]> git.proxmox.com Git - mirror_zfs-debian.git/blob - module/zfs/dmu_tx.c
projects / mirror_zfs-debian.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Imported Upstream version 0.6.2+git20140204
[mirror_zfs-debian.git] / module / zfs / dmu_tx.c
1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 */
26
27 #include <sys/dmu.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dbuf.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
33 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
34 #include <sys/dsl_pool.h>
35 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
36 #include <sys/spa.h>
37 #include <sys/sa.h>
38 #include <sys/sa_impl.h>
39 #include <sys/zfs_context.h>
40 #include <sys/varargs.h>
41
42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
43 uint64_t arg1, uint64_t arg2);
44
45 dmu_tx_stats_t dmu_tx_stats = {
46 { "dmu_tx_assigned", KSTAT_DATA_UINT64 },
47 { "dmu_tx_delay", KSTAT_DATA_UINT64 },
48 { "dmu_tx_error", KSTAT_DATA_UINT64 },
49 { "dmu_tx_suspended", KSTAT_DATA_UINT64 },
50 { "dmu_tx_group", KSTAT_DATA_UINT64 },
51 { "dmu_tx_how", KSTAT_DATA_UINT64 },
52 { "dmu_tx_memory_reserve", KSTAT_DATA_UINT64 },
53 { "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64 },
54 { "dmu_tx_memory_inflight", KSTAT_DATA_UINT64 },
55 { "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64 },
56 { "dmu_tx_dirty_delay", KSTAT_DATA_UINT64 },
57 { "dmu_tx_dirty_over_max", KSTAT_DATA_UINT64 },
58 { "dmu_tx_quota", KSTAT_DATA_UINT64 },
59 };
60
61 static kstat_t *dmu_tx_ksp;
62
63 dmu_tx_t *
64 dmu_tx_create_dd(dsl_dir_t *dd)
65 {
66 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_PUSHPAGE);
67 tx->tx_dir = dd;
68 if (dd != NULL)
69 tx->tx_pool = dd->dd_pool;
70 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
71 offsetof(dmu_tx_hold_t, txh_node));
72 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
73 offsetof(dmu_tx_callback_t, dcb_node));
74 tx->tx_start = gethrtime();
75 #ifdef DEBUG_DMU_TX
76 refcount_create(&tx->tx_space_written);
77 refcount_create(&tx->tx_space_freed);
78 #endif
79 return (tx);
80 }
81
82 dmu_tx_t *
83 dmu_tx_create(objset_t *os)
84 {
85 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
86 tx->tx_objset = os;
87 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
88 return (tx);
89 }
90
91 dmu_tx_t *
92 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
93 {
94 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
95
96 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
97 tx->tx_pool = dp;
98 tx->tx_txg = txg;
99 tx->tx_anyobj = TRUE;
100
101 return (tx);
102 }
103
104 int
105 dmu_tx_is_syncing(dmu_tx_t *tx)
106 {
107 return (tx->tx_anyobj);
108 }
109
110 int
111 dmu_tx_private_ok(dmu_tx_t *tx)
112 {
113 return (tx->tx_anyobj);
114 }
115
116 static dmu_tx_hold_t *
117 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
118 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
119 {
120 dmu_tx_hold_t *txh;
121 dnode_t *dn = NULL;
122 int err;
123
124 if (object != DMU_NEW_OBJECT) {
125 err = dnode_hold(os, object, tx, &dn);
126 if (err) {
127 tx->tx_err = err;
128 return (NULL);
129 }
130
131 if (err == 0 && tx->tx_txg != 0) {
132 mutex_enter(&dn->dn_mtx);
133 /*
134 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
135 * problem, but there's no way for it to happen (for
136 * now, at least).
137 */
138 ASSERT(dn->dn_assigned_txg == 0);
139 dn->dn_assigned_txg = tx->tx_txg;
140 (void) refcount_add(&dn->dn_tx_holds, tx);
141 mutex_exit(&dn->dn_mtx);
142 }
143 }
144
145 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_PUSHPAGE);
146 txh->txh_tx = tx;
147 txh->txh_dnode = dn;
148 #ifdef DEBUG_DMU_TX
149 txh->txh_type = type;
150 txh->txh_arg1 = arg1;
151 txh->txh_arg2 = arg2;
152 #endif
153 list_insert_tail(&tx->tx_holds, txh);
154
155 return (txh);
156 }
157
158 void
159 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
160 {
161 /*
162 * If we're syncing, they can manipulate any object anyhow, and
163 * the hold on the dnode_t can cause problems.
164 */
165 if (!dmu_tx_is_syncing(tx)) {
166 (void) dmu_tx_hold_object_impl(tx, os,
167 object, THT_NEWOBJECT, 0, 0);
168 }
169 }
170
171 static int
172 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
173 {
174 int err;
175 dmu_buf_impl_t *db;
176
177 rw_enter(&dn->dn_struct_rwlock, RW_READER);
178 db = dbuf_hold_level(dn, level, blkid, FTAG);
179 rw_exit(&dn->dn_struct_rwlock);
180 if (db == NULL)
181 return (SET_ERROR(EIO));
182 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
183 dbuf_rele(db, FTAG);
184 return (err);
185 }
186
187 static void
188 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
189 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
190 {
191 objset_t *os = dn->dn_objset;
192 dsl_dataset_t *ds = os->os_dsl_dataset;
193 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
194 dmu_buf_impl_t *parent = NULL;
195 blkptr_t *bp = NULL;
196 uint64_t space;
197
198 if (level >= dn->dn_nlevels || history[level] == blkid)
199 return;
200
201 history[level] = blkid;
202
203 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
204
205 if (db == NULL || db == dn->dn_dbuf) {
206 ASSERT(level != 0);
207 db = NULL;
208 } else {
209 ASSERT(DB_DNODE(db) == dn);
210 ASSERT(db->db_level == level);
211 ASSERT(db->db.db_size == space);
212 ASSERT(db->db_blkid == blkid);
213 bp = db->db_blkptr;
214 parent = db->db_parent;
215 }
216
217 freeable = (bp && (freeable ||
218 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
219
220 if (freeable)
221 txh->txh_space_tooverwrite += space;
222 else
223 txh->txh_space_towrite += space;
224 if (bp)
225 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
226
227 dmu_tx_count_twig(txh, dn, parent, level + 1,
228 blkid >> epbs, freeable, history);
229 }
230
231 /* ARGSUSED */
232 static void
233 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
234 {
235 dnode_t *dn = txh->txh_dnode;
236 uint64_t start, end, i;
237 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
238 int err = 0;
239 int l;
240
241 if (len == 0)
242 return;
243
244 min_bs = SPA_MINBLOCKSHIFT;
245 max_bs = SPA_MAXBLOCKSHIFT;
246 min_ibs = DN_MIN_INDBLKSHIFT;
247 max_ibs = DN_MAX_INDBLKSHIFT;
248
249 if (dn) {
250 uint64_t history[DN_MAX_LEVELS];
251 int nlvls = dn->dn_nlevels;
252 int delta;
253
254 /*
255 * For i/o error checking, read the first and last level-0
256 * blocks (if they are not aligned), and all the level-1 blocks.
