp,
1121 				 ripdata->check_algo);
1122 		if (cluster)
1123 			hammer2_cluster_unlock(cluster);
1124 		break;
1125 	case HAMMER2_COMP_AUTOZERO:
1126 		/*
1127 		 * Check for zero-fill only
1128 		 */
1129 		hammer2_zero_check_and_write(bp, trans, ip,
1130 				    ripdata, cparent, lbase,
1131 				    ioflag, pblksize, errorp,
1132 				    ripdata->check_algo);
1133 		break;
1134 	case HAMMER2_COMP_LZ4:
1135 	case HAMMER2_COMP_ZLIB:
1136 	default:
1137 		/*
1138 		 * Check for zero-fill and attempt compression.
1139 		 */
1140 		hammer2_compress_and_write(bp, trans, ip,
1141 					   ripdata, cparent,
1142 					   lbase, ioflag,
1143 					   pblksize, errorp,
1144 					   ripdata->comp_algo,
1145 					   ripdata->check_algo);
1146 		break;
1147 	}
1148 }
1149 
1150 /*
1151  * Generic function that will perform the compression in compression
1152  * write path. The compression algorithm is determined by the settings
1153  * obtained from inode.
1154  */
1155 static
1156 void
1157 hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
1158 	hammer2_inode_t *ip, const hammer2_inode_data_t *ripdata,
1159 	hammer2_cluster_t *cparent,
1160 	hammer2_key_t lbase, int ioflag, int pblksize,
1161 	int *errorp, int comp_algo, int check_algo)
1162 {
1163 	hammer2_cluster_t *cluster;
1164 	hammer2_chain_t *chain;
1165 	int comp_size;
1166 	int comp_block_size;
1167 	int i;
1168 	char *comp_buffer;
1169 
1170 	if (test_block_zeros(bp->b_data, pblksize)) {
1171 		zero_write(bp, trans, ip, ripdata, cparent, lbase, errorp);
1172 		return;
1173 	}
1174 
1175 	comp_size = 0;
1176 	comp_buffer = NULL;
1177 
1178 	KKASSERT(pblksize / 2 <= 32768);
1179 
1180 	if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
1181 		z_stream strm_compress;
1182 		int comp_level;
1183 		int ret;
1184 
1185 		switch(HAMMER2_DEC_ALGO(comp_algo)) {
1186 		case HAMMER2_COMP_LZ4:
1187 			comp_buffer = objcache_get(cache_buffer_write,
1188 						   M_INTWAIT);
1189 			comp_size = LZ4_compress_limitedOutput(
1190 					bp->b_data,
1191 					&comp_buffer[sizeof(int)],
1192 					pblksize,
1193 					pblksize / 2 - sizeof(int));
1194 			/*
1195 			 * We need to prefix with the size, LZ4
1196 			 * doesn't do it for us.  Add the related
1197 			 * overhead.
1198 			 */
1199 			*(int *)comp_buffer = comp_size;
1200 			if (comp_size)
1201 				comp_size += sizeof(int);
1202 			break;
1203 		case HAMMER2_COMP_ZLIB:
1204 			comp_level = HAMMER2_DEC_LEVEL(comp_algo);
1205 			if (comp_level == 0)
1206 				comp_level = 6;	/* default zlib compression */
1207 			else if (comp_level < 6)
1208 				comp_level = 6;
1209 			else if (comp_level > 9)
1210 				comp_level = 9;
1211 			ret = deflateInit(&strm_compress, comp_level);
1212 			if (ret != Z_OK) {
1213 				kprintf("HAMMER2 ZLIB: fatal error "
1214 					"on deflateInit.\n");
1215 			}
1216 
1217 			comp_buffer = objcache_get(cache_buffer_write,
1218 						   M_INTWAIT);
1219 			strm_compress.next_in = bp->b_data;
1220 			strm_compress.avail_in = pblksize;
1221 			strm_compress.next_out = comp_buffer;
1222 			strm_compress.avail_out = pblksize / 2;
1223 			ret = deflate(&strm_compress, Z_FINISH);
1224 			if (ret == Z_STREAM_END) {
1225 				comp_size = pblksize / 2 -
1226 					    strm_compress.avail_out;
1227 			} else {
1228 				comp_size = 0;
1229 			}
1230 			ret = deflateEnd(&strm_compress);
1231 			break;
1232 		default:
1233 			kprintf("Error: Unknown compression method.\n");
1234 			kprintf("Comp_method = %d.\n", comp_algo);
1235 			break;
1236 		}
1237 	}
1238 
1239 	if (comp_size == 0) {
1240 		/*
1241 		 * compression failed or turned off
1242 		 */
1243 		comp_block_size = pblksize;	/* safety */
1244 		if (++ip->comp_heuristic > 128)
1245 			ip->comp_heuristic = 8;
1246 	} else {
1247 		/*
1248 		 * compression succeeded
1249 		 */
1250 		ip->comp_heuristic = 0;
1251 		if (comp_size <= 1024) {
1252 			comp_block_size = 1024;
1253 		} else if (comp_size <= 2048) {
1254 			comp_block_size = 2048;
1255 		} else if (comp_size <= 4096) {
1256 			comp_block_size = 4096;
1257 		} else if (comp_size <= 8192) {
1258 			comp_block_size = 8192;
1259 		} else if (comp_size <= 16384) {
1260 			comp_block_size = 16384;
1261 		} else if (comp_size <= 32768) {
1262 			comp_block_size = 32768;
1263 		} else {
1264 			panic("hammer2: WRITE PATH: "
1265 			      "Weird comp_size value.");
1266 			/* NOT REACHED */
1267 			comp_block_size = pblksize;
1268 		}
1269 	}
1270 
1271 	cluster = hammer2_assign_physical(trans, ip, cparent,
1272 					  lbase, comp_block_size,
1273 					  errorp);
1274 	ripdata = NULL;
1275 
1276 	if (*errorp) {
1277 		kprintf("WRITE PATH: An error occurred while "
1278 			"assigning physical space.\n");
1279 		KKASSERT(cluster == NULL);
1280 		goto done;
1281 	}
1282 
1283 	for (i = 0; i < cluster->nchains; ++i) {
1284 		hammer2_inode_data_t *wipdata;
1285 		hammer2_io_t *dio;
1286 		char *bdata;
1287 
1288 		chain = cluster->array[i].chain;	/* XXX */
1289 		KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1290 
1291 		switch(chain->bref.type) {
1292 		case HAMMER2_BREF_TYPE_INODE:
1293 			wipdata = &hammer2_chain_wdata(chain)->ipdata;
1294 			KKASSERT(wipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1295 			KKASSERT(bp->b_loffset == 0);
1296 			bcopy(bp->b_data, wipdata->u.data,
1297 			      HAMMER2_EMBEDDED_BYTES);
1298 			break;
1299 		case HAMMER2_BREF_TYPE_DATA:
1300 			/*
1301 			 * Optimize out the read-before-write
1302 			 * if possible.
