HAMMER VFS - Fix deadlock which can occur under severe filesystem pressure
[dragonfly.git] / sys / vfs / hammer / hammer_inode.c
CommitLineData
427e5fc6 1/*
b84de5af 2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
427e5fc6
MD
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
69ce4424 34 * $DragonFly: src/sys/vfs/hammer/hammer_inode.c,v 1.114 2008/09/24 00:53:51 dillon Exp $
427e5fc6
MD
35 */
36
37#include "hammer.h"
869e8f55 38#include <vm/vm_extern.h>
427e5fc6 39
af209b0f 40static int hammer_unload_inode(struct hammer_inode *ip);
5fa5c92f 41static void hammer_free_inode(hammer_inode_t ip);
7a61b85d
MD
42static void hammer_flush_inode_core(hammer_inode_t ip,
43 hammer_flush_group_t flg, int flags);
af209b0f 44static int hammer_setup_child_callback(hammer_record_t rec, void *data);
7a61b85d 45#if 0
525aad3a 46static int hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
7a61b85d 47#endif
cc0758d0 48static int hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
7a61b85d
MD
49 hammer_flush_group_t flg);
50static int hammer_setup_parent_inodes_helper(hammer_record_t record,
cc0758d0 51 int depth, hammer_flush_group_t flg);
ccf6a64d 52static void hammer_inode_wakereclaims(hammer_inode_t ip);
e98f1b96
MD
53static struct hammer_inostats *hammer_inode_inostats(hammer_mount_t hmp,
54 pid_t pid);
b84de5af 55
0832c9bb
MD
56#ifdef DEBUG_TRUNCATE
57extern struct hammer_inode *HammerTruncIp;
58#endif
59
d113fda1 60/*
5fa5c92f 61 * RB-Tree support for inode structures
43c665ae
MD
62 */
63int
64hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
65{
66 if (ip1->obj_localization < ip2->obj_localization)
67 return(-1);
68 if (ip1->obj_localization > ip2->obj_localization)
69 return(1);
70 if (ip1->obj_id < ip2->obj_id)
71 return(-1);
72 if (ip1->obj_id > ip2->obj_id)
73 return(1);
74 if (ip1->obj_asof < ip2->obj_asof)
75 return(-1);
76 if (ip1->obj_asof > ip2->obj_asof)
77 return(1);
78 return(0);
79}
80
73896937
MD
81int
82hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
83{
84 if (ip1->redo_fifo_start < ip2->redo_fifo_start)
85 return(-1);
86 if (ip1->redo_fifo_start > ip2->redo_fifo_start)
87 return(1);
88 return(0);
89}
90
43c665ae 91/*
5fa5c92f 92 * RB-Tree support for inode structures / special LOOKUP_INFO
43c665ae
MD
93 */
94static int
95hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
96{
97 if (info->obj_localization < ip->obj_localization)
98 return(-1);
99 if (info->obj_localization > ip->obj_localization)
100 return(1);
101 if (info->obj_id < ip->obj_id)
102 return(-1);
103 if (info->obj_id > ip->obj_id)
104 return(1);
105 if (info->obj_asof < ip->obj_asof)
106 return(-1);
107 if (info->obj_asof > ip->obj_asof)
108 return(1);
109 return(0);
110}
111
112/*
113 * Used by hammer_scan_inode_snapshots() to locate all of an object's
114 * snapshots. Note that the asof field is not tested, which we can get
115 * away with because it is the lowest-priority field.
116 */
117static int
118hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
119{
120 hammer_inode_info_t info = data;
121
122 if (ip->obj_localization > info->obj_localization)
123 return(1);
124 if (ip->obj_localization < info->obj_localization)
125 return(-1);
126 if (ip->obj_id > info->obj_id)
127 return(1);
128 if (ip->obj_id < info->obj_id)
129 return(-1);
130 return(0);
131}
132
5fa5c92f 133/*
842e7a70
MD
134 * Used by hammer_unload_pseudofs() to locate all inodes associated with
135 * a particular PFS.
136 */
137static int
138hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
139{
140 u_int32_t localization = *(u_int32_t *)data;
141 if (ip->obj_localization > localization)
142 return(1);
143 if (ip->obj_localization < localization)
144 return(-1);
145 return(0);
146}
147
148/*
5fa5c92f
MD
149 * RB-Tree support for pseudofs structures
150 */
151static int
152hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
153{
154 if (p1->localization < p2->localization)
155 return(-1);
156 if (p1->localization > p2->localization)
157 return(1);
158 return(0);
159}
160
161
43c665ae
MD
162RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
163RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
164 hammer_inode_info_cmp, hammer_inode_info_t);
5fa5c92f
MD
165RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
166 hammer_pfs_rb_compare, u_int32_t, localization);
43c665ae
MD
167
168/*
d113fda1
MD
169 * The kernel is not actively referencing this vnode but is still holding
170 * it cached.
b84de5af
MD
171 *
172 * This is called from the frontend.
2247fe02
MD
173 *
174 * MPALMOSTSAFE
d113fda1 175 */
427e5fc6
MD
176int
177hammer_vop_inactive(struct vop_inactive_args *ap)
178{
66325755 179 struct hammer_inode *ip = VTOI(ap->a_vp);
b0aab9b9 180 hammer_mount_t hmp;
27ea2398 181
c0ade690
MD
182 /*
183 * Degenerate case
184 */
185 if (ip == NULL) {
66325755 186 vrecycle(ap->a_vp);
c0ade690
MD
187 return(0);
188 }
189
190 /*
4a2796f3
MD
191 * If the inode no longer has visibility in the filesystem try to
192 * recycle it immediately, even if the inode is dirty. Recycling
193 * it quickly allows the system to reclaim buffer cache and VM
194 * resources which can matter a lot in a heavily loaded system.
195 *
196 * This can deadlock in vfsync() if we aren't careful.
4e97774c
MD
197 *
198 * Do not queue the inode to the flusher if we still have visibility,
199 * otherwise namespace calls such as chmod will unnecessarily generate
200 * multiple inode updates.
c0ade690 201 */
4e97774c 202 if (ip->ino_data.nlinks == 0) {
b0aab9b9
MD
203 hmp = ip->hmp;
204 lwkt_gettoken(&hmp->fs_token);
2247fe02 205 hammer_inode_unloadable_check(ip, 0);
4e97774c
MD
206 if (ip->flags & HAMMER_INODE_MODMASK)
207 hammer_flush_inode(ip, 0);
b0aab9b9 208 lwkt_reltoken(&hmp->fs_token);
4a2796f3 209 vrecycle(ap->a_vp);
4e97774c 210 }
427e5fc6
MD
211 return(0);
212}
213
d113fda1
MD
214/*
215 * Release the vnode association. This is typically (but not always)
1f07f686 216 * the last reference on the inode.
d113fda1 217 *
1f07f686
MD
218 * Once the association is lost we are on our own with regards to
219 * flushing the inode.
e0092341
MD
220 *
221 * We must interlock ip->vp so hammer_get_vnode() can avoid races.
d113fda1 222 */
427e5fc6
MD
223int
224hammer_vop_reclaim(struct vop_reclaim_args *ap)
225{
427e5fc6 226 struct hammer_inode *ip;
7bc5b8c2 227 hammer_mount_t hmp;
427e5fc6
MD
228 struct vnode *vp;
229
230 vp = ap->a_vp;
c0ade690 231
a89aec1b 232 if ((ip = vp->v_data) != NULL) {
da2da375 233 hmp = ip->hmp;
b0aab9b9 234 lwkt_gettoken(&hmp->fs_token);
e0092341 235 hammer_lock_ex(&ip->lock);
a89aec1b
MD
236 vp->v_data = NULL;
237 ip->vp = NULL;
7bc5b8c2 238
4a2796f3 239 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
9f5097dc 240 ++hammer_count_reclaiming;
da2da375 241 ++hmp->inode_reclaims;
9f5097dc
MD
242 ip->flags |= HAMMER_INODE_RECLAIM;
243 }
e0092341 244 hammer_unlock(&ip->lock);
ec4e8497 245 hammer_rel_inode(ip, 1);
b0aab9b9 246 lwkt_reltoken(&hmp->fs_token);
a89aec1b 247 }
427e5fc6
MD
248 return(0);
249}
250
251/*
66325755
MD
252 * Return a locked vnode for the specified inode. The inode must be
253 * referenced but NOT LOCKED on entry and will remain referenced on
254 * return.
b84de5af
MD
255 *
256 * Called from the frontend.
66325755
MD
257 */
258int
e8599db1 259hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
66325755 260{
9f5097dc 261 hammer_mount_t hmp;
66325755
MD
262 struct vnode *vp;
263 int error = 0;
ea434b6f 264 u_int8_t obj_type;
66325755 265
9f5097dc
MD
266 hmp = ip->hmp;
267
66325755
MD
268 for (;;) {
269 if ((vp = ip->vp) == NULL) {
9f5097dc 270 error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
66325755
MD
271 if (error)
272 break;
8cd0a023
MD
273 hammer_lock_ex(&ip->lock);
274 if (ip->vp != NULL) {
275 hammer_unlock(&ip->lock);
ee23ac7d 276 vp = *vpp;
8cd0a023
MD
277 vp->v_type = VBAD;
278 vx_put(vp);
279 continue;
66325755 280 }
8cd0a023
MD
281 hammer_ref(&ip->lock);
282 vp = *vpp;
283 ip->vp = vp;
ea434b6f
MD
284
285 obj_type = ip->ino_data.obj_type;
286 vp->v_type = hammer_get_vnode_type(obj_type);
7a04d74f 287
ccf6a64d 288 hammer_inode_wakereclaims(ip);
9f5097dc 289
11ad5ade 290 switch(ip->ino_data.obj_type) {
7a04d74f
MD
291 case HAMMER_OBJTYPE_CDEV:
292 case HAMMER_OBJTYPE_BDEV:
9f5097dc 293 vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
7a04d74f
MD
294 addaliasu(vp, ip->ino_data.rmajor,
295 ip->ino_data.rminor);
296 break;
297 case HAMMER_OBJTYPE_FIFO:
9f5097dc 298 vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
7a04d74f 299 break;
42cd5131 300 case HAMMER_OBJTYPE_REGFILE:
42cd5131 301 break;
7a04d74f
MD
302 default:
303 break;
304 }
42c7d26b
MD
305
306 /*
307 * Only mark as the root vnode if the ip is not
308 * historical, otherwise the VFS cache will get
309 * confused. The other half of the special handling
310 * is in hammer_vop_nlookupdotdot().
ddfdf542 311 *
67863d04
MD
312 * Pseudo-filesystem roots can be accessed via
313 * non-root filesystem paths and setting VROOT may
314 * confuse the namecache. Set VPFSROOT instead.
42c7d26b
MD
315 */
316 if (ip->obj_id == HAMMER_OBJID_ROOT &&
67863d04
MD
317 ip->obj_asof == hmp->asof) {
318 if (ip->obj_localization == 0)
2247fe02 319 vsetflags(vp, VROOT);
67863d04 320 else
2247fe02 321 vsetflags(vp, VPFSROOT);
42c7d26b 322 }
7a04d74f 323
8cd0a023
MD
324 vp->v_data = (void *)ip;
325 /* vnode locked by getnewvnode() */
326 /* make related vnode dirty if inode dirty? */
327 hammer_unlock(&ip->lock);
b0d18f7d
MD
328 if (vp->v_type == VREG) {
329 vinitvmio(vp, ip->ino_data.size,
330 hammer_blocksize(ip->ino_data.size),
331 hammer_blockoff(ip->ino_data.size));
332 }
8cd0a023
MD
333 break;
334 }
335
336 /*
e0092341
MD
337 * Interlock vnode clearing. This does not prevent the
338 * vnode from going into a reclaimed state but it does
339 * prevent it from being destroyed or reused so the vget()
340 * will properly fail.
341 */
342 hammer_lock_ex(&ip->lock);
343 if ((vp = ip->vp) == NULL) {
344 hammer_unlock(&ip->lock);
345 continue;
346 }
347 vhold_interlocked(vp);
348 hammer_unlock(&ip->lock);
349
350 /*
8cd0a023
MD
351 * loop if the vget fails (aka races), or if the vp
352 * no longer matches ip->vp.
353 */
354 if (vget(vp, LK_EXCLUSIVE) == 0) {
e0092341
MD
355 if (vp == ip->vp) {
356 vdrop(vp);
8cd0a023 357 break;
e0092341 358 }
8cd0a023 359 vput(vp);
66325755 360 }
e0092341 361 vdrop(vp);
66325755 362 }
a89aec1b 363 *vpp = vp;
66325755
MD
364 return(error);
365}
366
367/*
43c665ae
MD
368 * Locate all copies of the inode for obj_id compatible with the specified
369 * asof, reference, and issue the related call-back. This routine is used
370 * for direct-io invalidation and does not create any new inodes.
371 */
372void
373hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
374 int (*callback)(hammer_inode_t ip, void *data),
375 void *data)
376{
377 hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
378 hammer_inode_info_cmp_all_history,
379 callback, iinfo);
380}
381
382/*
8cd0a023
MD
383 * Acquire a HAMMER inode. The returned inode is not locked. These functions
384 * do not attach or detach the related vnode (use hammer_get_vnode() for
385 * that).
d113fda1
MD
386 *
387 * The flags argument is only applied for newly created inodes, and only
388 * certain flags are inherited.
b84de5af
MD
389 *
390 * Called from the frontend.
66325755
MD
391 */
392struct hammer_inode *
bcac4bbb 393hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
adf01747 394 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
ddfdf542 395 int flags, int *errorp)
66325755 396{
36f82b23 397 hammer_mount_t hmp = trans->hmp;
39d8fd63 398 struct hammer_node_cache *cachep;
427e5fc6 399 struct hammer_inode_info iinfo;
8cd0a023 400 struct hammer_cursor cursor;
427e5fc6 401 struct hammer_inode *ip;
427e5fc6 402
5fa5c92f 403
427e5fc6
MD
404 /*
405 * Determine if we already have an inode cached. If we do then
406 * we are golden.
82010f9f
MD
407 *
408 * If we find an inode with no vnode we have to mark the
409 * transaction such that hammer_inode_waitreclaims() is
410 * called later on to avoid building up an infinite number
411 * of inodes. Otherwise we can continue to * add new inodes
412 * faster then they can be disposed of, even with the tsleep
413 * delay.
4c286c36
MD
414 *
415 * If we find a dummy inode we return a failure so dounlink
416 * (which does another lookup) doesn't try to mess with the
417 * link count. hammer_vop_nresolve() uses hammer_get_dummy_inode()
418 * to ref dummy inodes.
427e5fc6 419 */
66325755 420 iinfo.obj_id = obj_id;
7f7c1f84 421 iinfo.obj_asof = asof;
ddfdf542 422 iinfo.obj_localization = localization;
427e5fc6
MD
423loop:
424 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
425 if (ip) {
4c286c36
MD
426 if (ip->flags & HAMMER_INODE_DUMMY) {
427 *errorp = ENOENT;
428 return(NULL);
429 }
8cd0a023 430 hammer_ref(&ip->lock);
66325755
MD
431 *errorp = 0;
432 return(ip);
427e5fc6
MD
433 }
434
3897d7e9
MD
435 /*
436 * Allocate a new inode structure and deal with races later.
437 */
bac808fe 438 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
b3deaf57 439 ++hammer_count_inodes;
9f5097dc 440 ++hmp->count_inodes;
66325755 441 ip->obj_id = obj_id;
27ea2398 442 ip->obj_asof = iinfo.obj_asof;
ddfdf542 443 ip->obj_localization = localization;
66325755 444 ip->hmp = hmp;
d113fda1 445 ip->flags = flags & HAMMER_INODE_RO;
bcac4bbb
MD
446 ip->cache[0].ip = ip;
447 ip->cache[1].ip = ip;
39d8fd63
MD
448 ip->cache[2].ip = ip;
449 ip->cache[3].ip = ip;
5fa5c92f 450 if (hmp->ronly)
d113fda1 451 ip->flags |= HAMMER_INODE_RO;
a9d52b76
MD
452 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
453 0x7FFFFFFFFFFFFFFFLL;
8cd0a023 454 RB_INIT(&ip->rec_tree);
1f07f686 455 TAILQ_INIT(&ip->target_list);
5fa5c92f 456 hammer_ref(&ip->lock);
427e5fc6
MD
457
458 /*
ea434b6f
MD
459 * Locate the on-disk inode. If this is a PFS root we always
460 * access the current version of the root inode and (if it is not
461 * a master) always access information under it with a snapshot
462 * TID.