257 */
258 if (dn->dn_maxblkid == 0) {
259 delta = dn->dn_datablksz;
260 start = (off < dn->dn_datablksz) ? 0 : 1;
261 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
262 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
263 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
264 if (err)
265 goto out;
266 delta -= off;
267 }
268 } else {
269 zio_t *zio = zio_root(dn->dn_objset->os_spa,
270 NULL, NULL, ZIO_FLAG_CANFAIL);
271
272 /* first level-0 block */
273 start = off >> dn->dn_datablkshift;
274 if (P2PHASE(off, dn->dn_datablksz) ||
275 len < dn->dn_datablksz) {
276 err = dmu_tx_check_ioerr(zio, dn, 0, start);
277 if (err)
278 goto out;
279 }
280
281 /* last level-0 block */
282 end = (off+len-1) >> dn->dn_datablkshift;
283 if (end != start && end <= dn->dn_maxblkid &&
284 P2PHASE(off+len, dn->dn_datablksz)) {
285 err = dmu_tx_check_ioerr(zio, dn, 0, end);
286 if (err)
287 goto out;
288 }
289
290 /* level-1 blocks */
291 if (nlvls > 1) {
292 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
293 for (i = (start>>shft)+1; i < end>>shft; i++) {
294 err = dmu_tx_check_ioerr(zio, dn, 1, i);
295 if (err)
296 goto out;
297 }
298 }
299
300 err = zio_wait(zio);
301 if (err)
302 goto out;
303 delta = P2NPHASE(off, dn->dn_datablksz);
304 }
305
306 min_ibs = max_ibs = dn->dn_indblkshift;
307 if (dn->dn_maxblkid > 0) {
308 /*
309 * The blocksize can't change,
310 * so we can make a more precise estimate.
311 */
312 ASSERT(dn->dn_datablkshift != 0);
313 min_bs = max_bs = dn->dn_datablkshift;
314 }
315
316 /*
317 * If this write is not off the end of the file
318 * we need to account for overwrites/unref.
319 */
320 if (start <= dn->dn_maxblkid) {
321 for (l = 0; l < DN_MAX_LEVELS; l++)
322 history[l] = -1ULL;
323 }
324 while (start <= dn->dn_maxblkid) {
325 dmu_buf_impl_t *db;
326
327 rw_enter(&dn->dn_struct_rwlock, RW_READER);
328 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
329 rw_exit(&dn->dn_struct_rwlock);
330
331 if (err) {
332 txh->txh_tx->tx_err = err;
333 return;
334 }
335
336 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
337 history);
338 dbuf_rele(db, FTAG);
339 if (++start > end) {
340 /*
341 * Account for new indirects appearing
342 * before this IO gets assigned into a txg.
343 */
344 bits = 64 - min_bs;
345 epbs = min_ibs - SPA_BLKPTRSHIFT;
346 for (bits -= epbs * (nlvls - 1);
347 bits >= 0; bits -= epbs)
348 txh->txh_fudge += 1ULL << max_ibs;
349 goto out;
350 }
351 off += delta;
352 if (len >= delta)
353 len -= delta;
354 delta = dn->dn_datablksz;
355 }
356 }
357
358 /*
359 * 'end' is the last thing we will access, not one past.
360 * This way we won't overflow when accessing the last byte.
361 */
362 start = P2ALIGN(off, 1ULL << max_bs);
363 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
364 txh->txh_space_towrite += end - start + 1;
365
366 start >>= min_bs;
367 end >>= min_bs;
368
369 epbs = min_ibs - SPA_BLKPTRSHIFT;
370
371 /*
372 * The object contains at most 2^(64 - min_bs) blocks,
373 * and each indirect level maps 2^epbs.
374 */
375 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
376 start >>= epbs;
377 end >>= epbs;
378 ASSERT3U(end, >=, start);
379 txh->txh_space_towrite += (end - start + 1) << max_ibs;
380 if (start != 0) {
381 /*
382 * We also need a new blkid=0 indirect block
383 * to reference any existing file data.
384 */
385 txh->txh_space_towrite += 1ULL << max_ibs;
386 }
387 }
388
389 out:
390 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
391 2 * DMU_MAX_ACCESS)
392 err = SET_ERROR(EFBIG);
393
394 if (err)
395 txh->txh_tx->tx_err = err;
396 }
397
398 static void
399 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
400 {
401 dnode_t *dn = txh->txh_dnode;
402 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
403 uint64_t space = mdn->dn_datablksz +
404 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
405
406 if (dn && dn->dn_dbuf->db_blkptr &&
407 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
408 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
409 txh->txh_space_tooverwrite += space;
410 txh->txh_space_tounref += space;
411 } else {
412 txh->txh_space_towrite += space;
413 if (dn && dn->dn_dbuf->db_blkptr)
414 txh->txh_space_tounref += space;
415 }
416 }
417
418 void
419 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
420 {
421 dmu_tx_hold_t *txh;
422
423 ASSERT(tx->tx_txg == 0);
424 ASSERT(len < DMU_MAX_ACCESS);
425 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
426
427 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
428 object, THT_WRITE, off, len);
429 if (txh == NULL)
430 return;
431
432 dmu_tx_count_write(txh, off, len);
433 dmu_tx_count_dnode(txh);
434 }
435
436 static void
437 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
438 {
439 uint64_t blkid, nblks, lastblk;
440 uint64_t space = 0, unref = 0, skipped = 0;
441 dnode_t *dn = txh->txh_dnode;
442 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
443 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
444 int epbs;
445 uint64_t l0span = 0, nl1blks = 0;
446
447 if (dn->dn_nlevels == 0)
448 return;
449
450 /*
451 * The struct_rwlock protects us against dn_nlevels
452 * changing, in case (against all odds) we manage to dirty &
453 * sync out the changes after we check for being dirty.
