hammer2 - Performance work
[dragonfly.git] / sys / vfs / hammer2 / hammer2_vfsops.c
1 /*
2  * Copyright (c) 2011-2018 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  *
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in
16  *    the documentation and/or other materials provided with the
17  *    distribution.
18  * 3. Neither the name of The DragonFly Project nor the names of its
19  *    contributors may be used to endorse or promote products derived
20  *    from this software without specific, prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/nlookup.h>
39 #include <sys/vnode.h>
40 #include <sys/mount.h>
41 #include <sys/fcntl.h>
42 #include <sys/buf.h>
43 #include <sys/uuid.h>
44 #include <sys/vfsops.h>
45 #include <sys/sysctl.h>
46 #include <sys/socket.h>
47 #include <sys/objcache.h>
48
49 #include <sys/proc.h>
50 #include <sys/namei.h>
51 #include <sys/mountctl.h>
52 #include <sys/dirent.h>
53 #include <sys/uio.h>
54
55 #include <sys/mutex.h>
56 #include <sys/mutex2.h>
57
58 #include "hammer2.h"
59 #include "hammer2_disk.h"
60 #include "hammer2_mount.h"
61 #include "hammer2_lz4.h"
62
63 #include "zlib/hammer2_zlib.h"
64
65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
66
67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
68
69 struct hammer2_sync_info {
70         int error;
71         int waitfor;
72         int pass;
73 };
74
75 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
76 static struct hammer2_mntlist hammer2_mntlist;
77
78 struct hammer2_pfslist hammer2_pfslist;
79 struct hammer2_pfslist hammer2_spmplist;
80 struct lock hammer2_mntlk;
81
82 int hammer2_supported_version = HAMMER2_VOL_VERSION_DEFAULT;
83 int hammer2_debug;
84 long hammer2_debug_inode;
85 int hammer2_cluster_meta_read = 1;      /* physical read-ahead */
86 int hammer2_cluster_data_read = 4;      /* physical read-ahead */
87 int hammer2_cluster_write = 0;          /* physical write clustering */
88 int hammer2_dedup_enable = 1;
89 int hammer2_always_compress = 0;        /* always try to compress */
90 int hammer2_inval_enable = 0;
91 int hammer2_flush_pipe = 100;
92 int hammer2_dio_count;
93 int hammer2_dio_limit = 256;
94 int hammer2_bulkfree_tps = 5000;
95 int hammer2_worker_rmask = 3;
96 long hammer2_chain_allocs;
97 long hammer2_chain_frees;
98 long hammer2_limit_dirty_chains;
99 long hammer2_limit_dirty_inodes;
100 long hammer2_count_modified_chains;
101 long hammer2_iod_invals;
102 long hammer2_iod_file_read;
103 long hammer2_iod_meta_read;
104 long hammer2_iod_indr_read;
105 long hammer2_iod_fmap_read;
106 long hammer2_iod_volu_read;
107 long hammer2_iod_file_write;
108 long hammer2_iod_file_wembed;
109 long hammer2_iod_file_wzero;
110 long hammer2_iod_file_wdedup;
111 long hammer2_iod_meta_write;
112 long hammer2_iod_indr_write;
113 long hammer2_iod_fmap_write;
114 long hammer2_iod_volu_write;
115 long hammer2_iod_inode_creates;
116 long hammer2_iod_inode_deletes;
117
118 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
119 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
120                 "Buffer used for compression.");
121
122 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
123 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
124                 "Buffer used for decompression.");
125
126 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
127
128 SYSCTL_INT(_vfs_hammer2, OID_AUTO, supported_version, CTLFLAG_RD,
129            &hammer2_supported_version, 0, "");
130 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
131            &hammer2_debug, 0, "");
132 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, debug_inode, CTLFLAG_RW,
133            &hammer2_debug_inode, 0, "");
134 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_meta_read, CTLFLAG_RW,
135            &hammer2_cluster_meta_read, 0, "");
136 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_data_read, CTLFLAG_RW,
137            &hammer2_cluster_data_read, 0, "");
138 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
139            &hammer2_cluster_write, 0, "");
140 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
141            &hammer2_dedup_enable, 0, "");
142 SYSCTL_INT(_vfs_hammer2, OID_AUTO, always_compress, CTLFLAG_RW,
143            &hammer2_always_compress, 0, "");
144 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
145            &hammer2_inval_enable, 0, "");
146 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
147            &hammer2_flush_pipe, 0, "");
148 SYSCTL_INT(_vfs_hammer2, OID_AUTO, worker_rmask, CTLFLAG_RW,
149            &hammer2_worker_rmask, 0, "");
150 SYSCTL_INT(_vfs_hammer2, OID_AUTO, bulkfree_tps, CTLFLAG_RW,
151            &hammer2_bulkfree_tps, 0, "");
152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
153            &hammer2_chain_allocs, 0, "");
154 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
155            &hammer2_chain_frees, 0, "");
156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
157            &hammer2_limit_dirty_chains, 0, "");
158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_inodes, CTLFLAG_RW,
159            &hammer2_limit_dirty_inodes, 0, "");
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
161            &hammer2_count_modified_chains, 0, "");
162 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
163            &hammer2_dio_count, 0, "");
164 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_limit, CTLFLAG_RW,
165            &hammer2_dio_limit, 0, "");
166
167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
168            &hammer2_iod_invals, 0, "");
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
170            &hammer2_iod_file_read, 0, "");
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
172            &hammer2_iod_meta_read, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
174            &hammer2_iod_indr_read, 0, "");
175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
176            &hammer2_iod_fmap_read, 0, "");
177 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
178            &hammer2_iod_volu_read, 0, "");
179
180 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
181            &hammer2_iod_file_write, 0, "");
182 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
183            &hammer2_iod_file_wembed, 0, "");
184 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
185            &hammer2_iod_file_wzero, 0, "");
186 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
187            &hammer2_iod_file_wdedup, 0, "");
188 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
189            &hammer2_iod_meta_write, 0, "");
190 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
191            &hammer2_iod_indr_write, 0, "");
192 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
193            &hammer2_iod_fmap_write, 0, "");
194 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
195            &hammer2_iod_volu_write, 0, "");
196 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_creates, CTLFLAG_RW,
197            &hammer2_iod_inode_creates, 0, "");
198 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_inode_deletes, CTLFLAG_RW,
199            &hammer2_iod_inode_deletes, 0, "");
200
201 long hammer2_process_icrc32;
202 long hammer2_process_xxhash64;
203 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_icrc32, CTLFLAG_RW,
204            &hammer2_process_icrc32, 0, "");
205 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, process_xxhash64, CTLFLAG_RW,
206            &hammer2_process_xxhash64, 0, "");
207
208 static int hammer2_vfs_init(struct vfsconf *conf);
209 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
210 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
211                                 struct ucred *cred);
212 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
213                                 struct vnode *, struct ucred *);
214 static int hammer2_recovery(hammer2_dev_t *hmp);
215 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
216 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
217 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
218                                 struct ucred *cred);
219 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
220                                 struct ucred *cred);
221 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
222                                 struct fid *fhp, struct vnode **vpp);
223 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
224 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
225                                 int *exflagsp, struct ucred **credanonp);
226 static void hammer2_vfs_modifying(struct mount *mp);
227
228 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
229 #if 0
230 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
231 #endif
232
233 static void hammer2_update_pmps(hammer2_dev_t *hmp);
234
235 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
236 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
237                                 hammer2_dev_t *hmp);
238 static int hammer2_fixup_pfses(hammer2_dev_t *hmp);
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         .vfs_modifying  = hammer2_vfs_modifying
257 };
258
259 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
260
261 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
262 MODULE_VERSION(hammer2, 1);
263
264 static
265 int
266 hammer2_vfs_init(struct vfsconf *conf)
267 {
268         static struct objcache_malloc_args margs_read;
269         static struct objcache_malloc_args margs_write;
270         static struct objcache_malloc_args margs_vop;
271
272         int error;
273
274         error = 0;
275         kmalloc_raise_limit(M_HAMMER2, 0);      /* unlimited */
276
277         /*
278          * A large DIO cache is needed to retain dedup enablement masks.
279          * The bulkfree code clears related masks as part of the disk block
280          * recycling algorithm, preventing it from being used for a later
281          * dedup.
282          *
283          * NOTE: A large buffer cache can actually interfere with dedup
284          *       operation because we dedup based on media physical buffers
285          *       and not logical buffers.  Try to make the DIO case large
286          *       enough to avoid this problem, but also cap it.
287          */
288         hammer2_dio_limit = nbuf * 2;
289         if (hammer2_dio_limit > 100000)
290                 hammer2_dio_limit = 100000;
291
292         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
293                 error = EINVAL;
294         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
295                 error = EINVAL;
296         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
297                 error = EINVAL;
298
299         if (error)
300                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
301         
302         margs_read.objsize = 65536;
303         margs_read.mtype = M_HAMMER2_DEBUFFER;
304         
305         margs_write.objsize = 32768;
306         margs_write.mtype = M_HAMMER2_CBUFFER;
307
308         margs_vop.objsize = sizeof(hammer2_xop_t);
309         margs_vop.mtype = M_HAMMER2;
310         
311         /*
312          * Note thaht for the XOPS cache we want backing store allocations
313          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
314          * confusion), so use the backing store function that does it.  This
315          * means that initial XOPS objects are zerod but REUSED objects are
316          * not.  So we are responsible for cleaning the object up sufficiently
317          * for our needs before objcache_put()ing it back (typically just the
318          * FIFO indices).
319          */
320         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
321                                 0, 1, NULL, NULL, NULL,
322                                 objcache_malloc_alloc,
323                                 objcache_malloc_free,
324                                 &margs_read);
325         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
326                                 0, 1, NULL, NULL, NULL,
327                                 objcache_malloc_alloc,
328                                 objcache_malloc_free,
329                                 &margs_write);
330         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
331                                 0, 1, NULL, NULL, NULL,
332                                 objcache_malloc_alloc_zero,
333                                 objcache_malloc_free,
334                                 &margs_vop);
335
336
337         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
338         TAILQ_INIT(&hammer2_mntlist);
339         TAILQ_INIT(&hammer2_pfslist);
340         TAILQ_INIT(&hammer2_spmplist);
341
342         hammer2_limit_dirty_chains = maxvnodes / 10;
343         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
344                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
345         if (hammer2_limit_dirty_chains < 1000)
346                 hammer2_limit_dirty_chains = 1000;
347
348         hammer2_limit_dirty_inodes = maxvnodes / 25;
349         if (hammer2_limit_dirty_inodes < 100)
350                 hammer2_limit_dirty_inodes = 100;
351         if (hammer2_limit_dirty_inodes > HAMMER2_LIMIT_DIRTY_INODES)
352                 hammer2_limit_dirty_inodes = HAMMER2_LIMIT_DIRTY_INODES;
353
354         return (error);
355 }
356
357 static
358 int
359 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
360 {
361         objcache_destroy(cache_buffer_read);
362         objcache_destroy(cache_buffer_write);
363         objcache_destroy(cache_xops);
364         return 0;
365 }
366
367 /*
368  * Core PFS allocator.  Used to allocate or reference the pmp structure
369  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
370  * The pmp can be passed in or loaded by this function using the chain and
371  * inode data.
372  *
373  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
374  * transactions.  Note that synchronization does not use this field.
375  * (typically frontend operations and synchronization cannot run on the
376  * same PFS node at the same time).
377  *
378  * XXX check locking
379  */
380 hammer2_pfs_t *
381 hammer2_pfsalloc(hammer2_chain_t *chain,
382                  const hammer2_inode_data_t *ripdata,
383                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
384 {
385         hammer2_pfs_t *pmp;
386         hammer2_inode_t *iroot;
387         int count;
388         int i;
389         int j;
390
391         pmp = NULL;
392
393         /*
394          * Locate or create the PFS based on the cluster id.  If ripdata
395          * is NULL this is a spmp which is unique and is always allocated.
396          *
397          * If the device is mounted in local mode all PFSs are considered
398          * independent and not part of any cluster (for debugging only).
399          */
400         if (ripdata) {
401                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
402                         if (force_local != pmp->force_local)
403                                 continue;
404                         if (force_local == NULL &&
405                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
406                                  sizeof(pmp->pfs_clid)) == 0) {
407                                         break;
408                         } else if (force_local && pmp->pfs_names[0] &&
409                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
410                                         break;
411                         }
412                 }
413         }
414
415         if (pmp == NULL) {
416                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
417                 pmp->force_local = force_local;
418                 hammer2_trans_manage_init(pmp);
419                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
420                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
421                 lockinit(&pmp->lock, "pfslk", 0, 0);
422                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
423                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
424                 spin_init(&pmp->xop_spin, "h2xop");
425                 spin_init(&pmp->lru_spin, "h2lru");
426                 RB_INIT(&pmp->inum_tree);
427                 TAILQ_INIT(&pmp->syncq);
428                 TAILQ_INIT(&pmp->depq);
429                 TAILQ_INIT(&pmp->lru_list);
430                 spin_init(&pmp->list_spin, "h2pfsalloc_list");
431
432                 /*
433                  * Distribute backend operations to threads
434                  */
435                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
436                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
437
438                 /*
439                  * Save the last media transaction id for the flusher.  Set
440                  * initial 
441                  */
442                 if (ripdata) {
443                         pmp->pfs_clid = ripdata->meta.pfs_clid;
444                         TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
445                 } else {
446                         pmp->flags |= HAMMER2_PMPF_SPMP;
447                         TAILQ_INSERT_TAIL(&hammer2_spmplist, pmp, mntentry);
448                 }
449
450                 /*
451                  * The synchronization thread may start too early, make
452                  * sure it stays frozen until we are ready to let it go.
