]> 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
Fletcher4: Incremental updates and ctx calculation
[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) 2012, 2016 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,
336 FALSE, FALSE, FTAG, &db);
337 rw_exit(&dn->dn_struct_rwlock);
338
339 if (err) {
340 txh->txh_tx->tx_err = err;
341 return;
342 }
343
344 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
345 history);
346 dbuf_rele(db, FTAG);
347 if (++start > end) {
348 /*
349 * Account for new indirects appearing
350 * before this IO gets assigned into a txg.
351 */
352 bits = 64 - min_bs;
353 epbs = min_ibs - SPA_BLKPTRSHIFT;
354 for (bits -= epbs * (nlvls - 1);
355 bits >= 0; bits -= epbs)
356 txh->txh_fudge += 1ULL << max_ibs;
357 goto out;
358 }
359 off += delta;
360 if (len >= delta)
361 len -= delta;
362 delta = dn->dn_datablksz;
363 }
364 }
365
366 /*
367 * 'end' is the last thing we will access, not one past.
368 * This way we won't overflow when accessing the last byte.
369 */
370 start = P2ALIGN(off, 1ULL << max_bs);
371 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
372 txh->txh_space_towrite += end - start + 1;
373
374 start >>= min_bs;
375 end >>= min_bs;
376
377 epbs = min_ibs - SPA_BLKPTRSHIFT;
378
379 /*
380 * The object contains at most 2^(64 - min_bs) blocks,
381 * and each indirect level maps 2^epbs.
382 */
383 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
384 start >>= epbs;
385 end >>= epbs;
386 ASSERT3U(end, >=, start);
387 txh->txh_space_towrite += (end - start + 1) << max_ibs;
388 if (start != 0) {
389 /*
390 * We also need a new blkid=0 indirect block
391 * to reference any existing file data.
392 */
393 txh->txh_space_towrite += 1ULL << max_ibs;
394 }
395 }
396
397 out:
398 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
399 2 * DMU_MAX_ACCESS)
400 err = SET_ERROR(EFBIG);
401
402 if (err)
403 txh->txh_tx->tx_err = err;
404 }
405
406 static void
407 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
408 {
409 dnode_t *dn = txh->txh_dnode;
410 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
411 uint64_t space = mdn->dn_datablksz +
412 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
413
414 if (dn && dn->dn_dbuf->db_blkptr &&
415 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
416 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
417 txh->txh_space_tooverwrite += space;
418 txh->txh_space_tounref += space;
419 } else {
420 txh->txh_space_towrite += space;
421 if (dn && dn->dn_dbuf->db_blkptr)
422 txh->txh_space_tounref += space;
423 }
424 }
425
426 void
427 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
428 {
429 dmu_tx_hold_t *txh;
430
431 ASSERT(tx->tx_txg == 0);
432 ASSERT(len <= DMU_MAX_ACCESS);
433 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
434
435 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
436 object, THT_WRITE, off, len);
437 if (txh == NULL)
438 return;
439
440 dmu_tx_count_write(txh, off, len);
441 dmu_tx_count_dnode(txh);
442 }
443
444 static void
445 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
446 {
447 uint64_t blkid, nblks, lastblk;
448 uint64_t space = 0, unref = 0, skipped = 0;
449 dnode_t *dn = txh->txh_dnode;
450 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
451 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
452 int epbs;
453 uint64_t l0span = 0, nl1blks = 0;
454
455 if (dn->dn_nlevels == 0)
456 return;
457
458 /*
459 * The struct_rwlock protects us against dn_nlevels
460 * changing, in case (against all odds) we manage to dirty &
461 * sync out the changes after we check for being dirty.
