?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
hammer2 - Consolidate backend rename ops
[dragonfly.git] / sys / vfs / hammer2 / hammer2_vfsops.c
blob4069198eff86378fa6e3b286469920efb0d8e3f3
1 /*
2 * Copyright (c) 2011-2015 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 *
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
17 * distribution.
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * SUCH DAMAGE.
34 */
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/nlookup.h>
39 #include <sys/vnode.h>
40 #include <sys/mount.h>
41 #include <sys/fcntl.h>
42 #include <sys/buf.h>
43 #include <sys/uuid.h>
44 #include <sys/vfsops.h>
45 #include <sys/sysctl.h>
46 #include <sys/socket.h>
47 #include <sys/objcache.h>
48
49 #include <sys/proc.h>
50 #include <sys/namei.h>
51 #include <sys/mountctl.h>
52 #include <sys/dirent.h>
53 #include <sys/uio.h>
54
55 #include <sys/mutex.h>
56 #include <sys/mutex2.h>
57
58 #include "hammer2.h"
59 #include "hammer2_disk.h"
60 #include "hammer2_mount.h"
61 #include "hammer2_lz4.h"
62
63 #include "zlib/hammer2_zlib.h"
64
65 #define REPORT_REFS_ERRORS 1 /* XXX remove me */
66
67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
68
69 struct hammer2_sync_info {
70 int error;
71 int waitfor;
72 };
73
74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
75 static struct hammer2_mntlist hammer2_mntlist;
76
77 struct hammer2_pfslist hammer2_pfslist;
78 struct lock hammer2_mntlk;
79
80 int hammer2_debug;
81 int hammer2_cluster_read = 4; /* physical read-ahead */
82 int hammer2_cluster_write = 0; /* bdwrite() so later inval works */
83 int hammer2_dedup_enable = 1;
84 int hammer2_inval_enable = 0;
85 int hammer2_flush_pipe = 100;
86 int hammer2_synchronous_flush = 1;
87 int hammer2_dio_count;
88 long hammer2_chain_allocs;
89 long hammer2_chain_frees;
90 long hammer2_limit_dirty_chains;
91 long hammer2_count_modified_chains;
92 long hammer2_iod_invals;
93 long hammer2_iod_file_read;
94 long hammer2_iod_meta_read;
95 long hammer2_iod_indr_read;
96 long hammer2_iod_fmap_read;
97 long hammer2_iod_volu_read;
98 long hammer2_iod_file_write;
99 long hammer2_iod_file_wembed;
100 long hammer2_iod_file_wzero;
101 long hammer2_iod_file_wdedup;
102 long hammer2_iod_meta_write;
103 long hammer2_iod_indr_write;
104 long hammer2_iod_fmap_write;
105 long hammer2_iod_volu_write;
106
107 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
108 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
109 "Buffer used for compression.");
110
111 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
112 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
113 "Buffer used for decompression.");
114
115 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
116
117 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
118 &hammer2_debug, 0, "");
119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW,
120 &hammer2_cluster_read, 0, "");
121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
122 &hammer2_cluster_write, 0, "");
123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
124 &hammer2_dedup_enable, 0, "");
125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
126 &hammer2_inval_enable, 0, "");
127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
128 &hammer2_flush_pipe, 0, "");
129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
130 &hammer2_synchronous_flush, 0, "");
131 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
132 &hammer2_chain_allocs, 0, "");
133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
134 &hammer2_chain_frees, 0, "");
135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
136 &hammer2_limit_dirty_chains, 0, "");
137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
138 &hammer2_count_modified_chains, 0, "");
139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
140 &hammer2_dio_count, 0, "");
141
142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
143 &hammer2_iod_invals, 0, "");
144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
145 &hammer2_iod_file_read, 0, "");
146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
147 &hammer2_iod_meta_read, 0, "");
148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
149 &hammer2_iod_indr_read, 0, "");
150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
151 &hammer2_iod_fmap_read, 0, "");
152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
153 &hammer2_iod_volu_read, 0, "");
154
155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
156 &hammer2_iod_file_write, 0, "");
157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
158 &hammer2_iod_file_wembed, 0, "");
159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
160 &hammer2_iod_file_wzero, 0, "");
161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
162 &hammer2_iod_file_wdedup, 0, "");
163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
164 &hammer2_iod_meta_write, 0, "");
165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
166 &hammer2_iod_indr_write, 0, "");
167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
168 &hammer2_iod_fmap_write, 0, "");
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
170 &hammer2_iod_volu_write, 0, "");
171
172 long hammer2_check_icrc32;
173 long hammer2_check_xxhash64;
174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
175 &hammer2_check_icrc32, 0, "");
176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
177 &hammer2_check_xxhash64, 0, "");
178
179 static int hammer2_vfs_init(struct vfsconf *conf);
180 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
181 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
182 struct ucred *cred);
183 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
184 struct vnode *, struct ucred *);
185 static int hammer2_recovery(hammer2_dev_t *hmp);
186 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
187 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
188 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
189 struct ucred *cred);
190 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
191 struct ucred *cred);
192 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
193 struct fid *fhp, struct vnode **vpp);
194 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
195 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
196 int *exflagsp, struct ucred **credanonp);
197
198 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
199 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
200
201 static void hammer2_update_pmps(hammer2_dev_t *hmp);
202
203 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
204 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
205 hammer2_dev_t *hmp);
206
207 /*
208 * HAMMER2 vfs operations.
209 */
210 static struct vfsops hammer2_vfsops = {
211 .vfs_init = hammer2_vfs_init,
212 .vfs_uninit = hammer2_vfs_uninit,
213 .vfs_sync = hammer2_vfs_sync,
214 .vfs_mount = hammer2_vfs_mount,
215 .vfs_unmount = hammer2_vfs_unmount,
216 .vfs_root = hammer2_vfs_root,
217 .vfs_statfs = hammer2_vfs_statfs,
218 .vfs_statvfs = hammer2_vfs_statvfs,
219 .vfs_vget = hammer2_vfs_vget,
220 .vfs_vptofh = hammer2_vfs_vptofh,
221 .vfs_fhtovp = hammer2_vfs_fhtovp,
222 .vfs_checkexp = hammer2_vfs_checkexp
223 };
224
225 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
226
227 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
228 MODULE_VERSION(hammer2, 1);
229
230 static
231 int
232 hammer2_vfs_init(struct vfsconf *conf)
233 {
234 static struct objcache_malloc_args margs_read;
235 static struct objcache_malloc_args margs_write;
236 static struct objcache_malloc_args margs_vop;
237
238 int error;
239
240 error = 0;
241
242 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
243 error = EINVAL;
244 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
245 error = EINVAL;
246 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
247 error = EINVAL;
248
249 if (error)
250 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
251
252 margs_read.objsize = 65536;
253 margs_read.mtype = M_HAMMER2_DEBUFFER;
254
255 margs_write.objsize = 32768;
256 margs_write.mtype = M_HAMMER2_CBUFFER;
257
258 margs_vop.objsize = sizeof(hammer2_xop_t);
259 margs_vop.mtype = M_HAMMER2;
260
261 /*
262 * Note thaht for the XOPS cache we want backing store allocations
263 * to use M_ZERO. This is not allowed in objcache_get() (to avoid
264 * confusion), so use the backing store function that does it. This
265 * means that initial XOPS objects are zerod but REUSED objects are
266 * not. So we are responsible for cleaning the object up sufficiently
267 * for our needs before objcache_put()ing it back (typically just the
268 * FIFO indices).
269 */
270 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
271 0, 1, NULL, NULL, NULL,
272 objcache_malloc_alloc,
273 objcache_malloc_free,
274 &margs_read);
275 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
276 0, 1, NULL, NULL, NULL,
277 objcache_malloc_alloc,
278 objcache_malloc_free,
279 &margs_write);
280 cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
281 0, 1, NULL, NULL, NULL,
282 objcache_malloc_alloc_zero,
283 objcache_malloc_free,
284 &margs_vop);
285
286
287 lockinit(&hammer2_mntlk, "mntlk", 0, 0);
288 TAILQ_INIT(&hammer2_mntlist);
289 TAILQ_INIT(&hammer2_pfslist);
290
291 hammer2_limit_dirty_chains = maxvnodes / 10;
292 if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
293 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
294
295 return (error);
296 }
297
298 static
299 int
300 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
301 {
302 objcache_destroy(cache_buffer_read);
303 objcache_destroy(cache_buffer_write);
304 objcache_destroy(cache_xops);
305 return 0;
306 }
307
308 /*
309 * Core PFS allocator. Used to allocate or reference the pmp structure
310 * for PFS cluster mounts and the spmp structure for media (hmp) structures.
311 * The pmp can be passed in or loaded by this function using the chain and
312 * inode data.
313 *
314 * pmp->modify_tid tracks new modify_tid transaction ids for front-end
315 * transactions. Note that synchronization does not use this field.
316 * (typically frontend operations and synchronization cannot run on the
317 * same PFS node at the same time).
318 *
319 * XXX check locking
320 */
321 hammer2_pfs_t *
322 hammer2_pfsalloc(hammer2_chain_t *chain,
323 const hammer2_inode_data_t *ripdata,
324 hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
325 {
326 hammer2_pfs_t *pmp;
327 hammer2_inode_t *iroot;
328 int count;
329 int i;
330 int j;
331
332 pmp = NULL;
333
334 /*
335 * Locate or create the PFS based on the cluster id. If ripdata
336 * is NULL this is a spmp which is unique and is always allocated.
337 *
338 * If the device is mounted in local mode all PFSs are considered
339 * independent and not part of any cluster (for debugging only).