453                  * XXX
454                  */
455                 /*
456                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
457                                          HAMMER2_THREAD_REMASTER;
458                 */
459         }
460
461         /*
462          * Create the PFS's root inode and any missing XOP helper threads.
463          */
464         if ((iroot = pmp->iroot) == NULL) {
465                 iroot = hammer2_inode_get(pmp, NULL, 1, -1);
466                 if (ripdata)
467                         iroot->meta = ripdata->meta;
468                 pmp->iroot = iroot;
469                 hammer2_inode_ref(iroot);
470                 hammer2_inode_unlock(iroot);
471         }
472
473         /*
474          * Stop here if no chain is passed in.
475          */
476         if (chain == NULL)
477                 goto done;
478
479         /*
480          * When a chain is passed in we must add it to the PFS's root
481          * inode, update pmp->pfs_types[], and update the syncronization
482          * threads.
483          *
484          * When forcing local mode, mark the PFS as a MASTER regardless.
485          *
486          * At the moment empty spots can develop due to removals or failures.
487          * Ultimately we want to re-fill these spots but doing so might
488          * confused running code. XXX
489          */
490         hammer2_inode_ref(iroot);
491         hammer2_mtx_ex(&iroot->lock);
492         j = iroot->cluster.nchains;
493
494         if (j == HAMMER2_MAXCLUSTER) {
495                 kprintf("hammer2_mount: cluster full!\n");
496                 /* XXX fatal error? */
497         } else {
498                 KKASSERT(chain->pmp == NULL);
499                 chain->pmp = pmp;
500                 hammer2_chain_ref(chain);
501                 iroot->cluster.array[j].chain = chain;
502                 if (force_local)
503                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
504                 else
505                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
506                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
507                 pmp->pfs_hmps[j] = chain->hmp;
508                 hammer2_spin_ex(&pmp->inum_spin);
509                 pmp->pfs_iroot_blocksets[j] = chain->data->ipdata.u.blockset;
510                 hammer2_spin_unex(&pmp->inum_spin);
511
512                 /*
513                  * If the PFS is already mounted we must account
514                  * for the mount_count here.
515                  */
516                 if (pmp->mp)
517                         ++chain->hmp->mount_count;
518
519                 /*
520                  * May have to fixup dirty chain tracking.  Previous
521                  * pmp was NULL so nothing to undo.
522                  */
523                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
524                         hammer2_pfs_memory_inc(pmp);
525                 ++j;
526         }
527         iroot->cluster.nchains = j;
528
529         /*
530          * Update nmasters from any PFS inode which is part of the cluster.
531          * It is possible that this will result in a value which is too
532          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
533          * override this value later on.
534          *
535          * (This informs us of masters that might not currently be
536          *  discoverable by this mount).
537          */
538         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
539                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
540         }
541
542         /*
543          * Count visible masters.  Masters are usually added with
544          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
545          * are more (XXX and must update the master inodes).
546          */
547         count = 0;
548         for (i = 0; i < iroot->cluster.nchains; ++i) {
549                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
550                         ++count;
551         }
552         if (pmp->pfs_nmasters < count)
553                 pmp->pfs_nmasters = count;
554
555         /*
556          * Create missing synchronization and support threads.
557          *
558          * Single-node masters (including snapshots) have nothing to
559          * synchronize and do not require this thread.
560          *
561          * Multi-node masters or any number of soft masters, slaves, copy,
562          * or other PFS types need the thread.
563          *
564          * Each thread is responsible for its particular cluster index.
565          * We use independent threads so stalls or mismatches related to
566          * any given target do not affect other targets.
567          */
568         for (i = 0; i < iroot->cluster.nchains; ++i) {
569                 /*
570                  * Single-node masters (including snapshots) have nothing
571                  * to synchronize and will make direct xops support calls,
572                  * thus they do not require this thread.
573                  *
574                  * Note that there can be thousands of snapshots.  We do not
575                  * want to create thousands of threads.
576                  */
577                 if (pmp->pfs_nmasters <= 1 &&
578                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
579                         continue;
580                 }
581
582                 /*
583                  * Sync support thread
584                  */
585                 if (pmp->sync_thrs[i].td == NULL) {
586                         hammer2_thr_create(&pmp->sync_thrs[i], pmp, NULL,
587                                            "h2nod", i, -1,
588                                            hammer2_primary_sync_thread);
589                 }
590         }
591
592         /*
593          * Create missing Xop threads
594          *
595          * NOTE: We create helper threads for all mounted PFSs or any
596          *       PFSs with 2+ nodes (so the sync thread can update them,
597          *       even if not mounted).
598          */
599         if (pmp->mp || iroot->cluster.nchains >= 2)
600                 hammer2_xop_helper_create(pmp);
601
602         hammer2_mtx_unlock(&iroot->lock);
603         hammer2_inode_drop(iroot);
604 done:
605         return pmp;
606 }
607
608 /*
609  * Deallocate an element of a probed PFS.  If destroying and this is a
610  * MASTER, adjust nmasters.
611  *
612  * This function does not physically destroy the PFS element in its device
613  * under the super-root  (see hammer2_ioctl_pfs_delete()).
614  */
615 void
616 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
617 {
618         hammer2_inode_t *iroot;
619         hammer2_chain_t *chain;
620         int j;
621
622         /*
623          * Cleanup our reference on iroot.  iroot is (should) not be needed
624          * by the flush code.
625          */
626         iroot = pmp->iroot;
627         if (iroot) {
628                 /*
629                  * Stop synchronizing
630                  *
631                  * XXX flush after acquiring the iroot lock.
632                  * XXX clean out the cluster index from all inode structures.
633                  */
634                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
635
636                 /*
637                  * Remove the cluster index from the group.  If destroying
638                  * the PFS and this is a master, adjust pfs_nmasters.
639                  */
640                 hammer2_mtx_ex(&iroot->lock);
641                 chain = iroot->cluster.array[clindex].chain;
642                 iroot->cluster.array[clindex].chain = NULL;
643
644                 switch(pmp->pfs_types[clindex]) {
645                 case HAMMER2_PFSTYPE_MASTER:
646                         if (destroying && pmp->pfs_nmasters > 0)
647                                 --pmp->pfs_nmasters;
648                         /* XXX adjust ripdata->meta.pfs_nmasters */
649                         break;
650                 default:
651                         break;
652                 }
653                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
654
655                 hammer2_mtx_unlock(&iroot->lock);
656
657                 /*
658                  * Release the chain.
659                  */
660                 if (chain) {
661                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
662                         hammer2_chain_drop(chain);
663                 }
664
665                 /*
666                  * Terminate all XOP threads for the cluster index.
667                  */
668                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
669                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
670         }
671 }
672
673 /*
674  * Destroy a PFS, typically only occurs after the last mount on a device
675  * has gone away.
676  */
677 static void
678 hammer2_pfsfree(hammer2_pfs_t *pmp)
679 {
680         hammer2_inode_t *iroot;
681         hammer2_chain_t *chain;
682         int chains_still_present = 0;
683         int i;
684         int j;
685
686         /*
687          * Cleanup our reference on iroot.  iroot is (should) not be needed
688          * by the flush code.
689          */
690         if (pmp->flags & HAMMER2_PMPF_SPMP)
691                 TAILQ_REMOVE(&hammer2_spmplist, pmp, mntentry);
692         else
693                 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
694
695         /*
696          * Cleanup chains remaining on LRU list.
697          */
698         hammer2_spin_ex(&pmp->lru_spin);
699         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
700                 KKASSERT(chain->flags & HAMMER2_CHAIN_ONLRU);
701                 atomic_add_int(&pmp->lru_count, -1);
702                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU);
703                 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node);
704                 hammer2_chain_ref(chain);
705                 hammer2_spin_unex(&pmp->lru_spin);
706                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
707                 hammer2_chain_drop(chain);
708                 hammer2_spin_ex(&pmp->lru_spin);
709         }
710         hammer2_spin_unex(&pmp->lru_spin);
711
712         /*
713          * Clean up iroot
714          */
715         iroot = pmp->iroot;
716         if (iroot) {
717                 for (i = 0; i < iroot->cluster.nchains; ++i) {
718                         hammer2_thr_delete(&pmp->sync_thrs[i]);
719                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
720                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
721                         chain = iroot->cluster.array[i].chain;
722                         if (chain && !RB_EMPTY(&chain->core.rbtree)) {
723                                 kprintf("hammer2: Warning pmp %p still "
724                                         "has active chains\n", pmp);
725                                 chains_still_present = 1;
726                         }
727                 }
728 #if REPORT_REFS_ERRORS
729                 if (iroot->refs != 1)
730                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
731                                 iroot, iroot->refs);
732 #else
733                 KKASSERT(iroot->refs == 1);
734 #endif
735                 /* ref for iroot */
736                 hammer2_inode_drop(iroot);
737                 pmp->iroot = NULL;
738         }
739
740         /*
741          * Free remaining pmp resources
742          */
743         if (chains_still_present) {
744                 kprintf("hammer2: cannot free pmp %p, still in use\n", pmp);
745         } else {
746                 kmalloc_destroy(&pmp->mmsg);
747                 kmalloc_destroy(&pmp->minode);
748                 kfree(pmp, M_HAMMER2);
749         }
750 }
751
752 /*
753  * Remove all references to hmp from the pfs list.  Any PFS which becomes
754  * empty is terminated and freed.
755  *
756  * XXX inefficient.
757  */
758 static void
759 hammer2_pfsfree_scan(hammer2_dev_t *hmp, int which)
760 {
761         hammer2_pfs_t *pmp;
762         hammer2_inode_t *iroot;
763         hammer2_chain_t *rchain;
764         int i;
765         int j;
766         struct hammer2_pfslist *wlist;
767
768         if (which == 0)
769                 wlist = &hammer2_pfslist;
770         else
771                 wlist = &hammer2_spmplist;
772 again:
773         TAILQ_FOREACH(pmp, wlist, mntentry) {
774                 if ((iroot = pmp->iroot) == NULL)
775                         continue;
776
777                 /*
778                  * Determine if this PFS is affected.  If it is we must
779                  * freeze all management threads and lock its iroot.
780                  *
781                  * Freezing a management thread forces it idle, operations
782                  * in-progress will be aborted and it will have to start
783                  * over again when unfrozen, or exit if told to exit.
784                  */
785                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
786                         if (pmp->pfs_hmps[i] == hmp)
787                                 break;
788                 }
789                 if (i == HAMMER2_MAXCLUSTER)
790                         continue;
791
792                 hammer2_vfs_sync_pmp(pmp, MNT_WAIT);
793
794                 /*
795                  * Make sure all synchronization threads are locked
796                  * down.
797                  */
798                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
799                         if (pmp->pfs_hmps[i] == NULL)
800                                 continue;
801                         hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
802                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
803                                 hammer2_thr_freeze_async(
804                                         &pmp->xop_groups[j].thrs[i]);
805                         }
806                 }
807                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
808                         if (pmp->pfs_hmps[i] == NULL)
809                                 continue;
810                         hammer2_thr_freeze(&pmp->sync_thrs[i]);
811                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
812                                 hammer2_thr_freeze(
813                                         &pmp->xop_groups[j].thrs[i]);
814                         }
815                 }
816
817                 /*
818                  * Lock the inode and clean out matching chains.
819                  * Note that we cannot use hammer2_inode_lock_*()
820                  * here because that would attempt to validate the
821                  * cluster that we are in the middle of ripping
822                  * apart.
823                  *
824                  * WARNING! We are working directly on the inodes
825                  *          embedded cluster.
826                  */
827                 hammer2_mtx_ex(&iroot->lock);
828
829                 /*
830                  * Remove the chain from matching elements of the PFS.
831                  */
832                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
833                         if (pmp->pfs_hmps[i] != hmp)
834                                 continue;
835                         hammer2_thr_delete(&pmp->sync_thrs[i]);
836                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
837                                 hammer2_thr_delete(
838                                         &pmp->xop_groups[j].thrs[i]);
839                         }
840                         rchain = iroot->cluster.array[i].chain;
841                         iroot->cluster.array[i].chain = NULL;
842                         pmp->pfs_types[i] = 0;
843                         if (pmp->pfs_names[i]) {
844                                 kfree(pmp->pfs_names[i], M_HAMMER2);
845                                 pmp->pfs_names[i] = NULL;
846                         }
847                         if (rchain) {
848                                 hammer2_chain_drop(rchain);
849                                 /* focus hint */
850                                 if (iroot->cluster.focus == rchain)
851                                         iroot->cluster.focus = NULL;
852                         }
853                         pmp->pfs_hmps[i] = NULL;
854                 }
855                 hammer2_mtx_unlock(&iroot->lock);
856
857                 /*
858                  * Cleanup trailing chains.  Gaps may remain.
859                  */
860                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
861                         if (pmp->pfs_hmps[i])
862                                 break;
863                 }
864                 iroot->cluster.nchains = i + 1;
865
866                 /*
867                  * If the PMP has no elements remaining we can destroy it.
868                  * (this will transition management threads from frozen->exit).