1303 			 */
1304 			*errorp = hammer2_io_newnz(chain->hmp,
1305 						   chain->bref.data_off,
1306 						   chain->bytes,
1307 						   &dio);
1308 			if (*errorp) {
1309 				hammer2_io_brelse(&dio);
1310 				kprintf("hammer2: WRITE PATH: "
1311 					"dbp bread error\n");
1312 				break;
1313 			}
1314 			bdata = hammer2_io_data(dio, chain->bref.data_off);
1315 
1316 			/*
1317 			 * When loading the block make sure we don't
1318 			 * leave garbage after the compressed data.
1319 			 */
1320 			if (comp_size) {
1321 				chain->bref.methods =
1322 					HAMMER2_ENC_COMP(comp_algo) +
1323 					HAMMER2_ENC_CHECK(check_algo);
1324 				bcopy(comp_buffer, bdata, comp_size);
1325 				if (comp_size != comp_block_size) {
1326 					bzero(bdata + comp_size,
1327 					      comp_block_size - comp_size);
1328 				}
1329 			} else {
1330 				chain->bref.methods =
1331 					HAMMER2_ENC_COMP(
1332 						HAMMER2_COMP_NONE) +
1333 					HAMMER2_ENC_CHECK(check_algo);
1334 				bcopy(bp->b_data, bdata, pblksize);
1335 			}
1336 
1337 			/*
1338 			 * The flush code doesn't calculate check codes for
1339 			 * file data (doing so can result in excessive I/O),
1340 			 * so we do it here.
1341 			 */
1342 			hammer2_chain_setcheck(chain, bdata);
1343 
1344 			/*
1345 			 * Device buffer is now valid, chain is no longer in
1346 			 * the initial state.
1347 			 *
1348 			 * (No blockref table worries with file data)
1349 			 */
1350 			atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1351 
1352 			/* Now write the related bdp. */
1353 			if (ioflag & IO_SYNC) {
1354 				/*
1355 				 * Synchronous I/O requested.
1356 				 */
1357 				hammer2_io_bwrite(&dio);
1358 			/*
1359 			} else if ((ioflag & IO_DIRECT) &&
1360 				   loff + n == pblksize) {
1361 				hammer2_io_bdwrite(&dio);
1362 			*/
1363 			} else if (ioflag & IO_ASYNC) {
1364 				hammer2_io_bawrite(&dio);
1365 			} else {
1366 				hammer2_io_bdwrite(&dio);
1367 			}
1368 			break;
1369 		default:
1370 			panic("hammer2_write_bp: bad chain type %d\n",
1371 				chain->bref.type);
1372 			/* NOT REACHED */
1373 			break;
1374 		}
1375 	}
1376 done:
1377 	if (cluster)
1378 		hammer2_cluster_unlock(cluster);
1379 	if (comp_buffer)
1380 		objcache_put(cache_buffer_write, comp_buffer);
1381 }
1382 
1383 /*
1384  * Function that performs zero-checking and writing without compression,
1385  * it corresponds to default zero-checking path.
1386  */
1387 static
1388 void
1389 hammer2_zero_check_and_write(struct buf *bp, hammer2_trans_t *trans,
1390 	hammer2_inode_t *ip, const hammer2_inode_data_t *ripdata,
1391 	hammer2_cluster_t *cparent,
1392 	hammer2_key_t lbase, int ioflag, int pblksize, int *errorp,
1393 	int check_algo)
1394 {
1395 	hammer2_cluster_t *cluster;
1396 
1397 	if (test_block_zeros(bp->b_data, pblksize)) {
1398 		zero_write(bp, trans, ip, ripdata, cparent, lbase, errorp);
1399 	} else {
1400 		cluster = hammer2_assign_physical(trans, ip, cparent,
1401 						  lbase, pblksize, errorp);
1402 		hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp,
1403 				 check_algo);
1404 		if (cluster)
1405 			hammer2_cluster_unlock(cluster);
1406 	}
1407 }
1408 
1409 /*
1410  * A function to test whether a block of data contains only zeros,
1411  * returns TRUE (non-zero) if the block is all zeros.
1412  */
1413 static
1414 int
1415 test_block_zeros(const char *buf, size_t bytes)
1416 {
1417 	size_t i;
1418 
1419 	for (i = 0; i < bytes; i += sizeof(long)) {
1420 		if (*(const long *)(buf + i) != 0)
1421 			return (0);
1422 	}
1423 	return (1);
1424 }
1425 
1426 /*
1427  * Function to "write" a block that contains only zeros.
1428  */
1429 static
1430 void
1431 zero_write(struct buf *bp, hammer2_trans_t *trans,
1432 	   hammer2_inode_t *ip, const hammer2_inode_data_t *ripdata,
1433 	   hammer2_cluster_t *cparent,
1434 	   hammer2_key_t lbase, int *errorp __unused)
1435 {
1436 	hammer2_cluster_t *cluster;
1437 	hammer2_media_data_t *data;
1438 	hammer2_key_t key_dummy;
1439 	int ddflag;
1440 
1441 	cparent = hammer2_cluster_lookup_init(cparent, 0);
1442 	cluster = hammer2_cluster_lookup(cparent, &key_dummy, lbase, lbase,
1443 				     HAMMER2_LOOKUP_NODATA, &ddflag);
1444 	if (cluster) {
1445 		data = hammer2_cluster_wdata(cluster);
1446 
1447 		if (ddflag) {
1448 			KKASSERT(cluster->focus->flags &
1449 				 HAMMER2_CHAIN_MODIFIED);
1450 			bzero(data->ipdata.u.data, HAMMER2_EMBEDDED_BYTES);
1451 			hammer2_cluster_modsync(cluster);
1452 		} else {
1453 			hammer2_cluster_delete(trans, cparent, cluster,
1454 					       HAMMER2_DELETE_PERMANENT);
1455 		}
1456 		hammer2_cluster_unlock(cluster);
1457 	}
1458 	hammer2_cluster_lookup_done(cparent);
1459 }
1460 
1461 /*
1462  * Function to write the data as it is, without performing any sort of
1463  * compression. This function is used in path without compression and
1464  * default zero-checking path.