462 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
463 */
464 rw_enter(&dn->dn_struct_rwlock, RW_READER);
465 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
466 if (dn->dn_maxblkid == 0) {
467 if (off == 0 && len >= dn->dn_datablksz) {
468 blkid = 0;
469 nblks = 1;
470 } else {
471 rw_exit(&dn->dn_struct_rwlock);
472 return;
473 }
474 } else {
475 blkid = off >> dn->dn_datablkshift;
476 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
477
478 if (blkid > dn->dn_maxblkid) {
479 rw_exit(&dn->dn_struct_rwlock);
480 return;
481 }
482 if (blkid + nblks > dn->dn_maxblkid)
483 nblks = dn->dn_maxblkid - blkid + 1;
484
485 }
486 l0span = nblks; /* save for later use to calc level > 1 overhead */
487 if (dn->dn_nlevels == 1) {
488 int i;
489 for (i = 0; i < nblks; i++) {
490 blkptr_t *bp = dn->dn_phys->dn_blkptr;
491 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
492 bp += blkid + i;
493 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
494 dprintf_bp(bp, "can free old%s", "");
495 space += bp_get_dsize(spa, bp);
496 }
497 unref += BP_GET_ASIZE(bp);
498 }
499 nl1blks = 1;
500 nblks = 0;
501 }
502
503 lastblk = blkid + nblks - 1;
504 while (nblks) {
505 dmu_buf_impl_t *dbuf;
506 uint64_t ibyte, new_blkid;
507 int epb = 1 << epbs;
508 int err, i, blkoff, tochk;
509 blkptr_t *bp;
510
511 ibyte = blkid << dn->dn_datablkshift;
512 err = dnode_next_offset(dn,
513 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
514 new_blkid = ibyte >> dn->dn_datablkshift;
515 if (err == ESRCH) {
516 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
517 break;
518 }
519 if (err) {
520 txh->txh_tx->tx_err = err;
521 break;
522 }
523 if (new_blkid > lastblk) {
524 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
525 break;
526 }
527
528 if (new_blkid > blkid) {
529 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
530 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
531 nblks -= new_blkid - blkid;
532 blkid = new_blkid;
533 }
534 blkoff = P2PHASE(blkid, epb);
535 tochk = MIN(epb - blkoff, nblks);
536
537 err = dbuf_hold_impl(dn, 1, blkid >> epbs,
538 FALSE, FALSE, FTAG, &dbuf);
539 if (err) {
540 txh->txh_tx->tx_err = err;
541 break;
542 }
543
544 txh->txh_memory_tohold += dbuf->db.db_size;
545
546 /*
547 * We don't check memory_tohold against DMU_MAX_ACCESS because
548 * memory_tohold is an over-estimation (especially the >L1
549 * indirect blocks), so it could fail. Callers should have
550 * already verified that they will not be holding too much
551 * memory.
552 */
553
554 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
555 if (err != 0) {
556 txh->txh_tx->tx_err = err;
557 dbuf_rele(dbuf, FTAG);
558 break;
559 }
560
561 bp = dbuf->db.db_data;
562 bp += blkoff;
563
564 for (i = 0; i < tochk; i++) {
565 if (dsl_dataset_block_freeable(ds, &bp[i],
566 bp[i].blk_birth)) {
567 dprintf_bp(&bp[i], "can free old%s", "");
568 space += bp_get_dsize(spa, &bp[i]);
569 }
570 unref += BP_GET_ASIZE(bp);
571 }
572 dbuf_rele(dbuf, FTAG);
573
574 ++nl1blks;
575 blkid += tochk;
576 nblks -= tochk;
577 }
578 rw_exit(&dn->dn_struct_rwlock);
579
580 /*
581 * Add in memory requirements of higher-level indirects.
582 * This assumes a worst-possible scenario for dn_nlevels and a
583 * worst-possible distribution of l1-blocks over the region to free.
584 */
585 {
586 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
587 int level = 2;
588 /*
589 * Here we don't use DN_MAX_LEVEL, but calculate it with the
590 * given datablkshift and indblkshift. This makes the
591 * difference between 19 and 8 on large files.
592 */
593 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
594 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
595
596 while (level++ < maxlevel) {
597 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
598 << dn->dn_indblkshift;
599 blkcnt = 1 + (blkcnt >> epbs);
600 }
601 }
602
603 /* account for new level 1 indirect blocks that might show up */
604 if (skipped > 0) {
605 txh->txh_fudge += skipped << dn->dn_indblkshift;
606 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
607 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
608 }
609 txh->txh_space_tofree += space;
610 txh->txh_space_tounref += unref;
611 }
612
613 /*
614 * This function marks the transaction as being a "net free". The end
615 * result is that refquotas will be disabled for this transaction, and
616 * this transaction will be able to use half of the pool space overhead
617 * (see dsl_pool_adjustedsize()). Therefore this function should only
618 * be called for transactions that we expect will not cause a net increase
619 * in the amount of space used (but it's OK if that is occasionally not true).