454 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
455 */
456 rw_enter(&dn->dn_struct_rwlock, RW_READER);
457 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
458 if (dn->dn_maxblkid == 0) {
459 if (off == 0 && len >= dn->dn_datablksz) {
460 blkid = 0;
461 nblks = 1;
462 } else {
463 rw_exit(&dn->dn_struct_rwlock);
464 return;
465 }
466 } else {
467 blkid = off >> dn->dn_datablkshift;
468 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
469
470 if (blkid > dn->dn_maxblkid) {
471 rw_exit(&dn->dn_struct_rwlock);
472 return;
473 }
474 if (blkid + nblks > dn->dn_maxblkid)
475 nblks = dn->dn_maxblkid - blkid + 1;
476
477 }
478 l0span = nblks; /* save for later use to calc level > 1 overhead */
479 if (dn->dn_nlevels == 1) {
480 int i;
481 for (i = 0; i < nblks; i++) {
482 blkptr_t *bp = dn->dn_phys->dn_blkptr;
483 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
484 bp += blkid + i;
485 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
486 dprintf_bp(bp, "can free old%s", "");
487 space += bp_get_dsize(spa, bp);
488 }
489 unref += BP_GET_ASIZE(bp);
490 }
491 nl1blks = 1;
492 nblks = 0;
493 }
494
495 lastblk = blkid + nblks - 1;
496 while (nblks) {
497 dmu_buf_impl_t *dbuf;
498 uint64_t ibyte, new_blkid;
499 int epb = 1 << epbs;
500 int err, i, blkoff, tochk;
501 blkptr_t *bp;
502
503 ibyte = blkid << dn->dn_datablkshift;
504 err = dnode_next_offset(dn,
505 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
506 new_blkid = ibyte >> dn->dn_datablkshift;
507 if (err == ESRCH) {
508 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
509 break;
510 }
511 if (err) {
512 txh->txh_tx->tx_err = err;
513 break;
514 }
515 if (new_blkid > lastblk) {
516 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
517 break;
518 }
519
520 if (new_blkid > blkid) {
521 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
522 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
523 nblks -= new_blkid - blkid;
524 blkid = new_blkid;
525 }
526 blkoff = P2PHASE(blkid, epb);
527 tochk = MIN(epb - blkoff, nblks);
528
529 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
530 if (err) {
531 txh->txh_tx->tx_err = err;
532 break;
533 }
534
535 txh->txh_memory_tohold += dbuf->db.db_size;
536
537 /*
538 * We don't check memory_tohold against DMU_MAX_ACCESS because
539 * memory_tohold is an over-estimation (especially the >L1
540 * indirect blocks), so it could fail. Callers should have
541 * already verified that they will not be holding too much
542 * memory.
543 */
544
545 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
546 if (err != 0) {
547 txh->txh_tx->tx_err = err;
548 dbuf_rele(dbuf, FTAG);
549 break;
550 }
551
552 bp = dbuf->db.db_data;
553 bp += blkoff;
554
555 for (i = 0; i < tochk; i++) {
556 if (dsl_dataset_block_freeable(ds, &bp[i],
557 bp[i].blk_birth)) {
558 dprintf_bp(&bp[i], "can free old%s", "");
559 space += bp_get_dsize(spa, &bp[i]);
560 }
561 unref += BP_GET_ASIZE(bp);
562 }
563 dbuf_rele(dbuf, FTAG);
564
565 ++nl1blks;
566 blkid += tochk;
567 nblks -= tochk;
568 }
569 rw_exit(&dn->dn_struct_rwlock);
570
571 /*
572 * Add in memory requirements of higher-level indirects.
573 * This assumes a worst-possible scenario for dn_nlevels and a
574 * worst-possible distribution of l1-blocks over the region to free.
575 */
576 {
577 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
578 int level = 2;
579 /*
580 * Here we don't use DN_MAX_LEVEL, but calculate it with the
581 * given datablkshift and indblkshift. This makes the
582 * difference between 19 and 8 on large files.
583 */
584 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
585 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
586
587 while (level++ < maxlevel) {
588 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
589 << dn->dn_indblkshift;
590 blkcnt = 1 + (blkcnt >> epbs);
591 }
592 }
593
594 /* account for new level 1 indirect blocks that might show up */
595 if (skipped > 0) {
596 txh->txh_fudge += skipped << dn->dn_indblkshift;
597 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
598 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
599 }
600 txh->txh_space_tofree += space;
601 txh->txh_space_tounref += unref;
602 }
603
604 void
605 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
606 {
607 dmu_tx_hold_t *txh;
608 dnode_t *dn;
609 int err;
610 zio_t *zio;
611
612 ASSERT(tx->tx_txg == 0);
613
614 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
615 object, THT_FREE, off, len);
616 if (txh == NULL)
617 return;
618 dn = txh->txh_dnode;
619 dmu_tx_count_dnode(txh);
620
621 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
622 return;
623 if (len == DMU_OBJECT_END)
624 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
625
626 dmu_tx_count_dnode(txh);
627
628 /*
629 * For i/o error checking, we read the first and last level-0
630 * blocks if they are not aligned, and all the level-1 blocks.
631 *
632 * Note: dbuf_free_range() assumes that we have not instantiated
633 * any level-0 dbufs that will be completely freed. Therefore we must
634 * exercise care to not read or count the first and last blocks
635 * if they are blocksize-aligned.
636 */
637 if (dn->dn_datablkshift == 0) {
638 if (off != 0 || len < dn->dn_datablksz)
639 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
640 } else {
641 /* first block will be modified if it is not aligned */
642 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
643 dmu_tx_count_write(txh, off, 1);
644 /* last block will be modified if it is not aligned */
645 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
646 dmu_tx_count_write(txh, off+len, 1);
647 }
648
649 /*
650 * Check level-1 blocks.
