]> 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 0.6.4.2
[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 #include <sys/trace_dmu.h>
42
43 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
44 uint64_t arg1, uint64_t arg2);
45
46 dmu_tx_stats_t dmu_tx_stats = {
47 { "dmu_tx_assigned", KSTAT_DATA_UINT64 },
48 { "dmu_tx_delay", KSTAT_DATA_UINT64 },
49 { "dmu_tx_error", KSTAT_DATA_UINT64 },
50 { "dmu_tx_suspended", KSTAT_DATA_UINT64 },
51 { "dmu_tx_group", KSTAT_DATA_UINT64 },
52 { "dmu_tx_memory_reserve", KSTAT_DATA_UINT64 },
53 { "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64 },
54 { "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64 },
55 { "dmu_tx_dirty_delay", KSTAT_DATA_UINT64 },
56 { "dmu_tx_dirty_over_max", KSTAT_DATA_UINT64 },
57 { "dmu_tx_quota", KSTAT_DATA_UINT64 },
58 };
59
60 static kstat_t *dmu_tx_ksp;
61
62 dmu_tx_t *
63 dmu_tx_create_dd(dsl_dir_t *dd)
64 {
65 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
66 tx->tx_dir = dd;
67 if (dd != NULL)
68 tx->tx_pool = dd->dd_pool;
69 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
70 offsetof(dmu_tx_hold_t, txh_node));
71 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
72 offsetof(dmu_tx_callback_t, dcb_node));
73 tx->tx_start = gethrtime();
74 #ifdef DEBUG_DMU_TX
75 refcount_create(&tx->tx_space_written);
76 refcount_create(&tx->tx_space_freed);
77 #endif
78 return (tx);
79 }
80
81 dmu_tx_t *
82 dmu_tx_create(objset_t *os)
83 {
84 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
85 tx->tx_objset = os;
86 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
87 return (tx);
88 }
89
90 dmu_tx_t *
91 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
92 {
93 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
94
95 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
96 tx->tx_pool = dp;
97 tx->tx_txg = txg;
98 tx->tx_anyobj = TRUE;
99
100 return (tx);
101 }
102
103 int
104 dmu_tx_is_syncing(dmu_tx_t *tx)
105 {
106 return (tx->tx_anyobj);
107 }
108
109 int
110 dmu_tx_private_ok(dmu_tx_t *tx)
111 {
112 return (tx->tx_anyobj);
113 }
114
115 static dmu_tx_hold_t *
116 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
117 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
118 {
119 dmu_tx_hold_t *txh;
120 dnode_t *dn = NULL;
121 int err;
122
123 if (object != DMU_NEW_OBJECT) {
124 err = dnode_hold(os, object, tx, &dn);
125 if (err) {
126 tx->tx_err = err;
127 return (NULL);
128 }
129
130 if (err == 0 && tx->tx_txg != 0) {
131 mutex_enter(&dn->dn_mtx);
132 /*
133 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
134 * problem, but there's no way for it to happen (for
135 * now, at least).
136 */
137 ASSERT(dn->dn_assigned_txg == 0);
138 dn->dn_assigned_txg = tx->tx_txg;
139 (void) refcount_add(&dn->dn_tx_holds, tx);
140 mutex_exit(&dn->dn_mtx);
141 }
142 }
143
144 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
145 txh->txh_tx = tx;
146 txh->txh_dnode = dn;
147 #ifdef DEBUG_DMU_TX
148 txh->txh_type = type;
149 txh->txh_arg1 = arg1;
150 txh->txh_arg2 = arg2;
151 #endif
152 list_insert_tail(&tx->tx_holds, txh);
153
154 return (txh);
155 }
156
157 void
158 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
159 {
160 /*
161 * If we're syncing, they can manipulate any object anyhow, and
162 * the hold on the dnode_t can cause problems.