620 */
621 void
622 dmu_tx_mark_netfree(dmu_tx_t *tx)
623 {
624 dmu_tx_hold_t *txh;
625
626 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
627 DMU_NEW_OBJECT, THT_FREE, 0, 0);
628
629 /*
630 * Pretend that this operation will free 1GB of space. This
631 * should be large enough to cancel out the largest write.
632 * We don't want to use something like UINT64_MAX, because that would
633 * cause overflows when doing math with these values (e.g. in
634 * dmu_tx_try_assign()).
635 */
636 txh->txh_space_tofree = txh->txh_space_tounref = 1024 * 1024 * 1024;
637 }
638
639 void
640 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
641 {
642 dmu_tx_hold_t *txh;
643 dnode_t *dn;
644 int err;
645 zio_t *zio;
646
647 ASSERT(tx->tx_txg == 0);
648
649 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
650 object, THT_FREE, off, len);
651 if (txh == NULL)
652 return;
653 dn = txh->txh_dnode;
654 dmu_tx_count_dnode(txh);
655
656 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
657 return;
658 if (len == DMU_OBJECT_END)
659 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
660
661 dmu_tx_count_dnode(txh);
662
663 /*
664 * For i/o error checking, we read the first and last level-0
665 * blocks if they are not aligned, and all the level-1 blocks.
666 *
667 * Note: dbuf_free_range() assumes that we have not instantiated
668 * any level-0 dbufs that will be completely freed. Therefore we must
669 * exercise care to not read or count the first and last blocks
670 * if they are blocksize-aligned.
671 */
672 if (dn->dn_datablkshift == 0) {
673 if (off != 0 || len < dn->dn_datablksz)
674 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
675 } else {
676 /* first block will be modified if it is not aligned */
677 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
678 dmu_tx_count_write(txh, off, 1);
679 /* last block will be modified if it is not aligned */
680 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
681 dmu_tx_count_write(txh, off+len, 1);
682 }
683
684 /*
685 * Check level-1 blocks.
686 */
687 if (dn->dn_nlevels > 1) {
688 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
689 SPA_BLKPTRSHIFT;
690 uint64_t start = off >> shift;
691 uint64_t end = (off + len) >> shift;
692 uint64_t i;
693
694 ASSERT(dn->dn_indblkshift != 0);
695
696 /*
697 * dnode_reallocate() can result in an object with indirect
698 * blocks having an odd data block size. In this case,
699 * just check the single block.
700 */
701 if (dn->dn_datablkshift == 0)
702 start = end = 0;
703
704 zio = zio_root(tx->tx_pool->dp_spa,
705 NULL, NULL, ZIO_FLAG_CANFAIL);
706 for (i = start; i <= end; i++) {
707 uint64_t ibyte = i << shift;
708 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
709 i = ibyte >> shift;
710 if (err == ESRCH || i > end)
711 break;
712 if (err) {
713 tx->tx_err = err;
714 return;
715 }
716
717 err = dmu_tx_check_ioerr(zio, dn, 1, i);
718 if (err) {
719 tx->tx_err = err;
720 return;
721 }
722 }
723 err = zio_wait(zio);
724 if (err) {
725 tx->tx_err = err;
726 return;
727 }
728 }
729
730 dmu_tx_count_free(txh, off, len);
731 }
732
733 void
734 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
735 {
736 dmu_tx_hold_t *txh;
737 dnode_t *dn;
738 dsl_dataset_phys_t *ds_phys;
739 uint64_t nblocks;
740 int epbs, err;
741
742 ASSERT(tx->tx_txg == 0);
743
744 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
745 object, THT_ZAP, add, (uintptr_t)name);
746 if (txh == NULL)
747 return;
748 dn = txh->txh_dnode;
749
750 dmu_tx_count_dnode(txh);
751
752 if (dn == NULL) {
753 /*
754 * We will be able to fit a new object's entries into one leaf
755 * block. So there will be at most 2 blocks total,
756 * including the header block.