869                  */
870                 if (iroot->cluster.nchains == 0) {
871                         /*
872                          * If this was the hmp's spmp, we need to clean
873                          * a little more stuff out.
874                          */
875                         if (hmp->spmp == pmp) {
876                                 hmp->spmp = NULL;
877                                 hmp->vchain.pmp = NULL;
878                                 hmp->fchain.pmp = NULL;
879                         }
880
881                         /*
882                          * Free the pmp and restart the loop
883                          */
884                         KKASSERT(TAILQ_EMPTY(&pmp->syncq));
885                         KKASSERT(TAILQ_EMPTY(&pmp->depq));
886                         hammer2_pfsfree(pmp);
887                         goto again;
888                 }
889
890                 /*
891                  * If elements still remain we need to set the REMASTER
892                  * flag and unfreeze it.
893                  */
894                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
895                         if (pmp->pfs_hmps[i] == NULL)
896                                 continue;
897                         hammer2_thr_remaster(&pmp->sync_thrs[i]);
898                         hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
899                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
900                                 hammer2_thr_remaster(
901                                         &pmp->xop_groups[j].thrs[i]);
902                                 hammer2_thr_unfreeze(
903                                         &pmp->xop_groups[j].thrs[i]);
904                         }
905                 }
906         }
907 }
908
909 /*
910  * Mount or remount HAMMER2 fileystem from physical media
911  *
912  *      mountroot
913  *              mp              mount point structure
914  *              path            NULL
915  *              data            <unused>
916  *              cred            <unused>
917  *
918  *      mount
919  *              mp              mount point structure
920  *              path            path to mount point
921  *              data            pointer to argument structure in user space
922  *                      volume  volume path (device@LABEL form)
923  *                      hflags  user mount flags
924  *              cred            user credentials
925  *
926  * RETURNS:     0       Success
927  *              !0      error number
928  */
929 static
930 int
931 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
932                   struct ucred *cred)
933 {
934         struct hammer2_mount_info info;
935         hammer2_pfs_t *pmp;
936         hammer2_pfs_t *spmp;
937         hammer2_dev_t *hmp;
938         hammer2_dev_t *force_local;
939         hammer2_key_t key_next;
940         hammer2_key_t key_dummy;
941         hammer2_key_t lhc;
942         struct vnode *devvp;
943         struct nlookupdata nd;
944         hammer2_chain_t *parent;
945         hammer2_chain_t *chain;
946         const hammer2_inode_data_t *ripdata;
947         hammer2_blockref_t bref;
948         struct file *fp;
949         char devstr[MNAMELEN];
950         size_t size;
951         size_t done;
952         char *dev;
953         char *label;
954         int ronly = 1;
955         int error;
956         int i;
957
958         hmp = NULL;
959         pmp = NULL;
960         dev = NULL;
961         label = NULL;
962         devvp = NULL;
963
964         if (path == NULL) {
965                 /*
966                  * Root mount
967                  */
968                 bzero(&info, sizeof(info));
969                 info.cluster_fd = -1;
970                 ksnprintf(devstr, sizeof(devstr), "%s",
971                           mp->mnt_stat.f_mntfromname);
972                 kprintf("hammer2_mount: root '%s'\n", devstr);
973                 done = strlen(devstr) + 1;
974         } else {
975                 /*
976                  * Non-root mount or updating a mount
977                  */
978                 error = copyin(data, &info, sizeof(info));
979                 if (error)
980                         return (error);
981
982                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
983                 if (error)
984                         return (error);
985                 kprintf("hammer2_mount: '%s'\n", devstr);
986         }
987
988         /*
989          * Extract device and label, automatically mount @BOOT, @ROOT, or @DATA
990          * if no label specified, based on the partition id.  Error out if no
991          * label or device (with partition id) is specified.  This is strictly
992          * a convenience to match the default label created by newfs_hammer2,
993          * our preference is that a label always be specified.
994          *
995          * NOTE: We allow 'mount @LABEL <blah>'... that is, a mount command
996          *       that does not specify a device, as long as some H2 label
997          *       has already been mounted from that device.  This makes
998          *       mounting snapshots a lot easier.
999          */
1000         dev = devstr;
1001         label = strchr(devstr, '@');
1002         if (label && ((label + 1) - dev) > done) {
1003                 kprintf("hammer2: mount: bad label %s/%zd\n",
1004                         devstr, done);
1005                 return (EINVAL);
1006         }
1007         if (label == NULL || label[1] == 0) {
1008                 char slice;
1009
1010                 if (label == NULL)
1011                         label = devstr + strlen(devstr);
1012                 else
1013                         *label = '\0';          /* clean up trailing @ */
1014
1015                 slice = label[-1];
1016                 switch(slice) {
1017                 case 'a':
1018                         label = "BOOT";
1019                         break;
1020                 case 'd':
1021                         label = "ROOT";
1022                         break;
1023                 default:
1024                         label = "DATA";
1025                         break;
1026                 }
1027         } else {
1028                 *label = '\0';
1029                 label++;
1030         }
1031
1032         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\" rdonly=%d\n",
1033                 dev, label, (mp->mnt_flag & MNT_RDONLY));
1034
1035         if (mp->mnt_flag & MNT_UPDATE) {
1036                 /*
1037                  * Update mount.  Note that pmp->iroot->cluster is
1038                  * an inode-embedded cluster and thus cannot be
1039                  * directly locked.
1040                  *
1041                  * XXX HAMMER2 needs to implement NFS export via
1042                  *     mountctl.
1043                  */
1044                 hammer2_cluster_t *cluster;
1045
1046                 pmp = MPTOPMP(mp);
1047                 pmp->hflags = info.hflags;
1048                 cluster = &pmp->iroot->cluster;
1049                 for (i = 0; i < cluster->nchains; ++i) {
1050                         if (cluster->array[i].chain == NULL)
1051                                 continue;
1052                         hmp = cluster->array[i].chain->hmp;
1053                         devvp = hmp->devvp;
1054                         error = hammer2_remount(hmp, mp, path,
1055                                                 devvp, cred);
1056                         if (error)
1057                                 break;
1058                 }
1059
1060                 return error;
1061         }
1062
1063         /*
1064          * HMP device mount
1065          *
1066          * If a path is specified and dev is not an empty string, lookup the
1067          * name and verify that it referes to a block device.
1068          *
1069          * If a path is specified and dev is an empty string we fall through
1070          * and locate the label in the hmp search.
1071          */
1072         if (path && *dev != 0) {
1073                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
1074                 if (error == 0)
1075                         error = nlookup(&nd);
1076                 if (error == 0)
1077                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
1078                 nlookup_done(&nd);
1079         } else if (path == NULL) {
1080                 /* root mount */
1081                 cdev_t cdev = kgetdiskbyname(dev);
1082                 error = bdevvp(cdev, &devvp);
1083                 if (error)
1084                         kprintf("hammer2: cannot find '%s'\n", dev);
1085         } else {
1086                 /*
1087                  * We will locate the hmp using the label in the hmp loop.
1088                  */
1089                 error = 0;
1090         }
1091
1092         /*
1093          * Make sure its a block device.  Do not check to see if it is
1094          * already mounted until we determine that its a fresh H2 device.
1095          */
1096         if (error == 0 && devvp) {
1097                 vn_isdisk(devvp, &error);
1098         }
1099
1100         /*
1101          * Determine if the device has already been mounted.  After this
1102          * check hmp will be non-NULL if we are doing the second or more
1103          * hammer2 mounts from the same device.
1104          */
1105         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1106         if (devvp) {
1107                 /*
1108                  * Match the device.  Due to the way devfs works,
1109                  * we may not be able to directly match the vnode pointer,
1110                  * so also check to see if the underlying device matches.
1111                  */
1112                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1113                         if (hmp->devvp == devvp)
1114                                 break;
1115                         if (devvp->v_rdev &&
1116                             hmp->devvp->v_rdev == devvp->v_rdev) {
1117                                 break;
1118                         }
1119                 }
1120
1121                 /*
1122                  * If no match this may be a fresh H2 mount, make sure
1123                  * the device is not mounted on anything else.
1124                  */
1125                 if (hmp == NULL)
1126                         error = vfs_mountedon(devvp);
1127         } else if (error == 0) {
1128                 /*
1129                  * Match the label to a pmp already probed.
1130                  */
1131                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
1132                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
1133                                 if (pmp->pfs_names[i] &&
1134                                     strcmp(pmp->pfs_names[i], label) == 0) {
1135                                         hmp = pmp->pfs_hmps[i];
1136                                         break;
1137                                 }
1138                         }
1139                         if (hmp)
1140                                 break;
1141                 }
1142                 if (hmp == NULL)
1143                         error = ENOENT;
1144         }
1145
1146         /*
1147          * Open the device if this isn't a secondary mount and construct
1148          * the H2 device mount (hmp).
1149          */
1150         if (hmp == NULL) {
1151                 hammer2_chain_t *schain;
1152                 hammer2_xid_t xid;
1153                 hammer2_xop_head_t xop;
1154
1155                 if (error == 0 && vcount(devvp) > 0) {
1156                         kprintf("Primary device already has references\n");
1157                         error = EBUSY;
1158                 }
1159
1160                 /*
1161                  * Now open the device
1162                  */
1163                 if (error == 0) {
1164                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1165                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1166                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1167                         if (error == 0) {
1168                                 error = VOP_OPEN(devvp,
1169                                              (ronly ? FREAD : FREAD | FWRITE),
1170                                              FSCRED, NULL);
1171                         }
1172                         vn_unlock(devvp);
1173                 }
1174                 if (error && devvp) {
1175                         vrele(devvp);
1176                         devvp = NULL;
1177                 }
1178                 if (error) {
1179                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1180                         return error;
1181                 }
1182                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1183                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1184                 hmp->ronly = ronly;
1185                 hmp->devvp = devvp;
1186                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1187                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1188                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1189                 RB_INIT(&hmp->iotree);
1190                 spin_init(&hmp->io_spin, "h2mount_io");
1191                 spin_init(&hmp->list_spin, "h2mount_list");
1192
1193                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1194                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1195                 lockinit(&hmp->bflock, "h2bflk", 0, 0);
1196
1197                 /*
1198                  * vchain setup. vchain.data is embedded.
1199                  * vchain.refs is initialized and will never drop to 0.
1200                  *
1201                  * NOTE! voldata is not yet loaded.
1202                  */
1203                 hmp->vchain.hmp = hmp;
1204                 hmp->vchain.refs = 1;
1205                 hmp->vchain.data = (void *)&hmp->voldata;
1206                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1207                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1208                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1209
1210                 hammer2_chain_core_init(&hmp->vchain);
1211                 /* hmp->vchain.u.xxx is left NULL */
1212
1213                 /*
1214                  * fchain setup.  fchain.data is embedded.
1215                  * fchain.refs is initialized and will never drop to 0.
1216                  *
1217                  * The data is not used but needs to be initialized to
1218                  * pass assertion muster.  We use this chain primarily
1219                  * as a placeholder for the freemap's top-level RBTREE
1220                  * so it does not interfere with the volume's topology
1221                  * RBTREE.
1222                  */
1223                 hmp->fchain.hmp = hmp;
1224                 hmp->fchain.refs = 1;
1225                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1226                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1227                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1228                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1229                 hmp->fchain.bref.methods =
1230                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1231                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1232
1233                 hammer2_chain_core_init(&hmp->fchain);
1234                 /* hmp->fchain.u.xxx is left NULL */
1235
1236                 /*
1237                  * Install the volume header and initialize fields from
1238                  * voldata.
1239                  */
1240                 error = hammer2_install_volume_header(hmp);
1241                 if (error) {
1242                         hammer2_unmount_helper(mp, NULL, hmp);
1243                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1244                         hammer2_vfs_unmount(mp, MNT_FORCE);
1245                         return error;
1246                 }
1247
1248                 /*
1249                  * Really important to get these right or the flush and
1250                  * teardown code will get confused.
1251                  */
1252                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1253                 spmp = hmp->spmp;
1254                 spmp->pfs_hmps[0] = hmp;
1255
1256                 /*
1257                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1258                  * is inherited from the volume header.
1259                  */
1260                 xid = 0;
1261                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1262                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1263                 hmp->vchain.pmp = spmp;
1264                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1265                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1266                 hmp->fchain.pmp = spmp;
1267
1268                 /*
1269                  * First locate the super-root inode, which is key 0
1270                  * relative to the volume header's blockset.
1271                  *
1272                  * Then locate the root inode by scanning the directory keyspace
1273                  * represented by the label.
1274                  */
1275                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1276                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1277                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1278                                       &error, 0);
1279                 hammer2_chain_lookup_done(parent);
1280                 if (schain == NULL) {
1281                         kprintf("hammer2_mount: invalid super-root\n");
1282                         hammer2_unmount_helper(mp, NULL, hmp);
1283                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1284                         hammer2_vfs_unmount(mp, MNT_FORCE);
1285                         return EINVAL;
1286                 }
1287                 if (schain->error) {
1288                         kprintf("hammer2_mount: error %s reading super-root\n",
1289                                 hammer2_error_str(schain->error));
1290                         hammer2_chain_unlock(schain);
1291                         hammer2_chain_drop(schain);
1292                         schain = NULL;
1293                         hammer2_unmount_helper(mp, NULL, hmp);
1294                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1295                         hammer2_vfs_unmount(mp, MNT_FORCE);
1296                         return EINVAL;
1297                 }
1298
1299                 /*
1300                  * The super-root always uses an inode_tid of 1 when
1301                  * creating PFSs.