340 */
341 if (ripdata) {
342 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
343 if (force_local != pmp->force_local)
344 continue;
345 if (force_local == NULL &&
346 bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
347 sizeof(pmp->pfs_clid)) == 0) {
348 break;
349 } else if (force_local && pmp->pfs_names[0] &&
350 strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
351 break;
352 }
353 }
354 }
355
356 if (pmp == NULL) {
357 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
358 pmp->force_local = force_local;
359 hammer2_trans_manage_init(pmp);
360 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
361 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
362 lockinit(&pmp->lock, "pfslk", 0, 0);
363 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
364 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
365 spin_init(&pmp->xop_spin, "h2xop");
366 spin_init(&pmp->lru_spin, "h2lru");
367 RB_INIT(&pmp->inum_tree);
368 TAILQ_INIT(&pmp->sideq);
369 TAILQ_INIT(&pmp->lru_list);
370 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
371
372 /*
373 * Distribute backend operations to threads
374 */
375 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
376 hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
377
378 /*
379 * Save the last media transaction id for the flusher. Set
380 * initial
381 */
382 if (ripdata)
383 pmp->pfs_clid = ripdata->meta.pfs_clid;
384 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
385
386 /*
387 * The synchronization thread may start too early, make
388 * sure it stays frozen until we are ready to let it go.
389 * XXX
390 */
391 /*
392 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
393 HAMMER2_THREAD_REMASTER;
394 */
395 }
396
397 /*
398 * Create the PFS's root inode and any missing XOP helper threads.
399 */
400 if ((iroot = pmp->iroot) == NULL) {
401 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
402 if (ripdata)
403 iroot->meta = ripdata->meta;
404 pmp->iroot = iroot;
405 hammer2_inode_ref(iroot);
406 hammer2_inode_unlock(iroot);
407 }
408
409 /*
410 * Stop here if no chain is passed in.
411 */
412 if (chain == NULL)
413 goto done;
414
415 /*
416 * When a chain is passed in we must add it to the PFS's root
417 * inode, update pmp->pfs_types[], and update the syncronization
418 * threads.
419 *
420 * When forcing local mode, mark the PFS as a MASTER regardless.
421 *
422 * At the moment empty spots can develop due to removals or failures.
423 * Ultimately we want to re-fill these spots but doing so might
424 * confused running code. XXX
425 */
426 hammer2_inode_ref(iroot);
427 hammer2_mtx_ex(&iroot->lock);
428 j = iroot->cluster.nchains;
429
430 kprintf("add PFS to pmp %p[%d]\n", pmp, j);
431
432 if (j == HAMMER2_MAXCLUSTER) {
433 kprintf("hammer2_mount: cluster full!\n");
434 /* XXX fatal error? */
435 } else {
436 KKASSERT(chain->pmp == NULL);
437 chain->pmp = pmp;
438 hammer2_chain_ref(chain);
439 iroot->cluster.array[j].chain = chain;
440 if (force_local)
441 pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
442 else
443 pmp->pfs_types[j] = ripdata->meta.pfs_type;
444 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
445 pmp->pfs_hmps[j] = chain->hmp;
446
447 /*
448 * If the PFS is already mounted we must account
449 * for the mount_count here.
450 */
451 if (pmp->mp)
452 ++chain->hmp->mount_count;
453
454 /*
455 * May have to fixup dirty chain tracking. Previous
456 * pmp was NULL so nothing to undo.
457 */
458 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
459 hammer2_pfs_memory_inc(pmp);
460 ++j;
461 }
462 iroot->cluster.nchains = j;
463
464 /*
465 * Update nmasters from any PFS inode which is part of the cluster.
466 * It is possible that this will result in a value which is too
467 * high. MASTER PFSs are authoritative for pfs_nmasters and will
468 * override this value later on.
469 *
470 * (This informs us of masters that might not currently be
471 * discoverable by this mount).
472 */
473 if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
474 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
475 }
476
477 /*
478 * Count visible masters. Masters are usually added with
479 * ripdata->meta.pfs_nmasters set to 1. This detects when there
480 * are more (XXX and must update the master inodes).
481 */
482 count = 0;
483 for (i = 0; i < iroot->cluster.nchains; ++i) {
484 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
485 ++count;
486 }
487 if (pmp->pfs_nmasters < count)
488 pmp->pfs_nmasters = count;
489
490 /*
491 * Create missing synchronization and support threads.
492 *
493 * Single-node masters (including snapshots) have nothing to
494 * synchronize and do not require this thread.
495 *
496 * Multi-node masters or any number of soft masters, slaves, copy,
497 * or other PFS types need the thread.
498 *
499 * Each thread is responsible for its particular cluster index.
500 * We use independent threads so stalls or mismatches related to
501 * any given target do not affect other targets.
502 */
503 for (i = 0; i < iroot->cluster.nchains; ++i) {
504 /*
505 * Single-node masters (including snapshots) have nothing
506 * to synchronize and will make direct xops support calls,
507 * thus they do not require this thread.
508 *
509 * Note that there can be thousands of snapshots. We do not
510 * want to create thousands of threads.
511 */
512 if (pmp->pfs_nmasters <= 1 &&
513 pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
514 continue;
515 }
516
517 /*
518 * Sync support thread
519 */
520 if (pmp->sync_thrs[i].td == NULL) {
521 hammer2_thr_create(&pmp->sync_thrs[i], pmp,
522 "h2nod", i, -1,
523 hammer2_primary_sync_thread);
524 }
525 }
526
527 /*
528 * Create missing Xop threads
529 *
530 * NOTE: We create helper threads for all mounted PFSs or any
531 * PFSs with 2+ nodes (so the sync thread can update them,
532 * even if not mounted).
533 */
534 if (pmp->mp || iroot->cluster.nchains >= 2)
535 hammer2_xop_helper_create(pmp);
536
537 hammer2_mtx_unlock(&iroot->lock);
538 hammer2_inode_drop(iroot);
539 done:
540 return pmp;
541 }
542
543 /*
544 * Deallocate an element of a probed PFS. If destroying and this is a
545 * MASTER, adjust nmasters.
546 *
547 * This function does not physically destroy the PFS element in its device
548 * under the super-root (see hammer2_ioctl_pfs_delete()).
549 */
550 void
551 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
552 {
553 hammer2_inode_t *iroot;
554 hammer2_chain_t *chain;
555 int j;
556
557 /*
558 * Cleanup our reference on iroot. iroot is (should) not be needed
559 * by the flush code.
560 */
561 iroot = pmp->iroot;
562 if (iroot) {
563 /*
564 * Stop synchronizing
565 *
566 * XXX flush after acquiring the iroot lock.
567 * XXX clean out the cluster index from all inode structures.
568 */
569 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
570
571 /*
572 * Remove the cluster index from the group. If destroying
573 * the PFS and this is a master, adjust pfs_nmasters.
574 */
575 hammer2_mtx_ex(&iroot->lock);
576 chain = iroot->cluster.array[clindex].chain;
577 iroot->cluster.array[clindex].chain = NULL;
578
579 switch(pmp->pfs_types[clindex]) {
580 case HAMMER2_PFSTYPE_MASTER:
581 if (destroying && pmp->pfs_nmasters > 0)
582 --pmp->pfs_nmasters;
583 /* XXX adjust ripdata->meta.pfs_nmasters */
584 break;
585 default:
586 break;
587 }
588 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
589
590 hammer2_mtx_unlock(&iroot->lock);
591
592 /*
593 * Release the chain.
594 */
595 if (chain) {
596 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
597 hammer2_chain_drop(chain);
598 }
599
600 /*
601 * Terminate all XOP threads for the cluster index.
602 */
603 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
604 hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
605 }
606 }
607
608 /*
609 * Destroy a PFS, typically only occurs after the last mount on a device
610 * has gone away.
611 */
612 static void
613 hammer2_pfsfree(hammer2_pfs_t *pmp)
614 {
615 hammer2_inode_t *iroot;
616 hammer2_chain_t *chain;
617 int i;
618 int j;
619
620 /*
621 * Cleanup our reference on iroot. iroot is (should) not be needed
622 * by the flush code.
623 */
624 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
625
626 iroot = pmp->iroot;
627 if (iroot) {
628 for (i = 0; i < iroot->cluster.nchains; ++i) {
629 hammer2_thr_delete(&pmp->sync_thrs[i]);
630 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
631 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
632 }
633 #if REPORT_REFS_ERRORS
634 if (pmp->iroot->refs != 1)
635 kprintf("PMP->IROOT %p REFS WRONG %d\n",
636 pmp->iroot, pmp->iroot->refs);
637 #else
638 KKASSERT(pmp->iroot->refs == 1);
639 #endif
640 /* ref for pmp->iroot */
641 hammer2_inode_drop(pmp->iroot);
642 pmp->iroot = NULL;
643 }
644
645 /*
646 * Cleanup chains remaining on LRU list.
647 */
648 kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
649 while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
650 hammer2_chain_ref(chain);
651 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
652 hammer2_chain_drop(chain);
653 }
654
655 /*
656 * Free remaining pmp resources
657 */
658 kmalloc_destroy(&pmp->mmsg);
659 kmalloc_destroy(&pmp->minode);
660
661 kfree(pmp, M_HAMMER2);
662 }
663
664 /*
665 * Remove all references to hmp from the pfs list. Any PFS which becomes
666 * empty is terminated and freed.
667 *
668 * XXX inefficient.
669 */
670 static void
671 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
672 {
673 hammer2_pfs_t *pmp;
674 hammer2_inode_t *iroot;
675 hammer2_chain_t *rchain;
676 int didfreeze;
677 int i;
678 int j;
679
680 again:
681 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
682 if ((iroot = pmp->iroot) == NULL)
683 continue;
684 if (hmp->spmp == pmp) {
685 kprintf("unmount hmp %p remove spmp %p\n",
686 hmp, pmp);
687 hmp->spmp = NULL;
688 }
689
690 /*
691 * Determine if this PFS is affected. If it is we must
692 * freeze all management threads and lock its iroot.
