]> git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/dmu_tx.c
projects / mirror_zfs.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Illumos 5027 - zfs large block support
[mirror_zfs.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 = highbit64(txh->txh_tx->tx_objset->os_recordsize) - 1;
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 } else {
314 /*
315 * The blocksize can increase up to the recordsize,
316 * or if it is already more than the recordsize,
317 * up to the next power of 2.
318 */
319 min_bs = highbit64(dn->dn_datablksz - 1);
320 max_bs = MAX(max_bs, highbit64(dn->dn_datablksz - 1));
321 }
322
323 /*
324 * If this write is not off the end of the file
325 * we need to account for overwrites/unref.
326 */
327 if (start <= dn->dn_maxblkid) {
328 for (l = 0; l < DN_MAX_LEVELS; l++)
329 history[l] = -1ULL;
330 }
331 while (start <= dn->dn_maxblkid) {
332 dmu_buf_impl_t *db;
333
334 rw_enter(&dn->dn_struct_rwlock, RW_READER);
335 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
336 rw_exit(&dn->dn_struct_rwlock);
337
338 if (err) {
339 txh->txh_tx->tx_err = err;
340 return;
341 }
342
343 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
344 history);
345 dbuf_rele(db, FTAG);
346 if (++start > end) {
347 /*
348 * Account for new indirects appearing
349 * before this IO gets assigned into a txg.
350 */
351 bits = 64 - min_bs;
352 epbs = min_ibs - SPA_BLKPTRSHIFT;
353 for (bits -= epbs * (nlvls - 1);
354 bits >= 0; bits -= epbs)
355 txh->txh_fudge += 1ULL << max_ibs;
356 goto out;
357 }
358 off += delta;
359 if (len >= delta)
360 len -= delta;
361 delta = dn->dn_datablksz;
362 }
363 }
364
365 /*
366 * 'end' is the last thing we will access, not one past.
367 * This way we won't overflow when accessing the last byte.
368 */
369 start = P2ALIGN(off, 1ULL << max_bs);
370 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
371 txh->txh_space_towrite += end - start + 1;
372
373 start >>= min_bs;
374 end >>= min_bs;
375
376 epbs = min_ibs - SPA_BLKPTRSHIFT;
377
378 /*
379 * The object contains at most 2^(64 - min_bs) blocks,
380 * and each indirect level maps 2^epbs.
381 */
382 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
383 start >>= epbs;
384 end >>= epbs;
385 ASSERT3U(end, >=, start);
386 txh->txh_space_towrite += (end - start + 1) << max_ibs;
387 if (start != 0) {
388 /*
389 * We also need a new blkid=0 indirect block
390 * to reference any existing file data.
391 */
392 txh->txh_space_towrite += 1ULL << max_ibs;
393 }
394 }
395
396 out:
397 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
398 2 * DMU_MAX_ACCESS)
399 err = SET_ERROR(EFBIG);
400
401 if (err)
402 txh->txh_tx->tx_err = err;
403 }
404
405 static void
406 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
407 {
408 dnode_t *dn = txh->txh_dnode;
409 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
410 uint64_t space = mdn->dn_datablksz +
411 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
412
413 if (dn && dn->dn_dbuf->db_blkptr &&
414 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
415 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
416 txh->txh_space_tooverwrite += space;
417 txh->txh_space_tounref += space;
418 } else {
419 txh->txh_space_towrite += space;
420 if (dn && dn->dn_dbuf->db_blkptr)
421 txh->txh_space_tounref += space;
422 }
423 }
424
425 void
426 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
427 {
428 dmu_tx_hold_t *txh;
429
430 ASSERT(tx->tx_txg == 0);
431 ASSERT(len <= DMU_MAX_ACCESS);
432 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
433
434 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
435 object, THT_WRITE, off, len);
436 if (txh == NULL)
437 return;
438
439 dmu_tx_count_write(txh, off, len);
440 dmu_tx_count_dnode(txh);
441 }
442
443 static void
444 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
445 {
446 uint64_t blkid, nblks, lastblk;
447 uint64_t space = 0, unref = 0, skipped = 0;
448 dnode_t *dn = txh->txh_dnode;
449 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
450 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
451 int epbs;
452 uint64_t l0span = 0, nl1blks = 0;
453
454 if (dn->dn_nlevels == 0)
455 return;
456
457 /*
458 * The struct_rwlock protects us against dn_nlevels
459 * changing, in case (against all odds) we manage to dirty &
460 * sync out the changes after we check for being dirty.