1302                  */
1303                 spmp->inode_tid = 1;
1304                 spmp->modify_tid = schain->bref.modify_tid + 1;
1305
1306                 /*
1307                  * Sanity-check schain's pmp and finish initialization.
1308                  * Any chain belonging to the super-root topology should
1309                  * have a NULL pmp (not even set to spmp).
1310                  */
1311                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1312                 KKASSERT(schain->pmp == NULL);
1313                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1314
1315                 /*
1316                  * Replace the dummy spmp->iroot with a real one.  It's
1317                  * easier to just do a wholesale replacement than to try
1318                  * to update the chain and fixup the iroot fields.
1319                  *
1320                  * The returned inode is locked with the supplied cluster.
1321                  */
1322                 hammer2_dummy_xop_from_chain(&xop, schain);
1323                 hammer2_inode_drop(spmp->iroot);
1324                 spmp->iroot = NULL;
1325                 spmp->iroot = hammer2_inode_get(spmp, &xop, -1, -1);
1326                 spmp->spmp_hmp = hmp;
1327                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1328                 spmp->pfs_hmps[0] = hmp;
1329                 hammer2_inode_ref(spmp->iroot);
1330                 hammer2_inode_unlock(spmp->iroot);
1331                 hammer2_cluster_unlock(&xop.cluster);
1332                 hammer2_chain_drop(schain);
1333                 /* do not call hammer2_cluster_drop() on an embedded cluster */
1334                 schain = NULL;  /* now invalid */
1335                 /* leave spmp->iroot with one ref */
1336
1337                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1338                         error = hammer2_recovery(hmp);
1339                         if (error == 0)
1340                                 error |= hammer2_fixup_pfses(hmp);
1341                         /* XXX do something with error */
1342                 }
1343                 hammer2_update_pmps(hmp);
1344                 hammer2_iocom_init(hmp);
1345                 hammer2_bulkfree_init(hmp);
1346
1347                 /*
1348                  * Ref the cluster management messaging descriptor.  The mount
1349                  * program deals with the other end of the communications pipe.
1350                  *
1351                  * Root mounts typically do not supply one.
1352                  */
1353                 if (info.cluster_fd >= 0) {
1354                         fp = holdfp(curthread, info.cluster_fd, -1);
1355                         if (fp) {
1356                                 hammer2_cluster_reconnect(hmp, fp);
1357                         } else {
1358                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1359                         }
1360                 }
1361         } else {
1362                 spmp = hmp->spmp;
1363                 if (info.hflags & HMNT2_DEVFLAGS) {
1364                         kprintf("hammer2: Warning: mount flags pertaining "
1365                                 "to the whole device may only be specified "
1366                                 "on the first mount of the device: %08x\n",
1367                                 info.hflags & HMNT2_DEVFLAGS);
1368                 }
1369         }
1370
1371         /*
1372          * Force local mount (disassociate all PFSs from their clusters).
1373          * Used primarily for debugging.
1374          */
1375         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1376
1377         /*
1378          * Lookup the mount point under the media-localized super-root.
1379          * Scanning hammer2_pfslist doesn't help us because it represents
1380          * PFS cluster ids which can aggregate several named PFSs together.
1381          *
1382          * cluster->pmp will incorrectly point to spmp and must be fixed
1383          * up later on.
1384          */
1385         hammer2_inode_lock(spmp->iroot, 0);
1386         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1387         lhc = hammer2_dirhash(label, strlen(label));
1388         chain = hammer2_chain_lookup(&parent, &key_next,
1389                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1390                                      &error, 0);
1391         while (chain) {
1392                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1393                     strcmp(label, chain->data->ipdata.filename) == 0) {
1394                         break;
1395                 }
1396                 chain = hammer2_chain_next(&parent, chain, &key_next,
1397                                             key_next,
1398                                             lhc + HAMMER2_DIRHASH_LOMASK,
1399                                             &error, 0);
1400         }
1401         if (parent) {
1402                 hammer2_chain_unlock(parent);
1403                 hammer2_chain_drop(parent);
1404         }
1405         hammer2_inode_unlock(spmp->iroot);
1406
1407         /*
1408          * PFS could not be found?
1409          */
1410         if (chain == NULL) {
1411                 if (error)
1412                         kprintf("hammer2_mount: PFS label I/O error\n");
1413                 else
1414                         kprintf("hammer2_mount: PFS label not found\n");
1415                 hammer2_unmount_helper(mp, NULL, hmp);
1416                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1417                 hammer2_vfs_unmount(mp, MNT_FORCE);
1418
1419                 return EINVAL;
1420         }
1421
1422         /*
1423          * Acquire the pmp structure (it should have already been allocated
1424          * via hammer2_update_pmps() so do not pass cluster in to add to
1425          * available chains).
1426          *
1427          * Check if the cluster has already been mounted.  A cluster can
1428          * only be mounted once, use null mounts to mount additional copies.
1429          */
1430         if (chain->error) {
1431                 kprintf("hammer2_mount: PFS label I/O error\n");
1432         } else {
1433                 ripdata = &chain->data->ipdata;
1434                 bref = chain->bref;
1435                 pmp = hammer2_pfsalloc(NULL, ripdata,
1436                                        bref.modify_tid, force_local);
1437         }
1438         hammer2_chain_unlock(chain);
1439         hammer2_chain_drop(chain);
1440
1441         /*
1442          * Finish the mount
1443          */
1444         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1445
1446         if (pmp->mp) {
1447                 kprintf("hammer2_mount: PFS already mounted!\n");
1448                 hammer2_unmount_helper(mp, NULL, hmp);
1449                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1450                 hammer2_vfs_unmount(mp, MNT_FORCE);
1451
1452                 return EBUSY;
1453         }
1454
1455         pmp->hflags = info.hflags;
1456         mp->mnt_flag |= MNT_LOCAL;
1457         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1458         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1459  
1460         /*
1461          * required mount structure initializations
1462          */
1463         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1464         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1465  
1466         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1467         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1468  
1469         /*
1470          * Optional fields
1471          */
1472         mp->mnt_iosize_max = MAXPHYS;
1473
1474         /*
1475          * Connect up mount pointers.
1476          */
1477         hammer2_mount_helper(mp, pmp);
1478
1479         lockmgr(&hammer2_mntlk, LK_RELEASE);
1480
1481         /*
1482          * Finish setup
1483          */
1484         vfs_getnewfsid(mp);
1485         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1486         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1487         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1488
1489         if (path) {
1490                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1491                           MNAMELEN - 1, &size);
1492                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1493         } /* else root mount, already in there */
1494
1495         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1496         if (path) {
1497                 copyinstr(path, mp->mnt_stat.f_mntonname,
1498                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1499                           &size);
1500         } else {
1501                 /* root mount */
1502                 mp->mnt_stat.f_mntonname[0] = '/';
1503         }
1504
1505         /*
1506          * Initial statfs to prime mnt_stat.
1507          */
1508         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1509         
1510         return 0;
1511 }
1512
1513 /*
1514  * Scan PFSs under the super-root and create hammer2_pfs structures.
1515  */
1516 static
1517 void
1518 hammer2_update_pmps(hammer2_dev_t *hmp)
1519 {
1520         const hammer2_inode_data_t *ripdata;
1521         hammer2_chain_t *parent;
1522         hammer2_chain_t *chain;
1523         hammer2_blockref_t bref;
1524         hammer2_dev_t *force_local;
1525         hammer2_pfs_t *spmp;
1526         hammer2_pfs_t *pmp;
1527         hammer2_key_t key_next;
1528         int error;
1529
1530         /*
1531          * Force local mount (disassociate all PFSs from their clusters).
1532          * Used primarily for debugging.
1533          */
1534         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1535
1536         /*
1537          * Lookup mount point under the media-localized super-root.
1538          *
1539          * cluster->pmp will incorrectly point to spmp and must be fixed
1540          * up later on.
1541          */
1542         spmp = hmp->spmp;
1543         hammer2_inode_lock(spmp->iroot, 0);
1544         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1545         chain = hammer2_chain_lookup(&parent, &key_next,
1546                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1547                                          &error, 0);
1548         while (chain) {
1549                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1550                         continue;
1551                 if (chain->error) {
1552                         kprintf("I/O error scanning PFS labels\n");
1553                 } else {
1554                         ripdata = &chain->data->ipdata;
1555                         bref = chain->bref;
1556
1557                         pmp = hammer2_pfsalloc(chain, ripdata,
1558                                                bref.modify_tid, force_local);
1559                 }
1560                 chain = hammer2_chain_next(&parent, chain, &key_next,
1561                                            key_next, HAMMER2_KEY_MAX,
1562                                            &error, 0);
1563         }
1564         if (parent) {
1565                 hammer2_chain_unlock(parent);
1566                 hammer2_chain_drop(parent);
1567         }
1568         hammer2_inode_unlock(spmp->iroot);
1569 }
1570
1571 static
1572 int
1573 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1574                 struct vnode *devvp, struct ucred *cred)
1575 {
1576         int error;
1577
1578         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1579                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1580                 VOP_OPEN(devvp, FREAD | FWRITE, FSCRED, NULL);
1581                 vn_unlock(devvp);
1582                 error = hammer2_recovery(hmp);
1583                 if (error == 0)
1584                         error |= hammer2_fixup_pfses(hmp);
1585                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1586                 if (error == 0) {
1587                         VOP_CLOSE(devvp, FREAD, NULL);
1588                         hmp->ronly = 0;
1589                 } else {
1590                         VOP_CLOSE(devvp, FREAD | FWRITE, NULL);
1591                 }
1592                 vn_unlock(devvp);
1593         } else {
1594                 error = 0;
1595         }
1596         return error;
1597 }
1598
1599 static
1600 int
1601 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1602 {
1603         hammer2_pfs_t *pmp;
1604         int flags;
1605         int error = 0;
1606
1607         pmp = MPTOPMP(mp);
1608
1609         if (pmp == NULL)
1610                 return(0);
1611
1612         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1613
1614         /*
1615          * If mount initialization proceeded far enough we must flush
1616          * its vnodes and sync the underlying mount points.  Three syncs
1617          * are required to fully flush the filesystem (freemap updates lag
1618          * by one flush, and one extra for safety).
1619          */
1620         if (mntflags & MNT_FORCE)
1621                 flags = FORCECLOSE;
1622         else
1623                 flags = 0;
1624         if (pmp->iroot) {
1625                 error = vflush(mp, 0, flags);
1626                 if (error)
1627                         goto failed;
1628                 hammer2_vfs_sync(mp, MNT_WAIT);
1629                 hammer2_vfs_sync(mp, MNT_WAIT);
1630                 hammer2_vfs_sync(mp, MNT_WAIT);
1631         }
1632
1633         /*
1634          * Cleanup the frontend support XOPS threads
1635          */
1636         hammer2_xop_helper_cleanup(pmp);
1637
1638         if (pmp->mp)
1639                 hammer2_unmount_helper(mp, pmp, NULL);
1640
1641         error = 0;
1642 failed:
1643         lockmgr(&hammer2_mntlk, LK_RELEASE);
1644
1645         return (error);
1646 }
1647
1648 /*
1649  * Mount helper, hook the system mount into our PFS.
1650  * The mount lock is held.
1651  *
1652  * We must bump the mount_count on related devices for any
1653  * mounted PFSs.
1654  */
1655 static
1656 void
1657 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1658 {
1659         hammer2_cluster_t *cluster;
1660         hammer2_chain_t *rchain;
1661         int i;
1662
1663         mp->mnt_data = (qaddr_t)pmp;
1664         pmp->mp = mp;
1665
1666         /*
1667          * After pmp->mp is set we have to adjust hmp->mount_count.
1668          */
1669         cluster = &pmp->iroot->cluster;
1670         for (i = 0; i < cluster->nchains; ++i) {
1671                 rchain = cluster->array[i].chain;
1672                 if (rchain == NULL)
1673                         continue;
1674                 ++rchain->hmp->mount_count;
1675         }
1676
1677         /*
1678          * Create missing Xop threads
1679          */
1680         hammer2_xop_helper_create(pmp);
1681 }
1682
1683 /*
1684  * Mount helper, unhook the system mount from our PFS.
1685  * The mount lock is held.
1686  *
1687  * If hmp is supplied a mount responsible for being the first to open
1688  * the block device failed and the block device and all PFSs using the
1689  * block device must be cleaned up.
1690  *
1691  * If pmp is supplied multiple devices might be backing the PFS and each
1692  * must be disconnected.  This might not be the last PFS using some of the
1693  * underlying devices.  Also, we have to adjust our hmp->mount_count
1694  * accounting for the devices backing the pmp which is now undergoing an
1695  * unmount.
1696  */
1697 static
1698 void
1699 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1700 {
1701         hammer2_cluster_t *cluster;
1702         hammer2_chain_t *rchain;
1703         struct vnode *devvp;
1704         int dumpcnt;
1705         int ronly;
1706         int i;
1707
1708         /*
1709          * If no device supplied this is a high-level unmount and we have to
1710          * to disconnect the mount, adjust mount_count, and locate devices
1711          * that might now have no mounts.
1712          */
1713         if (pmp) {
1714                 KKASSERT(hmp == NULL);
1715                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1716                 pmp->mp = NULL;
1717                 mp->mnt_data = NULL;
1718
1719                 /*
1720                  * After pmp->mp is cleared we have to account for
1721                  * mount_count.