1465  */
1466 static
1467 void
1468 hammer2_write_bp(hammer2_cluster_t *cluster, struct buf *bp, int ioflag,
1469 				int pblksize, int *errorp, int check_algo)
1470 {
1471 	hammer2_chain_t *chain;
1472 	hammer2_inode_data_t *wipdata;
1473 	hammer2_io_t *dio;
1474 	char *bdata;
1475 	int error;
1476 	int i;
1477 
1478 	error = 0;	/* XXX TODO below */
1479 
1480 	for (i = 0; i < cluster->nchains; ++i) {
1481 		chain = cluster->array[i].chain;	/* XXX */
1482 		KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1483 
1484 		switch(chain->bref.type) {
1485 		case HAMMER2_BREF_TYPE_INODE:
1486 			wipdata = &hammer2_chain_wdata(chain)->ipdata;
1487 			KKASSERT(wipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1488 			KKASSERT(bp->b_loffset == 0);
1489 			bcopy(bp->b_data, wipdata->u.data,
1490 			      HAMMER2_EMBEDDED_BYTES);
1491 			error = 0;
1492 			break;
1493 		case HAMMER2_BREF_TYPE_DATA:
1494 			error = hammer2_io_newnz(chain->hmp,
1495 						 chain->bref.data_off,
1496 						 chain->bytes, &dio);
1497 			if (error) {
1498 				hammer2_io_bqrelse(&dio);
1499 				kprintf("hammer2: WRITE PATH: "
1500 					"dbp bread error\n");
1501 				break;
1502 			}
1503 			bdata = hammer2_io_data(dio, chain->bref.data_off);
1504 
1505 			chain->bref.methods = HAMMER2_ENC_COMP(
1506 							HAMMER2_COMP_NONE) +
1507 					      HAMMER2_ENC_CHECK(check_algo);
1508 			bcopy(bp->b_data, bdata, chain->bytes);
1509 
1510 			/*
1511 			 * The flush code doesn't calculate check codes for
1512 			 * file data (doing so can result in excessive I/O),
1513 			 * so we do it here.
1514 			 */
1515 			hammer2_chain_setcheck(chain, bdata);
1516 
1517 			/*
1518 			 * Device buffer is now valid, chain is no longer in
1519 			 * the initial state.
1520 			 *
1521 			 * (No blockref table worries with file data)
1522 			 */
1523 			atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1524 
1525 			if (ioflag & IO_SYNC) {
1526 				/*
1527 				 * Synchronous I/O requested.
1528 				 */
1529 				hammer2_io_bwrite(&dio);
1530 			/*
1531 			} else if ((ioflag & IO_DIRECT) &&
1532 				   loff + n == pblksize) {
1533 				hammer2_io_bdwrite(&dio);
1534 			*/
1535 			} else if (ioflag & IO_ASYNC) {
1536 				hammer2_io_bawrite(&dio);
1537 			} else {
1538 				hammer2_io_bdwrite(&dio);
1539 			}
1540 			break;
1541 		default:
1542 			panic("hammer2_write_bp: bad chain type %d\n",
1543 			      chain->bref.type);
1544 			/* NOT REACHED */
1545 			error = 0;
1546 			break;
1547 		}
1548 		KKASSERT(error == 0);	/* XXX TODO */
1549 	}
1550 	*errorp = error;
1551 }
1552 
1553 static
1554 int
1555 hammer2_remount(hammer2_mount_t *hmp, struct mount *mp, char *path,
1556 		struct vnode *devvp, struct ucred *cred)
1557 {
1558 	int error;
1559 
1560 	if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1561 		error = hammer2_recovery(hmp);
1562 	} else {
1563 		error = 0;
1564 	}
1565 	return error;
1566 }
1567 
1568 static
1569 int
1570 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1571 {
1572 	hammer2_pfsmount_t *pmp;
1573 	hammer2_mount_t *hmp;
1574 	hammer2_chain_t *rchain;
1575 	hammer2_cluster_t *cluster;
1576 	int flags;
1577 	int error = 0;
1578 	int i;
1579 
1580 	pmp = MPTOPMP(mp);
1581 
1582 	if (pmp == NULL)
1583 		return(0);
1584 
1585 	lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1586 	TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
1587 
1588 	/*
1589 	 * If mount initialization proceeded far enough we must flush
1590 	 * its vnodes.
1591 	 */
1592 	if (mntflags & MNT_FORCE)
1593 		flags = FORCECLOSE;
1594 	else
1595 		flags = 0;
1596 	if (pmp->iroot) {
1597 		error = vflush(mp, 0, flags);
1598 		if (error)
1599 			goto failed;
1600 	}
1601 
1602 	if (pmp->wthread_td) {
1603 		hammer2_mtx_ex(&pmp->wthread_mtx);
1604 		pmp->wthread_destroy = 1;
1605 		wakeup(&pmp->wthread_bioq);
1606 		while (pmp->wthread_destroy != -1) {
1607 			mtxsleep(&pmp->wthread_destroy,
1608 				&pmp->wthread_mtx, 0,
1609 				"umount-sleep",	0);
1610 		}
1611 		hammer2_mtx_unlock(&pmp->wthread_mtx);
1612 		pmp->wthread_td = NULL;
1613 	}
1614 
1615 	/*
1616 	 * Cleanup our reference on ihidden.
1617 	 */
1618 	if (pmp->ihidden) {
1619 		hammer2_inode_drop(pmp->ihidden);
1620 		pmp->ihidden = NULL;
1621 	}
1622 
1623 	/*
1624 	 * Cleanup our reference on iroot.  iroot is (should) not be needed
1625 	 * by the flush code.
1626 	 */
1627 	if (pmp->iroot) {
1628 		cluster = &pmp->iroot->cluster;
1629 		for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1630 			rchain = pmp->iroot->cluster.array[i].chain;
1631 			if (rchain == NULL)
1632 				continue;
1633 			hmp = rchain->hmp;
1634 			hammer2_vfs_unmount_hmp1(mp, hmp);
1635 
1636 			atomic_clear_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
1637 #if REPORT_REFS_ERRORS
1638 			if (rchain->refs != 1)
1639 				kprintf("PMP->RCHAIN %p REFS WRONG %d\n",
1640 					rchain, rchain->refs);
1641 #else
1642 			KKASSERT(rchain->refs == 1);
1643 #endif
1644 			hammer2_chain_drop(rchain);
1645 			cluster->array[i].chain = NULL;
1646 			hammer2_vfs_unmount_hmp2(mp, hmp);
1647 		}
1648 		cluster->focus = NULL;
1649 
1650 #if REPORT_REFS_ERRORS
1651 		if (pmp->iroot->refs != 1)
1652 			kprintf("PMP->IROOT %p REFS WRONG %d\n",
1653 				pmp->iroot, pmp->iroot->refs);
1654 #else
1655 		KKASSERT(pmp->iroot->refs == 1);
1656 #endif
1657 		/* ref for pmp->iroot */
1658 		hammer2_inode_drop(pmp->iroot);
1659 		pmp->iroot = NULL;
1660 	}
1661 
1662 	pmp->mp = NULL;
1663 	mp->mnt_data = NULL;
1664 
1665 	kmalloc_destroy(&pmp->mmsg);
1666 	kmalloc_destroy(&pmp->minode);
1667 
1668 	kfree(pmp, M_HAMMER2);
1669 	error = 0;
1670 
1671 failed:
1672 	lockmgr(&hammer2_mntlk, LK_RELEASE);
1673 
1674 	return (error);
1675 }
1676 
1677 static
1678 void
1679 hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp)
1680 {
1681 	hammer2_mount_exlock(hmp);
1682 	--hmp->pmp_count;
1683 
1684 	kprintf("hammer2_unmount hmp=%p pmpcnt=%d\n", hmp, hmp->pmp_count);
1685 
1686 	/*
1687 	 * Cycle the volume data lock as a safety (probably not needed any
1688 	 * more).  To ensure everything is out we need to flush at least
1689 	 * three times.  (1) The running of the unlinkq can dirty the
1690 	 * filesystem, (2) A normal flush can dirty the freemap, and
1691 	 * (3) ensure that the freemap is fully synchronized.