163 */
164 if (!dmu_tx_is_syncing(tx)) {
165 (void) dmu_tx_hold_object_impl(tx, os,
166 object, THT_NEWOBJECT, 0, 0);
167 }
168 }
169
170 static int
171 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
172 {
173 int err;
174 dmu_buf_impl_t *db;
175
176 rw_enter(&dn->dn_struct_rwlock, RW_READER);
177 db = dbuf_hold_level(dn, level, blkid, FTAG);
178 rw_exit(&dn->dn_struct_rwlock);
179 if (db == NULL)
180 return (SET_ERROR(EIO));
181 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
182 dbuf_rele(db, FTAG);
183 return (err);
184 }
185
186 static void
187 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
188 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
189 {
190 objset_t *os = dn->dn_objset;
191 dsl_dataset_t *ds = os->os_dsl_dataset;
192 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
193 dmu_buf_impl_t *parent = NULL;
194 blkptr_t *bp = NULL;
195 uint64_t space;
196
197 if (level >= dn->dn_nlevels || history[level] == blkid)
198 return;
199
200 history[level] = blkid;
201
202 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
203
204 if (db == NULL || db == dn->dn_dbuf) {
205 ASSERT(level != 0);
206 db = NULL;
207 } else {
208 ASSERT(DB_DNODE(db) == dn);
209 ASSERT(db->db_level == level);
210 ASSERT(db->db.db_size == space);
211 ASSERT(db->db_blkid == blkid);
212 bp = db->db_blkptr;
213 parent = db->db_parent;
214 }
215
216 freeable = (bp && (freeable ||
217 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
218
219 if (freeable)
220 txh->txh_space_tooverwrite += space;
221 else
222 txh->txh_space_towrite += space;
223 if (bp)
224 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
225
226 dmu_tx_count_twig(txh, dn, parent, level + 1,
227 blkid >> epbs, freeable, history);
228 }
229
230 /* ARGSUSED */
231 static void
232 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
233 {
234 dnode_t *dn = txh->txh_dnode;
235 uint64_t start, end, i;
236 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
237 int err = 0;
238 int l;
239
240 if (len == 0)
241 return;
242
243 min_bs = SPA_MINBLOCKSHIFT;
244 max_bs = SPA_MAXBLOCKSHIFT;
245 min_ibs = DN_MIN_INDBLKSHIFT;
246 max_ibs = DN_MAX_INDBLKSHIFT;
247
248 if (dn) {
249 uint64_t history[DN_MAX_LEVELS];
250 int nlvls = dn->dn_nlevels;
251 int delta;
252
253 /*
254 * For i/o error checking, read the first and last level-0
255 * blocks (if they are not aligned), and all the level-1 blocks.
256 */
257 if (dn->dn_maxblkid == 0) {
258 delta = dn->dn_datablksz;
259 start = (off < dn->dn_datablksz) ? 0 : 1;
260 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
261 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
262 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
263 if (err)
264 goto out;
265 delta -= off;
266 }
267 } else {
268 zio_t *zio = zio_root(dn->dn_objset->os_spa,
269 NULL, NULL, ZIO_FLAG_CANFAIL);
270
271 /* first level-0 block */
272 start = off >> dn->dn_datablkshift;
273 if (P2PHASE(off, dn->dn_datablksz) ||
274 len < dn->dn_datablksz) {
275 err = dmu_tx_check_ioerr(zio, dn, 0, start);
276 if (err)
277 goto out;
278 }
279
280 /* last level-0 block */
281 end = (off+len-1) >> dn->dn_datablkshift;
282 if (end != start && end <= dn->dn_maxblkid &&
283 P2PHASE(off+len, dn->dn_datablksz)) {
284 err = dmu_tx_check_ioerr(zio, dn, 0, end);
285 if (err)
286 goto out;
287 }
288
289 /* level-1 blocks */
290 if (nlvls > 1) {
291 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
292 for (i = (start>>shft)+1; i < end>>shft; i++) {
293 err = dmu_tx_check_ioerr(zio, dn, 1, i);
294 if (err)
295 goto out;
296 }
297 }
298
299 err = zio_wait(zio);
300 if (err)
301 goto out;
302 delta = P2NPHASE(off, dn->dn_datablksz);
303 }
304
305 min_ibs = max_ibs = dn->dn_indblkshift;
306 if (dn->dn_maxblkid > 0) {
307 /*
308 * The blocksize can't change,
309 * so we can make a more precise estimate.