461 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
462 */
463 rw_enter(&dn->dn_struct_rwlock, RW_READER);
464 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
465 if (dn->dn_maxblkid == 0) {
466 if (off == 0 && len >= dn->dn_datablksz) {
467 blkid = 0;
468 nblks = 1;
469 } else {
470 rw_exit(&dn->dn_struct_rwlock);
471 return;
472 }
473 } else {
474 blkid = off >> dn->dn_datablkshift;
475 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
476
477 if (blkid > dn->dn_maxblkid) {
478 rw_exit(&dn->dn_struct_rwlock);
479 return;
480 }
481 if (blkid + nblks > dn->dn_maxblkid)
482 nblks = dn->dn_maxblkid - blkid + 1;
483
484 }
485 l0span = nblks; /* save for later use to calc level > 1 overhead */
486 if (dn->dn_nlevels == 1) {
487 int i;
488 for (i = 0; i < nblks; i++) {
489 blkptr_t *bp = dn->dn_phys->dn_blkptr;
490 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
491 bp += blkid + i;
492 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
493 dprintf_bp(bp, "can free old%s", "");
494 space += bp_get_dsize(spa, bp);
495 }
496 unref += BP_GET_ASIZE(bp);
497 }
498 nl1blks = 1;
499 nblks = 0;
500 }
501
502 lastblk = blkid + nblks - 1;
503 while (nblks) {
504 dmu_buf_impl_t *dbuf;
505 uint64_t ibyte, new_blkid;
506 int epb = 1 << epbs;
507 int err, i, blkoff, tochk;
508 blkptr_t *bp;
509
510 ibyte = blkid << dn->dn_datablkshift;
511 err = dnode_next_offset(dn,
512 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
513 new_blkid = ibyte >> dn->dn_datablkshift;
514 if (err == ESRCH) {
515 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
516 break;
517 }
518 if (err) {
519 txh->txh_tx->tx_err = err;
520 break;
521 }
522 if (new_blkid > lastblk) {
523 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
524 break;
525 }
526
527 if (new_blkid > blkid) {
528 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
529 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
530 nblks -= new_blkid - blkid;
531 blkid = new_blkid;
532 }
533 blkoff = P2PHASE(blkid, epb);
534 tochk = MIN(epb - blkoff, nblks);
535
536 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
537 if (err) {
538 txh->txh_tx->tx_err = err;
539 break;
540 }
541
542 txh->txh_memory_tohold += dbuf->db.db_size;
543
544 /*
545 * We don't check memory_tohold against DMU_MAX_ACCESS because
546 * memory_tohold is an over-estimation (especially the >L1
547 * indirect blocks), so it could fail. Callers should have
548 * already verified that they will not be holding too much
549 * memory.
550 */
551
552 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
553 if (err != 0) {
554 txh->txh_tx->tx_err = err;
555 dbuf_rele(dbuf, FTAG);
556 break;
557 }
558
559 bp = dbuf->db.db_data;
560 bp += blkoff;
561
562 for (i = 0; i < tochk; i++) {
563 if (dsl_dataset_block_freeable(ds, &bp[i],
564 bp[i].blk_birth)) {
565 dprintf_bp(&bp[i], "can free old%s", "");
566 space += bp_get_dsize(spa, &bp[i]);
567 }
568 unref += BP_GET_ASIZE(bp);
569 }
570 dbuf_rele(dbuf, FTAG);
571
572 ++nl1blks;
573 blkid += tochk;
574 nblks -= tochk;
575 }
576 rw_exit(&dn->dn_struct_rwlock);
577
578 /*
579 * Add in memory requirements of higher-level indirects.
580 * This assumes a worst-possible scenario for dn_nlevels and a
581 * worst-possible distribution of l1-blocks over the region to free.
582 */
583 {
584 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
585 int level = 2;
586 /*
587 * Here we don't use DN_MAX_LEVEL, but calculate it with the
588 * given datablkshift and indblkshift. This makes the
589 * difference between 19 and 8 on large files.
590 */
591 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
592 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
593
594 while (level++ < maxlevel) {
595 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
596 << dn->dn_indblkshift;
597 blkcnt = 1 + (blkcnt >> epbs);
598 }
599 }
600
601 /* account for new level 1 indirect blocks that might show up */
602 if (skipped > 0) {
603 txh->txh_fudge += skipped << dn->dn_indblkshift;
604 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
605 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
606 }
607 txh->txh_space_tofree += space;
608 txh->txh_space_tounref += unref;
609 }
610
611 void
612 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
613 {
614 dmu_tx_hold_t *txh;
615 dnode_t *dn;
616 int err;
617 zio_t *zio;
618
619 ASSERT(tx->tx_txg == 0);
620
621 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
622 object, THT_FREE, off, len);
623 if (txh == NULL)
624 return;
625 dn = txh->txh_dnode;
626 dmu_tx_count_dnode(txh);
627
628 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
629 return;
630 if (len == DMU_OBJECT_END)
631 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
632
633 dmu_tx_count_dnode(txh);
634
635 /*
636 * For i/o error checking, we read the first and last level-0
637 * blocks if they are not aligned, and all the level-1 blocks.
638 *
639 * Note: dbuf_free_range() assumes that we have not instantiated
640 * any level-0 dbufs that will be completely freed. Therefore we must
641 * exercise care to not read or count the first and last blocks
642 * if they are blocksize-aligned.
643 */
644 if (dn->dn_datablkshift == 0) {
645 if (off != 0 || len < dn->dn_datablksz)
646 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
647 } else {
648 /* first block will be modified if it is not aligned */
649 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
650 dmu_tx_count_write(txh, off, 1);
651 /* last block will be modified if it is not aligned */
652 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
653 dmu_tx_count_write(txh, off+len, 1);
654 }
655
656 /*
657 * Check level-1 blocks.