693 *
694 * Freezing a management thread forces it idle, operations
695 * in-progress will be aborted and it will have to start
696 * over again when unfrozen, or exit if told to exit.
697 */
698 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
699 if (pmp->pfs_hmps[i] == hmp)
700 break;
701 }
702 if (i != HAMMER2_MAXCLUSTER) {
703 /*
704 * Make sure all synchronization threads are locked
705 * down.
706 */
707 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
708 if (pmp->pfs_hmps[i] == NULL)
709 continue;
710 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
711 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
712 hammer2_thr_freeze_async(
713 &pmp->xop_groups[j].thrs[i]);
714 }
715 }
716 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
717 if (pmp->pfs_hmps[i] == NULL)
718 continue;
719 hammer2_thr_freeze(&pmp->sync_thrs[i]);
720 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
721 hammer2_thr_freeze(
722 &pmp->xop_groups[j].thrs[i]);
723 }
724 }
725
726 /*
727 * Lock the inode and clean out matching chains.
728 * Note that we cannot use hammer2_inode_lock_*()
729 * here because that would attempt to validate the
730 * cluster that we are in the middle of ripping
731 * apart.
732 *
733 * WARNING! We are working directly on the inodes
734 * embedded cluster.
735 */
736 hammer2_mtx_ex(&iroot->lock);
737
738 /*
739 * Remove the chain from matching elements of the PFS.
740 */
741 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
742 if (pmp->pfs_hmps[i] != hmp)
743 continue;
744 hammer2_thr_delete(&pmp->sync_thrs[i]);
745 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
746 hammer2_thr_delete(
747 &pmp->xop_groups[j].thrs[i]);
748 }
749 rchain = iroot->cluster.array[i].chain;
750 iroot->cluster.array[i].chain = NULL;
751 pmp->pfs_types[i] = 0;
752 if (pmp->pfs_names[i]) {
753 kfree(pmp->pfs_names[i], M_HAMMER2);
754 pmp->pfs_names[i] = NULL;
755 }
756 if (rchain) {
757 hammer2_chain_drop(rchain);
758 /* focus hint */
759 if (iroot->cluster.focus == rchain)
760 iroot->cluster.focus = NULL;
761 }
762 pmp->pfs_hmps[i] = NULL;
763 }
764 hammer2_mtx_unlock(&iroot->lock);
765 didfreeze = 1; /* remaster, unfreeze down below */
766 } else {
767 didfreeze = 0;
768 }
769
770 /*
771 * Cleanup trailing chains. Gaps may remain.
772 */
773 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
774 if (pmp->pfs_hmps[i])
775 break;
776 }
777 iroot->cluster.nchains = i + 1;
778
779 /*
780 * If the PMP has no elements remaining we can destroy it.
781 * (this will transition management threads from frozen->exit).
782 */
783 if (iroot->cluster.nchains == 0) {
784 kprintf("unmount hmp %p last ref to PMP=%p\n",
785 hmp, pmp);
786 hammer2_pfsfree(pmp);
787 goto again;
788 }
789
790 /*
791 * If elements still remain we need to set the REMASTER
792 * flag and unfreeze it.
793 */
794 if (didfreeze) {
795 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
796 if (pmp->pfs_hmps[i] == NULL)
797 continue;
798 hammer2_thr_remaster(&pmp->sync_thrs[i]);
799 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
800 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
801 hammer2_thr_remaster(
802 &pmp->xop_groups[j].thrs[i]);
803 hammer2_thr_unfreeze(
804 &pmp->xop_groups[j].thrs[i]);
805 }
806 }
807 }
808 }
809 }
810
811 /*
812 * Mount or remount HAMMER2 fileystem from physical media
813 *
814 * mountroot
815 * mp mount point structure
816 * path NULL
817 * data <unused>
818 * cred <unused>
819 *
820 * mount
821 * mp mount point structure
822 * path path to mount point
823 * data pointer to argument structure in user space
824 * volume volume path (device@LABEL form)
825 * hflags user mount flags
826 * cred user credentials
827 *
828 * RETURNS: 0 Success
829 * !0 error number
830 */
831 static
832 int
833 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
834 struct ucred *cred)
835 {
836 struct hammer2_mount_info info;
837 hammer2_pfs_t *pmp;
838 hammer2_pfs_t *spmp;
839 hammer2_dev_t *hmp;
840 hammer2_dev_t *force_local;
841 hammer2_key_t key_next;
842 hammer2_key_t key_dummy;
843 hammer2_key_t lhc;
844 struct vnode *devvp;
845 struct nlookupdata nd;
846 hammer2_chain_t *parent;
847 hammer2_chain_t *chain;
848 hammer2_cluster_t *cluster;
849 const hammer2_inode_data_t *ripdata;
850 hammer2_blockref_t bref;
851 struct file *fp;
852 char devstr[MNAMELEN];
853 size_t size;
854 size_t done;
855 char *dev;
856 char *label;
857 int ronly = 1;
858 int error;
859 int cache_index;
860 int i;
861
862 hmp = NULL;
863 pmp = NULL;
864 dev = NULL;
865 label = NULL;
866 devvp = NULL;
867 cache_index = -1;
868
869 kprintf("hammer2_mount\n");
870
871 if (path == NULL) {
872 /*
873 * Root mount
874 */
875 bzero(&info, sizeof(info));
876 info.cluster_fd = -1;
877 ksnprintf(devstr, sizeof(devstr), "%s",
878 mp->mnt_stat.f_mntfromname);
879 kprintf("hammer2_mount: root '%s'\n", devstr);
880 } else {
881 /*
882 * Non-root mount or updating a mount
883 */
884 error = copyin(data, &info, sizeof(info));
885 if (error)
886 return (error);
887
888 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
889 if (error)
890 return (error);
891 }
892
893 /* Extract device and label */
894 dev = devstr;
895 label = strchr(devstr, '@');
896 if (label == NULL ||
897 ((label + 1) - dev) > done) {
898 return (EINVAL);
899 }
900 *label = '\0';
901 label++;
902 if (*label == '\0')
903 return (EINVAL);
904
905 kprintf("hammer2_mount: dev=\"%s\" label=\"%s\"\n",
906 dev, label);
907
908 if (mp->mnt_flag & MNT_UPDATE) {
909 /*
910 * Update mount. Note that pmp->iroot->cluster is
911 * an inode-embedded cluster and thus cannot be
912 * directly locked.
913 *
914 * XXX HAMMER2 needs to implement NFS export via
915 * mountctl.
916 */
917 pmp = MPTOPMP(mp);
918 pmp->hflags = info.hflags;
919 cluster = &pmp->iroot->cluster;
920 for (i = 0; i < cluster->nchains; ++i) {
921 if (cluster->array[i].chain == NULL)
922 continue;
923 hmp = cluster->array[i].chain->hmp;
924 devvp = hmp->devvp;
925 error = hammer2_remount(hmp, mp, path,
926 devvp, cred);
927 if (error)
928 break;
929 }
930
931 return error;
932 }
933
934 /*
935 * HMP device mount
936 *
937 * If a path is specified and dev is not an empty string, lookup the
938 * name and verify that it referes to a block device.
939 *
940 * If a path is specified and dev is an empty string we fall through
941 * and locate the label in the hmp search.
942 */
943 if (path && *dev != 0) {
944 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
945 if (error == 0)
946 error = nlookup(&nd);
947 if (error == 0)
948 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
949 nlookup_done(&nd);
950 } else if (path == NULL) {
951 /* root mount */
952 cdev_t cdev = kgetdiskbyname(dev);
953 error = bdevvp(cdev, &devvp);
954 if (error)
955 kprintf("hammer2: cannot find '%s'\n", dev);
956 } else {
957 /*
958 * We will locate the hmp using the label in the hmp loop.
959 */
960 error = 0;
961 }
962
963 if (error == 0 && devvp) {
964 if (vn_isdisk(devvp, &error))
965 error = vfs_mountedon(devvp);
966 }
967
968 /*
969 * Determine if the device has already been mounted. After this
970 * check hmp will be non-NULL if we are doing the second or more
971 * hammer2 mounts from the same device.
972 */
973 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
974 if (devvp) {
975 /*
976 * Match the device
977 */
978 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
979 if (hmp->devvp == devvp)
980 break;
981 }
982 } else if (error == 0) {
983 /*
984 * Match the label to a pmp already probed.
985 */
986 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
987 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
988 if (pmp->pfs_names[i] &&
989 strcmp(pmp->pfs_names[i], label) == 0) {
990 hmp = pmp->pfs_hmps[i];
991 break;
992 }
993 }
994 if (hmp)
995 break;
996 }
997 if (hmp == NULL)
998 error = ENOENT;
999 }
1000
1001 /*
1002 * Open the device if this isn't a secondary mount and construct
1003 * the H2 device mount (hmp).