39d8fd63
MD
463 *
464 * We cache recent inode lookups in this directory in dip->cache[2].
465 * If we can't find it we assume the inode we are looking for is
466 * close to the directory inode.
427e5fc6 467 */
6a37e7e4 468retry:
39d8fd63
MD
469 cachep = NULL;
470 if (dip) {
471 if (dip->cache[2].node)
472 cachep = &dip->cache[2];
473 else
474 cachep = &dip->cache[0];
475 }
476 hammer_init_cursor(trans, &cursor, cachep, NULL);
5a930e66 477 cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
8cd0a023
MD
478 cursor.key_beg.obj_id = ip->obj_id;
479 cursor.key_beg.key = 0;
d5530d22 480 cursor.key_beg.create_tid = 0;
8cd0a023
MD
481 cursor.key_beg.delete_tid = 0;
482 cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
483 cursor.key_beg.obj_type = 0;
ea434b6f 484
d5530d22 485 cursor.asof = iinfo.obj_asof;
11ad5ade 486 cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
d5530d22 487 HAMMER_CURSOR_ASOF;
8cd0a023
MD
488
489 *errorp = hammer_btree_lookup(&cursor);
6a37e7e4
MD
490 if (*errorp == EDEADLK) {
491 hammer_done_cursor(&cursor);
492 goto retry;
493 }
427e5fc6
MD
494
495 /*
496 * On success the B-Tree lookup will hold the appropriate
497 * buffer cache buffers and provide a pointer to the requested
d113fda1
MD
498 * information. Copy the information to the in-memory inode
499 * and cache the B-Tree node to improve future operations.
427e5fc6 500 */
66325755 501 if (*errorp == 0) {
11ad5ade 502 ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
40043e7f 503 ip->ino_data = cursor.data->inode;
bcac4bbb
MD
504
505 /*
506 * cache[0] tries to cache the location of the object inode.
507 * The assumption is that it is near the directory inode.
508 *
509 * cache[1] tries to cache the location of the object data.
39d8fd63
MD
510 * We might have something in the governing directory from
511 * scan optimizations (see the strategy code in
512 * hammer_vnops.c).
513 *
514 * We update dip->cache[2], if possible, with the location
515 * of the object inode for future directory shortcuts.
bcac4bbb
MD
516 */
517 hammer_cache_node(&ip->cache[0], cursor.node);
39d8fd63
MD
518 if (dip) {
519 if (dip->cache[3].node) {
520 hammer_cache_node(&ip->cache[1],
521 dip->cache[3].node);
522 }
523 hammer_cache_node(&dip->cache[2], cursor.node);
524 }
cb51be26
MD
525
526 /*
527 * The file should not contain any data past the file size
a9d52b76 528 * stored in the inode. Setting save_trunc_off to the
cb51be26
MD
529 * file size instead of max reduces B-Tree lookup overheads
530 * on append by allowing the flusher to avoid checking for
531 * record overwrites.
532 */
a9d52b76 533 ip->save_trunc_off = ip->ino_data.size;
5fa5c92f
MD
534
535 /*
536 * Locate and assign the pseudofs management structure to
537 * the inode.
538 */
539 if (dip && dip->obj_localization == ip->obj_localization) {
540 ip->pfsm = dip->pfsm;
541 hammer_ref(&ip->pfsm->lock);
542 } else {
ea434b6f
MD
543 ip->pfsm = hammer_load_pseudofs(trans,
544 ip->obj_localization,
545 errorp);
546 *errorp = 0; /* ignore ENOENT */
5fa5c92f 547 }
427e5fc6 548 }
427e5fc6
MD
549
550 /*
cb51be26
MD
551 * The inode is placed on the red-black tree and will be synced to
552 * the media when flushed or by the filesystem sync. If this races
553 * another instantiation/lookup the insertion will fail.
427e5fc6 554 */
66325755 555 if (*errorp == 0) {
427e5fc6 556 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
5fa5c92f 557 hammer_free_inode(ip);
b3deaf57 558 hammer_done_cursor(&cursor);
427e5fc6
MD
559 goto loop;
560 }
c0ade690 561 ip->flags |= HAMMER_INODE_ONDISK;
427e5fc6 562 } else {
e63644f0
MD
563 if (ip->flags & HAMMER_INODE_RSV_INODES) {
564 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
9f5097dc 565 --hmp->rsv_inodes;
e63644f0 566 }
e63644f0 567
5fa5c92f 568 hammer_free_inode(ip);
66325755 569 ip = NULL;
427e5fc6 570 }
b3deaf57 571 hammer_done_cursor(&cursor);
e98f1b96
MD
572
573 /*
574 * NEWINODE is only set if the inode becomes dirty later,
575 * setting it here just leads to unnecessary stalls.
576 *
577 * trans->flags |= HAMMER_TRANSF_NEWINODE;
578 */
66325755
MD
579 return (ip);
580}
581
8cd0a023 582/*
4c286c36
MD
583 * Get a dummy inode to placemark a broken directory entry.
584 */
585struct hammer_inode *
586hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
587 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
588 int flags, int *errorp)
589{
590 hammer_mount_t hmp = trans->hmp;
591 struct hammer_inode_info iinfo;
592 struct hammer_inode *ip;
593
594 /*
595 * Determine if we already have an inode cached. If we do then
596 * we are golden.
597 *
598 * If we find an inode with no vnode we have to mark the
599 * transaction such that hammer_inode_waitreclaims() is
600 * called later on to avoid building up an infinite number
601 * of inodes. Otherwise we can continue to * add new inodes
602 * faster then they can be disposed of, even with the tsleep
603 * delay.
604 *
605 * If we find a non-fake inode we return an error. Only fake
606 * inodes can be returned by this routine.
607 */
608 iinfo.obj_id = obj_id;
609 iinfo.obj_asof = asof;
610 iinfo.obj_localization = localization;
611loop:
612 *errorp = 0;
613 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
614 if (ip) {
615 if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
616 *errorp = ENOENT;
617 return(NULL);
618 }
619 hammer_ref(&ip->lock);
620 return(ip);
621 }
622
623 /*
624 * Allocate a new inode structure and deal with races later.
625 */
626 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
627 ++hammer_count_inodes;
628 ++hmp->count_inodes;
629 ip->obj_id = obj_id;
630 ip->obj_asof = iinfo.obj_asof;
631 ip->obj_localization = localization;
632 ip->hmp = hmp;
633 ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
634 ip->cache[0].ip = ip;
635 ip->cache[1].ip = ip;
39d8fd63
MD
636 ip->cache[2].ip = ip;
637 ip->cache[3].ip = ip;
4c286c36
MD
638 ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
639 0x7FFFFFFFFFFFFFFFLL;
640 RB_INIT(&ip->rec_tree);
641 TAILQ_INIT(&ip->target_list);
642 hammer_ref(&ip->lock);
643
644 /*
645 * Populate the dummy inode. Leave everything zero'd out.
646 *
647 * (ip->ino_leaf and ip->ino_data)
648 *
649 * Make the dummy inode a FIFO object which most copy programs
650 * will properly ignore.
651 */
652 ip->save_trunc_off = ip->ino_data.size;
653 ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;
654
655 /*
656 * Locate and assign the pseudofs management structure to
657 * the inode.
658 */
659 if (dip && dip->obj_localization == ip->obj_localization) {
660 ip->pfsm = dip->pfsm;
661 hammer_ref(&ip->pfsm->lock);
662 } else {
663 ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
664 errorp);
665 *errorp = 0; /* ignore ENOENT */
666 }
667
668 /*
669 * The inode is placed on the red-black tree and will be synced to
670 * the media when flushed or by the filesystem sync. If this races
671 * another instantiation/lookup the insertion will fail.
672 *
673 * NOTE: Do not set HAMMER_INODE_ONDISK. The inode is a fake.
674 */
675 if (*errorp == 0) {
676 if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
677 hammer_free_inode(ip);
678 goto loop;
679 }
680 } else {
681 if (ip->flags & HAMMER_INODE_RSV_INODES) {
682 ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
683 --hmp->rsv_inodes;
684 }
685 hammer_free_inode(ip);
686 ip = NULL;
687 }
688 trans->flags |= HAMMER_TRANSF_NEWINODE;
689 return (ip);
690}
691
692/*
39d8fd63
MD
693 * Return a referenced inode only if it is in our inode cache.
694 *
695 * Dummy inodes do not count.
696 */
697struct hammer_inode *
698hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
699 hammer_tid_t asof, u_int32_t localization)
700{
701 hammer_mount_t hmp = trans->hmp;
702 struct hammer_inode_info iinfo;
703 struct hammer_inode *ip;
704
705 iinfo.obj_id = obj_id;
706 iinfo.obj_asof = asof;
707 iinfo.obj_localization = localization;
5a64efa1 708
39d8fd63
MD
709 ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
710 if (ip) {
711 if (ip->flags & HAMMER_INODE_DUMMY)
712 ip = NULL;
713 else
714 hammer_ref(&ip->lock);
715 }
716 return(ip);
717}
718
719/*
8cd0a023 720 * Create a new filesystem object, returning the inode in *ipp. The
ea434b6f 721 * returned inode will be referenced. The inode is created in-memory.
8cd0a023 722 *
ea434b6f
MD
723 * If pfsm is non-NULL the caller wishes to create the root inode for
724 * a master PFS.
8cd0a023
MD
725 */
726int
a89aec1b 727hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
5a64efa1
MD
728 struct ucred *cred,
729 hammer_inode_t dip, const char *name, int namelen,
ea434b6f 730 hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
66325755 731{
a89aec1b
MD
732 hammer_mount_t hmp;
733 hammer_inode_t ip;
6b4f890b 734 uid_t xuid;
5a930e66 735 int error;
5a64efa1
MD
736 int64_t namekey;
737 u_int32_t dummy;
66325755 738
8cd0a023 739 hmp = trans->hmp;
5a930e66 740
bac808fe 741 ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
b3deaf57 742 ++hammer_count_inodes;
9f5097dc 743 ++hmp->count_inodes;
82010f9f 744 trans->flags |= HAMMER_TRANSF_NEWINODE;
5a930e66 745
ea434b6f
MD
746 if (pfsm) {
747 KKASSERT(pfsm->localization != 0);
5a930e66 748 ip->obj_id = HAMMER_OBJID_ROOT;
ea434b6f
MD
749 ip->obj_localization = pfsm->localization;
750 } else {
751 KKASSERT(dip != NULL);
5a64efa1
MD
752 namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
753 ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
ea434b6f
MD
754 ip->obj_localization = dip->obj_localization;
755 }
5a930e66 756
8cd0a023 757 KKASSERT(ip->obj_id != 0);
7f7c1f84 758 ip->obj_asof = hmp->asof;
8cd0a023 759 ip->hmp = hmp;
b84de5af 760 ip->flush_state = HAMMER_FST_IDLE;
ddfdf542
MD
761 ip->flags = HAMMER_INODE_DDIRTY |
762 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
bcac4bbb
MD
763 ip->cache[0].ip = ip;
764 ip->cache[1].ip = ip;
39d8fd63
MD
765 ip->cache[2].ip = ip;
766 ip->cache[3].ip = ip;
8cd0a023 767
a5fddc16 768 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
a9d52b76 769 /* ip->save_trunc_off = 0; (already zero) */
8cd0a023 770 RB_INIT(&ip->rec_tree);
1f07f686 771 TAILQ_INIT(&ip->target_list);
8cd0a023 772
bcac4bbb 773 ip->ino_data.atime = trans->time;
11ad5ade
MD
774 ip->ino_data.mtime = trans->time;
775 ip->ino_data.size = 0;
776 ip->ino_data.nlinks = 0;
e63644f0
MD
777
778 /*
779 * A nohistory designator on the parent directory is inherited by
5a930e66
MD
780 * the child. We will do this even for pseudo-fs creation... the
781 * sysad can turn it off.
e63644f0 782 */
ea434b6f
MD
783 if (dip) {
784 ip->ino_data.uflags = dip->ino_data.uflags &
785 (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
786 }
e63644f0 787
11ad5ade 788 ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
5a930e66
MD
789 ip->ino_leaf.base.localization = ip->obj_localization +
790 HAMMER_LOCALIZE_INODE;
11ad5ade
MD
791 ip->ino_leaf.base.obj_id = ip->obj_id;
792 ip->ino_leaf.base.key = 0;
793 ip->ino_leaf.base.create_tid = 0;
794 ip->ino_leaf.base.delete_tid = 0;
795 ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
796 ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);
797
798 ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
8cd0a023
MD
799 ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
800 ip->ino_data.mode = vap->va_mode;
b84de5af 801 ip->ino_data.ctime = trans->time;
5a930e66
MD
802
803 /*
beec5dc4
MD
804 * If we are running version 2 or greater directory entries are
805 * inode-localized instead of data-localized.
5e435c92
MD
806 */
807 if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
808 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
beec5dc4
MD
809 ip->ino_data.cap_flags |=
810 HAMMER_INODE_CAP_DIR_LOCAL_INO;
5e435c92
MD
811 }
812 }
89cba4df
MD
813 if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) {
814 if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
815 ip->ino_data.cap_flags |=
816 HAMMER_INODE_CAP_DIRHASH_ALG1;
817 }
818 }
5e435c92
MD
819
820 /*
5a930e66
MD
821 * Setup the ".." pointer. This only needs to be done for directories
822 * but we do it for all objects as a recovery aid.
ea434b6f
MD
823 */
824 if (dip)
825 ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
826#if 0
827 /*
5a930e66 828 * The parent_obj_localization field only applies to pseudo-fs roots.
ea434b6f
MD
829 * XXX this is no longer applicable, PFSs are no longer directly
830 * tied into the parent's directory structure.
5a930e66 831 */
5a930e66
MD
832 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
833 ip->obj_id == HAMMER_OBJID_ROOT) {
834 ip->ino_data.ext.obj.parent_obj_localization =
835 dip->obj_localization;
836 }
ea434b6f 837#endif
6b4f890b 838
11ad5ade 839 switch(ip->ino_leaf.base.obj_type) {
7a04d74f
MD
840 case HAMMER_OBJTYPE_CDEV:
841 case HAMMER_OBJTYPE_BDEV:
842 ip->ino_data.rmajor = vap->va_rmajor;
843 ip->ino_data.rminor = vap->va_rminor;
844 break;
845 default:
846 break;
847 }
848
6b4f890b
MD
849 /*
850 * Calculate default uid/gid and overwrite with information from
851 * the vap.