1692 	 *
1693 	 * The next mount's recovery scan can clean everything up but we want
1694 	 * to leave the filesystem in a 100% clean state on a normal unmount.
1695 	 */
1696 	hammer2_voldata_lock(hmp);
1697 	hammer2_voldata_unlock(hmp);
1698 	if (mp->mnt_data) {
1699 		hammer2_vfs_sync(mp, MNT_WAIT);
1700 		hammer2_vfs_sync(mp, MNT_WAIT);
1701 		hammer2_vfs_sync(mp, MNT_WAIT);
1702 	}
1703 
1704 	/*
1705 	 * XXX chain depend deadlock?
1706 	 */
1707 	hammer2_iocom_uninit(hmp);
1708 
1709 	if (hmp->pmp_count == 0) {
1710 		if ((hmp->vchain.flags | hmp->fchain.flags) &
1711 		    HAMMER2_CHAIN_FLUSH_MASK) {
1712 			kprintf("hammer2_unmount: chains left over "
1713 				"after final sync\n");
1714 			kprintf("    vchain %08x\n", hmp->vchain.flags);
1715 			kprintf("    fchain %08x\n", hmp->fchain.flags);
1716 
1717 			if (hammer2_debug & 0x0010)
1718 				Debugger("entered debugger");
1719 		}
1720 	}
1721 }
1722 
1723 static
1724 void
1725 hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp)
1726 {
1727 	hammer2_pfsmount_t *spmp;
1728 	struct vnode *devvp;
1729 	int dumpcnt;
1730 	int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1731 
1732 	/*
1733 	 * If no PFS's left drop the master hammer2_mount for the
1734 	 * device.
1735 	 */
1736 	if (hmp->pmp_count == 0) {
1737 		/*
1738 		 * Clean up SPMP and the super-root inode
1739 		 */
1740 		spmp = hmp->spmp;
1741 		if (spmp) {
1742 			if (spmp->iroot) {
1743 				hammer2_inode_drop(spmp->iroot);
1744 				spmp->iroot = NULL;
1745 			}
1746 			hmp->spmp = NULL;
1747 			kmalloc_destroy(&spmp->mmsg);
1748 			kmalloc_destroy(&spmp->minode);
1749 			kfree(spmp, M_HAMMER2);
1750 		}
1751 
1752 		/*
1753 		 * Finish up with the device vnode
1754 		 */
1755 		if ((devvp = hmp->devvp) != NULL) {
1756 			vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1757 			vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1758 			hmp->devvp = NULL;
1759 			VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1760 			vn_unlock(devvp);
1761 			vrele(devvp);
1762 			devvp = NULL;
1763 		}
1764 
1765 		/*
1766 		 * Clear vchain/fchain flags that might prevent final cleanup
1767 		 * of these chains.
1768 		 */
1769 		if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1770 			atomic_clear_int(&hmp->vchain.flags,
1771 					 HAMMER2_CHAIN_MODIFIED);
1772 			hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1773 			hammer2_chain_drop(&hmp->vchain);
1774 		}
1775 		if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1776 			atomic_clear_int(&hmp->vchain.flags,
1777 					 HAMMER2_CHAIN_UPDATE);
1778 			hammer2_chain_drop(&hmp->vchain);
1779 		}
1780 
1781 		if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1782 			atomic_clear_int(&hmp->fchain.flags,
1783 					 HAMMER2_CHAIN_MODIFIED);
1784 			hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1785 			hammer2_chain_drop(&hmp->fchain);
1786 		}
1787 		if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1788 			atomic_clear_int(&hmp->fchain.flags,
1789 					 HAMMER2_CHAIN_UPDATE);
1790 			hammer2_chain_drop(&hmp->fchain);
1791 		}
1792 
1793 		/*
1794 		 * Final drop of embedded freemap root chain to
1795 		 * clean up fchain.core (fchain structure is not
1796 		 * flagged ALLOCATED so it is cleaned out and then
1797 		 * left to rot).
1798 		 */
1799 		hammer2_chain_drop(&hmp->fchain);
1800 
1801 		/*
1802 		 * Final drop of embedded volume root chain to clean
1803 		 * up vchain.core (vchain structure is not flagged
1804 		 * ALLOCATED so it is cleaned out and then left to
1805 		 * rot).
1806 		 */
1807 		dumpcnt = 50;
1808 		hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1809 		dumpcnt = 50;
1810 		hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1811 		hammer2_mount_unlock(hmp);
1812 		hammer2_chain_drop(&hmp->vchain);
1813 
1814 		hammer2_io_cleanup(hmp, &hmp->iotree);
1815 		if (hmp->iofree_count) {
1816 			kprintf("io_cleanup: %d I/O's left hanging\n",
1817 				hmp->iofree_count);
1818 		}
1819 
1820 		TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1821 		kmalloc_destroy(&hmp->mchain);
1822 		kfree(hmp, M_HAMMER2);
1823 	} else {
1824 		hammer2_mount_unlock(hmp);
1825 	}
1826 }
1827 
1828 static
1829 int
1830 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1831 	     ino_t ino, struct vnode **vpp)
1832 {
1833 	kprintf("hammer2_vget\n");
1834 	return (EOPNOTSUPP);
1835 }
1836 
1837 static
1838 int
1839 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1840 {
1841 	hammer2_pfsmount_t *pmp;
1842 	hammer2_cluster_t *cparent;
1843 	int error;
1844 	struct vnode *vp;
1845 
1846 	pmp = MPTOPMP(mp);
1847 	if (pmp->iroot == NULL) {
1848 		*vpp = NULL;
1849 		error = EINVAL;
1850 	} else {
1851 		cparent = hammer2_inode_lock_sh(pmp->iroot);
1852 		vp = hammer2_igetv(pmp->iroot, cparent, &error);
1853 		hammer2_inode_unlock_sh(pmp->iroot, cparent);
1854 		*vpp = vp;
1855 		if (vp == NULL)
1856 			kprintf("vnodefail\n");
1857 	}
1858 
1859 	return (error);
1860 }
1861 
1862 /*
1863  * Filesystem status
1864  *
1865  * XXX incorporate ipdata->inode_quota and data_quota
1866  */
1867 static
1868 int
1869 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1870 {
1871 	hammer2_pfsmount_t *pmp;
1872 	hammer2_mount_t *hmp;
1873 
1874 	pmp = MPTOPMP(mp);
1875 	KKASSERT(pmp->iroot->cluster.nchains >= 1);
1876 	hmp = pmp->iroot->cluster.focus->hmp;	/* XXX */
1877 
1878 	mp->mnt_stat.f_files = pmp->inode_count;
1879 	mp->mnt_stat.f_ffree = 0;
1880 	mp->mnt_stat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1881 	mp->mnt_stat.f_bfree =  hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1882 	mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1883 
1884 	*sbp = mp->mnt_stat;
1885 	return (0);
1886 }
1887 
1888 static
1889 int
1890 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1891 {
1892 	hammer2_pfsmount_t *pmp;
1893 	hammer2_mount_t *hmp;
1894 
1895 	pmp = MPTOPMP(mp);
1896 	KKASSERT(pmp->iroot->cluster.nchains >= 1);
1897 	hmp = pmp->iroot->cluster.focus->hmp;	/* XXX */
1898 
1899 	mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1900 	mp->mnt_vstat.f_files = pmp->inode_count;
1901 	mp->mnt_vstat.f_ffree = 0;
1902 	mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1903 	mp->mnt_vstat.f_bfree =  hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1904 	mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1905 
1906 	*sbp = mp->mnt_vstat;
1907 	return (0);
1908 }
1909 
1910 /*
1911  * Mount-time recovery (RW mounts)
1912  *
1913  * Updates to the free block table are allowed to lag flushes by one
1914  * transaction.  In case of a crash, then on a fresh mount we must do an
1915  * incremental scan of the last committed transaction id and make sure that
1916  * all related blocks have been marked allocated.