1004 */
1005 if (hmp == NULL) {
1006 hammer2_chain_t *schain;
1007 hammer2_xid_t xid;
1008
1009 if (error == 0 && vcount(devvp) > 0)
1010 error = EBUSY;
1011
1012 /*
1013 * Now open the device
1014 */
1015 if (error == 0) {
1016 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1017 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1018 error = vinvalbuf(devvp, V_SAVE, 0, 0);
1019 if (error == 0) {
1020 error = VOP_OPEN(devvp,
1021 ronly ? FREAD : FREAD | FWRITE,
1022 FSCRED, NULL);
1023 }
1024 vn_unlock(devvp);
1025 }
1026 if (error && devvp) {
1027 vrele(devvp);
1028 devvp = NULL;
1029 }
1030 if (error) {
1031 lockmgr(&hammer2_mntlk, LK_RELEASE);
1032 return error;
1033 }
1034 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1035 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1036 hmp->ronly = ronly;
1037 hmp->devvp = devvp;
1038 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1039 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1040 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1041 RB_INIT(&hmp->iotree);
1042 spin_init(&hmp->io_spin, "hm2mount_io");
1043 spin_init(&hmp->list_spin, "hm2mount_list");
1044 TAILQ_INIT(&hmp->flushq);
1045
1046 lockinit(&hmp->vollk, "h2vol", 0, 0);
1047 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1048
1049 /*
1050 * vchain setup. vchain.data is embedded.
1051 * vchain.refs is initialized and will never drop to 0.
1052 *
1053 * NOTE! voldata is not yet loaded.
1054 */
1055 hmp->vchain.hmp = hmp;
1056 hmp->vchain.refs = 1;
1057 hmp->vchain.data = (void *)&hmp->voldata;
1058 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1059 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1060 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1061
1062 hammer2_chain_core_init(&hmp->vchain);
1063 /* hmp->vchain.u.xxx is left NULL */
1064
1065 /*
1066 * fchain setup. fchain.data is embedded.
1067 * fchain.refs is initialized and will never drop to 0.
1068 *
1069 * The data is not used but needs to be initialized to
1070 * pass assertion muster. We use this chain primarily
1071 * as a placeholder for the freemap's top-level RBTREE
1072 * so it does not interfere with the volume's topology
1073 * RBTREE.
1074 */
1075 hmp->fchain.hmp = hmp;
1076 hmp->fchain.refs = 1;
1077 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1078 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1079 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1080 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1081 hmp->fchain.bref.methods =
1082 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1083 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1084
1085 hammer2_chain_core_init(&hmp->fchain);
1086 /* hmp->fchain.u.xxx is left NULL */
1087
1088 /*
1089 * Install the volume header and initialize fields from
1090 * voldata.
1091 */
1092 error = hammer2_install_volume_header(hmp);
1093 if (error) {
1094 hammer2_unmount_helper(mp, NULL, hmp);
1095 lockmgr(&hammer2_mntlk, LK_RELEASE);
1096 hammer2_vfs_unmount(mp, MNT_FORCE);
1097 return error;
1098 }
1099
1100 /*
1101 * Really important to get these right or flush will get
1102 * confused.
1103 */
1104 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1105 kprintf("alloc spmp %p tid %016jx\n",
1106 hmp->spmp, hmp->voldata.mirror_tid);
1107 spmp = hmp->spmp;
1108
1109 /*
1110 * Dummy-up vchain and fchain's modify_tid. mirror_tid
1111 * is inherited from the volume header.
1112 */
1113 xid = 0;
1114 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1115 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1116 hmp->vchain.pmp = spmp;
1117 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1118 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1119 hmp->fchain.pmp = spmp;
1120
1121 /*
1122 * First locate the super-root inode, which is key 0
1123 * relative to the volume header's blockset.
1124 *
1125 * Then locate the root inode by scanning the directory keyspace
1126 * represented by the label.
1127 */
1128 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1129 schain = hammer2_chain_lookup(&parent, &key_dummy,
1130 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1131 &cache_index, 0);
1132 hammer2_chain_lookup_done(parent);
1133 if (schain == NULL) {
1134 kprintf("hammer2_mount: invalid super-root\n");
1135 hammer2_unmount_helper(mp, NULL, hmp);
1136 lockmgr(&hammer2_mntlk, LK_RELEASE);
1137 hammer2_vfs_unmount(mp, MNT_FORCE);
1138 return EINVAL;
1139 }
1140 if (schain->error) {
1141 kprintf("hammer2_mount: error %s reading super-root\n",
1142 hammer2_error_str(schain->error));
1143 hammer2_chain_unlock(schain);
1144 hammer2_chain_drop(schain);
1145 schain = NULL;
1146 hammer2_unmount_helper(mp, NULL, hmp);
1147 lockmgr(&hammer2_mntlk, LK_RELEASE);
1148 hammer2_vfs_unmount(mp, MNT_FORCE);
1149 return EINVAL;
1150 }
1151
1152 /*
1153 * The super-root always uses an inode_tid of 1 when
1154 * creating PFSs.
1155 */
1156 spmp->inode_tid = 1;
1157 spmp->modify_tid = schain->bref.modify_tid + 1;
1158
1159 /*
1160 * Sanity-check schain's pmp and finish initialization.
1161 * Any chain belonging to the super-root topology should
1162 * have a NULL pmp (not even set to spmp).
1163 */
1164 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1165 KKASSERT(schain->pmp == NULL);
1166 spmp->pfs_clid = ripdata->meta.pfs_clid;
1167
1168 /*
1169 * Replace the dummy spmp->iroot with a real one. It's
1170 * easier to just do a wholesale replacement than to try
1171 * to update the chain and fixup the iroot fields.
1172 *
1173 * The returned inode is locked with the supplied cluster.
1174 */
1175 cluster = hammer2_cluster_from_chain(schain);
1176 hammer2_inode_drop(spmp->iroot);
1177 spmp->iroot = NULL;
1178 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1179 spmp->spmp_hmp = hmp;
1180 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1181 spmp->pfs_hmps[0] = hmp;
1182 hammer2_inode_ref(spmp->iroot);
1183 hammer2_inode_unlock(spmp->iroot);
1184 hammer2_cluster_unlock(cluster);
1185 hammer2_cluster_drop(cluster);
1186 schain = NULL;
1187 /* leave spmp->iroot with one ref */
1188
1189 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1190 error = hammer2_recovery(hmp);
1191 /* XXX do something with error */
1192 }
1193 hammer2_update_pmps(hmp);
1194 hammer2_iocom_init(hmp);
1195
1196 /*
1197 * Ref the cluster management messaging descriptor. The mount
1198 * program deals with the other end of the communications pipe.
1199 *
1200 * Root mounts typically do not supply one.
1201 */
1202 if (info.cluster_fd >= 0) {
1203 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1204 if (fp) {
1205 hammer2_cluster_reconnect(hmp, fp);
1206 } else {
1207 kprintf("hammer2_mount: bad cluster_fd!\n");
1208 }
1209 }
1210 } else {
1211 spmp = hmp->spmp;
1212 if (info.hflags & HMNT2_DEVFLAGS) {
1213 kprintf("hammer2: Warning: mount flags pertaining "
1214 "to the whole device may only be specified "
1215 "on the first mount of the device: %08x\n",
1216 info.hflags & HMNT2_DEVFLAGS);
1217 }
1218 }
1219
1220 /*
1221 * Force local mount (disassociate all PFSs from their clusters).
1222 * Used primarily for debugging.
1223 */
1224 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1225
1226 /*
1227 * Lookup the mount point under the media-localized super-root.
1228 * Scanning hammer2_pfslist doesn't help us because it represents
1229 * PFS cluster ids which can aggregate several named PFSs together.
1230 *
1231 * cluster->pmp will incorrectly point to spmp and must be fixed
1232 * up later on.
1233 */
1234 hammer2_inode_lock(spmp->iroot, 0);
1235 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1236 lhc = hammer2_dirhash(label, strlen(label));
1237 chain = hammer2_chain_lookup(&parent, &key_next,
1238 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1239 &cache_index, 0);
1240 while (chain) {
1241 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1242 strcmp(label, chain->data->ipdata.filename) == 0) {
1243 break;
1244 }
1245 chain = hammer2_chain_next(&parent, chain, &key_next,
1246 key_next,
1247 lhc + HAMMER2_DIRHASH_LOMASK,
1248 &cache_index, 0);
1249 }
1250 if (parent) {
1251 hammer2_chain_unlock(parent);
1252 hammer2_chain_drop(parent);
1253 }
1254 hammer2_inode_unlock(spmp->iroot);
1255
1256 /*
1257 * PFS could not be found?
1258 */
1259 if (chain == NULL) {
1260 kprintf("hammer2_mount: PFS label not found\n");
1261 hammer2_unmount_helper(mp, NULL, hmp);
1262 lockmgr(&hammer2_mntlk, LK_RELEASE);
1263 hammer2_vfs_unmount(mp, MNT_FORCE);
1264
1265 return EINVAL;
1266 }
1267
1268 /*
1269 * Acquire the pmp structure (it should have already been allocated
1270 * via hammer2_update_pmps() so do not pass cluster in to add to
1271 * available chains).
1272 *
1273 * Check if the cluster has already been mounted. A cluster can
1274 * only be mounted once, use null mounts to mount additional copies.
1275 */
1276 ripdata = &chain->data->ipdata;
1277 bref = chain->bref;
1278 pmp = hammer2_pfsalloc(NULL, ripdata,
1279 bref.modify_tid, force_local);
1280 hammer2_chain_unlock(chain);
1281 hammer2_chain_drop(chain);
1282
1283 /*
1284 * Finish the mount
1285 */
1286 kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1287
1288 if (pmp->mp) {
1289 kprintf("hammer2_mount: PFS already mounted!\n");
1290 hammer2_unmount_helper(mp, NULL, hmp);
1291 lockmgr(&hammer2_mntlk, LK_RELEASE);
1292 hammer2_vfs_unmount(mp, MNT_FORCE);
1293
1294 return EBUSY;
1295 }
1296
1297 pmp->hflags = info.hflags;
1298 mp->mnt_flag = MNT_LOCAL;
1299 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */
1300 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */
1301
1302 /*
1303 * required mount structure initializations
1304 */
1305 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1306 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1307
1308 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1309 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1310
1311 /*
1312 * Optional fields
1313 */
1314 mp->mnt_iosize_max = MAXPHYS;
1315
1316 /*
1317 * Connect up mount pointers.