658 */
659 if (dn->dn_nlevels > 1) {
660 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
661 SPA_BLKPTRSHIFT;
662 uint64_t start = off >> shift;
663 uint64_t end = (off + len) >> shift;
664 uint64_t i;
665
666 ASSERT(dn->dn_indblkshift != 0);
667
668 /*
669 * dnode_reallocate() can result in an object with indirect
670 * blocks having an odd data block size. In this case,
671 * just check the single block.
672 */
673 if (dn->dn_datablkshift == 0)
674 start = end = 0;
675
676 zio = zio_root(tx->tx_pool->dp_spa,
677 NULL, NULL, ZIO_FLAG_CANFAIL);
678 for (i = start; i <= end; i++) {
679 uint64_t ibyte = i << shift;
680 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
681 i = ibyte >> shift;
682 if (err == ESRCH)
683 break;
684 if (err) {
685 tx->tx_err = err;
686 return;
687 }
688
689 err = dmu_tx_check_ioerr(zio, dn, 1, i);
690 if (err) {
691 tx->tx_err = err;
692 return;
693 }
694 }
695 err = zio_wait(zio);
696 if (err) {
697 tx->tx_err = err;
698 return;
699 }
700 }
701
702 dmu_tx_count_free(txh, off, len);
703 }
704
705 void
706 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
707 {
708 dmu_tx_hold_t *txh;
709 dnode_t *dn;
710 dsl_dataset_phys_t *ds_phys;
711 uint64_t nblocks;
712 int epbs, err;
713
714 ASSERT(tx->tx_txg == 0);
715
716 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
717 object, THT_ZAP, add, (uintptr_t)name);
718 if (txh == NULL)
719 return;
720 dn = txh->txh_dnode;
721
722 dmu_tx_count_dnode(txh);
723
724 if (dn == NULL) {
725 /*
726 * We will be able to fit a new object's entries into one leaf
727 * block. So there will be at most 2 blocks total,
728 * including the header block.
729 */
730 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
731 return;
732 }
733
734 ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
735
736 if (dn->dn_maxblkid == 0 && !add) {
737 blkptr_t *bp;
738
739 /*
740 * If there is only one block (i.e. this is a micro-zap)
741 * and we are not adding anything, the accounting is simple.
742 */
743 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
744 if (err) {
745 tx->tx_err = err;
746 return;
747 }
748
749 /*
750 * Use max block size here, since we don't know how much
751 * the size will change between now and the dbuf dirty call.
752 */
753 bp = &dn->dn_phys->dn_blkptr[0];
754 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
755 bp, bp->blk_birth))
756 txh->txh_space_tooverwrite += MZAP_MAX_BLKSZ;
757 else
758 txh->txh_space_towrite += MZAP_MAX_BLKSZ;
759 if (!BP_IS_HOLE(bp))
760 txh->txh_space_tounref += MZAP_MAX_BLKSZ;
761 return;
762 }
763
764 if (dn->dn_maxblkid > 0 && name) {
765 /*
766 * access the name in this fat-zap so that we'll check
767 * for i/o errors to the leaf blocks, etc.
768 */
769 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
770 8, 0, NULL);
771 if (err == EIO) {
772 tx->tx_err = err;
773 return;
774 }
775 }
776
777 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
778 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
779
780 /*
781 * If the modified blocks are scattered to the four winds,
782 * we'll have to modify an indirect twig for each.
783 */
784 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
785 ds_phys = dsl_dataset_phys(dn->dn_objset->os_dsl_dataset);
786 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
787 if (ds_phys->ds_prev_snap_obj)
788 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
789 else
790 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
791 }
792
793 void
794 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
795 {
796 dmu_tx_hold_t *txh;
797
798 ASSERT(tx->tx_txg == 0);
799
800 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
801 object, THT_BONUS, 0, 0);
802 if (txh)
803 dmu_tx_count_dnode(txh);
804 }
805
806 void
807 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
808 {
809 dmu_tx_hold_t *txh;
810
811 ASSERT(tx->tx_txg == 0);
812
813 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
814 DMU_NEW_OBJECT, THT_SPACE, space, 0);
815 if (txh)
816 txh->txh_space_towrite += space;
817 }
818
819 int
820 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
821 {
822 dmu_tx_hold_t *txh;
823 int holds = 0;
824
825 /*
826 * By asserting that the tx is assigned, we're counting the
827 * number of dn_tx_holds, which is the same as the number of
828 * dn_holds. Otherwise, we'd be counting dn_holds, but
829 * dn_tx_holds could be 0.
830 */
831 ASSERT(tx->tx_txg != 0);
832
833 /* if (tx->tx_anyobj == TRUE) */
834 /* return (0); */
835
836 for (txh = list_head(&tx->tx_holds); txh;
837 txh = list_next(&tx->tx_holds, txh)) {
838 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
839 holds++;
840 }
841
842 return (holds);
843 }
844
845 #ifdef DEBUG_DMU_TX
846 void
847 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
848 {
849 dmu_tx_hold_t *txh;
850 int match_object = FALSE, match_offset = FALSE;
851 dnode_t *dn;
852
853 DB_DNODE_ENTER(db);
854 dn = DB_DNODE(db);
855 ASSERT(dn != NULL);
856 ASSERT(tx->tx_txg != 0);
857 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
858 ASSERT3U(dn->dn_object, ==, db->db.db_object);
859
860 if (tx->tx_anyobj) {
861 DB_DNODE_EXIT(db);
862 return;
863 }
864
865 /* XXX No checking on the meta dnode for now */
866 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
867 DB_DNODE_EXIT(db);
868 return;
869 }
870
871 for (txh = list_head(&tx->tx_holds); txh;
872 txh = list_next(&tx->tx_holds, txh)) {
873 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
874 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
875 match_object = TRUE;
876 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
877 int datablkshift = dn->dn_datablkshift ?