852 */
ea434b6f
MD
853 if (dip) {
854 xuid = hammer_to_unix_xid(&dip->ino_data.uid);
855 xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
856 xuid, cred, &vap->va_mode);
857 } else {
858 xuid = 0;
859 }
6b4f890b
MD
860 ip->ino_data.mode = vap->va_mode;
861
8cd0a023
MD
862 if (vap->va_vaflags & VA_UID_UUID_VALID)
863 ip->ino_data.uid = vap->va_uid_uuid;
6b4f890b 864 else if (vap->va_uid != (uid_t)VNOVAL)
7538695e
MD
865 hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
866 else
6b4f890b 867 hammer_guid_to_uuid(&ip->ino_data.uid, xuid);
7538695e 868
8cd0a023
MD
869 if (vap->va_vaflags & VA_GID_UUID_VALID)
870 ip->ino_data.gid = vap->va_gid_uuid;
6b4f890b 871 else if (vap->va_gid != (gid_t)VNOVAL)
8cd0a023 872 hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
ea434b6f 873 else if (dip)
7538695e 874 ip->ino_data.gid = dip->ino_data.gid;
8cd0a023
MD
875
876 hammer_ref(&ip->lock);
5fa5c92f 877
ea434b6f
MD
878 if (pfsm) {
879 ip->pfsm = pfsm;
880 hammer_ref(&pfsm->lock);
881 error = 0;
882 } else if (dip->obj_localization == ip->obj_localization) {
5fa5c92f
MD
883 ip->pfsm = dip->pfsm;
884 hammer_ref(&ip->pfsm->lock);
885 error = 0;
886 } else {
ea434b6f
MD
887 ip->pfsm = hammer_load_pseudofs(trans,
888 ip->obj_localization,
889 &error);
890 error = 0; /* ignore ENOENT */
5fa5c92f
MD
891 }
892
893 if (error) {
894 hammer_free_inode(ip);
895 ip = NULL;
896 } else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
973c11b9
MD
897 panic("hammer_create_inode: duplicate obj_id %llx",
898 (long long)ip->obj_id);
5fa5c92f
MD
899 /* not reached */
900 hammer_free_inode(ip);
8cd0a023
MD
901 }
902 *ipp = ip;
5fa5c92f
MD
903 return(error);
904}
905
906/*
907 * Final cleanup / freeing of an inode structure
908 */
909static void
910hammer_free_inode(hammer_inode_t ip)
911{
bac808fe
MD
912 struct hammer_mount *hmp;
913
914 hmp = ip->hmp;
250aec18 915 KKASSERT(hammer_oneref(&ip->lock));
5fa5c92f
MD
916 hammer_uncache_node(&ip->cache[0]);
917 hammer_uncache_node(&ip->cache[1]);
39d8fd63
MD
918 hammer_uncache_node(&ip->cache[2]);
919 hammer_uncache_node(&ip->cache[3]);
ccf6a64d 920 hammer_inode_wakereclaims(ip);
5fa5c92f
MD
921 if (ip->objid_cache)
922 hammer_clear_objid(ip);
923 --hammer_count_inodes;
bac808fe 924 --hmp->count_inodes;
5fa5c92f 925 if (ip->pfsm) {
bac808fe 926 hammer_rel_pseudofs(hmp, ip->pfsm);
5fa5c92f
MD
927 ip->pfsm = NULL;
928 }
bac808fe 929 kfree(ip, hmp->m_inodes);
5fa5c92f
MD
930 ip = NULL;
931}
932
933/*
ea434b6f
MD
934 * Retrieve pseudo-fs data. NULL will never be returned.
935 *
936 * If an error occurs *errorp will be set and a default template is returned,
937 * otherwise *errorp is set to 0. Typically when an error occurs it will
938 * be ENOENT.
5fa5c92f 939 */
ea434b6f
MD
940hammer_pseudofs_inmem_t
941hammer_load_pseudofs(hammer_transaction_t trans,
942 u_int32_t localization, int *errorp)
5fa5c92f
MD
943{
944 hammer_mount_t hmp = trans->hmp;
ea434b6f 945 hammer_inode_t ip;
5fa5c92f
MD
946 hammer_pseudofs_inmem_t pfsm;
947 struct hammer_cursor cursor;
5fa5c92f
MD
948 int bytes;
949
950retry:
ea434b6f 951 pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
5fa5c92f 952 if (pfsm) {
5fa5c92f 953 hammer_ref(&pfsm->lock);
ea434b6f
MD
954 *errorp = 0;
955 return(pfsm);
956 }
957
958 /*
959 * PFS records are stored in the root inode (not the PFS root inode,
960 * but the real root). Avoid an infinite recursion if loading
961 * the PFS for the real root.
962 */
963 if (localization) {
964 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
965 HAMMER_MAX_TID,
966 HAMMER_DEF_LOCALIZATION, 0, errorp);
967 } else {
968 ip = NULL;
5fa5c92f
MD
969 }
970
bac808fe 971 pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
ea434b6f 972 pfsm->localization = localization;
a56cb012
MD
973 pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
974 pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;
5fa5c92f 975
ea434b6f
MD
976 hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
977 cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
5fa5c92f
MD
978 HAMMER_LOCALIZE_MISC;
979 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
980 cursor.key_beg.create_tid = 0;
981 cursor.key_beg.delete_tid = 0;
ea434b6f 982 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
5fa5c92f 983 cursor.key_beg.obj_type = 0;
ea434b6f 984 cursor.key_beg.key = localization;
5fa5c92f
MD
985 cursor.asof = HAMMER_MAX_TID;
986 cursor.flags |= HAMMER_CURSOR_ASOF;
987
ea434b6f
MD
988 if (ip)
989 *errorp = hammer_ip_lookup(&cursor);
990 else
991 *errorp = hammer_btree_lookup(&cursor);
992 if (*errorp == 0) {
993 *errorp = hammer_ip_resolve_data(&cursor);
994 if (*errorp == 0) {
842e7a70
MD
995 if (cursor.data->pfsd.mirror_flags &
996 HAMMER_PFSD_DELETED) {
997 *errorp = ENOENT;
998 } else {
999 bytes = cursor.leaf->data_len;
1000 if (bytes > sizeof(pfsm->pfsd))
1001 bytes = sizeof(pfsm->pfsd);
1002 bcopy(cursor.data, &pfsm->pfsd, bytes);
1003 }
5fa5c92f 1004 }
5fa5c92f 1005 }
5fa5c92f
MD
1006 hammer_done_cursor(&cursor);
1007
ea434b6f
MD
1008 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
1009 hammer_ref(&pfsm->lock);
1010 if (ip)
1011 hammer_rel_inode(ip, 0);
1012 if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
bac808fe 1013 kfree(pfsm, hmp->m_misc);
ea434b6f 1014 goto retry;
5fa5c92f 1015 }
ea434b6f 1016 return(pfsm);
5fa5c92f
MD
1017}
1018
1019/*
1020 * Store pseudo-fs data. The backend will automatically delete any prior
1021 * on-disk pseudo-fs data but we have to delete in-memory versions.
1022 */
1023int
ea434b6f 1024hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
5fa5c92f
MD
1025{
1026 struct hammer_cursor cursor;
5fa5c92f 1027 hammer_record_t record;
ea434b6f 1028 hammer_inode_t ip;
5fa5c92f
MD
1029 int error;
1030
ea434b6f
MD
1031 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1032 HAMMER_DEF_LOCALIZATION, 0, &error);
5fa5c92f 1033retry:
a56cb012 1034 pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
5fa5c92f
MD
1035 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1036 cursor.key_beg.localization = ip->obj_localization +
1037 HAMMER_LOCALIZE_MISC;
ea434b6f 1038 cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
5fa5c92f
MD
1039 cursor.key_beg.create_tid = 0;
1040 cursor.key_beg.delete_tid = 0;
ea434b6f 1041 cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
5fa5c92f 1042 cursor.key_beg.obj_type = 0;
ea434b6f 1043 cursor.key_beg.key = pfsm->localization;
5fa5c92f
MD
1044 cursor.asof = HAMMER_MAX_TID;
1045 cursor.flags |= HAMMER_CURSOR_ASOF;
1046
3214ade6
MD
1047 /*
1048 * Replace any in-memory version of the record.
1049 */
5fa5c92f
MD
1050 error = hammer_ip_lookup(&cursor);
1051 if (error == 0 && hammer_cursor_inmem(&cursor)) {
1052 record = cursor.iprec;
1053 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
1054 KKASSERT(cursor.deadlk_rec == NULL);
1055 hammer_ref(&record->lock);
1056 cursor.deadlk_rec = record;
1057 error = EDEADLK;
1058 } else {
1059 record->flags |= HAMMER_RECF_DELETED_FE;
1060 error = 0;
1061 }
1062 }
3214ade6
MD
1063
1064 /*
1065 * Allocate replacement general record. The backend flush will
1066 * delete any on-disk version of the record.
1067 */
5fa5c92f
MD
1068 if (error == 0 || error == ENOENT) {
1069 record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
1070 record->type = HAMMER_MEM_RECORD_GENERAL;
1071
1072 record->leaf.base.localization = ip->obj_localization +
1073 HAMMER_LOCALIZE_MISC;
ea434b6f
MD
1074 record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
1075 record->leaf.base.key = pfsm->localization;
5fa5c92f
MD
1076 record->leaf.data_len = sizeof(pfsm->pfsd);
1077 bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
1078 error = hammer_ip_add_record(trans, record);
1079 }
1080 hammer_done_cursor(&cursor);
1081 if (error == EDEADLK)
1082 goto retry;
ea434b6f
MD
1083 hammer_rel_inode(ip, 0);
1084 return(error);
1085}
1086
1087/*
1088 * Create a root directory for a PFS if one does not alredy exist.
4c038e17
MD
1089 *
1090 * The PFS root stands alone so we must also bump the nlinks count
1091 * to prevent it from being destroyed on release.
ea434b6f
MD
1092 */
1093int
1094hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1095 hammer_pseudofs_inmem_t pfsm)
1096{
1097 hammer_inode_t ip;
1098 struct vattr vap;
1099 int error;
1100
1101 ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
1102 pfsm->localization, 0, &error);
1103 if (ip == NULL) {
1104 vattr_null(&vap);
1105 vap.va_mode = 0755;
1106 vap.va_type = VDIR;
5a64efa1
MD
1107 error = hammer_create_inode(trans, &vap, cred,
1108 NULL, NULL, 0,
1109 pfsm, &ip);
4c038e17
MD
1110 if (error == 0) {
1111 ++ip->ino_data.nlinks;
e98f1b96 1112 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
4c038e17 1113 }
5fa5c92f 1114 }
ea434b6f
MD
1115 if (ip)
1116 hammer_rel_inode(ip, 0);
5fa5c92f
MD
1117 return(error);
1118}
1119
ea434b6f 1120/*
842e7a70
MD
1121 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
1122 * if we are unable to disassociate all the inodes.
1123 */
1124static
1125int
1126hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
1127{
1128 int res;
1129
1130 hammer_ref(&ip->lock);
250aec18 1131 if (hammer_isactive(&ip->lock) == 2 && ip->vp)
842e7a70 1132 vclean_unlocked(ip->vp);
250aec18 1133 if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
842e7a70
MD
1134 res = 0;
1135 else
1136 res = -1; /* stop, someone is using the inode */
1137 hammer_rel_inode(ip, 0);
1138 return(res);
1139}
1140
1141int
1142hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
1143{
1144 int res;
1145 int try;
1146
1147 for (try = res = 0; try < 4; ++try) {
1148 res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
1149 hammer_inode_pfs_cmp,
1150 hammer_unload_pseudofs_callback,
1151 &localization);
1152 if (res == 0 && try > 1)
1153 break;
1154 hammer_flusher_sync(trans->hmp);
1155 }
1156 if (res != 0)
1157 res = ENOTEMPTY;
1158 return(res);
1159}
1160
1161
1162/*
ea434b6f
MD
1163 * Release a reference on a PFS
1164 */
5fa5c92f
MD
1165void
1166hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
1167{
250aec18
MD
1168 hammer_rel(&pfsm->lock);
1169 if (hammer_norefs(&pfsm->lock)) {
5fa5c92f 1170 RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
bac808fe 1171 kfree(pfsm, hmp->m_misc);
5fa5c92f 1172 }
66325755
MD
1173}
1174
d113fda1
MD
1175/*
1176 * Called by hammer_sync_inode().
1177 */
1178static int
4e17f465 1179hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
c0ade690 1180{
4e17f465 1181 hammer_transaction_t trans = cursor->trans;
c0ade690
MD
1182 hammer_record_t record;
1183 int error;
06ad81ff 1184 int redirty;
c0ade690 1185
d26d0ae9 1186retry:
c0ade690
MD
1187 error = 0;
1188
869e8f55
MD
1189 /*
1190 * If the inode has a presence on-disk then locate it and mark
1191 * it deleted, setting DELONDISK.
1192 *
1193 * The record may or may not be physically deleted, depending on
1194 * the retention policy.
1195 */
76376933
MD
1196 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
1197 HAMMER_INODE_ONDISK) {
4e17f465 1198 hammer_normalize_cursor(cursor);
5a930e66
MD
1199 cursor->key_beg.localization = ip->obj_localization +
1200 HAMMER_LOCALIZE_INODE;
4e17f465
MD
1201 cursor->key_beg.obj_id = ip->obj_id;
1202 cursor->key_beg.key = 0;
1203 cursor->key_beg.create_tid = 0;
1204 cursor->key_beg.delete_tid = 0;
1205 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1206 cursor->key_beg.obj_type = 0;
1207 cursor->asof = ip->obj_asof;
1208 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
11ad5ade 1209 cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
4e17f465
MD
1210 cursor->flags |= HAMMER_CURSOR_BACKEND;
1211
1212 error = hammer_btree_lookup(cursor);
e8599db1
MD
1213 if (hammer_debug_inode)
1214 kprintf("IPDEL %p %08x %d", ip, ip->flags, error);
b84de5af 1215
c0ade690 1216 if (error == 0) {
e63644f0 1217 error = hammer_ip_delete_record(cursor, ip, trans->tid);
e8599db1
MD
1218 if (hammer_debug_inode)
1219 kprintf(" error %d\n", error);
1f07f686 1220 if (error == 0) {
195c19a1 1221 ip->flags |= HAMMER_INODE_DELONDISK;
1f07f686 1222 }
e8599db1 1223 if (cursor->node)
bcac4bbb 1224 hammer_cache_node(&ip->cache[0], cursor->node);
4e17f465
MD
1225 }
1226 if (error == EDEADLK) {
1227 hammer_done_cursor(cursor);
1228 error = hammer_init_cursor(trans, cursor,
1229 &ip->cache[0], ip);
e8599db1
MD
1230 if (hammer_debug_inode)
1231 kprintf("IPDED %p %d\n", ip, error);
4e17f465
MD
1232 if (error == 0)
1233 goto retry;
c0ade690 1234 }
c0ade690
MD
1235 }
1236
1237 /*
869e8f55
MD
1238 * Ok, write out the initial record or a new record (after deleting
1239 * the old one), unless the DELETED flag is set. This routine will
1240 * clear DELONDISK if it writes out a record.
76376933 1241 *
869e8f55
MD
1242 * Update our inode statistics if this is the first application of
1243 * the inode on-disk.
c0ade690 1244 */
869e8f55
MD
1245 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
1246 /*
7a61b85d
MD
1247 * Generate a record and write it to the media. We clean-up
1248 * the state before releasing so we do not have to set-up
1249 * a flush_group.
869e8f55 1250 */
11ad5ade 1251 record = hammer_alloc_mem_record(ip, 0);
930bf163 1252 record->type = HAMMER_MEM_RECORD_INODE;
1f07f686 1253 record->flush_state = HAMMER_FST_FLUSH;
11ad5ade
MD
1254 record->leaf = ip->sync_ino_leaf;
1255 record->leaf.base.create_tid = trans->tid;
1256 record->leaf.data_len = sizeof(ip->sync_ino_data);
dd94f1b1 1257 record->leaf.create_ts = trans->time32;
b84de5af 1258 record->data = (void *)&ip->sync_ino_data;
d36ec43b 1259 record->flags |= HAMMER_RECF_INTERLOCK_BE;
06ad81ff
MD
1260
1261 /*
1262 * If this flag is set we cannot sync the new file size
1263 * because we haven't finished related truncations. The
1264 * inode will be flushed in another flush group to finish
1265 * the job.
1266 */
1267 if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
1268 ip->sync_ino_data.size != ip->ino_data.size) {
1269 redirty = 1;
1270 ip->sync_ino_data.size = ip->ino_data.size;
1271 } else {
1272 redirty = 0;
1273 }
1274
4e17f465
MD
1275 for (;;) {
1276 error = hammer_ip_sync_record_cursor(cursor, record);
e8599db1
MD
1277 if (hammer_debug_inode)
1278 kprintf("GENREC %p rec %08x %d\n",
1279 ip, record->flags, error);
4e17f465
MD
1280 if (error != EDEADLK)
1281 break;
1282 hammer_done_cursor(cursor);
1283 error = hammer_init_cursor(trans, cursor,
1284 &ip->cache[0], ip);
e8599db1
MD
1285 if (hammer_debug_inode)
1286 kprintf("GENREC reinit %d\n", error);
4e17f465
MD
1287 if (error)
1288 break;
1289 }
d36ec43b
MD
1290
1291 /*
3214ade6
MD
1292 * Note: The record was never on the inode's record tree
1293 * so just wave our hands importantly and destroy it.
d36ec43b 1294 */
3214ade6 1295 record->flags |= HAMMER_RECF_COMMITTED;
d36ec43b 1296 record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
1f07f686 1297 record->flush_state = HAMMER_FST_IDLE;
3214ade6 1298 ++ip->rec_generation;
b3deaf57 1299 hammer_rel_mem_record(record);
d36ec43b 1300
869e8f55
MD
1301 /*
1302 * Finish up.