1722                  */
1723                 cluster = &pmp->iroot->cluster;
1724                 for (i = 0; i < cluster->nchains; ++i) {
1725                         rchain = cluster->array[i].chain;
1726                         if (rchain == NULL)
1727                                 continue;
1728                         --rchain->hmp->mount_count;
1729                         /* scrapping hmp now may invalidate the pmp */
1730                 }
1731 again:
1732                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1733                         if (hmp->mount_count == 0) {
1734                                 hammer2_unmount_helper(NULL, NULL, hmp);
1735                                 goto again;
1736                         }
1737                 }
1738                 return;
1739         }
1740
1741         /*
1742          * Try to terminate the block device.  We can't terminate it if
1743          * there are still PFSs referencing it.
1744          */
1745         if (hmp->mount_count)
1746                 return;
1747
1748         /*
1749          * Decomission the network before we start messing with the
1750          * device and PFS.
1751          */
1752         hammer2_iocom_uninit(hmp);
1753
1754         hammer2_bulkfree_uninit(hmp);
1755         hammer2_pfsfree_scan(hmp, 0);
1756 #if 0
1757         hammer2_dev_exlock(hmp);        /* XXX order */
1758 #endif
1759
1760         /*
1761          * Cycle the volume data lock as a safety (probably not needed any
1762          * more).  To ensure everything is out we need to flush at least
1763          * three times.  (1) The running of the sideq can dirty the
1764          * filesystem, (2) A normal flush can dirty the freemap, and
1765          * (3) ensure that the freemap is fully synchronized.
1766          *
1767          * The next mount's recovery scan can clean everything up but we want
1768          * to leave the filesystem in a 100% clean state on a normal unmount.
1769          */
1770 #if 0
1771         hammer2_voldata_lock(hmp);
1772         hammer2_voldata_unlock(hmp);
1773 #endif
1774
1775         /*
1776          * Flush whatever is left.  Unmounted but modified PFS's might still
1777          * have some dirty chains on them.
1778          */
1779         hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
1780         hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
1781
1782         if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1783                 hammer2_voldata_modify(hmp);
1784                 hammer2_flush(&hmp->fchain, HAMMER2_FLUSH_TOP |
1785                                             HAMMER2_FLUSH_ALL);
1786         }
1787         hammer2_chain_unlock(&hmp->fchain);
1788
1789         if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
1790                 hammer2_flush(&hmp->vchain, HAMMER2_FLUSH_TOP |
1791                                             HAMMER2_FLUSH_ALL);
1792         }
1793         hammer2_chain_unlock(&hmp->vchain);
1794
1795         if ((hmp->vchain.flags | hmp->fchain.flags) &
1796             HAMMER2_CHAIN_FLUSH_MASK) {
1797                 kprintf("hammer2_unmount: chains left over "
1798                         "after final sync\n");
1799                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1800                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1801
1802                 if (hammer2_debug & 0x0010)
1803                         Debugger("entered debugger");
1804         }
1805
1806         hammer2_pfsfree_scan(hmp, 1);
1807
1808         KKASSERT(hmp->spmp == NULL);
1809
1810         /*
1811          * Finish up with the device vnode
1812          */
1813         if ((devvp = hmp->devvp) != NULL) {
1814                 ronly = hmp->ronly;
1815                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1816                 kprintf("hammer2_unmount(A): devvp %s rbdirty %p ronly=%d\n",
1817                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree),
1818                         ronly);
1819                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1820                 kprintf("hammer2_unmount(B): devvp %s rbdirty %p\n",
1821                         hmp->devrepname, RB_ROOT(&devvp->v_rbdirty_tree));
1822                 hmp->devvp = NULL;
1823                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1824                 vn_unlock(devvp);
1825                 vrele(devvp);
1826                 devvp = NULL;
1827         }
1828
1829         /*
1830          * Clear vchain/fchain flags that might prevent final cleanup
1831          * of these chains.
1832          */
1833         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1834                 atomic_add_long(&hammer2_count_modified_chains, -1);
1835                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1836                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1837         }
1838         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1839                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1840         }
1841
1842         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1843                 atomic_add_long(&hammer2_count_modified_chains, -1);
1844                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1845                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1846         }
1847         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1848                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1849         }
1850
1851         /*
1852          * Final drop of embedded freemap root chain to
1853          * clean up fchain.core (fchain structure is not
1854          * flagged ALLOCATED so it is cleaned out and then
1855          * left to rot).
1856          */
1857         hammer2_chain_drop(&hmp->fchain);
1858
1859         /*
1860          * Final drop of embedded volume root chain to clean
1861          * up vchain.core (vchain structure is not flagged
1862          * ALLOCATED so it is cleaned out and then left to
1863          * rot).
1864          */
1865         dumpcnt = 50;
1866         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v', (u_int)-1);
1867         dumpcnt = 50;
1868         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f', (u_int)-1);
1869 #if 0
1870         hammer2_dev_unlock(hmp);
1871 #endif
1872         hammer2_chain_drop(&hmp->vchain);
1873
1874         hammer2_io_cleanup(hmp, &hmp->iotree);
1875         if (hmp->iofree_count) {
1876                 kprintf("io_cleanup: %d I/O's left hanging\n",
1877                         hmp->iofree_count);
1878         }
1879
1880         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1881         kmalloc_destroy(&hmp->mchain);
1882         kfree(hmp, M_HAMMER2);
1883 }
1884
1885 int
1886 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1887                  ino_t ino, struct vnode **vpp)
1888 {
1889         hammer2_xop_lookup_t *xop;
1890         hammer2_pfs_t *pmp;
1891         hammer2_inode_t *ip;
1892         hammer2_tid_t inum;
1893         int error;
1894
1895         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1896
1897         error = 0;
1898         pmp = MPTOPMP(mp);
1899
1900         /*
1901          * Easy if we already have it cached
1902          */
1903         ip = hammer2_inode_lookup(pmp, inum);
1904         if (ip) {
1905                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1906                 *vpp = hammer2_igetv(ip, &error);
1907                 hammer2_inode_unlock(ip);
1908                 hammer2_inode_drop(ip);         /* from lookup */
1909
1910                 return error;
1911         }
1912
1913         /*
1914          * Otherwise we have to find the inode
1915          */
1916         xop = hammer2_xop_alloc(pmp->iroot, 0);
1917         xop->lhc = inum;
1918         hammer2_xop_start(&xop->head, &hammer2_lookup_desc);
1919         error = hammer2_xop_collect(&xop->head, 0);
1920
1921         if (error == 0)
1922                 ip = hammer2_inode_get(pmp, &xop->head, -1, -1);
1923         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1924
1925         if (ip) {
1926                 *vpp = hammer2_igetv(ip, &error);
1927                 hammer2_inode_unlock(ip);
1928         } else {
1929                 *vpp = NULL;
1930                 error = ENOENT;
1931         }
1932         return (error);
1933 }
1934
1935 static
1936 int
1937 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1938 {
1939         hammer2_pfs_t *pmp;
1940         struct vnode *vp;
1941         int error;
1942
1943         pmp = MPTOPMP(mp);
1944         if (pmp->iroot == NULL) {
1945                 kprintf("hammer2 (%s): no root inode\n",
1946                         mp->mnt_stat.f_mntfromname);
1947                 *vpp = NULL;
1948                 return EINVAL;
1949         }
1950
1951         error = 0;
1952         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1953
1954         while (pmp->inode_tid == 0) {
1955                 hammer2_xop_ipcluster_t *xop;
1956                 const hammer2_inode_meta_t *meta;
1957
1958                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1959                 hammer2_xop_start(&xop->head, &hammer2_ipcluster_desc);
1960                 error = hammer2_xop_collect(&xop->head, 0);
1961
1962                 if (error == 0) {
1963                         meta = &hammer2_xop_gdata(&xop->head)->ipdata.meta;
1964                         pmp->iroot->meta = *meta;
1965                         pmp->inode_tid = meta->pfs_inum + 1;
1966                         hammer2_xop_pdata(&xop->head);
1967                         /* meta invalid */
1968
1969                         if (pmp->inode_tid < HAMMER2_INODE_START)
1970                                 pmp->inode_tid = HAMMER2_INODE_START;
1971                         pmp->modify_tid =
1972                                 xop->head.cluster.focus->bref.modify_tid + 1;
1973 #if 0
1974                         kprintf("PFS: Starting inode %jd\n",
1975                                 (intmax_t)pmp->inode_tid);
1976                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1977                                 pmp->inode_tid, pmp->modify_tid);
1978 #endif
1979                         wakeup(&pmp->iroot);
1980
1981                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1982
1983                         /*
1984                          * Prime the mount info.
1985                          */
1986                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1987                         break;
1988                 }
1989
1990                 /*
1991                  * Loop, try again
1992                  */
1993                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1994                 hammer2_inode_unlock(pmp->iroot);
1995                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1996                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1997                 if (error == EINTR)
1998                         break;
1999         }
2000
2001         if (error) {
2002                 hammer2_inode_unlock(pmp->iroot);
2003                 *vpp = NULL;
2004         } else {
2005                 vp = hammer2_igetv(pmp->iroot, &error);
2006                 hammer2_inode_unlock(pmp->iroot);
2007                 *vpp = vp;
2008         }
2009
2010         return (error);
2011 }
2012
2013 /*
2014  * Filesystem status
2015  *
2016  * XXX incorporate ipdata->meta.inode_quota and data_quota
2017  */
2018 static
2019 int
2020 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
2021 {
2022         hammer2_pfs_t *pmp;
2023         hammer2_dev_t *hmp;
2024         hammer2_blockref_t bref;
2025         struct statfs tmp;
2026         int i;
2027
2028         /*
2029          * NOTE: iroot might not have validated the cluster yet.
2030          */
2031         pmp = MPTOPMP(mp);
2032
2033         bzero(&tmp, sizeof(tmp));
2034
2035         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2036                 hmp = pmp->pfs_hmps[i];
2037                 if (hmp == NULL)
2038                         continue;
2039                 if (pmp->iroot->cluster.array[i].chain)
2040                         bref = pmp->iroot->cluster.array[i].chain->bref;
2041                 else
2042                         bzero(&bref, sizeof(bref));
2043
2044                 tmp.f_files = bref.embed.stats.inode_count;
2045                 tmp.f_ffree = 0;
2046                 tmp.f_blocks = hmp->voldata.allocator_size /
2047                                mp->mnt_vstat.f_bsize;
2048                 tmp.f_bfree = hmp->voldata.allocator_free /
2049                               mp->mnt_vstat.f_bsize;
2050                 tmp.f_bavail = tmp.f_bfree;
2051
2052                 if (cred && cred->cr_uid != 0) {
2053                         uint64_t adj;
2054
2055                         /* 5% */
2056                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2057                         tmp.f_blocks -= adj;
2058                         tmp.f_bfree -= adj;
2059                         tmp.f_bavail -= adj;
2060                 }
2061
2062                 mp->mnt_stat.f_blocks = tmp.f_blocks;
2063                 mp->mnt_stat.f_bfree = tmp.f_bfree;
2064                 mp->mnt_stat.f_bavail = tmp.f_bavail;
2065                 mp->mnt_stat.f_files = tmp.f_files;
2066                 mp->mnt_stat.f_ffree = tmp.f_ffree;
2067
2068                 *sbp = mp->mnt_stat;
2069         }
2070         return (0);
2071 }
2072
2073 static
2074 int
2075 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
2076 {
2077         hammer2_pfs_t *pmp;
2078         hammer2_dev_t *hmp;
2079         hammer2_blockref_t bref;
2080         struct statvfs tmp;
2081         int i;
2082
2083         /*
2084          * NOTE: iroot might not have validated the cluster yet.
2085          */
2086         pmp = MPTOPMP(mp);
2087         bzero(&tmp, sizeof(tmp));
2088
2089         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
2090                 hmp = pmp->pfs_hmps[i];
2091                 if (hmp == NULL)
2092                         continue;
2093                 if (pmp->iroot->cluster.array[i].chain)
2094                         bref = pmp->iroot->cluster.array[i].chain->bref;
2095                 else
2096                         bzero(&bref, sizeof(bref));
2097
2098                 tmp.f_files = bref.embed.stats.inode_count;
2099                 tmp.f_ffree = 0;
2100                 tmp.f_blocks = hmp->voldata.allocator_size /
2101                                mp->mnt_vstat.f_bsize;
2102                 tmp.f_bfree = hmp->voldata.allocator_free /
2103                               mp->mnt_vstat.f_bsize;
2104                 tmp.f_bavail = tmp.f_bfree;
2105
2106                 if (cred && cred->cr_uid != 0) {
2107                         uint64_t adj;
2108
2109                         /* 5% */
2110                         adj = hmp->free_reserved / mp->mnt_vstat.f_bsize;
2111                         tmp.f_blocks -= adj;
2112                         tmp.f_bfree -= adj;
2113                         tmp.f_bavail -= adj;
2114                 }
2115
2116                 mp->mnt_vstat.f_blocks = tmp.f_blocks;
2117                 mp->mnt_vstat.f_bfree = tmp.f_bfree;
2118                 mp->mnt_vstat.f_bavail = tmp.f_bavail;
2119                 mp->mnt_vstat.f_files = tmp.f_files;
2120                 mp->mnt_vstat.f_ffree = tmp.f_ffree;
2121
2122                 *sbp = mp->mnt_vstat;
2123         }
2124         return (0);
2125 }
2126
2127 /*
2128  * Mount-time recovery (RW mounts)
2129  *
2130  * Updates to the free block table are allowed to lag flushes by one
2131  * transaction.  In case of a crash, then on a fresh mount we must do an
2132  * incremental scan of the last committed transaction id and make sure that
2133  * all related blocks have been marked allocated.