1917  *
1918  * The super-root topology and each PFS has its own transaction id domain,
1919  * so we must track PFS boundary transitions.
1920  */
1921 struct hammer2_recovery_elm {
1922 	TAILQ_ENTRY(hammer2_recovery_elm) entry;
1923 	hammer2_chain_t *chain;
1924 	hammer2_tid_t sync_tid;
1925 };
1926 
1927 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1928 
1929 struct hammer2_recovery_info {
1930 	struct hammer2_recovery_list list;
1931 	int	depth;
1932 };
1933 
1934 static int hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1935 			hammer2_chain_t *parent,
1936 			struct hammer2_recovery_info *info,
1937 			hammer2_tid_t sync_tid);
1938 
1939 #define HAMMER2_RECOVERY_MAXDEPTH	10
1940 
1941 static
1942 int
1943 hammer2_recovery(hammer2_mount_t *hmp)
1944 {
1945 	hammer2_trans_t trans;
1946 	struct hammer2_recovery_info info;
1947 	struct hammer2_recovery_elm *elm;
1948 	hammer2_chain_t *parent;
1949 	hammer2_tid_t sync_tid;
1950 	int error;
1951 	int cumulative_error = 0;
1952 
1953 	hammer2_trans_init(&trans, hmp->spmp, 0);
1954 
1955 	sync_tid = 0;
1956 	TAILQ_INIT(&info.list);
1957 	info.depth = 0;
1958 	parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1959 	cumulative_error = hammer2_recovery_scan(&trans, hmp, parent,
1960 						 &info, sync_tid);
1961 	hammer2_chain_lookup_done(parent);
1962 
1963 	while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1964 		TAILQ_REMOVE(&info.list, elm, entry);
1965 		parent = elm->chain;
1966 		sync_tid = elm->sync_tid;
1967 		kfree(elm, M_HAMMER2);
1968 
1969 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
1970 					   HAMMER2_RESOLVE_NOREF);
1971 		error = hammer2_recovery_scan(&trans, hmp, parent,
1972 					      &info, sync_tid);
1973 		hammer2_chain_unlock(parent);
1974 		if (error)
1975 			cumulative_error = error;
1976 	}
1977 	hammer2_trans_done(&trans);
1978 
1979 	return cumulative_error;
1980 }
1981 
1982 static
1983 int
1984 hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1985 		      hammer2_chain_t *parent,
1986 		      struct hammer2_recovery_info *info,
1987 		      hammer2_tid_t sync_tid)
1988 {
1989 	const hammer2_inode_data_t *ripdata;
1990 	hammer2_chain_t *chain;
1991 	int cache_index;
1992 	int cumulative_error = 0;
1993 	int pfs_boundary = 0;
1994 	int error;
1995 
1996 	/*
1997 	 * Adjust freemap to ensure that the block(s) are marked allocated.
1998 	 */
1999 	if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2000 		hammer2_freemap_adjust(trans, hmp, &parent->bref,
2001 				       HAMMER2_FREEMAP_DORECOVER);
2002 	}
2003 
2004 	/*
2005 	 * Check type for recursive scan
2006 	 */
2007 	switch(parent->bref.type) {
2008 	case HAMMER2_BREF_TYPE_VOLUME:
2009 		/* data already instantiated */
2010 		break;
2011 	case HAMMER2_BREF_TYPE_INODE:
2012 		/*
2013 		 * Must instantiate data for DIRECTDATA test and also
2014 		 * for recursion.
2015 		 */
2016 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2017 		ripdata = &hammer2_chain_rdata(parent)->ipdata;
2018 		if (ripdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2019 			/* not applicable to recovery scan */
2020 			hammer2_chain_unlock(parent);
2021 			return 0;
2022 		}
2023 		if ((ripdata->op_flags & HAMMER2_OPFLAG_PFSROOT) &&
2024 		    info->depth != 0) {
2025 			pfs_boundary = 1;
2026 			sync_tid = parent->bref.mirror_tid - 1;
2027 		}
2028 		hammer2_chain_unlock(parent);
2029 		break;
2030 	case HAMMER2_BREF_TYPE_INDIRECT:
2031 		/*
2032 		 * Must instantiate data for recursion
2033 		 */
2034 		hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2035 		hammer2_chain_unlock(parent);
2036 		break;
2037 	case HAMMER2_BREF_TYPE_DATA:
2038 	case HAMMER2_BREF_TYPE_FREEMAP:
2039 	case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2040 	case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2041 		/* not applicable to recovery scan */
2042 		return 0;
2043 		break;
2044 	default:
2045 		return EDOM;
2046 	}
2047 
2048 	/*
2049 	 * Defer operation if depth limit reached or if we are crossing a
2050 	 * PFS boundary.
2051 	 */
2052 	if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH || pfs_boundary) {
2053 		struct hammer2_recovery_elm *elm;
2054 
2055 		elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2056 		elm->chain = parent;
2057 		elm->sync_tid = sync_tid;
2058 		hammer2_chain_ref(parent);
2059 		TAILQ_INSERT_TAIL(&info->list, elm, entry);
2060 		/* unlocked by caller */
2061 
2062 		return(0);
2063 	}
2064 
2065 
2066 	/*
2067 	 * Recursive scan of the last flushed transaction only.  We are
2068 	 * doing this without pmp assignments so don't leave the chains
2069 	 * hanging around after we are done with them.
2070 	 */
2071 	cache_index = 0;
2072 	chain = hammer2_chain_scan(parent, NULL, &cache_index,
2073 				   HAMMER2_LOOKUP_NODATA);
2074 	while (chain) {
2075 		atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2076 		if (chain->bref.mirror_tid >= sync_tid) {
2077 			++info->depth;
2078 			error = hammer2_recovery_scan(trans, hmp, chain,
2079 						      info, sync_tid);
2080 			--info->depth;
2081 			if (error)
2082 				cumulative_error = error;
2083 		}
2084 		chain = hammer2_chain_scan(parent, chain, &cache_index,
2085 					   HAMMER2_LOOKUP_NODATA);
2086 	}
2087 
2088 	return cumulative_error;
2089 }
2090 
2091 /*
2092  * Sync the entire filesystem; this is called from the filesystem syncer
2093  * process periodically and whenever a user calls sync(1) on the hammer
2094  * mountpoint.