1303 */
d26d0ae9 1304 if (error == 0) {
e8599db1
MD
1305 if (hammer_debug_inode)
1306 kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
11ad5ade 1307 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
9192654c 1308 HAMMER_INODE_SDIRTY |
ddfdf542
MD
1309 HAMMER_INODE_ATIME |
1310 HAMMER_INODE_MTIME);
b84de5af 1311 ip->flags &= ~HAMMER_INODE_DELONDISK;
06ad81ff
MD
1312 if (redirty)
1313 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1f07f686
MD
1314
1315 /*
1316 * Root volume count of inodes
1317 */
98da6d8c 1318 hammer_sync_lock_sh(trans);
d26d0ae9 1319 if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
e8599db1
MD
1320 hammer_modify_volume_field(trans,
1321 trans->rootvol,
1322 vol0_stat_inodes);
0b075555 1323 ++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
10a5d1ba 1324 hammer_modify_volume_done(trans->rootvol);
d26d0ae9 1325 ip->flags |= HAMMER_INODE_ONDISK;
e8599db1
MD
1326 if (hammer_debug_inode)
1327 kprintf("NOWONDISK %p\n", ip);
d26d0ae9 1328 }
98da6d8c 1329 hammer_sync_unlock(trans);
fbc6e32a 1330 }
c0ade690 1331 }
869e8f55
MD
1332
1333 /*
1334 * If the inode has been destroyed, clean out any left-over flags
1335 * that may have been set by the frontend.
1336 */
f90dde4c 1337 if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) {
11ad5ade 1338 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
9192654c 1339 HAMMER_INODE_SDIRTY |
ddfdf542
MD
1340 HAMMER_INODE_ATIME |
1341 HAMMER_INODE_MTIME);
f90dde4c 1342 }
c0ade690
MD
1343 return(error);
1344}
1345
a89aec1b 1346/*
ddfdf542
MD
1347 * Update only the itimes fields.
1348 *
1349 * ATIME can be updated without generating any UNDO. MTIME is updated
1350 * with UNDO so it is guaranteed to be synchronized properly in case of
1351 * a crash.
1352 *
1353 * Neither field is included in the B-Tree leaf element's CRC, which is how
1354 * we can get away with updating ATIME the way we do.
d113fda1
MD
1355 */
1356static int
4e17f465 1357hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
d113fda1 1358{
4e17f465 1359 hammer_transaction_t trans = cursor->trans;
d113fda1
MD
1360 int error;
1361
6a37e7e4 1362retry:
ddfdf542 1363 if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
d113fda1 1364 HAMMER_INODE_ONDISK) {
ddfdf542
MD
1365 return(0);
1366 }
4e17f465 1367
ddfdf542 1368 hammer_normalize_cursor(cursor);
5a930e66
MD
1369 cursor->key_beg.localization = ip->obj_localization +
1370 HAMMER_LOCALIZE_INODE;
ddfdf542
MD
1371 cursor->key_beg.obj_id = ip->obj_id;
1372 cursor->key_beg.key = 0;
1373 cursor->key_beg.create_tid = 0;
1374 cursor->key_beg.delete_tid = 0;
1375 cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
1376 cursor->key_beg.obj_type = 0;
1377 cursor->asof = ip->obj_asof;
1378 cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1379 cursor->flags |= HAMMER_CURSOR_ASOF;
1380 cursor->flags |= HAMMER_CURSOR_GET_LEAF;
1381 cursor->flags |= HAMMER_CURSOR_GET_DATA;
1382 cursor->flags |= HAMMER_CURSOR_BACKEND;
1383
1384 error = hammer_btree_lookup(cursor);
ddfdf542
MD
1385 if (error == 0) {
1386 hammer_cache_node(&ip->cache[0], cursor->node);
1387 if (ip->sync_flags & HAMMER_INODE_MTIME) {
10a5d1ba 1388 /*
ddfdf542
MD
1389 * Updating MTIME requires an UNDO. Just cover
1390 * both atime and mtime.
10a5d1ba 1391 */
98da6d8c 1392 hammer_sync_lock_sh(trans);
bcac4bbb
MD
1393 hammer_modify_buffer(trans, cursor->data_buffer,
1394 HAMMER_ITIMES_BASE(&cursor->data->inode),
1395 HAMMER_ITIMES_BYTES);
1396 cursor->data->inode.atime = ip->sync_ino_data.atime;
1397 cursor->data->inode.mtime = ip->sync_ino_data.mtime;
1398 hammer_modify_buffer_done(cursor->data_buffer);
98da6d8c 1399 hammer_sync_unlock(trans);
ddfdf542
MD
1400 } else if (ip->sync_flags & HAMMER_INODE_ATIME) {
1401 /*
1402 * Updating atime only can be done in-place with
1403 * no UNDO.
1404 */
98da6d8c 1405 hammer_sync_lock_sh(trans);
ddfdf542
MD
1406 hammer_modify_buffer(trans, cursor->data_buffer,
1407 NULL, 0);
1408 cursor->data->inode.atime = ip->sync_ino_data.atime;
1409 hammer_modify_buffer_done(cursor->data_buffer);
98da6d8c 1410 hammer_sync_unlock(trans);
d113fda1 1411 }
ddfdf542
MD
1412 ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
1413 }
1414 if (error == EDEADLK) {
1415 hammer_done_cursor(cursor);
1416 error = hammer_init_cursor(trans, cursor,
1417 &ip->cache[0], ip);
1418 if (error == 0)
1419 goto retry;
d113fda1
MD
1420 }
1421 return(error);
1422}
1423
1424/*
1f07f686 1425 * Release a reference on an inode, flush as requested.
b84de5af
MD
1426 *
1427 * On the last reference we queue the inode to the flusher for its final
1428 * disposition.
a89aec1b 1429 */
66325755 1430void
a89aec1b 1431hammer_rel_inode(struct hammer_inode *ip, int flush)
66325755 1432{
35a49944 1433 /*hammer_mount_t hmp = ip->hmp;*/
1f07f686 1434
f90dde4c
MD
1435 /*
1436 * Handle disposition when dropping the last ref.
1437 */
1f07f686 1438 for (;;) {
250aec18 1439 if (hammer_oneref(&ip->lock)) {
1f07f686
MD
1440 /*
1441 * Determine whether on-disk action is needed for
1442 * the inode's final disposition.
1443 */
e8599db1
MD
1444 KKASSERT(ip->vp == NULL);
1445 hammer_inode_unloadable_check(ip, 0);
4e17f465 1446 if (ip->flags & HAMMER_INODE_MODMASK) {
35a49944 1447 hammer_flush_inode(ip, 0);
250aec18 1448 } else if (hammer_oneref(&ip->lock)) {
1f07f686
MD
1449 hammer_unload_inode(ip);
1450 break;
1451 }
b84de5af 1452 } else {
4e17f465 1453 if (flush)
1f07f686 1454 hammer_flush_inode(ip, 0);
4e17f465 1455
1f07f686
MD
1456 /*
1457 * The inode still has multiple refs, try to drop
1458 * one ref.
1459 */
250aec18
MD
1460 KKASSERT(hammer_isactive(&ip->lock) >= 1);
1461 if (hammer_isactive(&ip->lock) > 1) {
1462 hammer_rel(&ip->lock);
1f07f686
MD
1463 break;
1464 }
b84de5af 1465 }
f90dde4c 1466 }
427e5fc6
MD
1467}
1468
1469/*
b84de5af
MD
1470 * Unload and destroy the specified inode. Must be called with one remaining
1471 * reference. The reference is disposed of.
8cd0a023 1472 *
cdb6e4e6 1473 * The inode must be completely clean.
27ea2398 1474 */
b84de5af 1475static int
ec4e8497 1476hammer_unload_inode(struct hammer_inode *ip)
27ea2398 1477{
9f5097dc
MD
1478 hammer_mount_t hmp = ip->hmp;
1479
250aec18
MD
1480 KASSERT(hammer_oneref(&ip->lock),
1481 ("hammer_unload_inode: %d refs\n", hammer_isactive(&ip->lock)));
8cd0a023 1482 KKASSERT(ip->vp == NULL);
f90dde4c
MD
1483 KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
1484 KKASSERT(ip->cursor_ip_refs == 0);
899eb297 1485 KKASSERT(hammer_notlocked(&ip->lock));
f90dde4c
MD
1486 KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);
1487
1488 KKASSERT(RB_EMPTY(&ip->rec_tree));
1f07f686 1489 KKASSERT(TAILQ_EMPTY(&ip->target_list));
f90dde4c 1490
73896937
MD
1491 if (ip->flags & HAMMER_INODE_RDIRTY) {
1492 RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
1493 ip->flags &= ~HAMMER_INODE_RDIRTY;
1494 }
9f5097dc 1495 RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);
f90dde4c 1496
5fa5c92f 1497 hammer_free_inode(ip);
27ea2398
MD
1498 return(0);
1499}
1500
27ea2398 1501/*
cdb6e4e6
MD
1502 * Called during unmounting if a critical error occured. The in-memory
1503 * inode and all related structures are destroyed.
1504 *
1505 * If a critical error did not occur the unmount code calls the standard
1506 * release and asserts that the inode is gone.
1507 */
1508int
1509hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
1510{
1511 hammer_record_t rec;
1512
1513 /*
1514 * Get rid of the inodes in-memory records, regardless of their
1515 * state, and clear the mod-mask.
1516 */
1517 while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
1518 TAILQ_REMOVE(&ip->target_list, rec, target_entry);
1519 rec->target_ip = NULL;
1520 if (rec->flush_state == HAMMER_FST_SETUP)
1521 rec->flush_state = HAMMER_FST_IDLE;
1522 }
1523 while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
1524 if (rec->flush_state == HAMMER_FST_FLUSH)
1525 --rec->flush_group->refs;
1526 else
1527 hammer_ref(&rec->lock);
250aec18 1528 KKASSERT(hammer_oneref(&rec->lock));
cdb6e4e6
MD
1529 rec->flush_state = HAMMER_FST_IDLE;
1530 rec->flush_group = NULL;
3214ade6
MD
1531 rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
1532 rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
1533 ++ip->rec_generation;
cdb6e4e6
MD
1534 hammer_rel_mem_record(rec);
1535 }
1536 ip->flags &= ~HAMMER_INODE_MODMASK;
1537 ip->sync_flags &= ~HAMMER_INODE_MODMASK;
1538 KKASSERT(ip->vp == NULL);
1539
1540 /*
1541 * Remove the inode from any flush group, force it idle. FLUSH
1542 * and SETUP states have an inode ref.
1543 */
1544 switch(ip->flush_state) {
1545 case HAMMER_FST_FLUSH:
ff003b11 1546 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
cdb6e4e6
MD
1547 --ip->flush_group->refs;
1548 ip->flush_group = NULL;
1549 /* fall through */
1550 case HAMMER_FST_SETUP:
250aec18 1551 hammer_rel(&ip->lock);
cdb6e4e6
MD
1552 ip->flush_state = HAMMER_FST_IDLE;
1553 /* fall through */
1554 case HAMMER_FST_IDLE:
1555 break;
1556 }
1557
1558 /*
1559 * There shouldn't be any associated vnode. The unload needs at
1560 * least one ref, if we do have a vp steal its ip ref.
1561 */
1562 if (ip->vp) {
1563 kprintf("hammer_destroy_inode_callback: Unexpected "
1564 "vnode association ip %p vp %p\n", ip, ip->vp);
1565 ip->vp->v_data = NULL;
1566 ip->vp = NULL;
1567 } else {
1568 hammer_ref(&ip->lock);
1569 }
1570 hammer_unload_inode(ip);
1571 return(0);
1572}
1573
1574/*
51c35492
MD
1575 * Called on mount -u when switching from RW to RO or vise-versa. Adjust
1576 * the read-only flag for cached inodes.
1577 *
1578 * This routine is called from a RB_SCAN().
1579 */
1580int
1581hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
1582{
1583 hammer_mount_t hmp = ip->hmp;
1584
1585 if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
1586 ip->flags |= HAMMER_INODE_RO;
1587 else
1588 ip->flags &= ~HAMMER_INODE_RO;
1589 return(0);
1590}
1591
1592/*
d113fda1
MD
1593 * A transaction has modified an inode, requiring updates as specified by
1594 * the passed flags.
7f7c1f84 1595 *
9192654c
MD
1596 * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
1597 * and not including size changes due to write-append
1598 * (but other size changes are included).
1599 * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
1600 * write-append.
1f07f686 1601 * HAMMER_INODE_XDIRTY: Dirty in-memory records
4e17f465 1602 * HAMMER_INODE_BUFS: Dirty buffer cache buffers
d113fda1 1603 * HAMMER_INODE_DELETED: Inode record/data must be deleted
ddfdf542 1604 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
427e5fc6 1605 */
66325755 1606void
e98f1b96 1607hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
427e5fc6 1608{
cdb6e4e6
MD
1609 /*
1610 * ronly of 0 or 2 does not trigger assertion.
1611 * 2 is a special error state
1612 */
1613 KKASSERT(ip->hmp->ronly != 1 ||
ddfdf542 1614 (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
9192654c 1615 HAMMER_INODE_SDIRTY |
ddfdf542
MD
1616 HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
1617 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
e63644f0
MD
1618 if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
1619 ip->flags |= HAMMER_INODE_RSV_INODES;
1620 ++ip->hmp->rsv_inodes;
1621 }
b84de5af 1622
e98f1b96
MD
1623 /*
1624 * Set the NEWINODE flag in the transaction if the inode
1625 * transitions to a dirty state. This is used to track
1626 * the load on the inode cache.
1627 */
1628 if (trans &&
1629 (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
1630 (flags & HAMMER_INODE_MODMASK)) {
1631 trans->flags |= HAMMER_TRANSF_NEWINODE;
1632 }
1633
b84de5af
MD
1634 ip->flags |= flags;
1635}
1636
1637/*
1f07f686 1638 * Request that an inode be flushed. This whole mess cannot block and may
7bc5b8c2
MD
1639 * recurse (if not synchronous). Once requested HAMMER will attempt to
1640 * actively flush the inode until the flush can be done.
b84de5af 1641 *
1f07f686
MD
1642 * The inode may already be flushing, or may be in a setup state. We can
1643 * place the inode in a flushing state if it is currently idle and flag it
1644 * to reflush if it is currently flushing.
7bc5b8c2 1645 *
4889cbd4
MD
1646 * Upon return if the inode could not be flushed due to a setup
1647 * dependancy, then it will be automatically flushed when the dependancy
1648 * is satisfied.
b84de5af
MD
1649 */
1650void
f90dde4c 1651hammer_flush_inode(hammer_inode_t ip, int flags)
b84de5af 1652{
7a61b85d
MD
1653 hammer_mount_t hmp;
1654 hammer_flush_group_t flg;
bf3b416b 1655 int good;
1f07f686
MD
1656
1657 /*
37646115
MD
1658 * fill_flush_group is the first flush group we may be able to
1659 * continue filling, it may be open or closed but it will always
1660 * be past the currently flushing (running) flg.
1661 *
1662 * next_flush_group is the next open flush group.
7a61b85d
MD
1663 */
1664 hmp = ip->hmp;
37646115 1665 while ((flg = hmp->fill_flush_group) != NULL) {
7b6ccb11 1666 KKASSERT(flg->running == 0);
37646115
MD
1667 if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
1668 flg->total_count <= hammer_autoflush) {
7b6ccb11 1669 break;
37646115
MD
1670 }
1671 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
7b6ccb11 1672 hammer_flusher_async(ip->hmp, flg);
7a61b85d
MD
1673 }
1674 if (flg == NULL) {
bac808fe 1675 flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
37646115
MD
1676 flg->seq = hmp->flusher.next++;
1677 if (hmp->next_flush_group == NULL)
1678 hmp->next_flush_group = flg;
1679 if (hmp->fill_flush_group == NULL)
1680 hmp->fill_flush_group = flg;
ff003b11 1681 RB_INIT(&flg->flush_tree);
7a61b85d
MD
1682 TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
1683 }
1684
1685 /*
1686 * Trivial 'nothing to flush' case. If the inode is in a SETUP
1f07f686
MD
1687 * state we have to put it back into an IDLE state so we can
1688 * drop the extra ref.