2134  *
2135  * The super-root topology and each PFS has its own transaction id domain,
2136  * so we must track PFS boundary transitions.
2137  */
2138 struct hammer2_recovery_elm {
2139         TAILQ_ENTRY(hammer2_recovery_elm) entry;
2140         hammer2_chain_t *chain;
2141         hammer2_tid_t sync_tid;
2142 };
2143
2144 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
2145
2146 struct hammer2_recovery_info {
2147         struct hammer2_recovery_list list;
2148         hammer2_tid_t   mtid;
2149         int     depth;
2150 };
2151
2152 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
2153                         hammer2_chain_t *parent,
2154                         struct hammer2_recovery_info *info,
2155                         hammer2_tid_t sync_tid);
2156
2157 #define HAMMER2_RECOVERY_MAXDEPTH       10
2158
2159 static
2160 int
2161 hammer2_recovery(hammer2_dev_t *hmp)
2162 {
2163         struct hammer2_recovery_info info;
2164         struct hammer2_recovery_elm *elm;
2165         hammer2_chain_t *parent;
2166         hammer2_tid_t sync_tid;
2167         hammer2_tid_t mirror_tid;
2168         int error;
2169
2170         hammer2_trans_init(hmp->spmp, 0);
2171
2172         sync_tid = hmp->voldata.freemap_tid;
2173         mirror_tid = hmp->voldata.mirror_tid;
2174
2175         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
2176         if (sync_tid >= mirror_tid) {
2177                 kprintf(" no recovery needed\n");
2178         } else {
2179                 kprintf(" freemap recovery %016jx-%016jx\n",
2180                         sync_tid + 1, mirror_tid);
2181         }
2182
2183         TAILQ_INIT(&info.list);
2184         info.depth = 0;
2185         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
2186         error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
2187         hammer2_chain_lookup_done(parent);
2188
2189         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
2190                 TAILQ_REMOVE(&info.list, elm, entry);
2191                 parent = elm->chain;
2192                 sync_tid = elm->sync_tid;
2193                 kfree(elm, M_HAMMER2);
2194
2195                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2196                 error |= hammer2_recovery_scan(hmp, parent, &info,
2197                                               hmp->voldata.freemap_tid);
2198                 hammer2_chain_unlock(parent);
2199                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
2200         }
2201
2202         hammer2_trans_done(hmp->spmp, 0);
2203
2204         return error;
2205 }
2206
2207 static
2208 int
2209 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
2210                       struct hammer2_recovery_info *info,
2211                       hammer2_tid_t sync_tid)
2212 {
2213         const hammer2_inode_data_t *ripdata;
2214         hammer2_chain_t *chain;
2215         hammer2_blockref_t bref;
2216         int tmp_error;
2217         int rup_error;
2218         int error;
2219         int first;
2220
2221         /*
2222          * Adjust freemap to ensure that the block(s) are marked allocated.
2223          */
2224         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2225                 hammer2_freemap_adjust(hmp, &parent->bref,
2226                                        HAMMER2_FREEMAP_DORECOVER);
2227         }
2228
2229         /*
2230          * Check type for recursive scan
2231          */
2232         switch(parent->bref.type) {
2233         case HAMMER2_BREF_TYPE_VOLUME:
2234                 /* data already instantiated */
2235                 break;
2236         case HAMMER2_BREF_TYPE_INODE:
2237                 /*
2238                  * Must instantiate data for DIRECTDATA test and also
2239                  * for recursion.
2240                  */
2241                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2242                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2243                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2244                         /* not applicable to recovery scan */
2245                         hammer2_chain_unlock(parent);
2246                         return 0;
2247                 }
2248                 hammer2_chain_unlock(parent);
2249                 break;
2250         case HAMMER2_BREF_TYPE_INDIRECT:
2251                 /*
2252                  * Must instantiate data for recursion
2253                  */
2254                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2255                 hammer2_chain_unlock(parent);
2256                 break;
2257         case HAMMER2_BREF_TYPE_DIRENT:
2258         case HAMMER2_BREF_TYPE_DATA:
2259         case HAMMER2_BREF_TYPE_FREEMAP:
2260         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2261         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2262                 /* not applicable to recovery scan */
2263                 return 0;
2264                 break;
2265         default:
2266                 return HAMMER2_ERROR_BADBREF;
2267         }
2268
2269         /*
2270          * Defer operation if depth limit reached or if we are crossing a
2271          * PFS boundary.
2272          */
2273         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2274                 struct hammer2_recovery_elm *elm;
2275
2276                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2277                 elm->chain = parent;
2278                 elm->sync_tid = sync_tid;
2279                 hammer2_chain_ref(parent);
2280                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2281                 /* unlocked by caller */
2282
2283                 return(0);
2284         }
2285
2286
2287         /*
2288          * Recursive scan of the last flushed transaction only.  We are
2289          * doing this without pmp assignments so don't leave the chains
2290          * hanging around after we are done with them.
2291          *
2292          * error        Cumulative error this level only
2293          * rup_error    Cumulative error for recursion
2294          * tmp_error    Specific non-cumulative recursion error
2295          */
2296         chain = NULL;
2297         first = 1;
2298         rup_error = 0;
2299         error = 0;
2300
2301         for (;;) {
2302                 error |= hammer2_chain_scan(parent, &chain, &bref,
2303                                             &first,
2304                                             HAMMER2_LOOKUP_NODATA);
2305
2306                 /*
2307                  * Problem during scan or EOF
2308                  */
2309                 if (error)
2310                         break;
2311
2312                 /*
2313                  * If this is a leaf
2314                  */
2315                 if (chain == NULL) {
2316                         if (bref.mirror_tid > sync_tid) {
2317                                 hammer2_freemap_adjust(hmp, &bref,
2318                                                      HAMMER2_FREEMAP_DORECOVER);
2319                         }
2320                         continue;
2321                 }
2322
2323                 /*
2324                  * This may or may not be a recursive node.
2325                  */
2326                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2327                 if (bref.mirror_tid > sync_tid) {
2328                         ++info->depth;
2329                         tmp_error = hammer2_recovery_scan(hmp, chain,
2330                                                            info, sync_tid);
2331                         --info->depth;
2332                 } else {
2333                         tmp_error = 0;
2334                 }
2335
2336                 /*
2337                  * Flush the recovery at the PFS boundary to stage it for
2338                  * the final flush of the super-root topology.
2339                  */
2340                 if (tmp_error == 0 &&
2341                     (bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2342                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2343                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2344                                              HAMMER2_FLUSH_ALL);
2345                 }
2346                 rup_error |= tmp_error;
2347         }
2348         return ((error | rup_error) & ~HAMMER2_ERROR_EOF);
2349 }
2350
2351 /*
2352  * This fixes up an error introduced in earlier H2 implementations where
2353  * moving a PFS inode into an indirect block wound up causing the
2354  * HAMMER2_BREF_FLAG_PFSROOT flag in the bref to get cleared.
2355  */
2356 static
2357 int
2358 hammer2_fixup_pfses(hammer2_dev_t *hmp)
2359 {
2360         const hammer2_inode_data_t *ripdata;
2361         hammer2_chain_t *parent;
2362         hammer2_chain_t *chain;
2363         hammer2_key_t key_next;
2364         hammer2_pfs_t *spmp;
2365         int error;
2366
2367         error = 0;
2368
2369         /*
2370          * Lookup mount point under the media-localized super-root.
2371          *
2372          * cluster->pmp will incorrectly point to spmp and must be fixed
2373          * up later on.
2374          */
2375         spmp = hmp->spmp;
2376         hammer2_inode_lock(spmp->iroot, 0);
2377         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
2378         chain = hammer2_chain_lookup(&parent, &key_next,
2379                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
2380                                          &error, 0);
2381         while (chain) {
2382                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
2383                         continue;
2384                 if (chain->error) {
2385                         kprintf("I/O error scanning PFS labels\n");
2386                         error |= chain->error;
2387                 } else if ((chain->bref.flags &
2388                             HAMMER2_BREF_FLAG_PFSROOT) == 0) {
2389                         int error2;
2390
2391                         ripdata = &chain->data->ipdata;
2392                         hammer2_trans_init(hmp->spmp, 0);
2393                         error2 = hammer2_chain_modify(chain,
2394                                                       chain->bref.modify_tid,
2395                                                       0, 0);
2396                         if (error2 == 0) {
2397                                 kprintf("hammer2: Correct mis-flagged PFS %s\n",
2398                                         ripdata->filename);
2399                                 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT;
2400                         } else {
2401                                 error |= error2;
2402                         }
2403                         hammer2_flush(chain, HAMMER2_FLUSH_TOP |
2404                                              HAMMER2_FLUSH_ALL);
2405                         hammer2_trans_done(hmp->spmp, 0);
2406                 }
2407                 chain = hammer2_chain_next(&parent, chain, &key_next,
2408                                            key_next, HAMMER2_KEY_MAX,
2409                                            &error, 0);
2410         }
2411         if (parent) {
2412                 hammer2_chain_unlock(parent);
2413                 hammer2_chain_drop(parent);
2414         }
2415         hammer2_inode_unlock(spmp->iroot);
2416
2417         return error;
2418 }
2419
2420 /*
2421  * Sync a mount point; this is called periodically on a per-mount basis from
2422  * the filesystem syncer, and whenever a user issues a sync.
2423  */
2424 int
2425 hammer2_vfs_sync(struct mount *mp, int waitfor)
2426 {
2427         int error;
2428
2429         error = hammer2_vfs_sync_pmp(MPTOPMP(mp), waitfor);
2430
2431         return error;
2432 }
2433
2434 /*
2435  * Because frontend operations lock vnodes before we get a chance to
2436  * lock the related inode, we can't just acquire a vnode lock without
2437  * risking a deadlock.  The frontend may be holding a vnode lock while
2438  * also blocked on our SYNCQ flag while trying to get the inode lock.
2439  *
2440  * To deal with this situation we can check the vnode lock situation
2441  * after locking the inode and perform a work-around.
2442  */
2443 int
2444 hammer2_vfs_sync_pmp(hammer2_pfs_t *pmp, int waitfor)
2445 {
2446         struct mount *mp;
2447         /*hammer2_xop_flush_t *xop;*/
2448         /*struct hammer2_sync_info info;*/
2449         hammer2_inode_t *ip;
2450         hammer2_depend_t *depend;
2451         hammer2_depend_t *depend_next;
2452         struct vnode *vp;
2453         uint32_t pass2;
2454         int error;
2455         int dorestart;
2456
2457         mp = pmp->mp;
2458
2459         /*
2460          * Move all inodes on sideq to syncq.  This will clear sideq.
2461          * This should represent all flushable inodes.  These inodes
2462          * will already have refs due to being on syncq or sideq.  We
2463          * must do this all at once with the spinlock held to ensure that
2464          * all inode dependencies are part of the same flush.
2465          *
2466          * We should be able to do this asynchronously from frontend
2467          * operations because we will be locking the inodes later on
2468          * to actually flush them, and that will partition any frontend
2469          * op using the same inode.  Either it has already locked the
2470          * inode and we will block, or it has not yet locked the inode
2471          * and it will block until we are finished flushing that inode.
2472          *
2473          * When restarting, only move the inodes flagged as PASS2 from
2474          * SIDEQ to SYNCQ.  PASS2 propagation by inode_lock4() and
2475          * inode_depend() are atomic with the spin-lock.
2476          */
2477         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2478 #ifdef HAMMER2_DEBUG_SYNC
2479         kprintf("FILESYSTEM SYNC BOUNDARY\n");
2480 #endif
2481         dorestart = 0;
2482
2483         /*
2484          * Move inodes from depq to syncq, releasing the related
2485          * depend structures.
2486          */
2487 restart:
2488 #ifdef HAMMER2_DEBUG_SYNC
2489         kprintf("FILESYSTEM SYNC RESTART (%d)\n", dorestart);
2490 #endif
2491         hammer2_trans_setflags(pmp, 0/*HAMMER2_TRANS_COPYQ*/);
2492         hammer2_trans_clearflags(pmp, HAMMER2_TRANS_RESCAN);
2493
2494         /*
2495          * Move inodes from depq to syncq.  When restarting, only depq's
2496          * marked pass2 are moved.
2497          */
2498         hammer2_spin_ex(&pmp->list_spin);
2499         depend_next = TAILQ_FIRST(&pmp->depq);
2500
2501         while ((depend = depend_next) != NULL) {
2502                 depend_next = TAILQ_NEXT(depend, entry);
2503                 if (dorestart && depend->pass2 == 0)
2504                         continue;
2505                 TAILQ_FOREACH(ip, &depend->sideq, entry) {
2506                         KKASSERT(ip->flags & HAMMER2_INODE_SIDEQ);
2507                         atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ);
2508                         atomic_clear_int(&ip->flags, HAMMER2_INODE_SIDEQ);
2509                         ip->depend = NULL;
2510                 }
2511                 TAILQ_CONCAT(&pmp->syncq, &depend->sideq, entry);
2512                 pmp->sideq_count -= depend->count;
2513                 depend->count = 0;
2514                 depend->pass2 = 0;
2515                 TAILQ_REMOVE(&pmp->depq, depend, entry);
2516         }
2517
2518         hammer2_spin_unex(&pmp->list_spin);
2519         hammer2_trans_clearflags(pmp, /*HAMMER2_TRANS_COPYQ |*/
2520                                       HAMMER2_TRANS_WAITING);
2521         dorestart = 0;
2522
2523         /*
2524          * sideq_count may have dropped enough to allow us to unstall
2525          * the frontend.