310 */
311 ASSERT(dn->dn_datablkshift != 0);
312 min_bs = max_bs = dn->dn_datablkshift;
313 }
314
315 /*
316 * If this write is not off the end of the file
317 * we need to account for overwrites/unref.
318 */
319 if (start <= dn->dn_maxblkid) {
320 for (l = 0; l < DN_MAX_LEVELS; l++)
321 history[l] = -1ULL;
322 }
323 while (start <= dn->dn_maxblkid) {
324 dmu_buf_impl_t *db;
325
326 rw_enter(&dn->dn_struct_rwlock, RW_READER);
327 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
328 rw_exit(&dn->dn_struct_rwlock);
329
330 if (err) {
331 txh->txh_tx->tx_err = err;
332 return;
333 }
334
335 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
336 history);
337 dbuf_rele(db, FTAG);
338 if (++start > end) {
339 /*
340 * Account for new indirects appearing
341 * before this IO gets assigned into a txg.
342 */
343 bits = 64 - min_bs;
344 epbs = min_ibs - SPA_BLKPTRSHIFT;
345 for (bits -= epbs * (nlvls - 1);
346 bits >= 0; bits -= epbs)
347 txh->txh_fudge += 1ULL << max_ibs;
348 goto out;
349 }
350 off += delta;
351 if (len >= delta)
352 len -= delta;
353 delta = dn->dn_datablksz;
354 }
355 }
356
357 /*
358 * 'end' is the last thing we will access, not one past.
359 * This way we won't overflow when accessing the last byte.
360 */
361 start = P2ALIGN(off, 1ULL << max_bs);
362 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
363 txh->txh_space_towrite += end - start + 1;
364
365 start >>= min_bs;
366 end >>= min_bs;
367
368 epbs = min_ibs - SPA_BLKPTRSHIFT;
369
370 /*
371 * The object contains at most 2^(64 - min_bs) blocks,
372 * and each indirect level maps 2^epbs.
373 */
374 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
375 start >>= epbs;
376 end >>= epbs;
377 ASSERT3U(end, >=, start);
378 txh->txh_space_towrite += (end - start + 1) << max_ibs;
379 if (start != 0) {
380 /*
381 * We also need a new blkid=0 indirect block
382 * to reference any existing file data.
383 */
384 txh->txh_space_towrite += 1ULL << max_ibs;
385 }
386 }
387
388 out:
389 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
390 2 * DMU_MAX_ACCESS)
391 err = SET_ERROR(EFBIG);
392
393 if (err)
394 txh->txh_tx->tx_err = err;
395 }
396
397 static void
398 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
399 {
400 dnode_t *dn = txh->txh_dnode;
401 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
402 uint64_t space = mdn->dn_datablksz +
403 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
404
405 if (dn && dn->dn_dbuf->db_blkptr &&
406 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
407 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
408 txh->txh_space_tooverwrite += space;
409 txh->txh_space_tounref += space;
410 } else {
411 txh->txh_space_towrite += space;
412 if (dn && dn->dn_dbuf->db_blkptr)
413 txh->txh_space_tounref += space;
414 }
415 }
416
417 void
418 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
419 {
420 dmu_tx_hold_t *txh;
421
422 ASSERT(tx->tx_txg == 0);
423 ASSERT(len <= DMU_MAX_ACCESS);
424 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
425
426 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
427 object, THT_WRITE, off, len);
428 if (txh == NULL)
429 return;
430
431 dmu_tx_count_write(txh, off, len);
432 dmu_tx_count_dnode(txh);
433 }
434
435 static void
436 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
437 {
438 uint64_t blkid, nblks, lastblk;
439 uint64_t space = 0, unref = 0, skipped = 0;
440 dnode_t *dn = txh->txh_dnode;
441 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
442 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
443 int epbs;
444 uint64_t l0span = 0, nl1blks = 0;
445
446 if (dn->dn_nlevels == 0)
447 return;
448
449 /*
450 * The struct_rwlock protects us against dn_nlevels
451 * changing, in case (against all odds) we manage to dirty &
452 * sync out the changes after we check for being dirty.