878 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
879 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
880 int shift = datablkshift + epbs * db->db_level;
881 uint64_t beginblk = shift >= 64 ? 0 :
882 (txh->txh_arg1 >> shift);
883 uint64_t endblk = shift >= 64 ? 0 :
884 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
885 uint64_t blkid = db->db_blkid;
886
887 /* XXX txh_arg2 better not be zero... */
888
889 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
890 txh->txh_type, beginblk, endblk);
891
892 switch (txh->txh_type) {
893 case THT_WRITE:
894 if (blkid >= beginblk && blkid <= endblk)
895 match_offset = TRUE;
896 /*
897 * We will let this hold work for the bonus
898 * or spill buffer so that we don't need to
899 * hold it when creating a new object.
900 */
901 if (blkid == DMU_BONUS_BLKID ||
902 blkid == DMU_SPILL_BLKID)
903 match_offset = TRUE;
904 /*
905 * They might have to increase nlevels,
906 * thus dirtying the new TLIBs. Or the
907 * might have to change the block size,
908 * thus dirying the new lvl=0 blk=0.
909 */
910 if (blkid == 0)
911 match_offset = TRUE;
912 break;
913 case THT_FREE:
914 /*
915 * We will dirty all the level 1 blocks in
916 * the free range and perhaps the first and
917 * last level 0 block.
918 */
919 if (blkid >= beginblk && (blkid <= endblk ||
920 txh->txh_arg2 == DMU_OBJECT_END))
921 match_offset = TRUE;
922 break;
923 case THT_SPILL:
924 if (blkid == DMU_SPILL_BLKID)
925 match_offset = TRUE;
926 break;
927 case THT_BONUS:
928 if (blkid == DMU_BONUS_BLKID)
929 match_offset = TRUE;
930 break;
931 case THT_ZAP:
932 match_offset = TRUE;
933 break;
934 case THT_NEWOBJECT:
935 match_object = TRUE;
936 break;
937 default:
938 cmn_err(CE_PANIC, "bad txh_type %d",
939 txh->txh_type);
940 }
941 }
942 if (match_object && match_offset) {
943 DB_DNODE_EXIT(db);
944 return;
945 }
946 }
947 DB_DNODE_EXIT(db);
948 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
949 (u_longlong_t)db->db.db_object, db->db_level,
950 (u_longlong_t)db->db_blkid);
951 }
952 #endif
953
954 /*
955 * If we can't do 10 iops, something is wrong. Let us go ahead
956 * and hit zfs_dirty_data_max.
957 */
958 hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */
959 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
960
961 /*
962 * We delay transactions when we've determined that the backend storage
963 * isn't able to accommodate the rate of incoming writes.
964 *
965 * If there is already a transaction waiting, we delay relative to when
966 * that transaction finishes waiting. This way the calculated min_time
967 * is independent of the number of threads concurrently executing
968 * transactions.
969 *
970 * If we are the only waiter, wait relative to when the transaction
971 * started, rather than the current time. This credits the transaction for
972 * "time already served", e.g. reading indirect blocks.
973 *
974 * The minimum time for a transaction to take is calculated as:
975 * min_time = scale * (dirty - min) / (max - dirty)
976 * min_time is then capped at zfs_delay_max_ns.
977 *
978 * The delay has two degrees of freedom that can be adjusted via tunables.
979 * The percentage of dirty data at which we start to delay is defined by
980 * zfs_delay_min_dirty_percent. This should typically be at or above
981 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
982 * delay after writing at full speed has failed to keep up with the incoming
983 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
984 * speaking, this variable determines the amount of delay at the midpoint of
985 * the curve.
986 *
987 * delay
988 * 10ms +-------------------------------------------------------------*+
989 * | *|
990 * 9ms + *+
991 * | *|
992 * 8ms + *+
993 * | * |
994 * 7ms + * +
995 * | * |
996 * 6ms + * +
997 * | * |
998 * 5ms + * +
999 * | * |
1000 * 4ms + * +
1001 * | * |
1002 * 3ms + * +
1003 * | * |
1004 * 2ms + (midpoint) * +
1005 * | | ** |
1006 * 1ms + v *** +
1007 * | zfs_delay_scale ----------> ******** |
1008 * 0 +-------------------------------------*********----------------+
1009 * 0% <- zfs_dirty_data_max -> 100%
1010 *
1011 * Note that since the delay is added to the outstanding time remaining on the
1012 * most recent transaction, the delay is effectively the inverse of IOPS.
1013 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
1014 * was chosen such that small changes in the amount of accumulated dirty data
1015 * in the first 3/4 of the curve yield relatively small differences in the
1016 * amount of delay.