2526          */
2527         hammer2_pfs_memory_inc(pmp);
2528         hammer2_pfs_memory_wakeup(pmp);
2529
2530         /*
2531          * Now run through all inodes on syncq.
2532          *
2533          * Flush transactions only interlock with other flush transactions.
2534          * Any conflicting frontend operations will block on the inode, but
2535          * may hold a vnode lock while doing so.
2536          */
2537         hammer2_spin_ex(&pmp->list_spin);
2538         while ((ip = TAILQ_FIRST(&pmp->syncq)) != NULL) {
2539                 /*
2540                  * Remove the inode from the SYNCQ, transfer the syncq ref
2541                  * to us.  We must clear SYNCQ to allow any potential
2542                  * front-end deadlock to proceed.  We must set PASS2 so
2543                  * the dependency code knows what to do.
2544                  */
2545                 pass2 = ip->flags;
2546                 cpu_ccfence();
2547                 if (atomic_cmpset_int(&ip->flags,
2548                               pass2,
2549                               (pass2 & ~(HAMMER2_INODE_SYNCQ |
2550                                          HAMMER2_INODE_SYNCQ_WAKEUP)) |
2551                               HAMMER2_INODE_SYNCQ_PASS2) == 0) {
2552                         continue;
2553                 }
2554                 TAILQ_REMOVE(&pmp->syncq, ip, entry);
2555                 hammer2_spin_unex(&pmp->list_spin);
2556                 if (pass2 & HAMMER2_INODE_SYNCQ_WAKEUP)
2557                         wakeup(&ip->flags);
2558
2559                 /*
2560                  * Relock the inode, and we inherit a ref from the above.
2561                  * We will check for a race after we acquire the vnode.
2562                  */
2563                 hammer2_mtx_ex(&ip->lock);
2564
2565                 /*
2566                  * We need the vp in order to vfsync() dirty buffers, so if
2567                  * one isn't attached we can skip it.
2568                  *
2569                  * Ordering the inode lock and then the vnode lock has the
2570                  * potential to deadlock.  If we had left SYNCQ set that could
2571                  * also deadlock us against the frontend even if we don't hold
2572                  * any locks, but the latter is not a problem now since we
2573                  * cleared it.  igetv will temporarily release the inode lock
2574                  * in a safe manner to work-around the deadlock.
2575                  *
2576                  * Unfortunately it is still possible to deadlock when the
2577                  * frontend obtains multiple inode locks, because all the
2578                  * related vnodes are already locked (nor can the vnode locks
2579                  * be released and reacquired without messing up RECLAIM and
2580                  * INACTIVE sequencing).
2581                  *
2582                  * The solution for now is to move the vp back onto SIDEQ
2583                  * and set dorestart, which will restart the flush after we
2584                  * exhaust the current SYNCQ.  Note that additional
2585                  * dependencies may build up, so we definitely need to move
2586                  * the whole SIDEQ back to SYNCQ when we restart.
2587                  */
2588                 vp = ip->vp;
2589                 if (vp) {
2590                         if (vget(vp, LK_EXCLUSIVE|LK_NOWAIT)) {
2591                                 /*
2592                                  * Failed to get the vnode, requeue the inode
2593                                  * (PASS2 is already set so it will be found
2594                                  * again on the restart).
2595                                  *
2596                                  * Then unlock, possibly sleep, and retry
2597                                  * later.  We sleep if PASS2 was *previously*
2598                                  * set, before we set it again above.
2599                                  */
2600                                 vp = NULL;
2601                                 dorestart = 1;
2602 #ifdef HAMMER2_DEBUG_SYNC
2603                                 kprintf("inum %ld (sync delayed by vnode)\n",
2604                                         (long)ip->meta.inum);
2605 #endif
2606                                 hammer2_inode_delayed_sideq(ip);
2607
2608                                 hammer2_mtx_unlock(&ip->lock);
2609                                 hammer2_inode_drop(ip);
2610
2611                                 if (pass2 & HAMMER2_INODE_SYNCQ_PASS2) {
2612                                         tsleep(&dorestart, 0, "h2syndel", 2);
2613                                 }
2614                                 hammer2_spin_ex(&pmp->list_spin);
2615                                 continue;
2616                         }
2617                 } else {
2618                         vp = NULL;
2619                 }
2620
2621                 /*
2622                  * If the inode wound up on a SIDEQ again it will already be
2623                  * prepped for another PASS2.  In this situation if we flush
2624                  * it now we will just wind up flushing it again in the same
2625                  * syncer run, so we might as well not flush it now.
2626                  */
2627                 if (ip->flags & HAMMER2_INODE_SIDEQ) {
2628                         hammer2_mtx_unlock(&ip->lock);
2629                         hammer2_inode_drop(ip);
2630                         if (vp)
2631                                 vput(vp);
2632                         dorestart = 1;
2633                         hammer2_spin_ex(&pmp->list_spin);
2634                         continue;
2635                 }
2636
2637                 /*
2638                  * Ok we have the inode exclusively locked and if vp is
2639                  * not NULL that will also be exclusively locked.  Do the
2640                  * meat of the flush.
2641                  *
2642                  * vp token needed for v_rbdirty_tree check / vclrisdirty
2643                  * sequencing.  Though we hold the vnode exclusively so
2644                  * we shouldn't need to hold the token also in this case.
2645                  */
2646                 if (vp) {
2647                         vfsync(vp, MNT_WAIT, 1, NULL, NULL);
2648                         bio_track_wait(&vp->v_track_write, 0, 0); /* XXX */
2649                 }
2650
2651                 /*
2652                  * If the inode has not yet been inserted into the tree
2653                  * we must do so.  Then sync and flush it.  The flush should
2654                  * update the parent.
2655                  */
2656                 if (ip->flags & HAMMER2_INODE_DELETING) {
2657 #ifdef HAMMER2_DEBUG_SYNC
2658                         kprintf("inum %ld destroy\n", (long)ip->meta.inum);
2659 #endif
2660                         hammer2_inode_chain_des(ip);
2661                         atomic_add_long(&hammer2_iod_inode_deletes, 1);
2662                 } else if (ip->flags & HAMMER2_INODE_CREATING) {
2663 #ifdef HAMMER2_DEBUG_SYNC
2664                         kprintf("inum %ld insert\n", (long)ip->meta.inum);
2665 #endif
2666                         hammer2_inode_chain_ins(ip);
2667                         atomic_add_long(&hammer2_iod_inode_creates, 1);
2668                 }
2669 #ifdef HAMMER2_DEBUG_SYNC
2670                 kprintf("inum %ld chain-sync\n", (long)ip->meta.inum);
2671 #endif
2672
2673                 /*
2674                  * Because I kinda messed up the design and index the inodes
2675                  * under the root inode, along side the directory entries,
2676                  * we can't flush the inode index under the iroot until the
2677                  * end.  If we do it now we might miss effects created by
2678                  * other inodes on the SYNCQ.
2679                  *
2680                  * Do a normal (non-FSSYNC) flush instead, which allows the
2681                  * vnode code to work the same.  We don't want to force iroot
2682                  * back onto the SIDEQ, and we also don't want the flush code
2683                  * to update pfs_iroot_blocksets until the final flush later.
2684                  *
2685                  * XXX at the moment this will likely result in a double-flush
2686                  * of the iroot chain.
2687                  */
2688                 hammer2_inode_chain_sync(ip);
2689                 if (ip == pmp->iroot) {
2690                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP);
2691                 } else {
2692                         hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2693                                                       HAMMER2_XOP_FSSYNC);
2694                 }
2695                 if (vp) {
2696                         lwkt_gettoken(&vp->v_token);
2697                         if ((ip->flags & (HAMMER2_INODE_MODIFIED |
2698                                           HAMMER2_INODE_RESIZED |
2699                                           HAMMER2_INODE_DIRTYDATA)) == 0 &&
2700                             RB_EMPTY(&vp->v_rbdirty_tree) &&
2701                             !bio_track_active(&vp->v_track_write)) {
2702                                 vclrisdirty(vp);
2703                         } else {
2704                                 hammer2_inode_delayed_sideq(ip);
2705                         }
2706                         lwkt_reltoken(&vp->v_token);
2707                         vput(vp);
2708                         vp = NULL;      /* safety */
2709                 }
2710                 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_PASS2);
2711                 hammer2_inode_unlock(ip);       /* unlock+drop */
2712                 /* ip pointer invalid */
2713
2714                 /*
2715                  * If the inode got dirted after we dropped our locks,
2716                  * it will have already been moved back to the SIDEQ.
2717                  */
2718                 hammer2_spin_ex(&pmp->list_spin);
2719         }
2720         hammer2_spin_unex(&pmp->list_spin);
2721         if (dorestart || (pmp->trans.flags & HAMMER2_TRANS_RESCAN)) {
2722 #ifdef HAMMER2_DEBUG_SYNC
2723                 kprintf("FILESYSTEM SYNC STAGE 1 RESTART\n");
2724                 /*tsleep(&dorestart, 0, "h2STG1-R", hz*20);*/
2725 #endif
2726                 dorestart = 1;
2727                 goto restart;
2728         }
2729 #ifdef HAMMER2_DEBUG_SYNC
2730         kprintf("FILESYSTEM SYNC STAGE 2 BEGIN\n");
2731         /*tsleep(&dorestart, 0, "h2STG2", hz*20);*/
2732 #endif
2733
2734         /*
2735          * We have to flush the PFS root last, even if it does not appear to
2736          * be dirty, because all the inodes in the PFS are indexed under it.
2737          * The normal flushing of iroot above would only occur if directory
2738          * entries under the root were changed.
2739          *
2740          * Specifying VOLHDR will cause an additionl flush of hmp->spmp
2741          * for the media making up the cluster.
2742          */
2743         if ((ip = pmp->iroot) != NULL) {
2744                 hammer2_inode_ref(ip);
2745                 hammer2_mtx_ex(&ip->lock);
2746                 hammer2_inode_chain_sync(ip);
2747                 hammer2_inode_chain_flush(ip, HAMMER2_XOP_INODE_STOP |
2748                                               HAMMER2_XOP_FSSYNC |
2749                                               HAMMER2_XOP_VOLHDR);
2750                 hammer2_inode_unlock(ip);       /* unlock+drop */
2751         }
2752 #ifdef HAMMER2_DEBUG_SYNC
2753         kprintf("FILESYSTEM SYNC STAGE 2 DONE\n");
2754 #endif
2755
2756         /*
2757          * device bioq sync
2758          */
2759         hammer2_bioq_sync(pmp);
2760
2761 #if 0
2762         info.pass = 1;
2763         info.waitfor = MNT_WAIT;
2764         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2765
2766         info.pass = 2;
2767         info.waitfor = MNT_WAIT;
2768         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2769 #endif
2770 #if 0
2771         /*
2772          * Generally speaking we now want to flush the media topology from
2773          * the iroot through to the inodes.  The flush stops at any inode
2774          * boundary, which allows the frontend to continue running concurrent
2775          * modifying operations on inodes (including kernel flushes of
2776          * buffers) without interfering with the main sync.
2777          *
2778          * Use the XOP interface to concurrently flush all nodes to
2779          * synchronize the PFSROOT subtopology to the media.  A standard
2780          * end-of-scan ENOENT error indicates cluster sufficiency.
2781          *
2782          * Note that this flush will not be visible on crash recovery until
2783          * we flush the super-root topology in the next loop.
2784          *
2785          * XXX For now wait for all flushes to complete.
2786          */
2787         if (mp && (ip = pmp->iroot) != NULL) {
2788                 /*
2789                  * If unmounting try to flush everything including any
2790                  * sub-trees under inodes, just in case there is dangling
2791                  * modified data, as a safety.  Otherwise just flush up to
2792                  * the inodes in this stage.