2095  *
2096  * Currently is actually called from the syncer! \o/
2097  *
2098  * This task will have to snapshot the state of the dirty inode chain.
2099  * From that, it will have to make sure all of the inodes on the dirty
2100  * chain have IO initiated. We make sure that io is initiated for the root
2101  * block.
2102  *
2103  * If waitfor is set, we wait for media to acknowledge the new rootblock.
2104  *
2105  * THINKS: side A vs side B, to have sync not stall all I/O?
2106  */
2107 int
2108 hammer2_vfs_sync(struct mount *mp, int waitfor)
2109 {
2110 	struct hammer2_sync_info info;
2111 	hammer2_inode_t *iroot;
2112 	hammer2_chain_t *chain;
2113 	hammer2_chain_t *parent;
2114 	hammer2_pfsmount_t *pmp;
2115 	hammer2_mount_t *hmp;
2116 	int flags;
2117 	int error;
2118 	int total_error;
2119 	int force_fchain;
2120 	int i;
2121 	int j;
2122 
2123 	pmp = MPTOPMP(mp);
2124 	iroot = pmp->iroot;
2125 	KKASSERT(iroot);
2126 	KKASSERT(iroot->pmp == pmp);
2127 
2128 	/*
2129 	 * We can't acquire locks on existing vnodes while in a transaction
2130 	 * without risking a deadlock.  This assumes that vfsync() can be
2131 	 * called without the vnode locked (which it can in DragonFly).
2132 	 * Otherwise we'd have to implement a multi-pass or flag the lock
2133 	 * failures and retry.
2134 	 *
2135 	 * The reclamation code interlocks with the sync list's token
2136 	 * (by removing the vnode from the scan list) before unlocking
2137 	 * the inode, giving us time to ref the inode.
2138 	 */
2139 	/*flags = VMSC_GETVP;*/
2140 	flags = 0;
2141 	if (waitfor & MNT_LAZY)
2142 		flags |= VMSC_ONEPASS;
2143 
2144 	/*
2145 	 * Start our flush transaction.  This does not return until all
2146 	 * concurrent transactions have completed and will prevent any
2147 	 * new transactions from running concurrently, except for the
2148 	 * buffer cache transactions.
2149 	 *
2150 	 * For efficiency do an async pass before making sure with a
2151 	 * synchronous pass on all related buffer cache buffers.  It
2152 	 * should theoretically not be possible for any new file buffers
2153 	 * to be instantiated during this sequence.
2154 	 */
2155 	hammer2_trans_init(&info.trans, pmp, HAMMER2_TRANS_ISFLUSH |
2156 					     HAMMER2_TRANS_PREFLUSH);
2157 	hammer2_run_unlinkq(&info.trans, pmp);
2158 
2159 	info.error = 0;
2160 	info.waitfor = MNT_NOWAIT;
2161 	vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2162 	info.waitfor = MNT_WAIT;
2163 	vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2164 
2165 	/*
2166 	 * Clear PREFLUSH.  This prevents (or asserts on) any new logical
2167 	 * buffer cache flushes which occur during the flush.  Device buffers
2168 	 * are not affected.
2169 	 */
2170 
2171 #if 0
2172 	if (info.error == 0 && (waitfor & MNT_WAIT)) {
2173 		info.waitfor = waitfor;
2174 		    vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2175 
2176 	}
2177 #endif
2178 	hammer2_bioq_sync(info.trans.pmp);
2179 	atomic_clear_int(&info.trans.flags, HAMMER2_TRANS_PREFLUSH);
2180 
2181 	total_error = 0;
2182 
2183 	/*
2184 	 * Flush all storage elements making up the cluster
2185 	 *
2186 	 * We must also flush any deleted siblings because the super-root
2187 	 * flush won't do it for us.  They all must be staged or the
2188 	 * super-root flush will not be able to update its block table
2189 	 * properly.
2190 	 *
2191 	 * XXX currently done serially instead of concurrently
2192 	 */
2193 	for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2194 		chain = iroot->cluster.array[i].chain;
2195 		if (chain) {
2196 			hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
2197 			hammer2_flush(&info.trans, chain);
2198 			hammer2_chain_unlock(chain);
2199 		}
2200 	}
2201 #if 0
2202 	hammer2_trans_done(&info.trans);
2203 #endif
2204 
2205 	/*
2206 	 * Flush all volume roots to synchronize PFS flushes with the
2207 	 * storage media.  Use a super-root transaction for each one.
2208 	 *
2209 	 * The flush code will detect super-root -> pfs-root chain
2210 	 * transitions using the last pfs-root flush.
2211 	 */
2212 	for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2213 		hammer2_chain_t *tmp;
2214 
2215 		chain = iroot->cluster.array[i].chain;
2216 		if (chain == NULL)
2217 			continue;
2218 
2219 		hmp = chain->hmp;
2220 
2221 		/*
2222 		 * We only have to flush each hmp once
2223 		 */
2224 		for (j = i - 1; j >= 0; --j) {
2225 			if ((tmp = iroot->cluster.array[j].chain) != NULL) {
2226 				if (tmp->hmp == hmp)
2227 					break;
2228 			}
2229 		}
2230 		if (j >= 0)
2231 			continue;
2232 		hammer2_trans_spmp(&info.trans, hmp->spmp);
2233 
2234 		/*
2235 		 * Force an update of the XID from the PFS root to the
2236 		 * topology root.  We couldn't do this from the PFS
2237 		 * transaction because a SPMP transaction is needed.
2238 		 * This does not modify blocks, instead what it does is
2239 		 * allow the flush code to find the transition point and
2240 		 * then update on the way back up.
2241 		 */
2242 		parent = chain->parent;
2243 		KKASSERT(chain->pmp != parent->pmp);
2244 		hammer2_chain_setflush(&info.trans, parent);
2245 
2246 		/*
2247 		 * Media mounts have two 'roots', vchain for the topology
2248 		 * and fchain for the free block table.  Flush both.
2249 		 *
2250 		 * Note that the topology and free block table are handled
2251 		 * independently, so the free block table can wind up being
2252 		 * ahead of the topology.  We depend on the bulk free scan
2253 		 * code to deal with any loose ends.
2254 		 */
2255 		hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2256 		hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2257 		if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2258 			/*
2259 			 * This will also modify vchain as a side effect,
2260 			 * mark vchain as modified now.