7a61b85d
MD
1689 *
1690 * If we have a parent dependancy we must still fall through
1691 * so we can run it.
1f07f686 1692 */
4e17f465 1693 if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
7a61b85d
MD
1694 if (ip->flush_state == HAMMER_FST_SETUP &&
1695 TAILQ_EMPTY(&ip->target_list)) {
1f07f686
MD
1696 ip->flush_state = HAMMER_FST_IDLE;
1697 hammer_rel_inode(ip, 0);
ec4e8497 1698 }
7a61b85d
MD
1699 if (ip->flush_state == HAMMER_FST_IDLE)
1700 return;
b84de5af 1701 }
42c7d26b 1702
1f07f686
MD
1703 /*
1704 * Our flush action will depend on the current state.
1705 */
1706 switch(ip->flush_state) {
1707 case HAMMER_FST_IDLE:
1708 /*
1709 * We have no dependancies and can flush immediately. Some
1710 * our children may not be flushable so we have to re-test
1711 * with that additional knowledge.
1712 */
7a61b85d 1713 hammer_flush_inode_core(ip, flg, flags);
1f07f686
MD
1714 break;
1715 case HAMMER_FST_SETUP:
1716 /*
1717 * Recurse upwards through dependancies via target_list
1718 * and start their flusher actions going if possible.
1719 *
1720 * 'good' is our connectivity. -1 means we have none and
1721 * can't flush, 0 means there weren't any dependancies, and
1722 * 1 means we have good connectivity.
1723 */
cc0758d0 1724 good = hammer_setup_parent_inodes(ip, 0, flg);
1f07f686 1725
1f07f686 1726 if (good >= 0) {
7b6ccb11
MD
1727 /*
1728 * We can continue if good >= 0. Determine how
1729 * many records under our inode can be flushed (and
1730 * mark them).
1731 */
7a61b85d 1732 hammer_flush_inode_core(ip, flg, flags);
1f07f686 1733 } else {
7b6ccb11 1734 /*
4889cbd4 1735 * Parent has no connectivity, tell it to flush
7b6ccb11 1736 * us as soon as it does.
4889cbd4
MD
1737 *
1738 * The REFLUSH flag is also needed to trigger
1739 * dependancy wakeups.
7b6ccb11 1740 */
4889cbd4
MD
1741 ip->flags |= HAMMER_INODE_CONN_DOWN |
1742 HAMMER_INODE_REFLUSH;
4e17f465
MD
1743 if (flags & HAMMER_FLUSH_SIGNAL) {
1744 ip->flags |= HAMMER_INODE_RESIGNAL;
7a61b85d 1745 hammer_flusher_async(ip->hmp, flg);
4e17f465 1746 }
1f07f686
MD
1747 }
1748 break;
7b6ccb11 1749 case HAMMER_FST_FLUSH:
1f07f686
MD
1750 /*
1751 * We are already flushing, flag the inode to reflush
1752 * if needed after it completes its current flush.
4889cbd4
MD
1753 *
1754 * The REFLUSH flag is also needed to trigger
1755 * dependancy wakeups.
1f07f686
MD
1756 */
1757 if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
1758 ip->flags |= HAMMER_INODE_REFLUSH;
4e17f465
MD
1759 if (flags & HAMMER_FLUSH_SIGNAL) {
1760 ip->flags |= HAMMER_INODE_RESIGNAL;
7a61b85d 1761 hammer_flusher_async(ip->hmp, flg);
4e17f465 1762 }
1f07f686
MD
1763 break;
1764 }
1765}
1766
1767/*
bf3b416b
MD
1768 * Scan ip->target_list, which is a list of records owned by PARENTS to our
1769 * ip which reference our ip.
1770 *
1771 * XXX This is a huge mess of recursive code, but not one bit of it blocks
1772 * so for now do not ref/deref the structures. Note that if we use the
1773 * ref/rel code later, the rel CAN block.
1774 */
1775static int
cc0758d0
MD
1776hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
1777 hammer_flush_group_t flg)
bf3b416b
MD
1778{
1779 hammer_record_t depend;
bf3b416b
MD
1780 int good;
1781 int r;
1782
cc0758d0
MD
1783 /*
1784 * If we hit our recursion limit and we have parent dependencies
1785 * We cannot continue. Returning < 0 will cause us to be flagged
1786 * for reflush. Returning -2 cuts off additional dependency checks
1787 * because they are likely to also hit the depth limit.
1788 *
1789 * We cannot return < 0 if there are no dependencies or there might
1790 * not be anything to wakeup (ip).
1791 */
1792 if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
1793 kprintf("HAMMER Warning: depth limit reached on "
1794 "setup recursion, inode %p %016llx\n",
1795 ip, (long long)ip->obj_id);
1796 return(-2);
1797 }
1798
1799 /*
1800 * Scan dependencies
1801 */
bf3b416b
MD
1802 good = 0;
1803 TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
cc0758d0 1804 r = hammer_setup_parent_inodes_helper(depend, depth, flg);
bf3b416b 1805 KKASSERT(depend->target_ip == ip);
bf3b416b
MD
1806 if (r < 0 && good == 0)
1807 good = -1;
1808 if (r > 0)
1809 good = 1;
cc0758d0
MD
1810
1811 /*
1812 * If we failed due to the recursion depth limit then stop
1813 * now.
1814 */
1815 if (r == -2)
1816 break;
bf3b416b
MD
1817 }
1818 return(good);
bf3b416b
MD
1819}
1820
1821/*
1822 * This helper function takes a record representing the dependancy between
1823 * the parent inode and child inode.
1824 *
1825 * record->ip = parent inode
1826 * record->target_ip = child inode
1827 *
1f07f686 1828 * We are asked to recurse upwards and convert the record from SETUP
bf3b416b 1829 * to FLUSH if possible.
1f07f686
MD
1830 *
1831 * Return 1 if the record gives us connectivity
1832 *
1833 * Return 0 if the record is not relevant
1834 *
1835 * Return -1 if we can't resolve the dependancy and there is no connectivity.
1836 */
1837static int
cc0758d0 1838hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
7a61b85d 1839 hammer_flush_group_t flg)
1f07f686 1840{
bf3b416b
MD
1841 hammer_mount_t hmp;
1842 hammer_inode_t pip;
1843 int good;
1f07f686
MD
1844
1845 KKASSERT(record->flush_state != HAMMER_FST_IDLE);
bf3b416b
MD
1846 pip = record->ip;
1847 hmp = pip->hmp;
1f07f686
MD
1848
1849 /*
1850 * If the record is already flushing, is it in our flush group?
1851 *
e8599db1
MD
1852 * If it is in our flush group but it is a general record or a
1853 * delete-on-disk, it does not improve our connectivity (return 0),
1854 * and if the target inode is not trying to destroy itself we can't
1855 * allow the operation yet anyway (the second return -1).
1f07f686
MD
1856 */
1857 if (record->flush_state == HAMMER_FST_FLUSH) {
7b6ccb11
MD
1858 /*
1859 * If not in our flush group ask the parent to reflush
1860 * us as soon as possible.
1861 */
7a61b85d 1862 if (record->flush_group != flg) {
bf3b416b 1863 pip->flags |= HAMMER_INODE_REFLUSH;
7b6ccb11 1864 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1f07f686 1865 return(-1);
f90dde4c 1866 }
7b6ccb11
MD
1867
1868 /*
1869 * If in our flush group everything is already set up,
1870 * just return whether the record will improve our
1871 * visibility or not.
1872 */
1f07f686
MD
1873 if (record->type == HAMMER_MEM_RECORD_ADD)
1874 return(1);
1875 return(0);
1876 }
1877
1878 /*
1879 * It must be a setup record. Try to resolve the setup dependancies
1880 * by recursing upwards so we can place ip on the flush list.
cc0758d0
MD
1881 *
1882 * Limit ourselves to 20 levels of recursion to avoid blowing out
1883 * the kernel stack. If we hit the recursion limit we can't flush
1884 * until the parent flushes. The parent will flush independantly
1885 * on its own and ultimately a deep recursion will be resolved.
1f07f686
MD
1886 */
1887 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1888
cc0758d0 1889 good = hammer_setup_parent_inodes(pip, depth + 1, flg);
1f07f686
MD
1890
1891 /*
7b6ccb11
MD
1892 * If good < 0 the parent has no connectivity and we cannot safely
1893 * flush the directory entry, which also means we can't flush our
cc0758d0
MD
1894 * ip. Flag us for downward recursion once the parent's
1895 * connectivity is resolved. Flag the parent for [re]flush or it
1896 * may not check for downward recursions.
1f07f686
MD
1897 */
1898 if (good < 0) {
cc0758d0 1899 pip->flags |= HAMMER_INODE_REFLUSH;
7b6ccb11 1900 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1f07f686
MD
1901 return(good);
1902 }
1903
1904 /*
1905 * We are go, place the parent inode in a flushing state so we can
1906 * place its record in a flushing state. Note that the parent
1907 * may already be flushing. The record must be in the same flush
1908 * group as the parent.
1909 */
bf3b416b 1910 if (pip->flush_state != HAMMER_FST_FLUSH)
7a61b85d 1911 hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
bf3b416b 1912 KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);
d006ca0c
MD
1913
1914 /*
1915 * It is possible for a rename to create a loop in the recursion
1916 * and revisit a record. This will result in the record being
1917 * placed in a flush state unexpectedly. This check deals with
1918 * the case.
1919 */
1920 if (record->flush_state == HAMMER_FST_FLUSH) {
1921 if (record->type == HAMMER_MEM_RECORD_ADD)
1922 return(1);
1923 return(0);
1924 }
1925
1f07f686
MD
1926 KKASSERT(record->flush_state == HAMMER_FST_SETUP);
1927
1928#if 0
1929 if (record->type == HAMMER_MEM_RECORD_DEL &&
869e8f55 1930 (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
1f07f686
MD
1931 /*
1932 * Regardless of flushing state we cannot sync this path if the
1933 * record represents a delete-on-disk but the target inode
1934 * is not ready to sync its own deletion.
1935 *
1936 * XXX need to count effective nlinks to determine whether
1937 * the flush is ok, otherwise removing a hardlink will
1938 * just leave the DEL record to rot.
1939 */
1940 record->target_ip->flags |= HAMMER_INODE_REFLUSH;
1941 return(-1);
1942 } else
1943#endif
7a61b85d 1944 if (pip->flush_group == flg) {
1f07f686 1945 /*
d7e278bb
MD
1946 * Because we have not calculated nlinks yet we can just
1947 * set records to the flush state if the parent is in
1948 * the same flush group as we are.
1f07f686 1949 */
7b6ccb11
MD
1950 record->flush_state = HAMMER_FST_FLUSH;
1951 record->flush_group = flg;
1952 ++record->flush_group->refs;
1953 hammer_ref(&record->lock);
1f07f686
MD
1954
1955 /*
7b6ccb11
MD
1956 * A general directory-add contributes to our visibility.
1957 *
1958 * Otherwise it is probably a directory-delete or
1959 * delete-on-disk record and does not contribute to our
1960 * visbility (but we can still flush it).
1f07f686 1961 */
7b6ccb11
MD
1962 if (record->type == HAMMER_MEM_RECORD_ADD)
1963 return(1);
1f07f686
MD
1964 return(0);
1965 } else {
1966 /*
7b6ccb11
MD
1967 * If the parent is not in our flush group we cannot
1968 * flush this record yet, there is no visibility.
1969 * We tell the parent to reflush and mark ourselves
1970 * so the parent knows it should flush us too.
1f07f686 1971 */
bf3b416b 1972 pip->flags |= HAMMER_INODE_REFLUSH;
7b6ccb11 1973 record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
1f07f686 1974 return(-1);
7f7c1f84 1975 }
c0ade690
MD
1976}
1977
1978/*
1f07f686 1979 * This is the core routine placing an inode into the FST_FLUSH state.
c0ade690 1980 */
b84de5af 1981static void
7a61b85d 1982hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
b84de5af 1983{
37646115 1984 hammer_mount_t hmp = ip->hmp;
1f07f686 1985 int go_count;
1f07f686 1986
4e17f465
MD
1987 /*
1988 * Set flush state and prevent the flusher from cycling into
1989 * the next flush group. Do not place the ip on the list yet.
1990 * Inodes not in the idle state get an extra reference.
1991 */
1f07f686
MD
1992 KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
1993 if (ip->flush_state == HAMMER_FST_IDLE)
1994 hammer_ref(&ip->lock);
1995 ip->flush_state = HAMMER_FST_FLUSH;
7a61b85d 1996 ip->flush_group = flg;
37646115
MD
1997 ++hmp->flusher.group_lock;
1998 ++hmp->count_iqueued;
af209b0f 1999 ++hammer_count_iqueued;
7a61b85d 2000 ++flg->total_count;
73896937 2001 hammer_redo_fifo_start_flush(ip);
b84de5af 2002
e0092341 2003#if 0
21fde338 2004 /*
e8599db1 2005 * We need to be able to vfsync/truncate from the backend.
e0092341
MD
2006 *
2007 * XXX Any truncation from the backend will acquire the vnode
2008 * independently.
e8599db1
MD
2009 */
2010 KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
2011 if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
2012 ip->flags |= HAMMER_INODE_VHELD;
2013 vref(ip->vp);
2014 }
e0092341 2015#endif
e8599db1
MD
2016
2017 /*
1f07f686
MD
2018 * Figure out how many in-memory records we can actually flush
2019 * (not including inode meta-data, buffers, etc).
ec4e8497 2020 */
d7e278bb 2021 KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
1f07f686 2022 if (flags & HAMMER_FLUSH_RECURSION) {
7a61b85d
MD
2023 /*
2024 * If this is a upwards recursion we do not want to
2025 * recurse down again!
2026 */
1f07f686 2027 go_count = 1;
d7e278bb 2028#if 0
312de84d 2029 } else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
7a61b85d
MD
2030 /*
2031 * No new records are added if we must complete a flush
2032 * from a previous cycle, but we do have to move the records
2033 * from the previous cycle to the current one.
2034 */
2035#if 0
525aad3a
MD
2036 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2037 hammer_syncgrp_child_callback, NULL);
7a61b85d 2038#endif
312de84d 2039 go_count = 1;
d7e278bb 2040#endif
1f07f686 2041 } else {
7a61b85d
MD
2042 /*
2043 * Normal flush, scan records and bring them into the flush.
2044 * Directory adds and deletes are usually skipped (they are
2045 * grouped with the related inode rather then with the
2046 * directory).
2047 *
2048 * go_count can be negative, which means the scan aborted
2049 * due to the flush group being over-full and we should
2050 * flush what we have.
2051 */
1f07f686
MD
2052 go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
2053 hammer_setup_child_callback, NULL);
2054 }
b84de5af
MD
2055
2056 /*
1f07f686
MD
2057 * This is a more involved test that includes go_count. If we
2058 * can't flush, flag the inode and return. If go_count is 0 we
2059 * were are unable to flush any records in our rec_tree and
2060 * must ignore the XDIRTY flag.
b84de5af 2061 */
1f07f686
MD
2062 if (go_count == 0) {
2063 if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
37646115 2064 --hmp->count_iqueued;
af209b0f
MD
2065 --hammer_count_iqueued;
2066
4889cbd4 2067 --flg->total_count;
1f07f686 2068 ip->flush_state = HAMMER_FST_SETUP;
7a61b85d 2069 ip->flush_group = NULL;
37646115
MD
2070 if (flags & HAMMER_FLUSH_SIGNAL) {
2071 ip->flags |= HAMMER_INODE_REFLUSH |
2072 HAMMER_INODE_RESIGNAL;
2073 } else {
2074 ip->flags |= HAMMER_INODE_REFLUSH;
2075 }
e0092341 2076#if 0
e8599db1
MD
2077 if (ip->flags & HAMMER_INODE_VHELD) {
2078 ip->flags &= ~HAMMER_INODE_VHELD;
2079 vrele(ip->vp);
2080 }
e0092341 2081#endif
4889cbd4
MD
2082
2083 /*
2084 * REFLUSH is needed to trigger dependancy wakeups
2085 * when an inode is in SETUP.
2086 */
2087 ip->flags |= HAMMER_INODE_REFLUSH;
37646115
MD
2088 if (--hmp->flusher.group_lock == 0)
2089 wakeup(&hmp->flusher.group_lock);
1f07f686
MD
2090 return;
2091 }
2092 }
b84de5af 2093
b84de5af
MD
2094 /*
2095 * Snapshot the state of the inode for the backend flusher.