453 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
454 */
455 rw_enter(&dn->dn_struct_rwlock, RW_READER);
456 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
457 if (dn->dn_maxblkid == 0) {
458 if (off == 0 && len >= dn->dn_datablksz) {
459 blkid = 0;
460 nblks = 1;
461 } else {
462 rw_exit(&dn->dn_struct_rwlock);
463 return;
464 }
465 } else {
466 blkid = off >> dn->dn_datablkshift;
467 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
468
469 if (blkid > dn->dn_maxblkid) {
470 rw_exit(&dn->dn_struct_rwlock);
471 return;
472 }
473 if (blkid + nblks > dn->dn_maxblkid)
474 nblks = dn->dn_maxblkid - blkid + 1;
475
476 }
477 l0span = nblks; /* save for later use to calc level > 1 overhead */
478 if (dn->dn_nlevels == 1) {
479 int i;
480 for (i = 0; i < nblks; i++) {
481 blkptr_t *bp = dn->dn_phys->dn_blkptr;
482 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
483 bp += blkid + i;
484 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
485 dprintf_bp(bp, "can free old%s", "");
486 space += bp_get_dsize(spa, bp);
487 }
488 unref += BP_GET_ASIZE(bp);
489 }
490 nl1blks = 1;
491 nblks = 0;
492 }
493
494 lastblk = blkid + nblks - 1;
495 while (nblks) {
496 dmu_buf_impl_t *dbuf;
497 uint64_t ibyte, new_blkid;
498 int epb = 1 << epbs;
499 int err, i, blkoff, tochk;
500 blkptr_t *bp;
501
502 ibyte = blkid << dn->dn_datablkshift;
503 err = dnode_next_offset(dn,
504 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
505 new_blkid = ibyte >> dn->dn_datablkshift;
506 if (err == ESRCH) {
507 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
508 break;
509 }
510 if (err) {
511 txh->txh_tx->tx_err = err;
512 break;
513 }
514 if (new_blkid > lastblk) {
515 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
516 break;
517 }
518
519 if (new_blkid > blkid) {
520 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
521 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
522 nblks -= new_blkid - blkid;
523 blkid = new_blkid;
524 }
525 blkoff = P2PHASE(blkid, epb);
526 tochk = MIN(epb - blkoff, nblks);
527
528 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
529 if (err) {
530 txh->txh_tx->tx_err = err;
531 break;
532 }
533
534 txh->txh_memory_tohold += dbuf->db.db_size;
535
536 /*
537 * We don't check memory_tohold against DMU_MAX_ACCESS because
538 * memory_tohold is an over-estimation (especially the >L1
539 * indirect blocks), so it could fail. Callers should have
540 * already verified that they will not be holding too much
541 * memory.
542 */
543
544 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
545 if (err != 0) {
546 txh->txh_tx->tx_err = err;
547 dbuf_rele(dbuf, FTAG);
548 break;
549 }
550
551 bp = dbuf->db.db_data;
552 bp += blkoff;
553
554 for (i = 0; i < tochk; i++) {
555 if (dsl_dataset_block_freeable(ds, &bp[i],
556 bp[i].blk_birth)) {
557 dprintf_bp(&bp[i], "can free old%s", "");
558 space += bp_get_dsize(spa, &bp[i]);
559 }
560 unref += BP_GET_ASIZE(bp);
561 }
562 dbuf_rele(dbuf, FTAG);
563
564 ++nl1blks;
565 blkid += tochk;
566 nblks -= tochk;
567 }
568 rw_exit(&dn->dn_struct_rwlock);
569
570 /*
571 * Add in memory requirements of higher-level indirects.
572 * This assumes a worst-possible scenario for dn_nlevels and a
573 * worst-possible distribution of l1-blocks over the region to free.
574 */
575 {
576 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
577 int level = 2;
578 /*
579 * Here we don't use DN_MAX_LEVEL, but calculate it with the
580 * given datablkshift and indblkshift. This makes the
581 * difference between 19 and 8 on large files.