651 */
652 if (dn->dn_nlevels > 1) {
653 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
654 SPA_BLKPTRSHIFT;
655 uint64_t start = off >> shift;
656 uint64_t end = (off + len) >> shift;
657 uint64_t i;
658
659 ASSERT(dn->dn_indblkshift != 0);
660
661 /*
662 * dnode_reallocate() can result in an object with indirect
663 * blocks having an odd data block size. In this case,
664 * just check the single block.
665 */
666 if (dn->dn_datablkshift == 0)
667 start = end = 0;
668
669 zio = zio_root(tx->tx_pool->dp_spa,
670 NULL, NULL, ZIO_FLAG_CANFAIL);
671 for (i = start; i <= end; i++) {
672 uint64_t ibyte = i << shift;
673 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
674 i = ibyte >> shift;
675 if (err == ESRCH)
676 break;
677 if (err) {
678 tx->tx_err = err;
679 return;
680 }
681
682 err = dmu_tx_check_ioerr(zio, dn, 1, i);
683 if (err) {
684 tx->tx_err = err;
685 return;
686 }
687 }
688 err = zio_wait(zio);
689 if (err) {
690 tx->tx_err = err;
691 return;
692 }
693 }
694
695 dmu_tx_count_free(txh, off, len);
696 }
697
698 void
699 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
700 {
701 dmu_tx_hold_t *txh;
702 dnode_t *dn;
703 uint64_t nblocks;
704 int epbs, err;
705
706 ASSERT(tx->tx_txg == 0);
707
708 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
709 object, THT_ZAP, add, (uintptr_t)name);
710 if (txh == NULL)
711 return;
712 dn = txh->txh_dnode;
713
714 dmu_tx_count_dnode(txh);
715
716 if (dn == NULL) {
717 /*
718 * We will be able to fit a new object's entries into one leaf
719 * block. So there will be at most 2 blocks total,
720 * including the header block.
721 */
722 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
723 return;
724 }
725
726 ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
727
728 if (dn->dn_maxblkid == 0 && !add) {
729 blkptr_t *bp;
730
731 /*
732 * If there is only one block (i.e. this is a micro-zap)
733 * and we are not adding anything, the accounting is simple.
734 */
735 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
736 if (err) {
737 tx->tx_err = err;
738 return;
739 }
740
741 /*
742 * Use max block size here, since we don't know how much
743 * the size will change between now and the dbuf dirty call.
744 */
745 bp = &dn->dn_phys->dn_blkptr[0];
746 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
747 bp, bp->blk_birth))
748 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
749 else
750 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
751 if (!BP_IS_HOLE(bp))
752 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
753 return;
754 }
755
756 if (dn->dn_maxblkid > 0 && name) {
757 /*
758 * access the name in this fat-zap so that we'll check
759 * for i/o errors to the leaf blocks, etc.
760 */
761 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
762 8, 0, NULL);
763 if (err == EIO) {
764 tx->tx_err = err;
765 return;
766 }
767 }
768
769 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
770 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
771
772 /*
773 * If the modified blocks are scattered to the four winds,
774 * we'll have to modify an indirect twig for each.
775 */
776 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
777 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
778 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
779 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
780 else
781 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
782 }
783
784 void
785 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
786 {
787 dmu_tx_hold_t *txh;
788
789 ASSERT(tx->tx_txg == 0);
790
791 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
792 object, THT_BONUS, 0, 0);
793 if (txh)
794 dmu_tx_count_dnode(txh);
795 }
796
797 void
798 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
799 {
800 dmu_tx_hold_t *txh;
801
802 ASSERT(tx->tx_txg == 0);
803
804 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
805 DMU_NEW_OBJECT, THT_SPACE, space, 0);
806 if (txh)
807 txh->txh_space_towrite += space;
808 }
809
810 int
811 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
812 {
813 dmu_tx_hold_t *txh;
814 int holds = 0;
815
816 /*
817 * By asserting that the tx is assigned, we're counting the
818 * number of dn_tx_holds, which is the same as the number of
819 * dn_holds. Otherwise, we'd be counting dn_holds, but
820 * dn_tx_holds could be 0.
821 */
822 ASSERT(tx->tx_txg != 0);
823
824 /* if (tx->tx_anyobj == TRUE) */
825 /* return (0); */
826
827 for (txh = list_head(&tx->tx_holds); txh;
828 txh = list_next(&tx->tx_holds, txh)) {
829 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
830 holds++;
831 }
832
833 return (holds);
834 }
835
836 #ifdef DEBUG_DMU_TX
837 void
838 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
839 {
840 dmu_tx_hold_t *txh;
841 int match_object = FALSE, match_offset = FALSE;
842 dnode_t *dn;
843
844 DB_DNODE_ENTER(db);
845 dn = DB_DNODE(db);
846 ASSERT(dn != NULL);
847 ASSERT(tx->tx_txg != 0);
848 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
849 ASSERT3U(dn->dn_object, ==, db->db.db_object);
850
851 if (tx->tx_anyobj) {
852 DB_DNODE_EXIT(db);
853 return;
854 }
855
856 /* XXX No checking on the meta dnode for now */
857 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
858 DB_DNODE_EXIT(db);
859 return;
860 }
861
862 for (txh = list_head(&tx->tx_holds); txh;
863 txh = list_next(&tx->tx_holds, txh)) {
864 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
865 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
866 match_object = TRUE;
867 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
868 int datablkshift = dn->dn_datablkshift ?
869 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
870 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
871 int shift = datablkshift + epbs * db->db_level;
872 uint64_t beginblk = shift >= 64 ? 0 :
873 (txh->txh_arg1 >> shift);
874 uint64_t endblk = shift >= 64 ? 0 :
875 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
876 uint64_t blkid = db->db_blkid;
877
878 /* XXX txh_arg2 better not be zero... */
879
880 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
881 txh->txh_type, beginblk, endblk);
882
883 switch (txh->txh_type) {
884 case THT_WRITE:
885 if (blkid >= beginblk && blkid <= endblk)
886 match_offset = TRUE;
887 /*
888 * We will let this hold work for the bonus
889 * or spill buffer so that we don't need to
890 * hold it when creating a new object.