2096 *
a9d52b76 2097 * We continue to retain save_trunc_off even when all truncations
cb51be26
MD
2098 * have been resolved as an optimization to determine if we can
2099 * skip the B-Tree lookup for overwrite deletions.
2100 *
1f07f686
MD
2101 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
2102 * and stays in ip->flags. Once set, it stays set until the
2103 * inode is destroyed.
b84de5af 2104 */
d7e278bb 2105 if (ip->flags & HAMMER_INODE_TRUNCATED) {
312de84d
MD
2106 KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
2107 ip->sync_trunc_off = ip->trunc_off;
2108 ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
2109 ip->flags &= ~HAMMER_INODE_TRUNCATED;
2110 ip->sync_flags |= HAMMER_INODE_TRUNCATED;
a9d52b76
MD
2111
2112 /*
2113 * The save_trunc_off used to cache whether the B-Tree
2114 * holds any records past that point is not used until
2115 * after the truncation has succeeded, so we can safely
2116 * set it now.
2117 */
2118 if (ip->save_trunc_off > ip->sync_trunc_off)
2119 ip->save_trunc_off = ip->sync_trunc_off;
2120 }
312de84d
MD
2121 ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
2122 ~HAMMER_INODE_TRUNCATED);
11ad5ade 2123 ip->sync_ino_leaf = ip->ino_leaf;
b84de5af 2124 ip->sync_ino_data = ip->ino_data;
312de84d 2125 ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
0832c9bb
MD
2126#ifdef DEBUG_TRUNCATE
2127 if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
2128 kprintf("truncateS %016llx\n", ip->sync_trunc_off);
2129#endif
b84de5af
MD
2130
2131 /*
4e17f465 2132 * The flusher list inherits our inode and reference.
b84de5af 2133 */
7a61b85d 2134 KKASSERT(flg->running == 0);
ff003b11 2135 RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
37646115
MD
2136 if (--hmp->flusher.group_lock == 0)
2137 wakeup(&hmp->flusher.group_lock);
1f07f686 2138
37646115
MD
2139 /*
2140 * Auto-flush the group if it grows too large. Make sure the
2141 * inode reclaim wait pipeline continues to work.
2142 */
2143 if (flg->total_count >= hammer_autoflush ||
2144 flg->total_count >= hammer_limit_reclaim / 4) {
2145 if (hmp->fill_flush_group == flg)
2146 hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
2147 hammer_flusher_async(hmp, flg);
0832c9bb 2148 }
b84de5af
MD
2149}
2150
ec4e8497 2151/*
1f07f686
MD
2152 * Callback for scan of ip->rec_tree. Try to include each record in our
2153 * flush. ip->flush_group has been set but the inode has not yet been
2154 * moved into a flushing state.
2155 *
2156 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
2157 * both inodes.
2158 *
2159 * We return 1 for any record placed or found in FST_FLUSH, which prevents
2160 * the caller from shortcutting the flush.
ec4e8497 2161 */
c0ade690 2162static int
1f07f686 2163hammer_setup_child_callback(hammer_record_t rec, void *data)
b84de5af 2164{
7a61b85d 2165 hammer_flush_group_t flg;
1f07f686
MD
2166 hammer_inode_t target_ip;
2167 hammer_inode_t ip;
2168 int r;
2169
2170 /*
3214ade6
MD
2171 * Records deleted or committed by the backend are ignored.
2172 * Note that the flush detects deleted frontend records at
2173 * multiple points to deal with races. This is just the first
2174 * line of defense. The only time HAMMER_RECF_DELETED_FE cannot
2175 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
2176 * messes up link-count calculations.
7bc5b8c2 2177 *
3214ade6
MD
2178 * NOTE: Don't get confused between record deletion and, say,
2179 * directory entry deletion. The deletion of a directory entry
2180 * which is on-media has nothing to do with the record deletion
2181 * flags.
1f07f686 2182 */
3214ade6
MD
2183 if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
2184 HAMMER_RECF_COMMITTED)) {
ecca949a 2185 if (rec->flush_state == HAMMER_FST_FLUSH) {
7a61b85d 2186 KKASSERT(rec->flush_group == rec->ip->flush_group);
ecca949a
MD
2187 r = 1;
2188 } else {
2189 r = 0;
2190 }
2191 return(r);
2192 }
1f07f686
MD
2193
2194 /*
2195 * If the record is in an idle state it has no dependancies and
2196 * can be flushed.
2197 */
2198 ip = rec->ip;
7a61b85d 2199 flg = ip->flush_group;
1f07f686
MD
2200 r = 0;
2201
2202 switch(rec->flush_state) {
2203 case HAMMER_FST_IDLE:
2204 /*
7a61b85d 2205 * The record has no setup dependancy, we can flush it.
1f07f686
MD
2206 */
2207 KKASSERT(rec->target_ip == NULL);
2208 rec->flush_state = HAMMER_FST_FLUSH;
7a61b85d
MD
2209 rec->flush_group = flg;
2210 ++flg->refs;
b84de5af 2211 hammer_ref(&rec->lock);
1f07f686
MD
2212 r = 1;
2213 break;
2214 case HAMMER_FST_SETUP:
2215 /*
7a61b85d
MD
2216 * The record has a setup dependancy. These are typically
2217 * directory entry adds and deletes. Such entries will be
7b6ccb11
MD
2218 * flushed when their inodes are flushed so we do not
2219 * usually have to add them to the flush here. However,
2220 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
2221 * it is asking us to flush this record (and it).
1f07f686
MD
2222 */
2223 target_ip = rec->target_ip;
2224 KKASSERT(target_ip != NULL);
2225 KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);
7a61b85d
MD
2226
2227 /*
2228 * If the target IP is already flushing in our group
5c8d05e2
MD
2229 * we could associate the record, but target_ip has
2230 * already synced ino_data to sync_ino_data and we
2231 * would also have to adjust nlinks. Plus there are
2232 * ordering issues for adds and deletes.
2233 *
2234 * Reflush downward if this is an ADD, and upward if
2235 * this is a DEL.
7a61b85d 2236 */
1f07f686 2237 if (target_ip->flush_state == HAMMER_FST_FLUSH) {
5c8d05e2
MD
2238 if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
2239 ip->flags |= HAMMER_INODE_REFLUSH;
2240 else
1f07f686 2241 target_ip->flags |= HAMMER_INODE_REFLUSH;
7a61b85d
MD
2242 break;
2243 }
2244
2245 /*
2246 * Target IP is not yet flushing. This can get complex
2247 * because we have to be careful about the recursion.
7b6ccb11
MD
2248 *
2249 * Directories create an issue for us in that if a flush
2250 * of a directory is requested the expectation is to flush
2251 * any pending directory entries, but this will cause the
2252 * related inodes to recursively flush as well. We can't
2253 * really defer the operation so just get as many as we
2254 * can and
7a61b85d 2255 */
7b6ccb11 2256#if 0
7a61b85d 2257 if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
7b6ccb11 2258 (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
7a61b85d 2259 /*
7b6ccb11
MD
2260 * We aren't reclaiming and the target ip was not
2261 * previously prevented from flushing due to this
2262 * record dependancy. Do not flush this record.
7a61b85d
MD
2263 */
2264 /*r = 0;*/
7b6ccb11
MD
2265 } else
2266#endif
2267 if (flg->total_count + flg->refs >
7a61b85d
MD
2268 ip->hmp->undo_rec_limit) {
2269 /*
2270 * Our flush group is over-full and we risk blowing
2271 * out the UNDO FIFO. Stop the scan, flush what we
2272 * have, then reflush the directory.
2273 *
2274 * The directory may be forced through multiple
2275 * flush groups before it can be completely
2276 * flushed.
2277 */
4889cbd4
MD
2278 ip->flags |= HAMMER_INODE_RESIGNAL |
2279 HAMMER_INODE_REFLUSH;
7a61b85d 2280 r = -1;
1f07f686
MD
2281 } else if (rec->type == HAMMER_MEM_RECORD_ADD) {
2282 /*
2283 * If the target IP is not flushing we can force
2284 * it to flush, even if it is unable to write out
2285 * any of its own records we have at least one in
2286 * hand that we CAN deal with.
2287 */
2288 rec->flush_state = HAMMER_FST_FLUSH;
7a61b85d
MD
2289 rec->flush_group = flg;
2290 ++flg->refs;
1f07f686 2291 hammer_ref(&rec->lock);
7a61b85d 2292 hammer_flush_inode_core(target_ip, flg,
1f07f686
MD
2293 HAMMER_FLUSH_RECURSION);
2294 r = 1;
2295 } else {
2296 /*
e8599db1
MD
2297 * General or delete-on-disk record.
2298 *
2299 * XXX this needs help. If a delete-on-disk we could
2300 * disconnect the target. If the target has its own
2301 * dependancies they really need to be flushed.
1f07f686
MD
2302 *
2303 * XXX
2304 */
2305 rec->flush_state = HAMMER_FST_FLUSH;
7a61b85d
MD
2306 rec->flush_group = flg;
2307 ++flg->refs;
1f07f686 2308 hammer_ref(&rec->lock);
7a61b85d 2309 hammer_flush_inode_core(target_ip, flg,
1f07f686
MD
2310 HAMMER_FLUSH_RECURSION);
2311 r = 1;
2312 }
2313 break;
2314 case HAMMER_FST_FLUSH:
2315 /*
e3c8589c
MD
2316 * The record could be part of a previous flush group if the
2317 * inode is a directory (the record being a directory entry).
2318 * Once the flush group was closed a hammer_test_inode()
2319 * function can cause a new flush group to be setup, placing
2320 * the directory inode itself in a new flush group.
2321 *
2322 * When associated with a previous flush group we count it
2323 * as if it were in our current flush group, since it will
2324 * effectively be flushed by the time we flush our current
2325 * flush group.
1f07f686 2326 */
e3c8589c
MD
2327 KKASSERT(
2328 rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
2329 rec->flush_group == flg);
1f07f686
MD
2330 r = 1;
2331 break;
b84de5af 2332 }
1f07f686 2333 return(r);
b84de5af
MD
2334}
2335
7a61b85d 2336#if 0
b84de5af 2337/*
525aad3a
MD
2338 * This version just moves records already in a flush state to the new
2339 * flush group and that is it.
2340 */
2341static int
2342hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
2343{
2344 hammer_inode_t ip = rec->ip;
2345
2346 switch(rec->flush_state) {
2347 case HAMMER_FST_FLUSH:
7a61b85d 2348 KKASSERT(rec->flush_group == ip->flush_group);
525aad3a
MD
2349 break;
2350 default:
2351 break;
2352 }
2353 return(0);
2354}
7a61b85d 2355#endif
525aad3a
MD
2356
2357/*
7a61b85d 2358 * Wait for a previously queued flush to complete.
cdb6e4e6
MD
2359 *
2360 * If a critical error occured we don't try to wait.
b84de5af
MD
2361 */
2362void
2363hammer_wait_inode(hammer_inode_t ip)
2364{
7a61b85d 2365 hammer_flush_group_t flg;
ddfdf542 2366
7a61b85d 2367 flg = NULL;
cdb6e4e6 2368 if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
cdb6e4e6
MD
2369 while (ip->flush_state != HAMMER_FST_IDLE &&
2370 (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
f153644d
MD
2371 if (ip->flush_state == HAMMER_FST_SETUP)
2372 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
526fce4d
MD
2373
2374 /*
2375 * If the inode was already being flushed its flg
2376 * may not have been queued to the backend. We have
2377 * to make sure it gets queued or we can wind up
2378 * blocked or deadlocked (particularly if we are
2379 * the vnlru thread).
2380 */
2381 KKASSERT(ip->flush_group);
2382 if (ip->flush_group->closed == 0) {
2383 kprintf("hammer: debug: forcing async "
2384 "flush ip %016jx\n",
2385 (intmax_t)ip->obj_id);
2386 hammer_flusher_async(ip->hmp, ip->flush_group);
2387 }
f153644d
MD
2388 if (ip->flush_state != HAMMER_FST_IDLE) {
2389 ip->flags |= HAMMER_INODE_FLUSHW;
2390 tsleep(&ip->flags, 0, "hmrwin", 0);
2391 }
cdb6e4e6 2392 }
b84de5af
MD
2393 }
2394}
2395
2396/*
2397 * Called by the backend code when a flush has been completed.
2398 * The inode has already been removed from the flush list.
2399 *
2400 * A pipelined flush can occur, in which case we must re-enter the
2401 * inode on the list and re-copy its fields.
2402 */
2403void
cdb6e4e6 2404hammer_flush_inode_done(hammer_inode_t ip, int error)
b84de5af 2405{
af209b0f
MD
2406 hammer_mount_t hmp;
2407 int dorel;
1955afa7 2408
b84de5af
MD
2409 KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
2410
af209b0f
MD
2411 hmp = ip->hmp;
2412
1f07f686 2413 /*
5c667a24
MD
2414 * Auto-reflush if the backend could not completely flush
2415 * the inode. This fixes a case where a deferred buffer flush
2416 * could cause fsync to return early.
2417 */
2418 if (ip->sync_flags & HAMMER_INODE_MODMASK)
2419 ip->flags |= HAMMER_INODE_REFLUSH;
2420
2421 /*
1f07f686 2422 * Merge left-over flags back into the frontend and fix the state.
a9d52b76 2423 * Incomplete truncations are retained by the backend.
1f07f686 2424 */
cdb6e4e6 2425 ip->error = error;
a9d52b76
MD
2426 ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
2427 ip->sync_flags &= HAMMER_INODE_TRUNCATED;
1f07f686
MD
2428
2429 /*
2430 * The backend may have adjusted nlinks, so if the adjusted nlinks
47f363f1 2431 * does not match the fronttend set the frontend's DDIRTY flag again.
1f07f686 2432 */
11ad5ade
MD
2433 if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
2434 ip->flags |= HAMMER_INODE_DDIRTY;
b84de5af 2435
b84de5af 2436 /*
a7e9bef1 2437 * Fix up the dirty buffer status.
4e17f465 2438 */
0832c9bb 2439 if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
1f07f686 2440 ip->flags |= HAMMER_INODE_BUFS;
1f07f686 2441 }
73896937 2442 hammer_redo_fifo_end_flush(ip);
1f07f686
MD
2443
2444 /*
2445 * Re-set the XDIRTY flag if some of the inode's in-memory records
2446 * could not be flushed.
2447 */
0832c9bb
MD
2448 KKASSERT((RB_EMPTY(&ip->rec_tree) &&
2449 (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
2450 (!RB_EMPTY(&ip->rec_tree) &&
2451 (ip->flags & HAMMER_INODE_XDIRTY) != 0));
4e17f465
MD
2452
2453 /*
2454 * Do not lose track of inodes which no longer have vnode
2455 * assocations, otherwise they may never get flushed again.
35a49944
MD
2456 *
2457 * The reflush flag can be set superfluously, causing extra pain
2458 * for no reason. If the inode is no longer modified it no longer
2459 * needs to be flushed.
4e17f465 2460 */
35a49944
MD
2461 if (ip->flags & HAMMER_INODE_MODMASK) {
2462 if (ip->vp == NULL)
2463 ip->flags |= HAMMER_INODE_REFLUSH;
2464 } else {
2465 ip->flags &= ~HAMMER_INODE_REFLUSH;
2466 }
4e17f465
MD
2467
2468 /*
7a61b85d 2469 * Adjust the flush state.
4e17f465 2470 */
06ad81ff 2471 if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
7a61b85d
MD
2472 /*
2473 * We were unable to flush out all our records, leave the
2474 * inode in a flush state and in the current flush group.
d7e278bb 2475 * The flush group will be re-run.
7a61b85d 2476 *
d7e278bb
MD
2477 * This occurs if the UNDO block gets too full or there is
2478 * too much dirty meta-data and allows the flusher to
2479 * finalize the UNDO block and then re-flush.
7a61b85d 2480 */
06ad81ff 2481 ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
af209b0f 2482 dorel = 0;
7a61b85d
MD
2483 } else {
2484 /*
2485 * Remove from the flush_group
2486 */
ff003b11 2487 RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
7a61b85d
MD
2488 ip->flush_group = NULL;
2489
e0092341 2490#if 0
7a61b85d
MD
2491 /*
2492 * Clean up the vnode ref and tracking counts.