582 */
583 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
584 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
585
586 while (level++ < maxlevel) {
587 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
588 << dn->dn_indblkshift;
589 blkcnt = 1 + (blkcnt >> epbs);
590 }
591 }
592
593 /* account for new level 1 indirect blocks that might show up */
594 if (skipped > 0) {
595 txh->txh_fudge += skipped << dn->dn_indblkshift;
596 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
597 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
598 }
599 txh->txh_space_tofree += space;
600 txh->txh_space_tounref += unref;
601 }
602
603 void
604 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
605 {
606 dmu_tx_hold_t *txh;
607 dnode_t *dn;
608 int err;
609 zio_t *zio;
610
611 ASSERT(tx->tx_txg == 0);
612
613 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
614 object, THT_FREE, off, len);
615 if (txh == NULL)
616 return;
617 dn = txh->txh_dnode;
618 dmu_tx_count_dnode(txh);
619
620 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
621 return;
622 if (len == DMU_OBJECT_END)
623 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
624
625 dmu_tx_count_dnode(txh);
626
627 /*
628 * For i/o error checking, we read the first and last level-0
629 * blocks if they are not aligned, and all the level-1 blocks.
630 *
631 * Note: dbuf_free_range() assumes that we have not instantiated
632 * any level-0 dbufs that will be completely freed. Therefore we must
633 * exercise care to not read or count the first and last blocks
634 * if they are blocksize-aligned.
635 */
636 if (dn->dn_datablkshift == 0) {
637 if (off != 0 || len < dn->dn_datablksz)
638 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
639 } else {
640 /* first block will be modified if it is not aligned */
641 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
642 dmu_tx_count_write(txh, off, 1);
643 /* last block will be modified if it is not aligned */
644 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
645 dmu_tx_count_write(txh, off+len, 1);
646 }
647
648 /*
649 * Check level-1 blocks.
650 */
651 if (dn->dn_nlevels > 1) {
652 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
653 SPA_BLKPTRSHIFT;
654 uint64_t start = off >> shift;
655 uint64_t end = (off + len) >> shift;
656 uint64_t i;
657
658 ASSERT(dn->dn_indblkshift != 0);
659
660 /*
661 * dnode_reallocate() can result in an object with indirect
662 * blocks having an odd data block size. In this case,
663 * just check the single block.
664 */
665 if (dn->dn_datablkshift == 0)
666 start = end = 0;
667
668 zio = zio_root(tx->tx_pool->dp_spa,
669 NULL, NULL, ZIO_FLAG_CANFAIL);
670 for (i = start; i <= end; i++) {
671 uint64_t ibyte = i << shift;
672 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
673 i = ibyte >> shift;
674 if (err == ESRCH)
675 break;
676 if (err) {
677 tx->tx_err = err;
678 return;
679 }
680
681 err = dmu_tx_check_ioerr(zio, dn, 1, i);
682 if (err) {
683 tx->tx_err = err;
684 return;
685 }
686 }
687 err = zio_wait(zio);
688 if (err) {
689 tx->tx_err = err;
690 return;
691 }
692 }
693
694 dmu_tx_count_free(txh, off, len);
695 }
696
697 void
698 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
699 {
700 dmu_tx_hold_t *txh;
701 dnode_t *dn;
702 uint64_t nblocks;
703 int epbs, err;
704
705 ASSERT(tx->tx_txg == 0);
706
707 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
708 object, THT_ZAP, add, (uintptr_t)name);
709 if (txh == NULL)
710 return;
711 dn = txh->txh_dnode;
712
713 dmu_tx_count_dnode(txh);
714
715 if (dn == NULL) {
716 /*
717 * We will be able to fit a new object's entries into one leaf
718 * block. So there will be at most 2 blocks total,
719 * including the header block.
720 */
721 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
722 return;
723 }
724
725 ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
726
727 if (dn->dn_maxblkid == 0 && !add) {
728 blkptr_t *bp;
729
730 /*
731 * If there is only one block (i.e. this is a micro-zap)
732 * and we are not adding anything, the accounting is simple.