891 */
892 if (blkid == DMU_BONUS_BLKID ||
893 blkid == DMU_SPILL_BLKID)
894 match_offset = TRUE;
895 /*
896 * They might have to increase nlevels,
897 * thus dirtying the new TLIBs. Or the
898 * might have to change the block size,
899 * thus dirying the new lvl=0 blk=0.
900 */
901 if (blkid == 0)
902 match_offset = TRUE;
903 break;
904 case THT_FREE:
905 /*
906 * We will dirty all the level 1 blocks in
907 * the free range and perhaps the first and
908 * last level 0 block.
909 */
910 if (blkid >= beginblk && (blkid <= endblk ||
911 txh->txh_arg2 == DMU_OBJECT_END))
912 match_offset = TRUE;
913 break;
914 case THT_SPILL:
915 if (blkid == DMU_SPILL_BLKID)
916 match_offset = TRUE;
917 break;
918 case THT_BONUS:
919 if (blkid == DMU_BONUS_BLKID)
920 match_offset = TRUE;
921 break;
922 case THT_ZAP:
923 match_offset = TRUE;
924 break;
925 case THT_NEWOBJECT:
926 match_object = TRUE;
927 break;
928 default:
929 ASSERT(!"bad txh_type");
930 }
931 }
932 if (match_object && match_offset) {
933 DB_DNODE_EXIT(db);
934 return;
935 }
936 }
937 DB_DNODE_EXIT(db);
938 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
939 (u_longlong_t)db->db.db_object, db->db_level,
940 (u_longlong_t)db->db_blkid);
941 }
942 #endif
943
944 /*
945 * If we can't do 10 iops, something is wrong. Let us go ahead
946 * and hit zfs_dirty_data_max.
947 */
948 hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */
949 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
950
951 /*
952 * We delay transactions when we've determined that the backend storage
953 * isn't able to accommodate the rate of incoming writes.
954 *
955 * If there is already a transaction waiting, we delay relative to when
956 * that transaction finishes waiting. This way the calculated min_time
957 * is independent of the number of threads concurrently executing
958 * transactions.
959 *
960 * If we are the only waiter, wait relative to when the transaction
961 * started, rather than the current time. This credits the transaction for
962 * "time already served", e.g. reading indirect blocks.
963 *
964 * The minimum time for a transaction to take is calculated as:
965 * min_time = scale * (dirty - min) / (max - dirty)
966 * min_time is then capped at zfs_delay_max_ns.
967 *
968 * The delay has two degrees of freedom that can be adjusted via tunables.
969 * The percentage of dirty data at which we start to delay is defined by
970 * zfs_delay_min_dirty_percent. This should typically be at or above
971 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
972 * delay after writing at full speed has failed to keep up with the incoming
973 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
974 * speaking, this variable determines the amount of delay at the midpoint of
975 * the curve.
976 *
977 * delay
978 * 10ms +-------------------------------------------------------------*+
979 * | *|
980 * 9ms + *+
981 * | *|
982 * 8ms + *+
983 * | * |
984 * 7ms + * +
985 * | * |
986 * 6ms + * +
987 * | * |
988 * 5ms + * +
989 * | * |
990 * 4ms + * +
991 * | * |
992 * 3ms + * +
993 * | * |
994 * 2ms + (midpoint) * +
995 * | | ** |
996 * 1ms + v *** +
997 * | zfs_delay_scale ----------> ******** |
998 * 0 +-------------------------------------*********----------------+
999 * 0% <- zfs_dirty_data_max -> 100%
1000 *
1001 * Note that since the delay is added to the outstanding time remaining on the
1002 * most recent transaction, the delay is effectively the inverse of IOPS.
1003 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
1004 * was chosen such that small changes in the amount of accumulated dirty data
1005 * in the first 3/4 of the curve yield relatively small differences in the
1006 * amount of delay.
1007 *
1008 * The effects can be easier to understand when the amount of delay is
1009 * represented on a log scale:
1010 *
1011 * delay
1012 * 100ms +-------------------------------------------------------------++
1013 * + +
1014 * | |
1015 * + *+
1016 * 10ms + *+
1017 * + ** +
1018 * | (midpoint) ** |
1019 * + | ** +
1020 * 1ms + v **** +
1021 * + zfs_delay_scale ----------> ***** +
1022 * | **** |
1023 * + **** +
1024 * 100us + ** +
1025 * + * +
1026 * | * |
1027 * + * +
1028 * 10us + * +
1029 * + +
1030 * | |
1031 * + +
1032 * +--------------------------------------------------------------+
1033 * 0% <- zfs_dirty_data_max -> 100%
1034 *
1035 * Note here that only as the amount of dirty data approaches its limit does
1036 * the delay start to increase rapidly. The goal of a properly tuned system
1037 * should be to keep the amount of dirty data out of that range by first
1038 * ensuring that the appropriate limits are set for the I/O scheduler to reach
1039 * optimal throughput on the backend storage, and then by changing the value
1040 * of zfs_delay_scale to increase the steepness of the curve.
1041 */
1042 static void
1043 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
1044 {
1045 dsl_pool_t *dp = tx->tx_pool;
1046 uint64_t delay_min_bytes =
1047 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
1048 hrtime_t wakeup, min_tx_time, now;
1049
1050 if (dirty <= delay_min_bytes)
1051 return;
1052
1053 /*
1054 * The caller has already waited until we are under the max.
1055 * We make them pass us the amount of dirty data so we don't
1056 * have to handle the case of it being >= the max, which could
1057 * cause a divide-by-zero if it's == the max.