1318 */
1319 hammer2_mount_helper(mp, pmp);
1320
1321 lockmgr(&hammer2_mntlk, LK_RELEASE);
1322
1323 /*
1324 * Finish setup
1325 */
1326 vfs_getnewfsid(mp);
1327 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1328 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1329 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1330
1331 if (path) {
1332 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1333 MNAMELEN - 1, &size);
1334 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1335 } /* else root mount, already in there */
1336
1337 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1338 if (path) {
1339 copyinstr(path, mp->mnt_stat.f_mntonname,
1340 sizeof(mp->mnt_stat.f_mntonname) - 1,
1341 &size);
1342 } else {
1343 /* root mount */
1344 mp->mnt_stat.f_mntonname[0] = '/';
1345 }
1346
1347 /*
1348 * Initial statfs to prime mnt_stat.
1349 */
1350 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1351
1352 return 0;
1353 }
1354
1355 /*
1356 * Scan PFSs under the super-root and create hammer2_pfs structures.
1357 */
1358 static
1359 void
1360 hammer2_update_pmps(hammer2_dev_t *hmp)
1361 {
1362 const hammer2_inode_data_t *ripdata;
1363 hammer2_chain_t *parent;
1364 hammer2_chain_t *chain;
1365 hammer2_blockref_t bref;
1366 hammer2_dev_t *force_local;
1367 hammer2_pfs_t *spmp;
1368 hammer2_pfs_t *pmp;
1369 hammer2_key_t key_next;
1370 int cache_index = -1;
1371
1372 /*
1373 * Force local mount (disassociate all PFSs from their clusters).
1374 * Used primarily for debugging.
1375 */
1376 force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1377
1378 /*
1379 * Lookup mount point under the media-localized super-root.
1380 *
1381 * cluster->pmp will incorrectly point to spmp and must be fixed
1382 * up later on.
1383 */
1384 spmp = hmp->spmp;
1385 hammer2_inode_lock(spmp->iroot, 0);
1386 parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1387 chain = hammer2_chain_lookup(&parent, &key_next,
1388 HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1389 &cache_index, 0);
1390 while (chain) {
1391 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1392 continue;
1393 ripdata = &chain->data->ipdata;
1394 bref = chain->bref;
1395 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1396
1397 pmp = hammer2_pfsalloc(chain, ripdata,
1398 bref.modify_tid, force_local);
1399 chain = hammer2_chain_next(&parent, chain, &key_next,
1400 key_next, HAMMER2_KEY_MAX,
1401 &cache_index, 0);
1402 }
1403 if (parent) {
1404 hammer2_chain_unlock(parent);
1405 hammer2_chain_drop(parent);
1406 }
1407 hammer2_inode_unlock(spmp->iroot);
1408 }
1409
1410 static
1411 int
1412 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1413 struct vnode *devvp, struct ucred *cred)
1414 {
1415 int error;
1416
1417 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1418 error = hammer2_recovery(hmp);
1419 } else {
1420 error = 0;
1421 }
1422 return error;
1423 }
1424
1425 static
1426 int
1427 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1428 {
1429 hammer2_pfs_t *pmp;
1430 int flags;
1431 int error = 0;
1432
1433 pmp = MPTOPMP(mp);
1434
1435 if (pmp == NULL)
1436 return(0);
1437
1438 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1439
1440 /*
1441 * If mount initialization proceeded far enough we must flush
1442 * its vnodes and sync the underlying mount points. Three syncs
1443 * are required to fully flush the filesystem (freemap updates lag
1444 * by one flush, and one extra for safety).
1445 */
1446 if (mntflags & MNT_FORCE)
1447 flags = FORCECLOSE;
1448 else
1449 flags = 0;
1450 if (pmp->iroot) {
1451 error = vflush(mp, 0, flags);
1452 if (error)
1453 goto failed;
1454 hammer2_vfs_sync(mp, MNT_WAIT);
1455 hammer2_vfs_sync(mp, MNT_WAIT);
1456 hammer2_vfs_sync(mp, MNT_WAIT);
1457 }
1458
1459 /*
1460 * Cleanup the frontend support XOPS threads
1461 */
1462 hammer2_xop_helper_cleanup(pmp);
1463
1464 if (pmp->mp)
1465 hammer2_unmount_helper(mp, pmp, NULL);
1466
1467 error = 0;
1468 failed:
1469 lockmgr(&hammer2_mntlk, LK_RELEASE);
1470
1471 return (error);
1472 }
1473
1474 /*
1475 * Mount helper, hook the system mount into our PFS.
1476 * The mount lock is held.
1477 *
1478 * We must bump the mount_count on related devices for any
1479 * mounted PFSs.
1480 */
1481 static
1482 void
1483 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1484 {
1485 hammer2_cluster_t *cluster;
1486 hammer2_chain_t *rchain;
1487 int i;
1488
1489 mp->mnt_data = (qaddr_t)pmp;
1490 pmp->mp = mp;
1491
1492 /*
1493 * After pmp->mp is set we have to adjust hmp->mount_count.
1494 */
1495 cluster = &pmp->iroot->cluster;
1496 for (i = 0; i < cluster->nchains; ++i) {
1497 rchain = cluster->array[i].chain;
1498 if (rchain == NULL)
1499 continue;
1500 ++rchain->hmp->mount_count;
1501 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1502 rchain->hmp, rchain->hmp->mount_count);
1503 }
1504
1505 /*
1506 * Create missing Xop threads
1507 */
1508 hammer2_xop_helper_create(pmp);
1509 }
1510
1511 /*
1512 * Mount helper, unhook the system mount from our PFS.
1513 * The mount lock is held.
1514 *
1515 * If hmp is supplied a mount responsible for being the first to open
1516 * the block device failed and the block device and all PFSs using the
1517 * block device must be cleaned up.
1518 *
1519 * If pmp is supplied multiple devices might be backing the PFS and each
1520 * must be disconnected. This might not be the last PFS using some of the
1521 * underlying devices. Also, we have to adjust our hmp->mount_count
1522 * accounting for the devices backing the pmp which is now undergoing an
1523 * unmount.
1524 */
1525 static
1526 void
1527 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1528 {
1529 hammer2_cluster_t *cluster;
1530 hammer2_chain_t *rchain;
1531 struct vnode *devvp;
1532 int dumpcnt;
1533 int ronly = 0;
1534 int i;
1535
1536 /*
1537 * If no device supplied this is a high-level unmount and we have to
1538 * to disconnect the mount, adjust mount_count, and locate devices
1539 * that might now have no mounts.
1540 */
1541 if (pmp) {
1542 KKASSERT(hmp == NULL);
1543 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1544 pmp->mp = NULL;
1545 mp->mnt_data = NULL;
1546
1547 /*
1548 * After pmp->mp is cleared we have to account for
1549 * mount_count.
1550 */
1551 cluster = &pmp->iroot->cluster;
1552 for (i = 0; i < cluster->nchains; ++i) {
1553 rchain = cluster->array[i].chain;
1554 if (rchain == NULL)
1555 continue;
1556 --rchain->hmp->mount_count;
1557 kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1558 rchain->hmp, rchain->hmp->mount_count);
1559 /* scrapping hmp now may invalidate the pmp */
1560 }
1561 again:
1562 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1563 if (hmp->mount_count == 0) {
1564 hammer2_unmount_helper(NULL, NULL, hmp);
1565 goto again;
1566 }
1567 }
1568 return;
1569 }
1570
1571 /*
1572 * Try to terminate the block device. We can't terminate it if
1573 * there are still PFSs referencing it.
1574 */
1575 kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1576 hmp, hmp->mount_count);
1577 if (hmp->mount_count)
1578 return;
1579
1580 hammer2_pfsfree_scan(hmp);
1581 hammer2_dev_exlock(hmp); /* XXX order */
1582
1583 /*
1584 * Cycle the volume data lock as a safety (probably not needed any
1585 * more). To ensure everything is out we need to flush at least
1586 * three times. (1) The running of the sideq can dirty the
1587 * filesystem, (2) A normal flush can dirty the freemap, and
1588 * (3) ensure that the freemap is fully synchronized.
1589 *
1590 * The next mount's recovery scan can clean everything up but we want
1591 * to leave the filesystem in a 100% clean state on a normal unmount.
1592 */
1593 #if 0
1594 hammer2_voldata_lock(hmp);
1595 hammer2_voldata_unlock(hmp);
1596 #endif
1597 hammer2_iocom_uninit(hmp);
1598
1599 if ((hmp->vchain.flags | hmp->fchain.flags) &
1600 HAMMER2_CHAIN_FLUSH_MASK) {
1601 kprintf("hammer2_unmount: chains left over "
1602 "after final sync\n");
1603 kprintf(" vchain %08x\n", hmp->vchain.flags);
1604 kprintf(" fchain %08x\n", hmp->fchain.flags);
1605
1606 if (hammer2_debug & 0x0010)
1607 Debugger("entered debugger");
1608 }
1609
1610 KKASSERT(hmp->spmp == NULL);
1611
1612 /*
1613 * Finish up with the device vnode
1614 */
1615 if ((devvp = hmp->devvp) != NULL) {
1616 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1617 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1618 hmp->devvp = NULL;
1619 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1620 vn_unlock(devvp);
1621 vrele(devvp);
1622 devvp = NULL;
1623 }
1624
1625 /*
1626 * Clear vchain/fchain flags that might prevent final cleanup
1627 * of these chains.