733 */
734 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
735 if (err) {
736 tx->tx_err = err;
737 return;
738 }
739
740 /*
741 * Use max block size here, since we don't know how much
742 * the size will change between now and the dbuf dirty call.
743 */
744 bp = &dn->dn_phys->dn_blkptr[0];
745 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
746 bp, bp->blk_birth))
747 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
748 else
749 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
750 if (!BP_IS_HOLE(bp))
751 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
752 return;
753 }
754
755 if (dn->dn_maxblkid > 0 && name) {
756 /*
757 * access the name in this fat-zap so that we'll check
758 * for i/o errors to the leaf blocks, etc.
759 */
760 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
761 8, 0, NULL);
762 if (err == EIO) {
763 tx->tx_err = err;
764 return;
765 }
766 }
767
768 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
769 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
770
771 /*
772 * If the modified blocks are scattered to the four winds,
773 * we'll have to modify an indirect twig for each.
774 */
775 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
776 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
777 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
778 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
779 else
780 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
781 }
782
783 void
784 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
785 {
786 dmu_tx_hold_t *txh;
787
788 ASSERT(tx->tx_txg == 0);
789
790 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
791 object, THT_BONUS, 0, 0);
792 if (txh)
793 dmu_tx_count_dnode(txh);
794 }
795
796 void
797 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
798 {
799 dmu_tx_hold_t *txh;
800
801 ASSERT(tx->tx_txg == 0);
802
803 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
804 DMU_NEW_OBJECT, THT_SPACE, space, 0);
805 if (txh)
806 txh->txh_space_towrite += space;
807 }
808
809 int
810 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
811 {
812 dmu_tx_hold_t *txh;
813 int holds = 0;
814
815 /*
816 * By asserting that the tx is assigned, we're counting the
817 * number of dn_tx_holds, which is the same as the number of
818 * dn_holds. Otherwise, we'd be counting dn_holds, but
819 * dn_tx_holds could be 0.
820 */
821 ASSERT(tx->tx_txg != 0);
822
823 /* if (tx->tx_anyobj == TRUE) */
824 /* return (0); */
825
826 for (txh = list_head(&tx->tx_holds); txh;
827 txh = list_next(&tx->tx_holds, txh)) {
828 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
829 holds++;
830 }
831
832 return (holds);
833 }
834
835 #ifdef DEBUG_DMU_TX
836 void
837 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
838 {
839 dmu_tx_hold_t *txh;
840 int match_object = FALSE, match_offset = FALSE;
841 dnode_t *dn;
842
843 DB_DNODE_ENTER(db);
844 dn = DB_DNODE(db);
845 ASSERT(dn != NULL);
846 ASSERT(tx->tx_txg != 0);
847 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
848 ASSERT3U(dn->dn_object, ==, db->db.db_object);
849
850 if (tx->tx_anyobj) {
851 DB_DNODE_EXIT(db);
852 return;
853 }
854
855 /* XXX No checking on the meta dnode for now */
856 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
857 DB_DNODE_EXIT(db);
858 return;
859 }
860
861 for (txh = list_head(&tx->tx_holds); txh;
862 txh = list_next(&tx->tx_holds, txh)) {
863 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
864 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
865 match_object = TRUE;
866 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
867 int datablkshift = dn->dn_datablkshift ?
868 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
869 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
870 int shift = datablkshift + epbs * db->db_level;
871 uint64_t beginblk = shift >= 64 ? 0 :
872 (txh->txh_arg1 >> shift);
873 uint64_t endblk = shift >= 64 ? 0 :
874 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
875 uint64_t blkid = db->db_blkid;
876
877 /* XXX txh_arg2 better not be zero... */
878
879 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
880 txh->txh_type, beginblk, endblk);
881
882 switch (txh->txh_type) {
883 case THT_WRITE:
884 if (blkid >= beginblk && blkid <= endblk)
885 match_offset = TRUE;
886 /*
887 * We will let this hold work for the bonus
888 * or spill buffer so that we don't need to
889 * hold it when creating a new object.