1017 *
1018 * The effects can be easier to understand when the amount of delay is
1019 * represented on a log scale:
1020 *
1021 * delay
1022 * 100ms +-------------------------------------------------------------++
1023 * + +
1024 * | |
1025 * + *+
1026 * 10ms + *+
1027 * + ** +
1028 * | (midpoint) ** |
1029 * + | ** +
1030 * 1ms + v **** +
1031 * + zfs_delay_scale ----------> ***** +
1032 * | **** |
1033 * + **** +
1034 * 100us + ** +
1035 * + * +
1036 * | * |
1037 * + * +
1038 * 10us + * +
1039 * + +
1040 * | |
1041 * + +
1042 * +--------------------------------------------------------------+
1043 * 0% <- zfs_dirty_data_max -> 100%
1044 *
1045 * Note here that only as the amount of dirty data approaches its limit does
1046 * the delay start to increase rapidly. The goal of a properly tuned system
1047 * should be to keep the amount of dirty data out of that range by first
1048 * ensuring that the appropriate limits are set for the I/O scheduler to reach
1049 * optimal throughput on the backend storage, and then by changing the value
1050 * of zfs_delay_scale to increase the steepness of the curve.
1051 */
1052 static void
1053 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
1054 {
1055 dsl_pool_t *dp = tx->tx_pool;
1056 uint64_t delay_min_bytes =
1057 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
1058 hrtime_t wakeup, min_tx_time, now;
1059
1060 if (dirty <= delay_min_bytes)
1061 return;
1062
1063 /*
1064 * The caller has already waited until we are under the max.
1065 * We make them pass us the amount of dirty data so we don't
1066 * have to handle the case of it being >= the max, which could
1067 * cause a divide-by-zero if it's == the max.
1068 */
1069 ASSERT3U(dirty, <, zfs_dirty_data_max);
1070
1071 now = gethrtime();
1072 min_tx_time = zfs_delay_scale *
1073 (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
1074 min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
1075 if (now > tx->tx_start + min_tx_time)
1076 return;
1077
1078 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1079 uint64_t, min_tx_time);
1080
1081 mutex_enter(&dp->dp_lock);
1082 wakeup = MAX(tx->tx_start + min_tx_time,
1083 dp->dp_last_wakeup + min_tx_time);
1084 dp->dp_last_wakeup = wakeup;
1085 mutex_exit(&dp->dp_lock);
1086
1087 zfs_sleep_until(wakeup);
1088 }
1089
1090 static int
1091 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1092 {
1093 dmu_tx_hold_t *txh;
1094 spa_t *spa = tx->tx_pool->dp_spa;
1095 uint64_t memory, asize, fsize, usize;
1096 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1097
1098 ASSERT0(tx->tx_txg);
1099
1100 if (tx->tx_err) {
1101 DMU_TX_STAT_BUMP(dmu_tx_error);
1102 return (tx->tx_err);
1103 }
1104
1105 if (spa_suspended(spa)) {
1106 DMU_TX_STAT_BUMP(dmu_tx_suspended);
1107
1108 /*
1109 * If the user has indicated a blocking failure mode
1110 * then return ERESTART which will block in dmu_tx_wait().
1111 * Otherwise, return EIO so that an error can get
1112 * propagated back to the VOP calls.
1113 *
1114 * Note that we always honor the txg_how flag regardless
1115 * of the failuremode setting.
1116 */
1117 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1118 txg_how != TXG_WAIT)
1119 return (SET_ERROR(EIO));
1120
1121 return (SET_ERROR(ERESTART));
1122 }
1123
1124 if (!tx->tx_waited &&
1125 dsl_pool_need_dirty_delay(tx->tx_pool)) {
1126 tx->tx_wait_dirty = B_TRUE;
1127 DMU_TX_STAT_BUMP(dmu_tx_dirty_delay);
1128 return (ERESTART);
1129 }
1130
1131 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1132 tx->tx_needassign_txh = NULL;
1133
1134 /*
1135 * NB: No error returns are allowed after txg_hold_open, but
1136 * before processing the dnode holds, due to the
1137 * dmu_tx_unassign() logic.
1138 */
1139
1140 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1141 for (txh = list_head(&tx->tx_holds); txh;
1142 txh = list_next(&tx->tx_holds, txh)) {
1143 dnode_t *dn = txh->txh_dnode;
1144 if (dn != NULL) {
1145 mutex_enter(&dn->dn_mtx);
1146 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1147 mutex_exit(&dn->dn_mtx);
1148 tx->tx_needassign_txh = txh;
1149 DMU_TX_STAT_BUMP(dmu_tx_group);
1150 return (SET_ERROR(ERESTART));
1151 }
1152 if (dn->dn_assigned_txg == 0)
1153 dn->dn_assigned_txg = tx->tx_txg;
1154 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1155 (void) refcount_add(&dn->dn_tx_holds, tx);
1156 mutex_exit(&dn->dn_mtx);
1157 }
1158 towrite += txh->txh_space_towrite;
1159 tofree += txh->txh_space_tofree;
1160 tooverwrite += txh->txh_space_tooverwrite;
1161 tounref += txh->txh_space_tounref;
1162 tohold += txh->txh_memory_tohold;
1163 fudge += txh->txh_fudge;
1164 }
1165
1166 /*
1167 * If a snapshot has been taken since we made our estimates,
1168 * assume that we won't be able to free or overwrite anything.