1058 */
1059 ASSERT3U(dirty, <, zfs_dirty_data_max);
1060
1061 now = gethrtime();
1062 min_tx_time = zfs_delay_scale *
1063 (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
1064 min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
1065 if (now > tx->tx_start + min_tx_time)
1066 return;
1067
1068 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1069 uint64_t, min_tx_time);
1070
1071 mutex_enter(&dp->dp_lock);
1072 wakeup = MAX(tx->tx_start + min_tx_time,
1073 dp->dp_last_wakeup + min_tx_time);
1074 dp->dp_last_wakeup = wakeup;
1075 mutex_exit(&dp->dp_lock);
1076
1077 zfs_sleep_until(wakeup);
1078 }
1079
1080 static int
1081 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1082 {
1083 dmu_tx_hold_t *txh;
1084 spa_t *spa = tx->tx_pool->dp_spa;
1085 uint64_t memory, asize, fsize, usize;
1086 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1087
1088 ASSERT0(tx->tx_txg);
1089
1090 if (tx->tx_err) {
1091 DMU_TX_STAT_BUMP(dmu_tx_error);
1092 return (tx->tx_err);
1093 }
1094
1095 if (spa_suspended(spa)) {
1096 DMU_TX_STAT_BUMP(dmu_tx_suspended);
1097
1098 /*
1099 * If the user has indicated a blocking failure mode
1100 * then return ERESTART which will block in dmu_tx_wait().
1101 * Otherwise, return EIO so that an error can get
1102 * propagated back to the VOP calls.
1103 *
1104 * Note that we always honor the txg_how flag regardless
1105 * of the failuremode setting.
1106 */
1107 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1108 txg_how != TXG_WAIT)
1109 return (SET_ERROR(EIO));
1110
1111 return (SET_ERROR(ERESTART));
1112 }
1113
1114 if (!tx->tx_waited &&
1115 dsl_pool_need_dirty_delay(tx->tx_pool)) {
1116 tx->tx_wait_dirty = B_TRUE;
1117 DMU_TX_STAT_BUMP(dmu_tx_dirty_delay);
1118 return (ERESTART);
1119 }
1120
1121 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1122 tx->tx_needassign_txh = NULL;
1123
1124 /*
1125 * NB: No error returns are allowed after txg_hold_open, but
1126 * before processing the dnode holds, due to the
1127 * dmu_tx_unassign() logic.
1128 */
1129
1130 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1131 for (txh = list_head(&tx->tx_holds); txh;
1132 txh = list_next(&tx->tx_holds, txh)) {
1133 dnode_t *dn = txh->txh_dnode;
1134 if (dn != NULL) {
1135 mutex_enter(&dn->dn_mtx);
1136 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1137 mutex_exit(&dn->dn_mtx);
1138 tx->tx_needassign_txh = txh;
1139 DMU_TX_STAT_BUMP(dmu_tx_group);
1140 return (SET_ERROR(ERESTART));
1141 }
1142 if (dn->dn_assigned_txg == 0)
1143 dn->dn_assigned_txg = tx->tx_txg;
1144 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1145 (void) refcount_add(&dn->dn_tx_holds, tx);
1146 mutex_exit(&dn->dn_mtx);
1147 }
1148 towrite += txh->txh_space_towrite;
1149 tofree += txh->txh_space_tofree;
1150 tooverwrite += txh->txh_space_tooverwrite;
1151 tounref += txh->txh_space_tounref;
1152 tohold += txh->txh_memory_tohold;
1153 fudge += txh->txh_fudge;
1154 }
1155
1156 /*
1157 * If a snapshot has been taken since we made our estimates,
1158 * assume that we won't be able to free or overwrite anything.
1159 */
1160 if (tx->tx_objset &&
1161 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1162 tx->tx_lastsnap_txg) {
1163 towrite += tooverwrite;
1164 tooverwrite = tofree = 0;
1165 }
1166
1167 /* needed allocation: worst-case estimate of write space */
1168 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1169 /* freed space estimate: worst-case overwrite + free estimate */
1170 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1171 /* convert unrefd space to worst-case estimate */
1172 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1173 /* calculate memory footprint estimate */
1174 memory = towrite + tooverwrite + tohold;
1175
1176 #ifdef DEBUG_DMU_TX
1177 /*
1178 * Add in 'tohold' to account for our dirty holds on this memory
1179 * XXX - the "fudge" factor is to account for skipped blocks that
1180 * we missed because dnode_next_offset() misses in-core-only blocks.
1181 */
1182 tx->tx_space_towrite = asize +
1183 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1184 tx->tx_space_tofree = tofree;
1185 tx->tx_space_tooverwrite = tooverwrite;
1186 tx->tx_space_tounref = tounref;
1187 #endif
1188
1189 if (tx->tx_dir && asize != 0) {
1190 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1191 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1192 if (err)
1193 return (err);
1194 }
1195
1196 DMU_TX_STAT_BUMP(dmu_tx_assigned);
1197
1198 return (0);
1199 }
1200
1201 static void
1202 dmu_tx_unassign(dmu_tx_t *tx)
1203 {
1204 dmu_tx_hold_t *txh;
1205
1206 if (tx->tx_txg == 0)
1207 return;
1208
1209 txg_rele_to_quiesce(&tx->tx_txgh);
1210
1211 /*
1212 * Walk the transaction's hold list, removing the hold on the
1213 * associated dnode, and notifying waiters if the refcount drops to 0.
1214 */
1215 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1216 txh = list_next(&tx->tx_holds, txh)) {
1217 dnode_t *dn = txh->txh_dnode;
1218
1219 if (dn == NULL)
1220 continue;
1221 mutex_enter(&dn->dn_mtx);
1222 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1223
1224 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1225 dn->dn_assigned_txg = 0;
1226 cv_broadcast(&dn->dn_notxholds);
1227 }
1228 mutex_exit(&dn->dn_mtx);
1229 }
1230
1231 txg_rele_to_sync(&tx->tx_txgh);
1232
1233 tx->tx_lasttried_txg = tx->tx_txg;
1234 tx->tx_txg = 0;
1235 }
1236
1237 /*
1238 * Assign tx to a transaction group. txg_how can be one of:
1239 *
1240 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1241 * a new one. This should be used when you're not holding locks.
1242 * It will only fail if we're truly out of space (or over quota).
1243 *
1244 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1245 * blocking, returns immediately with ERESTART. This should be used
1246 * whenever you're holding locks. On an ERESTART error, the caller
1247 * should drop locks, do a dmu_tx_wait(tx), and try again.
1248 *
1249 * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1250 * has already been called on behalf of this operation (though
1251 * most likely on a different tx).