890 */
891 if (blkid == DMU_BONUS_BLKID ||
892 blkid == DMU_SPILL_BLKID)
893 match_offset = TRUE;
894 /*
895 * They might have to increase nlevels,
896 * thus dirtying the new TLIBs. Or the
897 * might have to change the block size,
898 * thus dirying the new lvl=0 blk=0.
899 */
900 if (blkid == 0)
901 match_offset = TRUE;
902 break;
903 case THT_FREE:
904 /*
905 * We will dirty all the level 1 blocks in
906 * the free range and perhaps the first and
907 * last level 0 block.
908 */
909 if (blkid >= beginblk && (blkid <= endblk ||
910 txh->txh_arg2 == DMU_OBJECT_END))
911 match_offset = TRUE;
912 break;
913 case THT_SPILL:
914 if (blkid == DMU_SPILL_BLKID)
915 match_offset = TRUE;
916 break;
917 case THT_BONUS:
918 if (blkid == DMU_BONUS_BLKID)
919 match_offset = TRUE;
920 break;
921 case THT_ZAP:
922 match_offset = TRUE;
923 break;
924 case THT_NEWOBJECT:
925 match_object = TRUE;
926 break;
927 default:
928 cmn_err(CE_PANIC, "bad txh_type %d",
929 txh->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 int err;
1257
1258 ASSERT(tx->tx_txg == 0);
1259 ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1260 txg_how == TXG_WAITED);
1261 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1262
1263 if (txg_how == TXG_WAITED)
1264 tx->tx_waited = B_TRUE;
1265
1266 /* If we might wait, we must not hold the config lock. */
1267 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1268
1269 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1270 dmu_tx_unassign(tx);
1271
1272 if (err != ERESTART || txg_how != TXG_WAIT)
1273 return (err);
1274
1275 dmu_tx_wait(tx);
1276 }
1277
1278 txg_rele_to_quiesce(&tx->tx_txgh);
1279
1280 return (0);
1281 }
1282
1283 void
1284 dmu_tx_wait(dmu_tx_t *tx)
1285 {
1286 spa_t *spa = tx->tx_pool->dp_spa;
1287 dsl_pool_t *dp = tx->tx_pool;
1288 hrtime_t before;
1289
1290 ASSERT(tx->tx_txg == 0);
1291 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1292
1293 before = gethrtime();
1294
1295 if (tx->tx_wait_dirty) {
1296 uint64_t dirty;
1297
1298 /*
1299 * dmu_tx_try_assign() has determined that we need to wait
1300 * because we've consumed much or all of the dirty buffer
1301 * space.
1302 */
1303 mutex_enter(&dp->dp_lock);
1304 if (dp->dp_dirty_total >= zfs_dirty_data_max)
1305 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
1306 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1307 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1308 dirty = dp->dp_dirty_total;
1309 mutex_exit(&dp->dp_lock);
1310
1311 dmu_tx_delay(tx, dirty);
1312
1313 tx->tx_wait_dirty = B_FALSE;
1314
1315 /*
1316 * Note: setting tx_waited only has effect if the caller
1317 * used TX_WAIT. Otherwise they are going to destroy
1318 * this tx and try again. The common case, zfs_write(),
1319 * uses TX_WAIT.
1320 */
1321 tx->tx_waited = B_TRUE;
1322 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1323 /*
1324 * If the pool is suspended we need to wait until it
1325 * is resumed. Note that it's possible that the pool
1326 * has become active after this thread has tried to
1327 * obtain a tx. If that's the case then tx_lasttried_txg
1328 * would not have been set.
1329 */
1330 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1331 } else if (tx->tx_needassign_txh) {
1332 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1333
1334 mutex_enter(&dn->dn_mtx);
1335 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1336 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1337 mutex_exit(&dn->dn_mtx);
1338 tx->tx_needassign_txh = NULL;
1339 } else {
1340 /*
1341 * A dnode is assigned to the quiescing txg. Wait for its
1342 * transaction to complete.
1343 */
1344 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1345 }
1346
1347 spa_tx_assign_add_nsecs(spa, gethrtime() - before);
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_SLEEP);
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