1169 */
1170 if (tx->tx_objset &&
1171 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1172 tx->tx_lastsnap_txg) {
1173 towrite += tooverwrite;
1174 tooverwrite = tofree = 0;
1175 }
1176
1177 /* needed allocation: worst-case estimate of write space */
1178 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1179 /* freed space estimate: worst-case overwrite + free estimate */
1180 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1181 /* convert unrefd space to worst-case estimate */
1182 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1183 /* calculate memory footprint estimate */
1184 memory = towrite + tooverwrite + tohold;
1185
1186 #ifdef DEBUG_DMU_TX
1187 /*
1188 * Add in 'tohold' to account for our dirty holds on this memory
1189 * XXX - the "fudge" factor is to account for skipped blocks that
1190 * we missed because dnode_next_offset() misses in-core-only blocks.
1191 */
1192 tx->tx_space_towrite = asize +
1193 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1194 tx->tx_space_tofree = tofree;
1195 tx->tx_space_tooverwrite = tooverwrite;
1196 tx->tx_space_tounref = tounref;
1197 #endif
1198
1199 if (tx->tx_dir && asize != 0) {
1200 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1201 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1202 if (err)
1203 return (err);
1204 }
1205
1206 DMU_TX_STAT_BUMP(dmu_tx_assigned);
1207
1208 return (0);
1209 }
1210
1211 static void
1212 dmu_tx_unassign(dmu_tx_t *tx)
1213 {
1214 dmu_tx_hold_t *txh;
1215
1216 if (tx->tx_txg == 0)
1217 return;
1218
1219 txg_rele_to_quiesce(&tx->tx_txgh);
1220
1221 /*
1222 * Walk the transaction's hold list, removing the hold on the
1223 * associated dnode, and notifying waiters if the refcount drops to 0.
1224 */
1225 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1226 txh = list_next(&tx->tx_holds, txh)) {
1227 dnode_t *dn = txh->txh_dnode;
1228
1229 if (dn == NULL)
1230 continue;
1231 mutex_enter(&dn->dn_mtx);
1232 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1233
1234 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1235 dn->dn_assigned_txg = 0;
1236 cv_broadcast(&dn->dn_notxholds);
1237 }
1238 mutex_exit(&dn->dn_mtx);
1239 }
1240
1241 txg_rele_to_sync(&tx->tx_txgh);
1242
1243 tx->tx_lasttried_txg = tx->tx_txg;
1244 tx->tx_txg = 0;
1245 }
1246
1247 /*
1248 * Assign tx to a transaction group. txg_how can be one of:
1249 *
1250 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1251 * a new one. This should be used when you're not holding locks.
1252 * It will only fail if we're truly out of space (or over quota).
1253 *
1254 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1255 * blocking, returns immediately with ERESTART. This should be used
1256 * whenever you're holding locks. On an ERESTART error, the caller
1257 * should drop locks, do a dmu_tx_wait(tx), and try again.
1258 *
1259 * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1260 * has already been called on behalf of this operation (though
1261 * most likely on a different tx).
1262 */
1263 int
1264 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1265 {
1266 int err;
1267
1268 ASSERT(tx->tx_txg == 0);
1269 ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1270 txg_how == TXG_WAITED);
1271 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1272
1273 if (txg_how == TXG_WAITED)
1274 tx->tx_waited = B_TRUE;
1275
1276 /* If we might wait, we must not hold the config lock. */
1277 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1278
1279 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1280 dmu_tx_unassign(tx);
1281
1282 if (err != ERESTART || txg_how != TXG_WAIT)
1283 return (err);
1284
1285 dmu_tx_wait(tx);
1286 }
1287
1288 txg_rele_to_quiesce(&tx->tx_txgh);
1289
1290 return (0);
1291 }
1292
1293 void
1294 dmu_tx_wait(dmu_tx_t *tx)
1295 {
1296 spa_t *spa = tx->tx_pool->dp_spa;
1297 dsl_pool_t *dp = tx->tx_pool;
1298 hrtime_t before;
1299
1300 ASSERT(tx->tx_txg == 0);
1301 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1302
1303 before = gethrtime();
1304
1305 if (tx->tx_wait_dirty) {
1306 uint64_t dirty;
1307
1308 /*
1309 * dmu_tx_try_assign() has determined that we need to wait
1310 * because we've consumed much or all of the dirty buffer
1311 * space.
1312 */
1313 mutex_enter(&dp->dp_lock);
1314 if (dp->dp_dirty_total >= zfs_dirty_data_max)
1315 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
1316 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1317 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1318 dirty = dp->dp_dirty_total;
1319 mutex_exit(&dp->dp_lock);
1320
1321 dmu_tx_delay(tx, dirty);
1322
1323 tx->tx_wait_dirty = B_FALSE;
1324
1325 /*
1326 * Note: setting tx_waited only has effect if the caller
1327 * used TX_WAIT. Otherwise they are going to destroy
1328 * this tx and try again. The common case, zfs_write(),
1329 * uses TX_WAIT.
1330 */
1331 tx->tx_waited = B_TRUE;
1332 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1333 /*
1334 * If the pool is suspended we need to wait until it
1335 * is resumed. Note that it's possible that the pool
1336 * has become active after this thread has tried to
1337 * obtain a tx. If that's the case then tx_lasttried_txg
1338 * would not have been set.