2793                  */
2794                 kprintf("MP & IROOT\n");
2795 #ifdef HAMMER2_DEBUG_SYNC
2796                 kprintf("FILESYSTEM SYNC STAGE 3 IROOT BEGIN\n");
2797 #endif
2798                 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
2799                         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2800                                                     HAMMER2_XOP_VOLHDR |
2801                                                     HAMMER2_XOP_FSSYNC |
2802                                                     HAMMER2_XOP_INODE_STOP);
2803                 } else {
2804                         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
2805                                                     HAMMER2_XOP_INODE_STOP |
2806                                                     HAMMER2_XOP_VOLHDR |
2807                                                     HAMMER2_XOP_FSSYNC |
2808                                                     HAMMER2_XOP_INODE_STOP);
2809                 }
2810                 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc);
2811                 error = hammer2_xop_collect(&xop->head,
2812                                             HAMMER2_XOP_COLLECT_WAITALL);
2813                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2814 #ifdef HAMMER2_DEBUG_SYNC
2815                 kprintf("FILESYSTEM SYNC STAGE 3 IROOT END\n");
2816 #endif
2817                 if (error == HAMMER2_ERROR_ENOENT)
2818                         error = 0;
2819                 else
2820                         error = hammer2_error_to_errno(error);
2821         } else {
2822                 error = 0;
2823         }
2824 #endif
2825         error = 0;      /* XXX */
2826         hammer2_trans_done(pmp, HAMMER2_TRANS_ISFLUSH);
2827
2828         return (error);
2829 }
2830
2831 static
2832 int
2833 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2834 {
2835         hammer2_inode_t *ip;
2836
2837         KKASSERT(MAXFIDSZ >= 16);
2838         ip = VTOI(vp);
2839         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2840         fhp->fid_ext = 0;
2841         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2842         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2843
2844         return 0;
2845 }
2846
2847 static
2848 int
2849 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2850                struct fid *fhp, struct vnode **vpp)
2851 {
2852         hammer2_pfs_t *pmp;
2853         hammer2_tid_t inum;
2854         int error;
2855
2856         pmp = MPTOPMP(mp);
2857         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2858         if (vpp) {
2859                 if (inum == 1)
2860                         error = hammer2_vfs_root(mp, vpp);
2861                 else
2862                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2863         } else {
2864                 error = 0;
2865         }
2866         if (error)
2867                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2868         return error;
2869 }
2870
2871 static
2872 int
2873 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2874                  int *exflagsp, struct ucred **credanonp)
2875 {
2876         hammer2_pfs_t *pmp;
2877         struct netcred *np;
2878         int error;
2879
2880         pmp = MPTOPMP(mp);
2881         np = vfs_export_lookup(mp, &pmp->export, nam);
2882         if (np) {
2883                 *exflagsp = np->netc_exflags;
2884                 *credanonp = &np->netc_anon;
2885                 error = 0;
2886         } else {
2887                 error = EACCES;
2888         }
2889         return error;
2890 }
2891
2892 /*
2893  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2894  * header into the HMP
2895  *
2896  * XXX read four volhdrs and use the one with the highest TID whos CRC
2897  *     matches.
2898  *
2899  * XXX check iCRCs.
2900  *
2901  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2902  *     nonexistant locations.
2903  *
2904  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2905  */
2906 static
2907 int
2908 hammer2_install_volume_header(hammer2_dev_t *hmp)
2909 {
2910         hammer2_volume_data_t *vd;
2911         struct buf *bp;
2912         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2913         int error_reported;
2914         int error;
2915         int valid;
2916         int i;
2917
2918         error_reported = 0;
2919         error = 0;
2920         valid = 0;
2921         bp = NULL;
2922
2923         /*
2924          * There are up to 4 copies of the volume header (syncs iterate
2925          * between them so there is no single master).  We don't trust the
2926          * volu_size field so we don't know precisely how large the filesystem
2927          * is, so depend on the OS to return an error if we go beyond the
2928          * block device's EOF.
2929          */
2930         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2931                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2932                               HAMMER2_VOLUME_BYTES, &bp);
2933                 if (error) {
2934                         brelse(bp);
2935                         bp = NULL;
2936                         continue;
2937                 }
2938
2939                 vd = (struct hammer2_volume_data *) bp->b_data;
2940                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2941                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2942                         brelse(bp);
2943                         bp = NULL;
2944                         continue;
2945                 }
2946
2947                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2948                         /* XXX: Reversed-endianness filesystem */
2949                         kprintf("hammer2: reverse-endian filesystem detected");
2950                         brelse(bp);
2951                         bp = NULL;
2952                         continue;
2953                 }
2954
2955                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2956                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2957                                       HAMMER2_VOLUME_ICRC0_SIZE);
2958                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2959                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2960                                        HAMMER2_VOLUME_ICRC1_SIZE);
2961                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2962                         kprintf("hammer2 volume header crc "
2963                                 "mismatch copy #%d %08x/%08x\n",
2964                                 i, crc0, crc);
2965                         error_reported = 1;
2966                         brelse(bp);
2967                         bp = NULL;
2968                         continue;
2969                 }
2970                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2971                         valid = 1;
2972                         hmp->voldata = *vd;
2973                         hmp->volhdrno = i;
2974                 }
2975                 brelse(bp);
2976                 bp = NULL;
2977         }
2978         if (valid) {
2979                 hmp->volsync = hmp->voldata;
2980                 hmp->free_reserved = hmp->voldata.allocator_size / 20;
2981                 error = 0;
2982                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2983                         kprintf("hammer2: using volume header #%d\n",
2984                                 hmp->volhdrno);
2985                 }
2986         } else {
2987                 error = EINVAL;
2988                 kprintf("hammer2: no valid volume headers found!\n");
2989         }
2990         return (error);
2991 }
2992
2993 /*
2994  * This handles hysteresis on regular file flushes.  Because the BIOs are
2995  * routed to a thread it is possible for an excessive number to build up
2996  * and cause long front-end stalls long before the runningbuffspace limit
2997  * is hit, so we implement hammer2_flush_pipe to control the
2998  * hysteresis.
2999  *
3000  * This is a particular problem when compression is used.
3001  */
3002 void
3003 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
3004 {
3005         atomic_add_int(&pmp->count_lwinprog, 1);
3006 }
3007
3008 void
3009 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
3010 {
3011         int lwinprog;
3012
3013         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
3014         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
3015             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
3016                 atomic_clear_int(&pmp->count_lwinprog,
3017                                  HAMMER2_LWINPROG_WAITING);
3018                 wakeup(&pmp->count_lwinprog);
3019         }
3020         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
3021             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
3022                 atomic_clear_int(&pmp->count_lwinprog,
3023                                  HAMMER2_LWINPROG_WAITING0);
3024                 wakeup(&pmp->count_lwinprog);
3025         }
3026 }
3027
3028 void
3029 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
3030 {
3031         int lwinprog;
3032         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
3033                                     HAMMER2_LWINPROG_WAITING0;
3034
3035         for (;;) {
3036                 lwinprog = pmp->count_lwinprog;
3037                 cpu_ccfence();
3038                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3039                         break;
3040                 tsleep_interlock(&pmp->count_lwinprog, 0);
3041                 atomic_set_int(&pmp->count_lwinprog, lwflag);
3042                 lwinprog = pmp->count_lwinprog;
3043                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
3044                         break;
3045                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
3046         }
3047 }
3048
3049 /*
3050  * It is possible for an excessive number of dirty chains or dirty inodes
3051  * to build up.  When this occurs we start an asynchronous filesystem sync.
3052  * If the level continues to build up, we stall, waiting for it to drop,
3053  * with some hysteresis.
3054  *
3055  * We limit the stall to two seconds per call.
3056  *
3057  * This relies on the kernel calling hammer2_vfs_modifying() prior to
3058  * obtaining any vnode locks before making a modifying VOP call.
3059  */
3060 static void
3061 hammer2_vfs_modifying(struct mount *mp)
3062 {
3063         hammer2_pfs_memory_wait(MPTOPMP(mp));
3064 }
3065
3066 /*
3067  * Initiate an asynchronous filesystem sync and, with hysteresis,
3068  * stall if the internal data structure count becomes too bloated.
3069  */
3070 void
3071 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
3072 {
3073         uint32_t waiting;
3074         int loops;
3075
3076         if (pmp == NULL || pmp->mp == NULL)
3077                 return;
3078
3079         for (loops = 0; loops < 2; ++loops) {
3080                 waiting = pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK;
3081                 cpu_ccfence();
3082
3083                 /*
3084                  * Start the syncer running at 1/2 the limit
3085                  */
3086                 if (waiting > hammer2_limit_dirty_chains / 2 ||
3087                     pmp->sideq_count > hammer2_limit_dirty_inodes / 2) {
3088                         trigger_syncer(pmp->mp);
3089                 }
3090
3091                 /*
3092                  * Stall at the limit waiting for the counts to drop.
3093                  * This code will typically be woken up once the count
3094                  * drops below 3/4 the limit, or in one second.
3095                  */
3096                 if (waiting < hammer2_limit_dirty_chains &&
3097                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
3098                         break;
3099                 }
3100                 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
3101                 atomic_set_int(&pmp->inmem_dirty_chains,
3102                                HAMMER2_DIRTYCHAIN_WAITING);
3103                 if (waiting < hammer2_limit_dirty_chains &&
3104                     pmp->sideq_count < hammer2_limit_dirty_inodes) {
3105                         break;
3106                 }
3107                 trigger_syncer(pmp->mp);
3108                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED, "h2memw", hz);
3109 #if 0
3110                 limit = pmp->mp->mnt_nvnodelistsize / 10;
3111                 if (limit < hammer2_limit_dirty_chains)
3112                         limit = hammer2_limit_dirty_chains;
3113                 if (limit < 1000)
3114                         limit = 1000;
3115 #endif
3116         }
3117 }
3118
3119 void
3120 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
3121 {
3122         if (pmp) {
3123                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
3124         }
3125 }
3126
3127 void
3128 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
3129 {
3130         uint32_t waiting;
3131
3132         if (pmp) {
3133                 waiting = atomic_fetchadd_int(&pmp->inmem_dirty_chains, -1);
3134                 /* don't need --waiting to test flag */
3135
3136                 if ((waiting & HAMMER2_DIRTYCHAIN_WAITING) &&
3137                     (pmp->inmem_dirty_chains & HAMMER2_DIRTYCHAIN_MASK) <=
3138                     hammer2_limit_dirty_chains * 2 / 3 &&
3139                     pmp->sideq_count <= hammer2_limit_dirty_inodes * 2 / 3) {
3140                         atomic_clear_int(&pmp->inmem_dirty_chains,
3141                                          HAMMER2_DIRTYCHAIN_WAITING);
3142                         wakeup(&pmp->inmem_dirty_chains);
3143                 }
3144         }
3145 }
3146
3147 /*
3148  * Returns 0 if the filesystem has tons of free space
3149  * Returns 1 if the filesystem has less than 10% remaining
3150  * Returns 2 if the filesystem has less than 2%/5% (user/root) remaining.
3151  */
3152 int
3153 hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred)
3154 {
3155         hammer2_pfs_t *pmp;
3156         hammer2_dev_t *hmp;
3157         hammer2_off_t free_reserved;
3158         hammer2_off_t free_nominal;
3159         int i;
3160
3161         pmp = ip->pmp;
3162
3163         if (pmp->free_ticks == 0 || pmp->free_ticks != ticks) {
3164                 free_reserved = HAMMER2_SEGSIZE;
3165                 free_nominal = 0x7FFFFFFFFFFFFFFFLLU;
3166                 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
3167                         hmp = pmp->pfs_hmps[i];
3168                         if (hmp == NULL)
3169                                 continue;
3170                         if (pmp->pfs_types[i] != HAMMER2_PFSTYPE_MASTER &&
3171                             pmp->pfs_types[i] != HAMMER2_PFSTYPE_SOFT_MASTER)
3172                                 continue;
3173
3174                         if (free_nominal > hmp->voldata.allocator_free)
3175                                 free_nominal = hmp->voldata.allocator_free;
3176                         if (free_reserved < hmp->free_reserved)
3177                                 free_reserved = hmp->free_reserved;
3178                 }
3179
3180                 /*
3181                  * SMP races ok
3182                  */
3183                 pmp->free_reserved = free_reserved;
3184                 pmp->free_nominal = free_nominal;
3185                 pmp->free_ticks = ticks;
3186         } else {
3187                 free_reserved = pmp->free_reserved;
3188                 free_nominal = pmp->free_nominal;
3189         }
3190         if (cred && cred->cr_uid != 0) {
3191                 if ((int64_t)(free_nominal - bytes) <
3192                     (int64_t)free_reserved) {
3193                         return 2;
3194                 }
3195         } else {
3196                 if ((int64_t)(free_nominal - bytes) <
3197                     (int64_t)free_reserved / 2) {
3198                         return 2;
3199                 }
3200         }
3201         if ((int64_t)(free_nominal - bytes) < (int64_t)free_reserved * 2)
3202                 return 1;
3203         return 0;
3204 }
3205
3206 /*
3207  * Debugging
3208  */
3209 void
3210 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
3211                    u_int flags)
3212 {
3213         hammer2_chain_t *scan;
3214         hammer2_chain_t *parent;
3215
3216         --*countp;
3217         if (*countp == 0) {
3218                 kprintf("%*.*s...\n", tab, tab, "");
3219                 return;
3220         }
3221         if (*countp < 0)
3222                 return;
3223         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
3224                 tab, tab, "", pfx,
3225                 chain, chain->bref.type,
3226                 chain->bref.key, chain->bref.keybits,
3227                 chain->bref.mirror_tid);
3228
3229         kprintf("%*.*s      [%08x] (%s) refs=%d",
3230                 tab, tab, "",
3231                 chain->flags,
3232                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
3233                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
3234                 chain->refs);
3235
3236         parent = chain->parent;
3237         if (parent)
3238                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
3239                         tab, tab, "",
3240                         parent, parent->flags, parent->refs);
3241         if (RB_EMPTY(&chain->core.rbtree)) {
3242                 kprintf("\n");
3243         } else {
3244                 kprintf(" {\n");
3245                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree) {
3246                         if ((scan->flags & flags) || flags == (u_int)-1) {
3247                                 hammer2_dump_chain(scan, tab + 4, countp, 'a',
3248                                                    flags);
3249                         }
3250                 }
3251                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
3252                         kprintf("%*.*s}(%s)\n", tab, tab, "",
3253                                 chain->data->ipdata.filename);
3254                 else
3255                         kprintf("%*.*s}\n", tab, tab, "");
3256         }
3257 }