1628 */
1629 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1630 atomic_add_long(&hammer2_count_modified_chains, -1);
1631 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1632 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1633 }
1634 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1635 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1636 }
1637
1638 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1639 atomic_add_long(&hammer2_count_modified_chains, -1);
1640 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1641 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1642 }
1643 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1644 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1645 }
1646
1647 /*
1648 * Final drop of embedded freemap root chain to
1649 * clean up fchain.core (fchain structure is not
1650 * flagged ALLOCATED so it is cleaned out and then
1651 * left to rot).
1652 */
1653 hammer2_chain_drop(&hmp->fchain);
1654
1655 /*
1656 * Final drop of embedded volume root chain to clean
1657 * up vchain.core (vchain structure is not flagged
1658 * ALLOCATED so it is cleaned out and then left to
1659 * rot).
1660 */
1661 dumpcnt = 50;
1662 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1663 dumpcnt = 50;
1664 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1665 hammer2_dev_unlock(hmp);
1666 hammer2_chain_drop(&hmp->vchain);
1667
1668 hammer2_io_cleanup(hmp, &hmp->iotree);
1669 if (hmp->iofree_count) {
1670 kprintf("io_cleanup: %d I/O's left hanging\n",
1671 hmp->iofree_count);
1672 }
1673
1674 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1675 kmalloc_destroy(&hmp->mchain);
1676 kfree(hmp, M_HAMMER2);
1677 }
1678
1679 int
1680 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1681 ino_t ino, struct vnode **vpp)
1682 {
1683 hammer2_xop_lookup_t *xop;
1684 hammer2_pfs_t *pmp;
1685 hammer2_inode_t *ip;
1686 hammer2_tid_t inum;
1687 int error;
1688
1689 inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1690
1691 error = 0;
1692 pmp = MPTOPMP(mp);
1693
1694 /*
1695 * Easy if we already have it cached
1696 */
1697 ip = hammer2_inode_lookup(pmp, inum);
1698 if (ip) {
1699 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1700 *vpp = hammer2_igetv(ip, &error);
1701 hammer2_inode_unlock(ip);
1702 hammer2_inode_drop(ip); /* from lookup */
1703
1704 return error;
1705 }
1706
1707 /*
1708 * Otherwise we have to find the inode
1709 */
1710 xop = hammer2_xop_alloc(pmp->iroot, 0);
1711 xop->lhc = inum;
1712 hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1713 error = hammer2_xop_collect(&xop->head, 0);
1714
1715 if (error == 0) {
1716 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1717 kprintf("vget: no collect error but also no rdata\n");
1718 kprintf("xop %p\n", xop);
1719 while ((hammer2_debug & 0x80000) == 0) {
1720 tsleep(xop, PCATCH, "wait", hz * 10);
1721 }
1722 ip = NULL;
1723 } else {
1724 ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1725 }
1726 }
1727 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1728
1729 if (ip) {
1730 *vpp = hammer2_igetv(ip, &error);
1731 hammer2_inode_unlock(ip);
1732 } else {
1733 *vpp = NULL;
1734 error = ENOENT;
1735 }
1736 return (error);
1737 }
1738
1739 static
1740 int
1741 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1742 {
1743 hammer2_pfs_t *pmp;
1744 struct vnode *vp;
1745 int error;
1746
1747 pmp = MPTOPMP(mp);
1748 if (pmp->iroot == NULL) {
1749 *vpp = NULL;
1750 return EINVAL;
1751 }
1752
1753 error = 0;
1754 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1755
1756 while (pmp->inode_tid == 0) {
1757 hammer2_xop_ipcluster_t *xop;
1758 hammer2_inode_meta_t *meta;
1759
1760 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1761 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1762 error = hammer2_xop_collect(&xop->head, 0);
1763
1764 if (error == 0) {
1765 meta = &xop->head.cluster.focus->data->ipdata.meta;
1766 pmp->iroot->meta = *meta;
1767 pmp->inode_tid = meta->pfs_inum + 1;
1768 if (pmp->inode_tid < HAMMER2_INODE_START)
1769 pmp->inode_tid = HAMMER2_INODE_START;
1770 pmp->modify_tid =
1771 xop->head.cluster.focus->bref.modify_tid + 1;
1772 kprintf("PFS: Starting inode %jd\n",
1773 (intmax_t)pmp->inode_tid);
1774 kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1775 pmp->inode_tid, pmp->modify_tid);
1776 wakeup(&pmp->iroot);
1777
1778 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1779
1780 /*
1781 * Prime the mount info.
1782 */
1783 hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1784 break;
1785 }
1786
1787 /*
1788 * Loop, try again
1789 */
1790 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1791 hammer2_inode_unlock(pmp->iroot);
1792 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1793 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1794 if (error == EINTR)
1795 break;
1796 }
1797
1798 if (error) {
1799 hammer2_inode_unlock(pmp->iroot);
1800 *vpp = NULL;
1801 } else {
1802 vp = hammer2_igetv(pmp->iroot, &error);
1803 hammer2_inode_unlock(pmp->iroot);
1804 *vpp = vp;
1805 }
1806
1807 return (error);
1808 }
1809
1810 /*
1811 * Filesystem status
1812 *
1813 * XXX incorporate ipdata->meta.inode_quota and data_quota
1814 */
1815 static
1816 int
1817 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1818 {
1819 hammer2_pfs_t *pmp;
1820 hammer2_dev_t *hmp;
1821 hammer2_blockref_t bref;
1822 int i;
1823
1824 /*
1825 * NOTE: iroot might not have validated the cluster yet.
1826 */
1827 pmp = MPTOPMP(mp);
1828
1829 mp->mnt_stat.f_files = 0;
1830 mp->mnt_stat.f_ffree = 0;
1831 mp->mnt_stat.f_blocks = 0;
1832 mp->mnt_stat.f_bfree = 0;
1833 mp->mnt_stat.f_bavail = 0;
1834
1835 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1836 hmp = pmp->pfs_hmps[i];
1837 if (hmp == NULL)
1838 continue;
1839 if (pmp->iroot->cluster.array[i].chain)
1840 bref = pmp->iroot->cluster.array[i].chain->bref;
1841 else
1842 bzero(&bref, sizeof(bref));
1843
1844 mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1845 mp->mnt_stat.f_ffree = 0;
1846 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1847 mp->mnt_vstat.f_bsize;
1848 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1849 mp->mnt_vstat.f_bsize;
1850 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1851
1852 *sbp = mp->mnt_stat;
1853 }
1854 return (0);
1855 }
1856
1857 static
1858 int
1859 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1860 {
1861 hammer2_pfs_t *pmp;
1862 hammer2_dev_t *hmp;
1863 hammer2_blockref_t bref;
1864 int i;
1865
1866 /*
1867 * NOTE: iroot might not have validated the cluster yet.
1868 */
1869 pmp = MPTOPMP(mp);
1870
1871 mp->mnt_vstat.f_bsize = 0;
1872 mp->mnt_vstat.f_files = 0;
1873 mp->mnt_vstat.f_ffree = 0;
1874 mp->mnt_vstat.f_blocks = 0;
1875 mp->mnt_vstat.f_bfree = 0;
1876 mp->mnt_vstat.f_bavail = 0;
1877
1878 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1879 hmp = pmp->pfs_hmps[i];
1880 if (hmp == NULL)
1881 continue;
1882 if (pmp->iroot->cluster.array[i].chain)
1883 bref = pmp->iroot->cluster.array[i].chain->bref;
1884 else
1885 bzero(&bref, sizeof(bref));
1886
1887 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1888 mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
1889 mp->mnt_vstat.f_ffree = 0;
1890 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1891 mp->mnt_vstat.f_bsize;
1892 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1893 mp->mnt_vstat.f_bsize;
1894 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1895
1896 *sbp = mp->mnt_vstat;
1897 }
1898 return (0);
1899 }
1900
1901 /*
1902 * Mount-time recovery (RW mounts)
1903 *
1904 * Updates to the free block table are allowed to lag flushes by one
1905 * transaction. In case of a crash, then on a fresh mount we must do an
1906 * incremental scan of the last committed transaction id and make sure that
1907 * all related blocks have been marked allocated.
1908 *
1909 * The super-root topology and each PFS has its own transaction id domain,
1910 * so we must track PFS boundary transitions.
1911 */
1912 struct hammer2_recovery_elm {
1913 TAILQ_ENTRY(hammer2_recovery_elm) entry;
1914 hammer2_chain_t *chain;
1915 hammer2_tid_t sync_tid;
1916 };
1917
1918 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1919
1920 struct hammer2_recovery_info {
1921 struct hammer2_recovery_list list;
1922 hammer2_tid_t mtid;
1923 int depth;
1924 };
1925
1926 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1927 hammer2_chain_t *parent,
1928 struct hammer2_recovery_info *info,
1929 hammer2_tid_t sync_tid);
1930
1931 #define HAMMER2_RECOVERY_MAXDEPTH 10
1932
1933 static
1934 int
1935 hammer2_recovery(hammer2_dev_t *hmp)
1936 {
1937 struct hammer2_recovery_info info;
1938 struct hammer2_recovery_elm *elm;
1939 hammer2_chain_t *parent;
1940 hammer2_tid_t sync_tid;
1941 hammer2_tid_t mirror_tid;
1942 int error;
1943 int cumulative_error = 0;
1944
1945 hammer2_trans_init(hmp->spmp, 0);
1946
1947 sync_tid = hmp->voldata.freemap_tid;
1948 mirror_tid = hmp->voldata.mirror_tid;
1949
1950 kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1951 if (sync_tid >= mirror_tid) {
1952 kprintf(" no recovery needed\n");
1953 } else {
1954 kprintf(" freemap recovery %016jx-%016jx\n",
1955 sync_tid + 1, mirror_tid);
1956 }
1957
1958 TAILQ_INIT(&info.list);
1959 info.depth = 0;
1960 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1961 cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
1962 hammer2_chain_lookup_done(parent);
1963
1964 while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1965 TAILQ_REMOVE(&info.list, elm, entry);
1966 parent = elm->chain;
1967 sync_tid = elm->sync_tid;
1968 kfree(elm, M_HAMMER2);
1969
1970 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1971 error = hammer2_recovery_scan(hmp, parent, &info,
1972 hmp->voldata.freemap_tid);
1973 hammer2_chain_unlock(parent);
1974 hammer2_chain_drop(parent); /* drop elm->chain ref */
1975 if (error)
1976 cumulative_error = error;
1977 }
1978 hammer2_trans_done(hmp->spmp);
1979
1980 return cumulative_error;
1981 }
1982
1983 static
1984 int
1985 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1986 struct hammer2_recovery_info *info,
1987 hammer2_tid_t sync_tid)
1988 {
1989 const hammer2_inode_data_t *ripdata;
1990 hammer2_chain_t *chain;
1991 hammer2_blockref_t bref;
1992 int cache_index;
1993 int cumulative_error = 0;
1994 int error;
1995 int first;
1996
1997 /*
1998 * Adjust freemap to ensure that the block(s) are marked allocated.