2493 */
2494 if (ip->flags & HAMMER_INODE_VHELD) {
2495 ip->flags &= ~HAMMER_INODE_VHELD;
2496 vrele(ip->vp);
2497 }
e0092341 2498#endif
7a61b85d
MD
2499 --hmp->count_iqueued;
2500 --hammer_count_iqueued;
2501
2502 /*
2503 * And adjust the state.
2504 */
2505 if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
2506 ip->flush_state = HAMMER_FST_IDLE;
2507 dorel = 1;
2508 } else {
2509 ip->flush_state = HAMMER_FST_SETUP;
2510 dorel = 0;
2511 }
b84de5af 2512
7a61b85d
MD
2513 /*
2514 * If the frontend is waiting for a flush to complete,
2515 * wake it up.
2516 */
2517 if (ip->flags & HAMMER_INODE_FLUSHW) {
2518 ip->flags &= ~HAMMER_INODE_FLUSHW;
2519 wakeup(&ip->flags);
2520 }
af209b0f 2521
d7e278bb
MD
2522 /*
2523 * If the frontend made more changes and requested another
2524 * flush, then try to get it running.
2525 *
2526 * Reflushes are aborted when the inode is errored out.
2527 */
2528 if (ip->flags & HAMMER_INODE_REFLUSH) {
2529 ip->flags &= ~HAMMER_INODE_REFLUSH;
2530 if (ip->flags & HAMMER_INODE_RESIGNAL) {
2531 ip->flags &= ~HAMMER_INODE_RESIGNAL;
2532 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
2533 } else {
2534 hammer_flush_inode(ip, 0);
2535 }
0729c8c8 2536 }
4e17f465
MD
2537 }
2538
2539 /*
7b6ccb11
MD
2540 * If we have no parent dependancies we can clear CONN_DOWN
2541 */
2542 if (TAILQ_EMPTY(&ip->target_list))
2543 ip->flags &= ~HAMMER_INODE_CONN_DOWN;
2544
2545 /*
e63644f0
MD
2546 * If the inode is now clean drop the space reservation.
2547 */
2548 if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
2549 (ip->flags & HAMMER_INODE_RSV_INODES)) {
2550 ip->flags &= ~HAMMER_INODE_RSV_INODES;
af209b0f 2551 --hmp->rsv_inodes;
e63644f0
MD
2552 }
2553
1f07f686
MD
2554 if (dorel)
2555 hammer_rel_inode(ip, 0);
b84de5af
MD
2556}
2557
2558/*
2559 * Called from hammer_sync_inode() to synchronize in-memory records
2560 * to the media.
2561 */
2562static int
2563hammer_sync_record_callback(hammer_record_t record, void *data)
c0ade690 2564{
4e17f465
MD
2565 hammer_cursor_t cursor = data;
2566 hammer_transaction_t trans = cursor->trans;
6c1f89f4 2567 hammer_mount_t hmp = trans->hmp;
c0ade690
MD
2568 int error;
2569
b84de5af 2570 /*
1f07f686 2571 * Skip records that do not belong to the current flush.
b84de5af 2572 */
47637bff 2573 ++hammer_stats_record_iterations;
1f07f686 2574 if (record->flush_state != HAMMER_FST_FLUSH)
b84de5af 2575 return(0);
47637bff 2576
1f07f686
MD
2577#if 1
2578 if (record->flush_group != record->ip->flush_group) {
7a61b85d 2579 kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
fc73edd8
MD
2580 if (hammer_debug_critical)
2581 Debugger("blah2");
1f07f686
MD
2582 return(0);
2583 }
2584#endif
2585 KKASSERT(record->flush_group == record->ip->flush_group);
d36ec43b
MD
2586
2587 /*
2588 * Interlock the record using the BE flag. Once BE is set the
2589 * frontend cannot change the state of FE.
2590 *
2591 * NOTE: If FE is set prior to us setting BE we still sync the
2592 * record out, but the flush completion code converts it to
2593 * a delete-on-disk record instead of destroying it.
2594 */
4e17f465 2595 KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
d36ec43b
MD
2596 record->flags |= HAMMER_RECF_INTERLOCK_BE;
2597
2598 /*
3214ade6 2599 * The backend has already disposed of the record.
47637bff 2600 */
3214ade6 2601 if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
47637bff
MD
2602 error = 0;
2603 goto done;
2604 }
2605
2606 /*
98f7132d 2607 * If the whole inode is being deleting all on-disk records will
930bf163
MD
2608 * be deleted very soon, we can't sync any new records to disk
2609 * because they will be deleted in the same transaction they were
2610 * created in (delete_tid == create_tid), which will assert.
2611 *
2612 * XXX There may be a case with RECORD_ADD with DELETED_FE set
2613 * that we currently panic on.
98f7132d
MD
2614 */
2615 if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
930bf163 2616 switch(record->type) {
47637bff
MD
2617 case HAMMER_MEM_RECORD_DATA:
2618 /*
2619 * We don't have to do anything, if the record was
2620 * committed the space will have been accounted for
2621 * in the blockmap.
2622 */
2623 /* fall through */
930bf163 2624 case HAMMER_MEM_RECORD_GENERAL:
3214ade6
MD
2625 /*
2626 * Set deleted-by-backend flag. Do not set the
2627 * backend committed flag, because we are throwing
2628 * the record away.
2629 */
98f7132d 2630 record->flags |= HAMMER_RECF_DELETED_BE;
3214ade6 2631 ++record->ip->rec_generation;
930bf163
MD
2632 error = 0;
2633 goto done;
2634 case HAMMER_MEM_RECORD_ADD:
2635 panic("hammer_sync_record_callback: illegal add "
2636 "during inode deletion record %p", record);
2637 break; /* NOT REACHED */
2638 case HAMMER_MEM_RECORD_INODE:
2639 panic("hammer_sync_record_callback: attempt to "
2640 "sync inode record %p?", record);
2641 break; /* NOT REACHED */
2642 case HAMMER_MEM_RECORD_DEL:
2643 /*
2644 * Follow through and issue the on-disk deletion
98f7132d 2645 */
930bf163 2646 break;
98f7132d 2647 }
98f7132d
MD
2648 }
2649
2650 /*
7bc5b8c2
MD
2651 * If DELETED_FE is set special handling is needed for directory
2652 * entries. Dependant pieces related to the directory entry may
2653 * have already been synced to disk. If this occurs we have to
2654 * sync the directory entry and then change the in-memory record
2655 * from an ADD to a DELETE to cover the fact that it's been
2656 * deleted by the frontend.
2657 *
2658 * A directory delete covering record (MEM_RECORD_DEL) can never
2659 * be deleted by the frontend.
2660 *
2661 * Any other record type (aka DATA) can be deleted by the frontend.
2662 * XXX At the moment the flusher must skip it because there may
2663 * be another data record in the flush group for the same block,
2664 * meaning that some frontend data changes can leak into the backend's
2665 * synchronization point.
d36ec43b 2666 */
1f07f686 2667 if (record->flags & HAMMER_RECF_DELETED_FE) {
e8599db1 2668 if (record->type == HAMMER_MEM_RECORD_ADD) {
3214ade6
MD
2669 /*
2670 * Convert a front-end deleted directory-add to
2671 * a directory-delete entry later.
2672 */
e8599db1
MD
2673 record->flags |= HAMMER_RECF_CONVERT_DELETE;
2674 } else {
3214ade6
MD
2675 /*
2676 * Dispose of the record (race case). Mark as
2677 * deleted by backend (and not committed).
2678 */
e8599db1 2679 KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
7bc5b8c2 2680 record->flags |= HAMMER_RECF_DELETED_BE;
3214ade6 2681 ++record->ip->rec_generation;
7bc5b8c2
MD
2682 error = 0;
2683 goto done;
e8599db1 2684 }
1f07f686 2685 }
b84de5af
MD
2686
2687 /*
2688 * Assign the create_tid for new records. Deletions already
2689 * have the record's entire key properly set up.
2690 */
3214ade6 2691 if (record->type != HAMMER_MEM_RECORD_DEL) {
11ad5ade 2692 record->leaf.base.create_tid = trans->tid;
dd94f1b1 2693 record->leaf.create_ts = trans->time32;
3214ade6 2694 }
47f363f1
MD
2695
2696 /*
2697 * This actually moves the record to the on-media B-Tree. We
2698 * must also generate REDO_TERM entries in the UNDO/REDO FIFO
2699 * indicating that the related REDO_WRITE(s) have been committed.
2700 *
2701 * During recovery any REDO_TERM's within the nominal recovery span
2702 * are ignored since the related meta-data is being undone, causing
2703 * any matching REDO_WRITEs to execute. The REDO_TERMs outside
2704 * the nominal recovery span will match against REDO_WRITEs and
2705 * prevent them from being executed (because the meta-data has
2706 * already been synchronized).
2707 */
2708 if (record->flags & HAMMER_RECF_REDO) {
2709 KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
2710 hammer_generate_redo(trans, record->ip,
2711 record->leaf.base.key -
2712 record->leaf.data_len,
2713 HAMMER_REDO_TERM_WRITE,
2714 NULL,
2715 record->leaf.data_len);
2716 }
4e17f465
MD
2717 for (;;) {
2718 error = hammer_ip_sync_record_cursor(cursor, record);
2719 if (error != EDEADLK)
2720 break;
2721 hammer_done_cursor(cursor);
2722 error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
2723 record->ip);
2724 if (error)
2725 break;
2726 }
2727 record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
c0ade690 2728
cdb6e4e6 2729 if (error)
b3deaf57 2730 error = -error;
98f7132d 2731done:
d36ec43b 2732 hammer_flush_record_done(record, error);
6c1f89f4
MD
2733
2734 /*
2735 * Do partial finalization if we have built up too many dirty
2736 * buffers. Otherwise a buffer cache deadlock can occur when
2737 * doing things like creating tens of thousands of tiny files.
2738 *
842e7a70
MD
2739 * We must release our cursor lock to avoid a 3-way deadlock
2740 * due to the exclusive sync lock the finalizer must get.
c9ce54d6
MD
2741 *
2742 * WARNING: See warnings in hammer_unlock_cursor() function.
6c1f89f4 2743 */
842e7a70 2744 if (hammer_flusher_meta_limit(hmp)) {
982be4bf 2745 hammer_unlock_cursor(cursor);
6c1f89f4 2746 hammer_flusher_finalize(trans, 0);
982be4bf 2747 hammer_lock_cursor(cursor);
842e7a70 2748 }
6c1f89f4 2749
b3deaf57 2750 return(error);
c0ade690
MD
2751}
2752
2753/*
d7e278bb 2754 * Backend function called by the flusher to sync an inode to media.
c0ade690
MD
2755 */
2756int
02325004 2757hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
c0ade690 2758{
4e17f465 2759 struct hammer_cursor cursor;
cb51be26 2760 hammer_node_t tmp_node;
1f07f686
MD
2761 hammer_record_t depend;
2762 hammer_record_t next;
ec4e8497 2763 int error, tmp_error;
1f07f686 2764 u_int64_t nlinks;
c0ade690 2765
1f07f686 2766 if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
d113fda1 2767 return(0);
d113fda1 2768
02325004 2769 error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
4e17f465
MD
2770 if (error)
2771 goto done;
c0ade690
MD
2772
2773 /*
1f07f686
MD
2774 * Any directory records referencing this inode which are not in
2775 * our current flush group must adjust our nlink count for the
47f363f1 2776 * purposes of synchronizating to disk.
1f07f686
MD
2777 *
2778 * Records which are in our flush group can be unlinked from our
c4bae5fd
MD
2779 * inode now, potentially allowing the inode to be physically
2780 * deleted.
bf3b416b
MD
2781 *
2782 * This cannot block.
ec4e8497 2783 */
11ad5ade 2784 nlinks = ip->ino_data.nlinks;
1f07f686
MD
2785 next = TAILQ_FIRST(&ip->target_list);
2786 while ((depend = next) != NULL) {
2787 next = TAILQ_NEXT(depend, target_entry);
2788 if (depend->flush_state == HAMMER_FST_FLUSH &&
7a61b85d 2789 depend->flush_group == ip->flush_group) {
c4bae5fd
MD
2790 /*
2791 * If this is an ADD that was deleted by the frontend
2792 * the frontend nlinks count will have already been
2793 * decremented, but the backend is going to sync its
2794 * directory entry and must account for it. The
2795 * record will be converted to a delete-on-disk when
2796 * it gets synced.
2797 *
2798 * If the ADD was not deleted by the frontend we
2799 * can remove the dependancy from our target_list.
2800 */
2801 if (depend->flags & HAMMER_RECF_DELETED_FE) {
2802 ++nlinks;
2803 } else {
2804 TAILQ_REMOVE(&ip->target_list, depend,
2805 target_entry);
2806 depend->target_ip = NULL;
2807 }
1f07f686 2808 } else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
c4bae5fd 2809 /*
3214ade6
MD
2810 * Not part of our flush group and not deleted by
2811 * the front-end, adjust the link count synced to
2812 * the media (undo what the frontend did when it
2813 * queued the record).
c4bae5fd
MD
2814 */
2815 KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
1f07f686
MD
2816 switch(depend->type) {
2817 case HAMMER_MEM_RECORD_ADD:
2818 --nlinks;
2819 break;
2820 case HAMMER_MEM_RECORD_DEL:
2821 ++nlinks;
2822 break;
e8599db1
MD
2823 default:
2824 break;
1f07f686 2825 }
ec4e8497 2826 }
ec4e8497
MD
2827 }
2828
ec4e8497 2829 /*
1f07f686 2830 * Set dirty if we had to modify the link count.
c0ade690 2831 */
11ad5ade 2832 if (ip->sync_ino_data.nlinks != nlinks) {
1f07f686 2833 KKASSERT((int64_t)nlinks >= 0);
11ad5ade
MD
2834 ip->sync_ino_data.nlinks = nlinks;
2835 ip->sync_flags |= HAMMER_INODE_DDIRTY;
1f07f686 2836 }
b84de5af 2837
4e17f465 2838 /*
869e8f55
MD
2839 * If there is a trunction queued destroy any data past the (aligned)
2840 * truncation point. Userland will have dealt with the buffer
2841 * containing the truncation point for us.
2842 *
2843 * We don't flush pending frontend data buffers until after we've
cb51be26 2844 * dealt with the truncation.
1f07f686 2845 */
869e8f55 2846 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
b84de5af
MD
2847 /*
2848 * Interlock trunc_off. The VOP front-end may continue to
2849 * make adjustments to it while we are blocked.
2850 */
2851 off_t trunc_off;
2852 off_t aligned_trunc_off;
4a2796f3 2853 int blkmask;
c0ade690 2854
b84de5af 2855 trunc_off = ip->sync_trunc_off;
4a2796f3
MD
2856 blkmask = hammer_blocksize(trunc_off) - 1;
2857 aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;
b84de5af
MD
2858
2859 /*
2860 * Delete any whole blocks on-media. The front-end has
2861 * already cleaned out any partial block and made it
2862 * pending. The front-end may have updated trunc_off
47637bff 2863 * while we were blocked so we only use sync_trunc_off.
06ad81ff
MD
2864 *
2865 * This operation can blow out the buffer cache, EWOULDBLOCK
a9d52b76
MD
2866 * means we were unable to complete the deletion. The
2867 * deletion will update sync_trunc_off in that case.
b84de5af 2868 */
4e17f465 2869 error = hammer_ip_delete_range(&cursor, ip,
b84de5af 2870 aligned_trunc_off,
06ad81ff
MD
2871 0x7FFFFFFFFFFFFFFFLL, 2);
2872 if (error == EWOULDBLOCK) {
2873 ip->flags |= HAMMER_INODE_WOULDBLOCK;
2874 error = 0;
2875 goto defer_buffer_flush;
2876 }
2877
b84de5af 2878 if (error)
cdb6e4e6 2879 goto done;
47637bff
MD
2880
2881 /*
47f363f1
MD
2882 * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
2883 *
2884 * XXX we do this even if we did not previously generate
2885 * a REDO_TRUNC record. This operation may enclosed the
2886 * range for multiple prior truncation entries in the REDO
2887 * log.