2261 			 */
2262 			hammer2_voldata_modify(hmp);
2263 			chain = &hmp->fchain;
2264 			hammer2_flush(&info.trans, chain);
2265 			KKASSERT(chain == &hmp->fchain);
2266 		}
2267 		hammer2_chain_unlock(&hmp->fchain);
2268 		hammer2_chain_unlock(&hmp->vchain);
2269 
2270 		hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2271 		if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2272 			chain = &hmp->vchain;
2273 			hammer2_flush(&info.trans, chain);
2274 			KKASSERT(chain == &hmp->vchain);
2275 			force_fchain = 1;
2276 		} else {
2277 			force_fchain = 0;
2278 		}
2279 		hammer2_chain_unlock(&hmp->vchain);
2280 
2281 #if 0
2282 		hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2283 		if ((hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) ||
2284 		    force_fchain) {
2285 			/* this will also modify vchain as a side effect */
2286 			chain = &hmp->fchain;
2287 			hammer2_flush(&info.trans, chain);
2288 			KKASSERT(chain == &hmp->fchain);
2289 		}
2290 		hammer2_chain_unlock(&hmp->fchain);
2291 #endif
2292 
2293 		error = 0;
2294 
2295 		/*
2296 		 * We can't safely flush the volume header until we have
2297 		 * flushed any device buffers which have built up.
2298 		 *
2299 		 * XXX this isn't being incremental
2300 		 */
2301 		vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
2302 		error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
2303 		vn_unlock(hmp->devvp);
2304 
2305 		/*
2306 		 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
2307 		 * volume header needs synchronization via hmp->volsync.
2308 		 *
2309 		 * XXX synchronize the flag & data with only this flush XXX
2310 		 */
2311 		if (error == 0 &&
2312 		    (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
2313 			struct buf *bp;
2314 
2315 			/*
2316 			 * Synchronize the disk before flushing the volume
2317 			 * header.
2318 			 */
2319 			bp = getpbuf(NULL);
2320 			bp->b_bio1.bio_offset = 0;
2321 			bp->b_bufsize = 0;
2322 			bp->b_bcount = 0;
2323 			bp->b_cmd = BUF_CMD_FLUSH;
2324 			bp->b_bio1.bio_done = biodone_sync;
2325 			bp->b_bio1.bio_flags |= BIO_SYNC;
2326 			vn_strategy(hmp->devvp, &bp->b_bio1);
2327 			biowait(&bp->b_bio1, "h2vol");
2328 			relpbuf(bp, NULL);
2329 
2330 			/*
2331 			 * Then we can safely flush the version of the
2332 			 * volume header synchronized by the flush code.
2333 			 */
2334 			i = hmp->volhdrno + 1;
2335 			if (i >= HAMMER2_NUM_VOLHDRS)
2336 				i = 0;
2337 			if (i * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
2338 			    hmp->volsync.volu_size) {
2339 				i = 0;
2340 			}
2341 			kprintf("sync volhdr %d %jd\n",
2342 				i, (intmax_t)hmp->volsync.volu_size);
2343 			bp = getblk(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2344 				    HAMMER2_PBUFSIZE, 0, 0);
2345 			atomic_clear_int(&hmp->vchain.flags,
2346 					 HAMMER2_CHAIN_VOLUMESYNC);
2347 			bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
2348 			bawrite(bp);
2349 			hmp->volhdrno = i;
2350 		}
2351 		if (error)
2352 			total_error = error;
2353 
2354 #if 0
2355 		hammer2_trans_done(&info.trans);
2356 #endif
2357 	}
2358 	hammer2_trans_done(&info.trans);
2359 
2360 	return (total_error);
2361 }
2362 
2363 /*
2364  * Sync passes.
2365  */
2366 static int
2367 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2368 {
2369 	struct hammer2_sync_info *info = data;
2370 	hammer2_inode_t *ip;
2371 	int error;
2372 
2373 	/*
2374 	 *
2375 	 */
2376 	ip = VTOI(vp);
2377 	if (ip == NULL)
2378 		return(0);
2379 	if (vp->v_type == VNON || vp->v_type == VBAD) {
2380 		vclrisdirty(vp);
2381 		return(0);
2382 	}
2383 	if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2384 	    RB_EMPTY(&vp->v_rbdirty_tree)) {
2385 		vclrisdirty(vp);
2386 		return(0);
2387 	}
2388 
2389 	/*
2390 	 * VOP_FSYNC will start a new transaction so replicate some code
2391 	 * here to do it inline (see hammer2_vop_fsync()).
2392 	 *
2393 	 * WARNING: The vfsync interacts with the buffer cache and might
2394 	 *          block, we can't hold the inode lock at that time.
2395 	 *	    However, we MUST ref ip before blocking to ensure that
2396 	 *	    it isn't ripped out from under us (since we do not
2397 	 *	    hold a lock on the vnode).
2398 	 */
2399 	hammer2_inode_ref(ip);
2400 	atomic_clear_int(&ip->flags, HAMMER2_INODE_MODIFIED);
2401 	if (vp)
2402 		vfsync(vp, MNT_NOWAIT, 1, NULL, NULL);
2403 
2404 	hammer2_inode_drop(ip);
2405 #if 1
2406 	error = 0;
2407 	if (error)
2408 		info->error = error;
2409 #endif
2410 	return(0);
2411 }
2412 
2413 static
2414 int
2415 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2416 {
2417 	return (0);
2418 }
2419 
2420 static
2421 int
2422 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2423 	       struct fid *fhp, struct vnode **vpp)
2424 {
2425 	return (0);
2426 }
2427 
2428 static
2429 int
2430 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2431 		 int *exflagsp, struct ucred **credanonp)
2432 {
2433 	return (0);
2434 }
2435 
2436 /*
2437  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2438  * header into the HMP
2439  *
2440  * XXX read four volhdrs and use the one with the highest TID whos CRC
2441  *     matches.
2442  *
2443  * XXX check iCRCs.
2444  *
2445  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2446  *     nonexistant locations.
2447  *
2448  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2449  */
2450 static
2451 int
2452 hammer2_install_volume_header(hammer2_mount_t *hmp)
2453 {
2454 	hammer2_volume_data_t *vd;
2455 	struct buf *bp;
2456 	hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2457 	int error_reported;
2458 	int error;
2459 	int valid;
2460 	int i;
2461 
2462 	error_reported = 0;
2463 	error = 0;
2464 	valid = 0;
2465 	bp = NULL;
2466 
2467 	/*
2468 	 * There are up to 4 copies of the volume header (syncs iterate
2469 	 * between them so there is no single master).  We don't trust the
2470 	 * volu_size field so we don't know precisely how large the filesystem
2471 	 * is, so depend on the OS to return an error if we go beyond the
2472 	 * block device's EOF.