1999 */
2000 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2001 hammer2_freemap_adjust(hmp, &parent->bref,
2002 HAMMER2_FREEMAP_DORECOVER);
2003 }
2004
2005 /*
2006 * Check type for recursive scan
2007 */
2008 switch(parent->bref.type) {
2009 case HAMMER2_BREF_TYPE_VOLUME:
2010 /* data already instantiated */
2011 break;
2012 case HAMMER2_BREF_TYPE_INODE:
2013 /*
2014 * Must instantiate data for DIRECTDATA test and also
2015 * for recursion.
2016 */
2017 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2018 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2019 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2020 /* not applicable to recovery scan */
2021 hammer2_chain_unlock(parent);
2022 return 0;
2023 }
2024 hammer2_chain_unlock(parent);
2025 break;
2026 case HAMMER2_BREF_TYPE_INDIRECT:
2027 /*
2028 * Must instantiate data for recursion
2029 */
2030 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2031 hammer2_chain_unlock(parent);
2032 break;
2033 case HAMMER2_BREF_TYPE_DIRENT:
2034 case HAMMER2_BREF_TYPE_DATA:
2035 case HAMMER2_BREF_TYPE_FREEMAP:
2036 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2037 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2038 /* not applicable to recovery scan */
2039 return 0;
2040 break;
2041 default:
2042 return EDOM;
2043 }
2044
2045 /*
2046 * Defer operation if depth limit reached or if we are crossing a
2047 * PFS boundary.
2048 */
2049 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2050 struct hammer2_recovery_elm *elm;
2051
2052 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2053 elm->chain = parent;
2054 elm->sync_tid = sync_tid;
2055 hammer2_chain_ref(parent);
2056 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2057 /* unlocked by caller */
2058
2059 return(0);
2060 }
2061
2062
2063 /*
2064 * Recursive scan of the last flushed transaction only. We are
2065 * doing this without pmp assignments so don't leave the chains
2066 * hanging around after we are done with them.
2067 */
2068 cache_index = 0;
2069 chain = NULL;
2070 first = 1;
2071
2072 while (hammer2_chain_scan(parent, &chain, &bref,
2073 &first, &cache_index,
2074 HAMMER2_LOOKUP_NODATA) != NULL) {
2075 /*
2076 * If this is a leaf
2077 */
2078 if (chain == NULL) {
2079 if (bref.mirror_tid > sync_tid) {
2080 hammer2_freemap_adjust(hmp, &bref,
2081 HAMMER2_FREEMAP_DORECOVER);
2082 }
2083 continue;
2084 }
2085
2086 /*
2087 * This may or may not be a recursive node.
2088 */
2089 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2090 if (bref.mirror_tid > sync_tid) {
2091 ++info->depth;
2092 error = hammer2_recovery_scan(hmp, chain,
2093 info, sync_tid);
2094 --info->depth;
2095 if (error)
2096 cumulative_error = error;
2097 }
2098
2099 /*
2100 * Flush the recovery at the PFS boundary to stage it for
2101 * the final flush of the super-root topology.
2102 */
2103 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2104 (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2105 hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2106 }
2107 }
2108
2109 return cumulative_error;
2110 }
2111
2112 /*
2113 * Sync a mount point; this is called on a per-mount basis from the
2114 * filesystem syncer process periodically and whenever a user issues
2115 * a sync.
2116 */
2117 int
2118 hammer2_vfs_sync(struct mount *mp, int waitfor)
2119 {
2120 hammer2_xop_flush_t *xop;
2121 struct hammer2_sync_info info;
2122 hammer2_inode_t *iroot;
2123 hammer2_pfs_t *pmp;
2124 int flags;
2125 int error;
2126
2127 pmp = MPTOPMP(mp);
2128 iroot = pmp->iroot;
2129 KKASSERT(iroot);
2130 KKASSERT(iroot->pmp == pmp);
2131
2132 /*
2133 * We can't acquire locks on existing vnodes while in a transaction
2134 * without risking a deadlock. This assumes that vfsync() can be
2135 * called without the vnode locked (which it can in DragonFly).
2136 * Otherwise we'd have to implement a multi-pass or flag the lock
2137 * failures and retry.
2138 *
2139 * The reclamation code interlocks with the sync list's token
2140 * (by removing the vnode from the scan list) before unlocking
2141 * the inode, giving us time to ref the inode.
2142 */
2143 /*flags = VMSC_GETVP;*/
2144 flags = 0;
2145 if (waitfor & MNT_LAZY)
2146 flags |= VMSC_ONEPASS;
2147
2148 /*
2149 * Preflush the vnodes using a normal transaction before interlocking
2150 * with a flush transaction. We do this to try to run as much of
2151 * the compression as possible outside the flush transaction.
2152 *
2153 * For efficiency do an async pass before making sure with a
2154 * synchronous pass on all related buffer cache buffers.
2155 */
2156 hammer2_trans_init(pmp, 0);
2157 info.error = 0;
2158 info.waitfor = MNT_NOWAIT;
2159 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2160 info.waitfor = MNT_WAIT;
2161 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2162 hammer2_trans_done(pmp);
2163
2164 /*
2165 * Start our flush transaction. This does not return until all
2166 * concurrent transactions have completed and will prevent any
2167 * new transactions from running concurrently, except for the
2168 * buffer cache transactions.
2169 *
2170 * NOTE! It is still possible for the paging code to push pages
2171 * out via a UIO_NOCOPY hammer2_vop_write() during the main
2172 * flush.
2173 */
2174 hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2175 hammer2_inode_run_sideq(pmp);
2176
2177 info.error = 0;
2178 info.waitfor = MNT_NOWAIT;
2179 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2180 info.waitfor = MNT_WAIT;
2181 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2182 hammer2_bioq_sync(pmp);
2183
2184 /*
2185 * Use the XOP interface to concurrently flush all nodes to
2186 * synchronize the PFSROOT subtopology to the media. A standard
2187 * end-of-scan ENOENT error indicates cluster sufficiency.
2188 *
2189 * Note that this flush will not be visible on crash recovery until
2190 * we flush the super-root topology in the next loop.
2191 *
2192 * XXX For now wait for all flushes to complete.
2193 */
2194 if (iroot) {
2195 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2196 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2197 error = hammer2_xop_collect(&xop->head,
2198 HAMMER2_XOP_COLLECT_WAITALL);
2199 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2200 if (error == ENOENT)
2201 error = 0;
2202 } else {
2203 error = 0;
2204 }
2205 hammer2_trans_done(pmp);
2206
2207 return (error);
2208 }
2209
2210 /*
2211 * Sync passes.
2212 *
2213 * Note that we ignore the tranasction mtid we got above. Instead,
2214 * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2215 * transactions.
2216 */
2217 static int
2218 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2219 {
2220 struct hammer2_sync_info *info = data;
2221 hammer2_inode_t *ip;
2222 int error;
2223
2224 /*
2225 * Degenerate cases. Note that ip == NULL typically means the
2226 * syncer vnode itself and we don't want to vclrisdirty() in that
2227 * situation.
2228 */
2229 ip = VTOI(vp);
2230 if (ip == NULL) {
2231 return(0);
2232 }
2233 if (vp->v_type == VNON || vp->v_type == VBAD) {
2234 vclrisdirty(vp);
2235 return(0);
2236 }
2237
2238 /*
2239 * VOP_FSYNC will start a new transaction so replicate some code
2240 * here to do it inline (see hammer2_vop_fsync()).
2241 *
2242 * WARNING: The vfsync interacts with the buffer cache and might
2243 * block, we can't hold the inode lock at that time.
2244 * However, we MUST ref ip before blocking to ensure that
2245 * it isn't ripped out from under us (since we do not
2246 * hold a lock on the vnode).