1339 */
1340 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1341 } else if (tx->tx_needassign_txh) {
1342 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1343
1344 mutex_enter(&dn->dn_mtx);
1345 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1346 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1347 mutex_exit(&dn->dn_mtx);
1348 tx->tx_needassign_txh = NULL;
1349 } else {
1350 /*
1351 * A dnode is assigned to the quiescing txg. Wait for its
1352 * transaction to complete.
1353 */
1354 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1355 }
1356
1357 spa_tx_assign_add_nsecs(spa, gethrtime() - before);
1358 }
1359
1360 void
1361 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1362 {
1363 #ifdef DEBUG_DMU_TX
1364 if (tx->tx_dir == NULL || delta == 0)
1365 return;
1366
1367 if (delta > 0) {
1368 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1369 tx->tx_space_towrite);
1370 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1371 } else {
1372 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1373 }
1374 #endif
1375 }
1376
1377 void
1378 dmu_tx_commit(dmu_tx_t *tx)
1379 {
1380 dmu_tx_hold_t *txh;
1381
1382 ASSERT(tx->tx_txg != 0);
1383
1384 /*
1385 * Go through the transaction's hold list and remove holds on
1386 * associated dnodes, notifying waiters if no holds remain.
1387 */
1388 while ((txh = list_head(&tx->tx_holds))) {
1389 dnode_t *dn = txh->txh_dnode;
1390
1391 list_remove(&tx->tx_holds, txh);
1392 kmem_free(txh, sizeof (dmu_tx_hold_t));
1393 if (dn == NULL)
1394 continue;
1395 mutex_enter(&dn->dn_mtx);
1396 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1397
1398 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1399 dn->dn_assigned_txg = 0;
1400 cv_broadcast(&dn->dn_notxholds);
1401 }
1402 mutex_exit(&dn->dn_mtx);
1403 dnode_rele(dn, tx);
1404 }
1405
1406 if (tx->tx_tempreserve_cookie)
1407 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1408
1409 if (!list_is_empty(&tx->tx_callbacks))
1410 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1411
1412 if (tx->tx_anyobj == FALSE)
1413 txg_rele_to_sync(&tx->tx_txgh);
1414
1415 list_destroy(&tx->tx_callbacks);
1416 list_destroy(&tx->tx_holds);
1417 #ifdef DEBUG_DMU_TX
1418 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1419 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1420 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1421 refcount_destroy_many(&tx->tx_space_written,
1422 refcount_count(&tx->tx_space_written));
1423 refcount_destroy_many(&tx->tx_space_freed,
1424 refcount_count(&tx->tx_space_freed));
1425 #endif
1426 kmem_free(tx, sizeof (dmu_tx_t));
1427 }
1428
1429 void
1430 dmu_tx_abort(dmu_tx_t *tx)
1431 {
1432 dmu_tx_hold_t *txh;
1433
1434 ASSERT(tx->tx_txg == 0);
1435
1436 while ((txh = list_head(&tx->tx_holds))) {
1437 dnode_t *dn = txh->txh_dnode;
1438
1439 list_remove(&tx->tx_holds, txh);
1440 kmem_free(txh, sizeof (dmu_tx_hold_t));
1441 if (dn != NULL)
1442 dnode_rele(dn, tx);
1443 }
1444
1445 /*
1446 * Call any registered callbacks with an error code.
1447 */
1448 if (!list_is_empty(&tx->tx_callbacks))
1449 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1450
1451 list_destroy(&tx->tx_callbacks);
1452 list_destroy(&tx->tx_holds);
1453 #ifdef DEBUG_DMU_TX
1454 refcount_destroy_many(&tx->tx_space_written,
1455 refcount_count(&tx->tx_space_written));
1456 refcount_destroy_many(&tx->tx_space_freed,
1457 refcount_count(&tx->tx_space_freed));
1458 #endif
1459 kmem_free(tx, sizeof (dmu_tx_t));
1460 }
1461
1462 uint64_t
1463 dmu_tx_get_txg(dmu_tx_t *tx)
1464 {
1465 ASSERT(tx->tx_txg != 0);
1466 return (tx->tx_txg);
1467 }
1468
1469 dsl_pool_t *
1470 dmu_tx_pool(dmu_tx_t *tx)
1471 {
1472 ASSERT(tx->tx_pool != NULL);
1473 return (tx->tx_pool);
1474 }
1475
1476 void
1477 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1478 {
1479 dmu_tx_callback_t *dcb;
1480
1481 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1482
1483 dcb->dcb_func = func;
1484 dcb->dcb_data = data;
1485
1486 list_insert_tail(&tx->tx_callbacks, dcb);
1487 }
1488
1489 /*
1490 * Call all the commit callbacks on a list, with a given error code.
1491 */
1492 void
1493 dmu_tx_do_callbacks(list_t *cb_list, int error)
1494 {
1495 dmu_tx_callback_t *dcb;
1496
1497 while ((dcb = list_head(cb_list))) {
1498 list_remove(cb_list, dcb);
1499 dcb->dcb_func(dcb->dcb_data, error);
1500 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1501 }
1502 }
1503
1504 /*
1505 * Interface to hold a bunch of attributes.
1506 * used for creating new files.
1507 * attrsize is the total size of all attributes
1508 * to be added during object creation
1509 *
1510 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1511 */
1512
1513 /*
1514 * hold necessary attribute name for attribute registration.