2473 	 */
2474 	for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2475 		error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2476 			      HAMMER2_VOLUME_BYTES, &bp);
2477 		if (error) {
2478 			brelse(bp);
2479 			bp = NULL;
2480 			continue;
2481 		}
2482 
2483 		vd = (struct hammer2_volume_data *) bp->b_data;
2484 		if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2485 		    (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2486 			brelse(bp);
2487 			bp = NULL;
2488 			continue;
2489 		}
2490 
2491 		if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2492 			/* XXX: Reversed-endianness filesystem */
2493 			kprintf("hammer2: reverse-endian filesystem detected");
2494 			brelse(bp);
2495 			bp = NULL;
2496 			continue;
2497 		}
2498 
2499 		crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2500 		crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2501 				      HAMMER2_VOLUME_ICRC0_SIZE);
2502 		bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2503 		bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2504 				       HAMMER2_VOLUME_ICRC1_SIZE);
2505 		if ((crc0 != crc) || (bcrc0 != bcrc)) {
2506 			kprintf("hammer2 volume header crc "
2507 				"mismatch copy #%d %08x/%08x\n",
2508 				i, crc0, crc);
2509 			error_reported = 1;
2510 			brelse(bp);
2511 			bp = NULL;
2512 			continue;
2513 		}
2514 		if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2515 			valid = 1;
2516 			hmp->voldata = *vd;
2517 			hmp->volhdrno = i;
2518 		}
2519 		brelse(bp);
2520 		bp = NULL;
2521 	}
2522 	if (valid) {
2523 		hmp->volsync = hmp->voldata;
2524 		error = 0;
2525 		if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2526 			kprintf("hammer2: using volume header #%d\n",
2527 				hmp->volhdrno);
2528 		}
2529 	} else {
2530 		error = EINVAL;
2531 		kprintf("hammer2: no valid volume headers found!\n");
2532 	}
2533 	return (error);
2534 }
2535 
2536 /*
2537  * This handles hysteresis on regular file flushes.  Because the BIOs are
2538  * routed to a thread it is possible for an excessive number to build up
2539  * and cause long front-end stalls long before the runningbuffspace limit
2540  * is hit, so we implement hammer2_flush_pipe to control the
2541  * hysteresis.
2542  *
2543  * This is a particular problem when compression is used.
2544  */
2545 void
2546 hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp)
2547 {
2548 	atomic_add_int(&pmp->count_lwinprog, 1);
2549 }
2550 
2551 void
2552 hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp)
2553 {
2554 	int lwinprog;
2555 
2556 	lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2557 	if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2558 	    (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2559 		atomic_clear_int(&pmp->count_lwinprog,
2560 				 HAMMER2_LWINPROG_WAITING);
2561 		wakeup(&pmp->count_lwinprog);
2562 	}
2563 }
2564 
2565 void
2566 hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp)
2567 {
2568 	int lwinprog;
2569 
2570 	for (;;) {
2571 		lwinprog = pmp->count_lwinprog;
2572 		cpu_ccfence();
2573 		if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2574 			break;
2575 		tsleep_interlock(&pmp->count_lwinprog, 0);
2576 		atomic_set_int(&pmp->count_lwinprog, HAMMER2_LWINPROG_WAITING);
2577 		lwinprog = pmp->count_lwinprog;
2578 		if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2579 			break;
2580 		tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2581 	}
2582 }
2583 
2584 /*
2585  * Manage excessive memory resource use for chain and related
2586  * structures.
2587  */
2588 void
2589 hammer2_pfs_memory_wait(hammer2_pfsmount_t *pmp)
2590 {
2591 	uint32_t waiting;
2592 	uint32_t count;
2593 	uint32_t limit;
2594 #if 0
2595 	static int zzticks;
2596 #endif
2597 
2598 	/*
2599 	 * Atomic check condition and wait.  Also do an early speedup of
2600 	 * the syncer to try to avoid hitting the wait.
2601 	 */
2602 	for (;;) {
2603 		waiting = pmp->inmem_dirty_chains;
2604 		cpu_ccfence();
2605 		count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2606 
2607 		limit = pmp->mp->mnt_nvnodelistsize / 10;
2608 		if (limit < hammer2_limit_dirty_chains)
2609 			limit = hammer2_limit_dirty_chains;
2610 		if (limit < 1000)
2611 			limit = 1000;
2612 
2613 #if 0
2614 		if ((int)(ticks - zzticks) > hz) {
2615 			zzticks = ticks;
2616 			kprintf("count %ld %ld\n", count, limit);
2617 		}
2618 #endif
2619 
2620 		/*
2621 		 * Block if there are too many dirty chains present, wait
2622 		 * for the flush to clean some out.
2623 		 */
2624 		if (count > limit) {
2625 			tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2626 			if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2627 					       waiting,
2628 				       waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2629 				speedup_syncer(pmp->mp);
2630 				tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2631 				       "chnmem", hz);
2632 			}
2633 			continue;	/* loop on success or fail */
2634 		}
2635 
2636 		/*
2637 		 * Try to start an early flush before we are forced to block.
2638 		 */
2639 		if (count > limit * 7 / 10)
2640 			speedup_syncer(pmp->mp);
2641 		break;
2642 	}
2643 }
2644 
2645 void
2646 hammer2_pfs_memory_inc(hammer2_pfsmount_t *pmp)
2647 {
2648 	if (pmp) {
2649 		atomic_add_int(&pmp->inmem_dirty_chains, 1);
2650 	}
2651 }
2652 
2653 void
2654 hammer2_pfs_memory_wakeup(hammer2_pfsmount_t *pmp)
2655 {
2656 	uint32_t waiting;
2657 
2658 	if (pmp == NULL)
2659 		return;
2660 
2661 	for (;;) {
2662 		waiting = pmp->inmem_dirty_chains;
2663 		cpu_ccfence();
2664 		if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2665 				       waiting,
2666 				       (waiting - 1) &
2667 					~HAMMER2_DIRTYCHAIN_WAITING)) {
2668 			break;
2669 		}
2670 	}
2671 
2672 	if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2673 		wakeup(&pmp->inmem_dirty_chains);
2674 }
2675 
2676 /*
2677  * Debugging
2678  */
2679 void
2680 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2681 {
2682 	hammer2_chain_t *scan;
2683 	hammer2_chain_t *parent;
2684 
2685 	--*countp;
2686 	if (*countp == 0) {
2687 		kprintf("%*.*s...\n", tab, tab, "");
2688 		return;
2689 	}
2690 	if (*countp < 0)
2691 		return;
2692 	kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2693 		tab, tab, "", pfx,
2694 		chain, chain->bref.type,
2695 		chain->bref.key, chain->bref.keybits,
2696 		chain->bref.mirror_tid);
2697 
2698 	kprintf("%*.*s      [%08x] (%s) refs=%d\n",
2699 		tab, tab, "",
2700 		chain->flags,
2701 		((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2702 		chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2703 		chain->refs);
2704 
2705 	kprintf("%*.*s      core [%08x]",
2706 		tab, tab, "",
2707 		chain->core.flags);
2708 
2709 	parent = chain->parent;
2710 	if (parent)
2711 		kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2712 			tab, tab, "",
2713 			parent, parent->flags, parent->refs);
2714 	if (RB_EMPTY(&chain->core.rbtree)) {
2715 		kprintf("\n");
2716 	} else {
2717 		kprintf(" {\n");
2718 		RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2719 			hammer2_dump_chain(scan, tab + 4, countp, 'a');
2720 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2721 			kprintf("%*.*s}(%s)\n", tab, tab, "",
2722 				chain->data->ipdata.filename);
2723 		else
2724 			kprintf("%*.*s}\n", tab, tab, "");
2725 	}
2726 }
2727