1252 */
1253 int
1254 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1255 {
1256 hrtime_t before;
1257 int err;
1258
1259 ASSERT(tx->tx_txg == 0);
1260 ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1261 txg_how == TXG_WAITED);
1262 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1263
1264 before = gethrtime();
1265
1266 if (txg_how == TXG_WAITED)
1267 tx->tx_waited = B_TRUE;
1268
1269 /* If we might wait, we must not hold the config lock. */
1270 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1271
1272 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1273 dmu_tx_unassign(tx);
1274
1275 if (err != ERESTART || txg_how != TXG_WAIT)
1276 return (err);
1277
1278 dmu_tx_wait(tx);
1279 }
1280
1281 txg_rele_to_quiesce(&tx->tx_txgh);
1282
1283 spa_tx_assign_add_nsecs(tx->tx_pool->dp_spa, gethrtime() - before);
1284
1285 return (0);
1286 }
1287
1288 void
1289 dmu_tx_wait(dmu_tx_t *tx)
1290 {
1291 spa_t *spa = tx->tx_pool->dp_spa;
1292 dsl_pool_t *dp = tx->tx_pool;
1293
1294 ASSERT(tx->tx_txg == 0);
1295 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1296
1297 if (tx->tx_wait_dirty) {
1298 uint64_t dirty;
1299
1300 /*
1301 * dmu_tx_try_assign() has determined that we need to wait
1302 * because we've consumed much or all of the dirty buffer
1303 * space.
1304 */
1305 mutex_enter(&dp->dp_lock);
1306 if (dp->dp_dirty_total >= zfs_dirty_data_max)
1307 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
1308 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1309 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1310 dirty = dp->dp_dirty_total;
1311 mutex_exit(&dp->dp_lock);
1312
1313 dmu_tx_delay(tx, dirty);
1314
1315 tx->tx_wait_dirty = B_FALSE;
1316
1317 /*
1318 * Note: setting tx_waited only has effect if the caller
1319 * used TX_WAIT. Otherwise they are going to destroy
1320 * this tx and try again. The common case, zfs_write(),
1321 * uses TX_WAIT.
1322 */
1323 tx->tx_waited = B_TRUE;
1324 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1325 /*
1326 * If the pool is suspended we need to wait until it
1327 * is resumed. Note that it's possible that the pool
1328 * has become active after this thread has tried to
1329 * obtain a tx. If that's the case then tx_lasttried_txg
1330 * would not have been set.
1331 */
1332 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1333 } else if (tx->tx_needassign_txh) {
1334 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1335
1336 mutex_enter(&dn->dn_mtx);
1337 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1338 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1339 mutex_exit(&dn->dn_mtx);
1340 tx->tx_needassign_txh = NULL;
1341 } else {
1342 /*
1343 * A dnode is assigned to the quiescing txg. Wait for its
1344 * transaction to complete.
1345 */
1346 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1347 }
1348 }
1349
1350 void
1351 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1352 {
1353 #ifdef DEBUG_DMU_TX
1354 if (tx->tx_dir == NULL || delta == 0)
1355 return;
1356
1357 if (delta > 0) {
1358 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1359 tx->tx_space_towrite);
1360 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1361 } else {
1362 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1363 }
1364 #endif
1365 }
1366
1367 void
1368 dmu_tx_commit(dmu_tx_t *tx)
1369 {
1370 dmu_tx_hold_t *txh;
1371
1372 ASSERT(tx->tx_txg != 0);
1373
1374 /*
1375 * Go through the transaction's hold list and remove holds on
1376 * associated dnodes, notifying waiters if no holds remain.
1377 */
1378 while ((txh = list_head(&tx->tx_holds))) {
1379 dnode_t *dn = txh->txh_dnode;
1380
1381 list_remove(&tx->tx_holds, txh);
1382 kmem_free(txh, sizeof (dmu_tx_hold_t));
1383 if (dn == NULL)
1384 continue;
1385 mutex_enter(&dn->dn_mtx);
1386 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1387
1388 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1389 dn->dn_assigned_txg = 0;
1390 cv_broadcast(&dn->dn_notxholds);
1391 }
1392 mutex_exit(&dn->dn_mtx);
1393 dnode_rele(dn, tx);
1394 }
1395
1396 if (tx->tx_tempreserve_cookie)
1397 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1398
1399 if (!list_is_empty(&tx->tx_callbacks))
1400 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1401
1402 if (tx->tx_anyobj == FALSE)
1403 txg_rele_to_sync(&tx->tx_txgh);
1404
1405 list_destroy(&tx->tx_callbacks);
1406 list_destroy(&tx->tx_holds);
1407 #ifdef DEBUG_DMU_TX
1408 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1409 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1410 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1411 refcount_destroy_many(&tx->tx_space_written,
1412 refcount_count(&tx->tx_space_written));
1413 refcount_destroy_many(&tx->tx_space_freed,
1414 refcount_count(&tx->tx_space_freed));
1415 #endif
1416 kmem_free(tx, sizeof (dmu_tx_t));
1417 }
1418
1419 void
1420 dmu_tx_abort(dmu_tx_t *tx)
1421 {
1422 dmu_tx_hold_t *txh;
1423
1424 ASSERT(tx->tx_txg == 0);
1425
1426 while ((txh = list_head(&tx->tx_holds))) {
1427 dnode_t *dn = txh->txh_dnode;
1428
1429 list_remove(&tx->tx_holds, txh);
1430 kmem_free(txh, sizeof (dmu_tx_hold_t));
1431 if (dn != NULL)
1432 dnode_rele(dn, tx);
1433 }
1434
1435 /*
1436 * Call any registered callbacks with an error code.
1437 */
1438 if (!list_is_empty(&tx->tx_callbacks))
1439 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1440
1441 list_destroy(&tx->tx_callbacks);
1442 list_destroy(&tx->tx_holds);
1443 #ifdef DEBUG_DMU_TX
1444 refcount_destroy_many(&tx->tx_space_written,
1445 refcount_count(&tx->tx_space_written));
1446 refcount_destroy_many(&tx->tx_space_freed,
1447 refcount_count(&tx->tx_space_freed));
1448 #endif
1449 kmem_free(tx, sizeof (dmu_tx_t));
1450 }
1451
1452 uint64_t
1453 dmu_tx_get_txg(dmu_tx_t *tx)
1454 {
1455 ASSERT(tx->tx_txg != 0);
1456 return (tx->tx_txg);
1457 }
1458
1459 dsl_pool_t *
1460 dmu_tx_pool(dmu_tx_t *tx)
1461 {
1462 ASSERT(tx->tx_pool != NULL);
1463 return (tx->tx_pool);
1464 }
1465
1466 void
1467 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1468 {
1469 dmu_tx_callback_t *dcb;
1470
1471 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_PUSHPAGE);
1472
1473 dcb->dcb_func = func;
1474 dcb->dcb_data = data;
1475
1476 list_insert_tail(&tx->tx_callbacks, dcb);
1477 }
1478
1479 /*
1480 * Call all the commit callbacks on a list, with a given error code.