1515 * should be a very rare case where this is needed. If it does
1516 * happen it would only happen on the first write to the file system.
1517 */
1518 static void
1519 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1520 {
1521 int i;
1522
1523 if (!sa->sa_need_attr_registration)
1524 return;
1525
1526 for (i = 0; i != sa->sa_num_attrs; i++) {
1527 if (!sa->sa_attr_table[i].sa_registered) {
1528 if (sa->sa_reg_attr_obj)
1529 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1530 B_TRUE, sa->sa_attr_table[i].sa_name);
1531 else
1532 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1533 B_TRUE, sa->sa_attr_table[i].sa_name);
1534 }
1535 }
1536 }
1537
1538
1539 void
1540 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1541 {
1542 dnode_t *dn;
1543 dmu_tx_hold_t *txh;
1544
1545 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1546 THT_SPILL, 0, 0);
1547 if (txh == NULL)
1548 return;
1549
1550 dn = txh->txh_dnode;
1551
1552 if (dn == NULL)
1553 return;
1554
1555 /* If blkptr doesn't exist then add space to towrite */
1556 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1557 txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1558 } else {
1559 blkptr_t *bp;
1560
1561 bp = &dn->dn_phys->dn_spill;
1562 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1563 bp, bp->blk_birth))
1564 txh->txh_space_tooverwrite += SPA_OLD_MAXBLOCKSIZE;
1565 else
1566 txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1567 if (!BP_IS_HOLE(bp))
1568 txh->txh_space_tounref += SPA_OLD_MAXBLOCKSIZE;
1569 }
1570 }
1571
1572 void
1573 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1574 {
1575 sa_os_t *sa = tx->tx_objset->os_sa;
1576
1577 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1578
1579 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1580 return;
1581
1582 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1583 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1584 else {
1585 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1586 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1587 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1588 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1589 }
1590
1591 dmu_tx_sa_registration_hold(sa, tx);
1592
1593 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1594 return;
1595
1596 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1597 THT_SPILL, 0, 0);
1598 }
1599
1600 /*
1601 * Hold SA attribute
1602 *
1603 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1604 *
1605 * variable_size is the total size of all variable sized attributes
1606 * passed to this function. It is not the total size of all
1607 * variable size attributes that *may* exist on this object.
1608 */
1609 void
1610 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1611 {
1612 uint64_t object;
1613 sa_os_t *sa = tx->tx_objset->os_sa;
1614
1615 ASSERT(hdl != NULL);
1616
1617 object = sa_handle_object(hdl);
1618
1619 dmu_tx_hold_bonus(tx, object);
1620
1621 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1622 return;
1623
1624 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1625 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1626 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1627 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1628 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1629 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1630 }
1631
1632 dmu_tx_sa_registration_hold(sa, tx);
1633
1634 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1635 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1636
1637 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1638 ASSERT(tx->tx_txg == 0);
1639 dmu_tx_hold_spill(tx, object);
1640 } else {
1641 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1642 dnode_t *dn;
1643
1644 DB_DNODE_ENTER(db);
1645 dn = DB_DNODE(db);
1646 if (dn->dn_have_spill) {
1647 ASSERT(tx->tx_txg == 0);
1648 dmu_tx_hold_spill(tx, object);
1649 }
1650 DB_DNODE_EXIT(db);
1651 }
1652 }
1653
1654 void
1655 dmu_tx_init(void)
1656 {
1657 dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc",
1658 KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t),
1659 KSTAT_FLAG_VIRTUAL);
1660
1661 if (dmu_tx_ksp != NULL) {
1662 dmu_tx_ksp->ks_data = &dmu_tx_stats;
1663 kstat_install(dmu_tx_ksp);
1664 }
1665 }
1666
1667 void
1668 dmu_tx_fini(void)
1669 {
1670 if (dmu_tx_ksp != NULL) {
1671 kstat_delete(dmu_tx_ksp);
1672 dmu_tx_ksp = NULL;
1673 }
1674 }
1675
1676 #if defined(_KERNEL) && defined(HAVE_SPL)
1677 EXPORT_SYMBOL(dmu_tx_create);
1678 EXPORT_SYMBOL(dmu_tx_hold_write);
1679 EXPORT_SYMBOL(dmu_tx_hold_free);
1680 EXPORT_SYMBOL(dmu_tx_hold_zap);
1681 EXPORT_SYMBOL(dmu_tx_hold_bonus);
1682 EXPORT_SYMBOL(dmu_tx_abort);
1683 EXPORT_SYMBOL(dmu_tx_assign);
1684 EXPORT_SYMBOL(dmu_tx_wait);
1685 EXPORT_SYMBOL(dmu_tx_commit);
1686 EXPORT_SYMBOL(dmu_tx_get_txg);
1687 EXPORT_SYMBOL(dmu_tx_callback_register);
1688 EXPORT_SYMBOL(dmu_tx_do_callbacks);
1689 EXPORT_SYMBOL(dmu_tx_hold_spill);
1690 EXPORT_SYMBOL(dmu_tx_hold_sa_create);
1691 EXPORT_SYMBOL(dmu_tx_hold_sa);
1692 #endif
ZFS On Linux mirror for PVE