2247 */
2248 hammer2_inode_ref(ip);
2249 if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2250 !RB_EMPTY(&vp->v_rbdirty_tree)) {
2251 vfsync(vp, info->waitfor, 1, NULL, NULL);
2252 if (ip->flags & (HAMMER2_INODE_RESIZED |
2253 HAMMER2_INODE_MODIFIED)) {
2254 hammer2_inode_lock(ip, 0);
2255 hammer2_inode_chain_sync(ip);
2256 hammer2_inode_unlock(ip);
2257 }
2258 }
2259 if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2260 RB_EMPTY(&vp->v_rbdirty_tree)) {
2261 vclrisdirty(vp);
2262 }
2263
2264 hammer2_inode_drop(ip);
2265 #if 1
2266 error = 0;
2267 if (error)
2268 info->error = error;
2269 #endif
2270 return(0);
2271 }
2272
2273 static
2274 int
2275 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2276 {
2277 hammer2_inode_t *ip;
2278
2279 KKASSERT(MAXFIDSZ >= 16);
2280 ip = VTOI(vp);
2281 fhp->fid_len = offsetof(struct fid, fid_data[16]);
2282 fhp->fid_ext = 0;
2283 ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2284 ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2285
2286 return 0;
2287 }
2288
2289 static
2290 int
2291 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2292 struct fid *fhp, struct vnode **vpp)
2293 {
2294 hammer2_pfs_t *pmp;
2295 hammer2_tid_t inum;
2296 int error;
2297
2298 pmp = MPTOPMP(mp);
2299 inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2300 if (vpp) {
2301 if (inum == 1)
2302 error = hammer2_vfs_root(mp, vpp);
2303 else
2304 error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2305 } else {
2306 error = 0;
2307 }
2308 if (error)
2309 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2310 return error;
2311 }
2312
2313 static
2314 int
2315 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2316 int *exflagsp, struct ucred **credanonp)
2317 {
2318 hammer2_pfs_t *pmp;
2319 struct netcred *np;
2320 int error;
2321
2322 pmp = MPTOPMP(mp);
2323 np = vfs_export_lookup(mp, &pmp->export, nam);
2324 if (np) {
2325 *exflagsp = np->netc_exflags;
2326 *credanonp = &np->netc_anon;
2327 error = 0;
2328 } else {
2329 error = EACCES;
2330 }
2331 return error;
2332 }
2333
2334 /*
2335 * Support code for hammer2_vfs_mount(). Read, verify, and install the volume
2336 * header into the HMP
2337 *
2338 * XXX read four volhdrs and use the one with the highest TID whos CRC
2339 * matches.
2340 *
2341 * XXX check iCRCs.
2342 *
2343 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2344 * nonexistant locations.
2345 *
2346 * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2347 */
2348 static
2349 int
2350 hammer2_install_volume_header(hammer2_dev_t *hmp)
2351 {
2352 hammer2_volume_data_t *vd;
2353 struct buf *bp;
2354 hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2355 int error_reported;
2356 int error;
2357 int valid;
2358 int i;
2359
2360 error_reported = 0;
2361 error = 0;
2362 valid = 0;
2363 bp = NULL;
2364
2365 /*
2366 * There are up to 4 copies of the volume header (syncs iterate
2367 * between them so there is no single master). We don't trust the
2368 * volu_size field so we don't know precisely how large the filesystem
2369 * is, so depend on the OS to return an error if we go beyond the
2370 * block device's EOF.
2371 */
2372 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2373 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2374 HAMMER2_VOLUME_BYTES, &bp);
2375 if (error) {
2376 brelse(bp);
2377 bp = NULL;
2378 continue;
2379 }
2380
2381 vd = (struct hammer2_volume_data *) bp->b_data;
2382 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2383 (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2384 brelse(bp);
2385 bp = NULL;
2386 continue;
2387 }
2388
2389 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2390 /* XXX: Reversed-endianness filesystem */
2391 kprintf("hammer2: reverse-endian filesystem detected");
2392 brelse(bp);
2393 bp = NULL;
2394 continue;
2395 }
2396
2397 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2398 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2399 HAMMER2_VOLUME_ICRC0_SIZE);
2400 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2401 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2402 HAMMER2_VOLUME_ICRC1_SIZE);
2403 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2404 kprintf("hammer2 volume header crc "
2405 "mismatch copy #%d %08x/%08x\n",
2406 i, crc0, crc);
2407 error_reported = 1;
2408 brelse(bp);
2409 bp = NULL;
2410 continue;
2411 }
2412 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2413 valid = 1;
2414 hmp->voldata = *vd;
2415 hmp->volhdrno = i;
2416 }
2417 brelse(bp);
2418 bp = NULL;
2419 }
2420 if (valid) {
2421 hmp->volsync = hmp->voldata;
2422 error = 0;
2423 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2424 kprintf("hammer2: using volume header #%d\n",
2425 hmp->volhdrno);
2426 }
2427 } else {
2428 error = EINVAL;
2429 kprintf("hammer2: no valid volume headers found!\n");
2430 }
2431 return (error);
2432 }
2433
2434 /*
2435 * This handles hysteresis on regular file flushes. Because the BIOs are
2436 * routed to a thread it is possible for an excessive number to build up
2437 * and cause long front-end stalls long before the runningbuffspace limit
2438 * is hit, so we implement hammer2_flush_pipe to control the
2439 * hysteresis.
2440 *
2441 * This is a particular problem when compression is used.
2442 */
2443 void
2444 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2445 {
2446 atomic_add_int(&pmp->count_lwinprog, 1);
2447 }
2448
2449 void
2450 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2451 {
2452 int lwinprog;
2453
2454 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2455 if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2456 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2457 atomic_clear_int(&pmp->count_lwinprog,
2458 HAMMER2_LWINPROG_WAITING);
2459 wakeup(&pmp->count_lwinprog);
2460 }
2461 if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2462 (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2463 atomic_clear_int(&pmp->count_lwinprog,
2464 HAMMER2_LWINPROG_WAITING0);
2465 wakeup(&pmp->count_lwinprog);
2466 }
2467 }
2468
2469 void
2470 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2471 {
2472 int lwinprog;
2473 int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2474 HAMMER2_LWINPROG_WAITING0;
2475
2476 for (;;) {
2477 lwinprog = pmp->count_lwinprog;
2478 cpu_ccfence();
2479 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2480 break;
2481 tsleep_interlock(&pmp->count_lwinprog, 0);
2482 atomic_set_int(&pmp->count_lwinprog, lwflag);
2483 lwinprog = pmp->count_lwinprog;
2484 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2485 break;
2486 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2487 }
2488 }
2489
2490 /*
2491 * Manage excessive memory resource use for chain and related
2492 * structures.
2493 */
2494 void
2495 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2496 {
2497 uint32_t waiting;
2498 uint32_t count;
2499 uint32_t limit;
2500 #if 0
2501 static int zzticks;
2502 #endif
2503
2504 /*
2505 * Atomic check condition and wait. Also do an early speedup of
2506 * the syncer to try to avoid hitting the wait.
2507 */
2508 for (;;) {
2509 waiting = pmp->inmem_dirty_chains;
2510 cpu_ccfence();
2511 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2512
2513 limit = pmp->mp->mnt_nvnodelistsize / 10;
2514 if (limit < hammer2_limit_dirty_chains)
2515 limit = hammer2_limit_dirty_chains;
2516 if (limit < 1000)
2517 limit = 1000;
2518
2519 #if 0
2520 if ((int)(ticks - zzticks) > hz) {
2521 zzticks = ticks;
2522 kprintf("count %ld %ld\n", count, limit);
2523 }
2524 #endif
2525
2526 /*
2527 * Block if there are too many dirty chains present, wait
2528 * for the flush to clean some out.
2529 */
2530 if (count > limit) {
2531 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2532 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2533 waiting,
2534 waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2535 speedup_syncer(pmp->mp);
2536 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2537 "chnmem", hz);
2538 }
2539 continue; /* loop on success or fail */
2540 }
2541
2542 /*
2543 * Try to start an early flush before we are forced to block.
2544 */
2545 if (count > limit * 7 / 10)
2546 speedup_syncer(pmp->mp);
2547 break;
2548 }
2549 }
2550
2551 void
2552 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2553 {
2554 if (pmp) {
2555 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2556 }
2557 }
2558
2559 void
2560 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2561 {
2562 uint32_t waiting;
2563
2564 if (pmp == NULL)
2565 return;
2566
2567 for (;;) {
2568 waiting = pmp->inmem_dirty_chains;
2569 cpu_ccfence();
2570 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2571 waiting,
2572 (waiting - 1) &
2573 ~HAMMER2_DIRTYCHAIN_WAITING)) {
2574 break;
2575 }
2576 }
2577
2578 if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2579 wakeup(&pmp->inmem_dirty_chains);
2580 }
2581
2582 /*
2583 * Debugging
2584 */
2585 void
2586 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2587 {
2588 hammer2_chain_t *scan;
2589 hammer2_chain_t *parent;
2590
2591 --*countp;
2592 if (*countp == 0) {
2593 kprintf("%*.*s...\n", tab, tab, "");
2594 return;
2595 }
2596 if (*countp < 0)
2597 return;
2598 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2599 tab, tab, "", pfx,
2600 chain, chain->bref.type,
2601 chain->bref.key, chain->bref.keybits,
2602 chain->bref.mirror_tid);
2603
2604 kprintf("%*.*s [%08x] (%s) refs=%d",
2605 tab, tab, "",
2606 chain->flags,
2607 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2608 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
2609 chain->refs);
2610
2611 parent = chain->parent;
2612 if (parent)
2613 kprintf("\n%*.*s p=%p [pflags %08x prefs %d",
2614 tab, tab, "",
2615 parent, parent->flags, parent->refs);
2616 if (RB_EMPTY(&chain->core.rbtree)) {
2617 kprintf("\n");
2618 } else {
2619 kprintf(" {\n");
2620 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2621 hammer2_dump_chain(scan, tab + 4, countp, 'a');
2622 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2623 kprintf("%*.*s}(%s)\n", tab, tab, "",
2624 chain->data->ipdata.filename);
2625 else
2626 kprintf("%*.*s}\n", tab, tab, "");
2627 }
2628 }
DragonFly BSD