1481 */
1482 void
1483 dmu_tx_do_callbacks(list_t *cb_list, int error)
1484 {
1485 dmu_tx_callback_t *dcb;
1486
1487 while ((dcb = list_head(cb_list))) {
1488 list_remove(cb_list, dcb);
1489 dcb->dcb_func(dcb->dcb_data, error);
1490 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1491 }
1492 }
1493
1494 /*
1495 * Interface to hold a bunch of attributes.
1496 * used for creating new files.
1497 * attrsize is the total size of all attributes
1498 * to be added during object creation
1499 *
1500 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1501 */
1502
1503 /*
1504 * hold necessary attribute name for attribute registration.
1505 * should be a very rare case where this is needed. If it does
1506 * happen it would only happen on the first write to the file system.
1507 */
1508 static void
1509 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1510 {
1511 int i;
1512
1513 if (!sa->sa_need_attr_registration)
1514 return;
1515
1516 for (i = 0; i != sa->sa_num_attrs; i++) {
1517 if (!sa->sa_attr_table[i].sa_registered) {
1518 if (sa->sa_reg_attr_obj)
1519 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1520 B_TRUE, sa->sa_attr_table[i].sa_name);
1521 else
1522 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1523 B_TRUE, sa->sa_attr_table[i].sa_name);
1524 }
1525 }
1526 }
1527
1528
1529 void
1530 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1531 {
1532 dnode_t *dn;
1533 dmu_tx_hold_t *txh;
1534
1535 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1536 THT_SPILL, 0, 0);
1537 if (txh == NULL)
1538 return;
1539
1540 dn = txh->txh_dnode;
1541
1542 if (dn == NULL)
1543 return;
1544
1545 /* If blkptr doesn't exist then add space to towrite */
1546 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1547 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1548 } else {
1549 blkptr_t *bp;
1550
1551 bp = &dn->dn_phys->dn_spill;
1552 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1553 bp, bp->blk_birth))
1554 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1555 else
1556 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1557 if (!BP_IS_HOLE(bp))
1558 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1559 }
1560 }
1561
1562 void
1563 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1564 {
1565 sa_os_t *sa = tx->tx_objset->os_sa;
1566
1567 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1568
1569 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1570 return;
1571
1572 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1573 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1574 else {
1575 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1576 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1577 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1578 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1579 }
1580
1581 dmu_tx_sa_registration_hold(sa, tx);
1582
1583 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1584 return;
1585
1586 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1587 THT_SPILL, 0, 0);
1588 }
1589
1590 /*
1591 * Hold SA attribute
1592 *
1593 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1594 *
1595 * variable_size is the total size of all variable sized attributes
1596 * passed to this function. It is not the total size of all
1597 * variable size attributes that *may* exist on this object.
1598 */
1599 void
1600 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1601 {
1602 uint64_t object;
1603 sa_os_t *sa = tx->tx_objset->os_sa;
1604
1605 ASSERT(hdl != NULL);
1606
1607 object = sa_handle_object(hdl);
1608
1609 dmu_tx_hold_bonus(tx, object);
1610
1611 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1612 return;
1613
1614 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1615 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1616 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1617 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1618 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1619 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1620 }
1621
1622 dmu_tx_sa_registration_hold(sa, tx);
1623
1624 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1625 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1626
1627 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1628 ASSERT(tx->tx_txg == 0);
1629 dmu_tx_hold_spill(tx, object);
1630 } else {
1631 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1632 dnode_t *dn;
1633
1634 DB_DNODE_ENTER(db);
1635 dn = DB_DNODE(db);
1636 if (dn->dn_have_spill) {
1637 ASSERT(tx->tx_txg == 0);
1638 dmu_tx_hold_spill(tx, object);
1639 }
1640 DB_DNODE_EXIT(db);
1641 }
1642 }
1643
1644 void
1645 dmu_tx_init(void)
1646 {
1647 dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc",
1648 KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t),
1649 KSTAT_FLAG_VIRTUAL);
1650
1651 if (dmu_tx_ksp != NULL) {
1652 dmu_tx_ksp->ks_data = &dmu_tx_stats;
1653 kstat_install(dmu_tx_ksp);
1654 }
1655 }
1656
1657 void
1658 dmu_tx_fini(void)
1659 {
1660 if (dmu_tx_ksp != NULL) {
1661 kstat_delete(dmu_tx_ksp);
1662 dmu_tx_ksp = NULL;
1663 }
1664 }
1665
1666 #if defined(_KERNEL) && defined(HAVE_SPL)
1667 EXPORT_SYMBOL(dmu_tx_create);
1668 EXPORT_SYMBOL(dmu_tx_hold_write);
1669 EXPORT_SYMBOL(dmu_tx_hold_free);
1670 EXPORT_SYMBOL(dmu_tx_hold_zap);
1671 EXPORT_SYMBOL(dmu_tx_hold_bonus);
1672 EXPORT_SYMBOL(dmu_tx_abort);
1673 EXPORT_SYMBOL(dmu_tx_assign);
1674 EXPORT_SYMBOL(dmu_tx_wait);
1675 EXPORT_SYMBOL(dmu_tx_commit);
1676 EXPORT_SYMBOL(dmu_tx_get_txg);
1677 EXPORT_SYMBOL(dmu_tx_callback_register);
1678 EXPORT_SYMBOL(dmu_tx_do_callbacks);
1679 EXPORT_SYMBOL(dmu_tx_hold_spill);
1680 EXPORT_SYMBOL(dmu_tx_hold_sa_create);
1681 EXPORT_SYMBOL(dmu_tx_hold_sa);
1682 #endif
Mirror of ZFSOnLinux for Debian