2888 */
73896937
MD
2889 if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
2890 (ip->flags & HAMMER_INODE_RDIRTY)) {
47f363f1
MD
2891 hammer_generate_redo(trans, ip, aligned_trunc_off,
2892 HAMMER_REDO_TERM_TRUNC,
2893 NULL, 0);
2894 }
2895
2896 /*
47637bff 2897 * Clear the truncation flag on the backend after we have
47f363f1 2898 * completed the deletions. Backend data is now good again
47637bff 2899 * (including new records we are about to sync, below).
cb51be26
MD
2900 *
2901 * Leave sync_trunc_off intact. As we write additional
2902 * records the backend will update sync_trunc_off. This
2903 * tells the backend whether it can skip the overwrite
2904 * test. This should work properly even when the backend
2905 * writes full blocks where the truncation point straddles
2906 * the block because the comparison is against the base
2907 * offset of the record.
47637bff 2908 */
b84de5af 2909 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
cb51be26 2910 /* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
1f07f686
MD
2911 } else {
2912 error = 0;
f3b0f382
MD
2913 }
2914
1f07f686
MD
2915 /*
2916 * Now sync related records. These will typically be directory
6c1f89f4 2917 * entries, records tracking direct-writes, or delete-on-disk records.
1f07f686
MD
2918 */
2919 if (error == 0) {
2920 tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
4e17f465 2921 hammer_sync_record_callback, &cursor);
1f07f686
MD
2922 if (tmp_error < 0)
2923 tmp_error = -error;
2924 if (tmp_error)
2925 error = tmp_error;
2926 }
bcac4bbb 2927 hammer_cache_node(&ip->cache[1], cursor.node);
cb51be26
MD
2928
2929 /*
43c665ae
MD
2930 * Re-seek for inode update, assuming our cache hasn't been ripped
2931 * out from under us.
cb51be26
MD
2932 */
2933 if (error == 0) {
4c286c36 2934 tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
cb51be26 2935 if (tmp_node) {
5fa5c92f
MD
2936 hammer_cursor_downgrade(&cursor);
2937 hammer_lock_sh(&tmp_node->lock);
43c665ae
MD
2938 if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
2939 hammer_cursor_seek(&cursor, tmp_node, 0);
5fa5c92f 2940 hammer_unlock(&tmp_node->lock);
cb51be26
MD
2941 hammer_rel_node(tmp_node);
2942 }
2943 error = 0;
2944 }
1f07f686
MD
2945
2946 /*
869e8f55
MD
2947 * If we are deleting the inode the frontend had better not have
2948 * any active references on elements making up the inode.
a9d52b76
MD
2949 *
2950 * The call to hammer_ip_delete_clean() cleans up auxillary records
2951 * but not DB or DATA records. Those must have already been deleted
2952 * by the normal truncation mechanic.
1f07f686 2953 */
11ad5ade 2954 if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
869e8f55
MD
2955 RB_EMPTY(&ip->rec_tree) &&
2956 (ip->sync_flags & HAMMER_INODE_DELETING) &&
2957 (ip->flags & HAMMER_INODE_DELETED) == 0) {
2958 int count1 = 0;
1f07f686 2959
a9d52b76 2960 error = hammer_ip_delete_clean(&cursor, ip, &count1);
869e8f55 2961 if (error == 0) {
06ad81ff 2962 ip->flags |= HAMMER_INODE_DELETED;
869e8f55
MD
2963 ip->sync_flags &= ~HAMMER_INODE_DELETING;
2964 ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
2965 KKASSERT(RB_EMPTY(&ip->rec_tree));
1f07f686 2966
869e8f55
MD
2967 /*
2968 * Set delete_tid in both the frontend and backend
2969 * copy of the inode record. The DELETED flag handles
47f363f1 2970 * this, do not set DDIRTY.
869e8f55 2971 */
02325004
MD
2972 ip->ino_leaf.base.delete_tid = trans->tid;
2973 ip->sync_ino_leaf.base.delete_tid = trans->tid;
2974 ip->ino_leaf.delete_ts = trans->time32;
2975 ip->sync_ino_leaf.delete_ts = trans->time32;
dd94f1b1 2976
1f07f686 2977
869e8f55
MD
2978 /*
2979 * Adjust the inode count in the volume header
2980 */
02325004 2981 hammer_sync_lock_sh(trans);
f36a9737 2982 if (ip->flags & HAMMER_INODE_ONDISK) {
02325004
MD
2983 hammer_modify_volume_field(trans,
2984 trans->rootvol,
f36a9737
MD
2985 vol0_stat_inodes);
2986 --ip->hmp->rootvol->ondisk->vol0_stat_inodes;
02325004 2987 hammer_modify_volume_done(trans->rootvol);
f36a9737 2988 }
02325004 2989 hammer_sync_unlock(trans);
869e8f55 2990 }
1f07f686 2991 }
b84de5af 2992
b84de5af 2993 if (error)
cdb6e4e6
MD
2994 goto done;
2995 ip->sync_flags &= ~HAMMER_INODE_BUFS;
c0ade690 2996
06ad81ff 2997defer_buffer_flush:
c0ade690
MD
2998 /*
2999 * Now update the inode's on-disk inode-data and/or on-disk record.
b84de5af 3000 * DELETED and ONDISK are managed only in ip->flags.
06ad81ff
MD
3001 *
3002 * In the case of a defered buffer flush we still update the on-disk
3003 * inode to satisfy visibility requirements if there happen to be
3004 * directory dependancies.
c0ade690 3005 */
b84de5af 3006 switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
c0ade690
MD
3007 case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
3008 /*
3009 * If deleted and on-disk, don't set any additional flags.
3010 * the delete flag takes care of things.
869e8f55
MD
3011 *
3012 * Clear flags which may have been set by the frontend.
c0ade690 3013 */
ddfdf542 3014 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
9192654c 3015 HAMMER_INODE_SDIRTY |
ddfdf542 3016 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
869e8f55 3017 HAMMER_INODE_DELETING);
c0ade690
MD
3018 break;
3019 case HAMMER_INODE_DELETED:
3020 /*
3021 * Take care of the case where a deleted inode was never
3022 * flushed to the disk in the first place.
869e8f55
MD
3023 *
3024 * Clear flags which may have been set by the frontend.
c0ade690 3025 */
ddfdf542 3026 ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
9192654c 3027 HAMMER_INODE_SDIRTY |
ddfdf542 3028 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
869e8f55 3029 HAMMER_INODE_DELETING);
d26d0ae9 3030 while (RB_ROOT(&ip->rec_tree)) {
d36ec43b
MD
3031 hammer_record_t record = RB_ROOT(&ip->rec_tree);
3032 hammer_ref(&record->lock);
250aec18 3033 KKASSERT(hammer_oneref(&record->lock));
d36ec43b 3034 record->flags |= HAMMER_RECF_DELETED_BE;
3214ade6 3035 ++record->ip->rec_generation;
d36ec43b 3036 hammer_rel_mem_record(record);
d26d0ae9 3037 }
c0ade690
MD
3038 break;
3039 case HAMMER_INODE_ONDISK:
3040 /*
3041 * If already on-disk, do not set any additional flags.
3042 */
3043 break;
3044 default:
3045 /*
ddfdf542
MD
3046 * If not on-disk and not deleted, set DDIRTY to force
3047 * an initial record to be written.
b84de5af 3048 *
ddfdf542 3049 * Also set the create_tid in both the frontend and backend
b84de5af 3050 * copy of the inode record.
c0ade690 3051 */
02325004
MD
3052 ip->ino_leaf.base.create_tid = trans->tid;
3053 ip->ino_leaf.create_ts = trans->time32;
3054 ip->sync_ino_leaf.base.create_tid = trans->tid;
3055 ip->sync_ino_leaf.create_ts = trans->time32;
11ad5ade 3056 ip->sync_flags |= HAMMER_INODE_DDIRTY;
c0ade690
MD
3057 break;
3058 }
3059
3060 /*
47f363f1 3061 * If DDIRTY or SDIRTY is set, write out a new record.
9192654c
MD
3062 * If the inode is already on-disk the old record is marked as
3063 * deleted.
d113fda1
MD
3064 *
3065 * If DELETED is set hammer_update_inode() will delete the existing
3066 * record without writing out a new one.
3067 *
3068 * If *ONLY* the ITIMES flag is set we can update the record in-place.
c0ade690 3069 */
b84de5af 3070 if (ip->flags & HAMMER_INODE_DELETED) {
4e17f465 3071 error = hammer_update_inode(&cursor, ip);
b84de5af 3072 } else
9192654c 3073 if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
ddfdf542 3074 (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
4e17f465 3075 error = hammer_update_itimes(&cursor, ip);
d113fda1 3076 } else
9192654c
MD
3077 if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
3078 HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
4e17f465 3079 error = hammer_update_inode(&cursor, ip);
c0ade690 3080 }
4e17f465 3081done:
cdb6e4e6
MD
3082 if (error) {
3083 hammer_critical_error(ip->hmp, ip, error,
3084 "while syncing inode");
3085 }
4e17f465 3086 hammer_done_cursor(&cursor);
c0ade690 3087 return(error);
8cd0a023
MD
3088}
3089
1f07f686
MD
3090/*
3091 * This routine is called when the OS is no longer actively referencing
3092 * the inode (but might still be keeping it cached), or when releasing
3093 * the last reference to an inode.
3094 *
3095 * At this point if the inode's nlinks count is zero we want to destroy
3096 * it, which may mean destroying it on-media too.
3097 */
3bf2d80a 3098void
e8599db1 3099hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
1f07f686 3100{
e8599db1
MD
3101 struct vnode *vp;
3102
1f07f686 3103 /*
c4bae5fd
MD
3104 * Set the DELETING flag when the link count drops to 0 and the
3105 * OS no longer has any opens on the inode.
3106 *
3107 * The backend will clear DELETING (a mod flag) and set DELETED
3108 * (a state flag) when it is actually able to perform the
3109 * operation.
35a49944
MD
3110 *
3111 * Don't reflag the deletion if the flusher is currently syncing
3112 * one that was already flagged. A previously set DELETING flag
3113 * may bounce around flags and sync_flags until the operation is
3114 * completely done.
ef11b6e1
MD
3115 *
3116 * Do not attempt to modify a snapshot inode (one set to read-only).
1f07f686 3117 */
11ad5ade 3118 if (ip->ino_data.nlinks == 0 &&
ef11b6e1 3119 ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
e8599db1
MD
3120 ip->flags |= HAMMER_INODE_DELETING;
3121 ip->flags |= HAMMER_INODE_TRUNCATED;
3122 ip->trunc_off = 0;
3123 vp = NULL;
3124 if (getvp) {
3125 if (hammer_get_vnode(ip, &vp) != 0)
3126 return;
3127 }
29ce0677
MD
3128
3129 /*
29ce0677
MD
3130 * Final cleanup
3131 */
6362a262
MD
3132 if (ip->vp)
3133 nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0);
3134 if (getvp)
e8599db1 3135 vput(vp);
1f07f686 3136 }
1f07f686
MD
3137}
3138
3bf2d80a 3139/*
7b6ccb11
MD
3140 * After potentially resolving a dependancy the inode is tested
3141 * to determine whether it needs to be reflushed.
3bf2d80a 3142 */
1f07f686
MD
3143void
3144hammer_test_inode(hammer_inode_t ip)
3145{
3146 if (ip->flags & HAMMER_INODE_REFLUSH) {
3147 ip->flags &= ~HAMMER_INODE_REFLUSH;
3148 hammer_ref(&ip->lock);
3bf2d80a
MD
3149 if (ip->flags & HAMMER_INODE_RESIGNAL) {
3150 ip->flags &= ~HAMMER_INODE_RESIGNAL;
3151 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
3152 } else {
3153 hammer_flush_inode(ip, 0);
3154 }
1f07f686
MD
3155 hammer_rel_inode(ip, 0);
3156 }
3157}
3158
9f5097dc 3159/*
7bc5b8c2
MD
3160 * Clear the RECLAIM flag on an inode. This occurs when the inode is
3161 * reassociated with a vp or just before it gets freed.
af209b0f 3162 *
82010f9f 3163 * Pipeline wakeups to threads blocked due to an excessive number of
ccf6a64d
MD
3164 * detached inodes. This typically occurs when atime updates accumulate
3165 * while scanning a directory tree.
9f5097dc 3166 */
7bc5b8c2 3167static void
ccf6a64d 3168hammer_inode_wakereclaims(hammer_inode_t ip)
9f5097dc 3169{
7bc5b8c2 3170 struct hammer_reclaim *reclaim;
d99d6bf5 3171 hammer_mount_t hmp = ip->hmp;
d99d6bf5 3172
7bc5b8c2 3173 if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
d99d6bf5 3174 return;
3897d7e9 3175
7bc5b8c2
MD
3176 --hammer_count_reclaiming;
3177 --hmp->inode_reclaims;
3178 ip->flags &= ~HAMMER_INODE_RECLAIM;
9f5097dc 3179
37646115
MD
3180 if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
3181 KKASSERT(reclaim->count > 0);
3182 if (--reclaim->count == 0) {
82010f9f
MD
3183 TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
3184 wakeup(reclaim);
3185 }
9f5097dc
MD
3186 }
3187}
3188
4a2796f3
MD
3189/*
3190 * Setup our reclaim pipeline. We only let so many detached (and dirty)
ccf6a64d
MD
3191 * inodes build up before we start blocking. This routine is called
3192 * if a new inode is created or an inode is loaded from media.
4a2796f3
MD
3193 *
3194 * When we block we don't care *which* inode has finished reclaiming,
ccf6a64d 3195 * as lone as one does.
37646115
MD
3196 *
3197 * The reclaim pipeline is primary governed by the auto-flush which is
3198 * 1/4 hammer_limit_reclaim. We don't want to block if the count is
3199 * less than 1/2 hammer_limit_reclaim. From 1/2 to full count is
3200 * dynamically governed.
4a2796f3
MD
3201 */
3202void
e98f1b96 3203hammer_inode_waitreclaims(hammer_transaction_t trans)
4a2796f3 3204{
e98f1b96 3205 hammer_mount_t hmp = trans->hmp;
4a2796f3 3206 struct hammer_reclaim reclaim;
37646115 3207 int lower_limit;
4a2796f3 3208
e98f1b96 3209 /*
37646115
MD
3210 * Track inode load, delay if the number of reclaiming inodes is
3211 * between 2/4 and 4/4 hammer_limit_reclaim, depending.
e98f1b96
MD
3212 */
3213 if (curthread->td_proc) {
3214 struct hammer_inostats *stats;
e98f1b96
MD
3215
3216 stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
3217 ++stats->count;
3218
3219 if (stats->count > hammer_limit_reclaim / 2)
3220 stats->count = hammer_limit_reclaim / 2;
3221 lower_limit = hammer_limit_reclaim - stats->count;
37646115
MD
3222 if (hammer_debug_general & 0x10000) {
3223 kprintf("pid %5d limit %d\n",
3224 (int)curthread->td_proc->p_pid, lower_limit);
3225 }
e98f1b96 3226 } else {
37646115
MD
3227 lower_limit = hammer_limit_reclaim * 3 / 4;
3228 }
3229 if (hmp->inode_reclaims >= lower_limit) {
3230 reclaim.count = 1;
3231 TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
3232 tsleep(&reclaim, 0, "hmrrcm", hz);
3233 if (reclaim.count > 0)
3234 TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
e98f1b96 3235 }
82010f9f 3236}
4a2796f3 3237
858cc00a
MD
3238/*
3239 * Keep track of reclaim statistics on a per-pid basis using a loose
3240 * 4-way set associative hash table. Collisions inherit the count of
3241 * the previous entry.
3242 *
3243 * NOTE: We want to be careful here to limit the chain size. If the chain
3244 * size is too large a pid will spread its stats out over too many
3245 * entries under certain types of heavy filesystem activity and
3246 * wind up not delaying long enough.
3247 */
e98f1b96
MD
3248static
3249struct hammer_inostats *
3250hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
3251{
3252 struct hammer_inostats *stats;
3253 int delta;
3254 int chain;
fd094a8f 3255 static volatile int iterator; /* we don't care about MP races */
e98f1b96 3256
858cc00a
MD
3257 /*
3258 * Chain up to 4 times to find our entry.
3259 */
e98f1b96
MD