757 */
758 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
759 return;
760 }
761
762 ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
763
764 if (dn->dn_maxblkid == 0 && !add) {
765 blkptr_t *bp;
766
767 /*
768 * If there is only one block (i.e. this is a micro-zap)
769 * and we are not adding anything, the accounting is simple.
770 */
771 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
772 if (err) {
773 tx->tx_err = err;
774 return;
775 }
776
777 /*
778 * Use max block size here, since we don't know how much
779 * the size will change between now and the dbuf dirty call.
780 */
781 bp = &dn->dn_phys->dn_blkptr[0];
782 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
783 bp, bp->blk_birth))
784 txh->txh_space_tooverwrite += MZAP_MAX_BLKSZ;
785 else
786 txh->txh_space_towrite += MZAP_MAX_BLKSZ;
787 if (!BP_IS_HOLE(bp))
788 txh->txh_space_tounref += MZAP_MAX_BLKSZ;
789 return;
790 }
791
792 if (dn->dn_maxblkid > 0 && name) {
793 /*
794 * access the name in this fat-zap so that we'll check
795 * for i/o errors to the leaf blocks, etc.
796 */
797 err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
798 if (err == EIO) {
799 tx->tx_err = err;
800 return;
801 }
802 }
803
804 err = zap_count_write_by_dnode(dn, name, add,
805 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
806
807 /*
808 * If the modified blocks are scattered to the four winds,
809 * we'll have to modify an indirect twig for each.
810 */
811 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
812 ds_phys = dsl_dataset_phys(dn->dn_objset->os_dsl_dataset);
813 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
814 if (ds_phys->ds_prev_snap_obj)
815 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
816 else
817 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
818 }
819
820 void
821 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
822 {
823 dmu_tx_hold_t *txh;
824
825 ASSERT(tx->tx_txg == 0);
826
827 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
828 object, THT_BONUS, 0, 0);
829 if (txh)
830 dmu_tx_count_dnode(txh);
831 }
832
833 void
834 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
835 {
836 dmu_tx_hold_t *txh;
837
838 ASSERT(tx->tx_txg == 0);
839
840 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
841 DMU_NEW_OBJECT, THT_SPACE, space, 0);
842 if (txh)
843 txh->txh_space_towrite += space;
844 }
845
846 int
847 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
848 {
849 dmu_tx_hold_t *txh;
850 int holds = 0;
851
852 /*
853 * By asserting that the tx is assigned, we're counting the
854 * number of dn_tx_holds, which is the same as the number of
855 * dn_holds. Otherwise, we'd be counting dn_holds, but
856 * dn_tx_holds could be 0.
857 */
858 ASSERT(tx->tx_txg != 0);
859
860 /* if (tx->tx_anyobj == TRUE) */
861 /* return (0); */
862
863 for (txh = list_head(&tx->tx_holds); txh;
864 txh = list_next(&tx->tx_holds, txh)) {
865 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
866 holds++;
867 }
868
869 return (holds);
870 }
871
872 #ifdef DEBUG_DMU_TX
873 void
874 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
875 {
876 dmu_tx_hold_t *txh;
877 int match_object = FALSE, match_offset = FALSE;
878 dnode_t *dn;
879
880 DB_DNODE_ENTER(db);
881 dn = DB_DNODE(db);
882 ASSERT(dn != NULL);
883 ASSERT(tx->tx_txg != 0);
884 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
885 ASSERT3U(dn->dn_object, ==, db->db.db_object);
886
887 if (tx->tx_anyobj) {
888 DB_DNODE_EXIT(db);
889 return;
890 }
891
892 /* XXX No checking on the meta dnode for now */
893 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
894 DB_DNODE_EXIT(db);
895 return;
896 }
897
898 for (txh = list_head(&tx->tx_holds); txh;
899 txh = list_next(&tx->tx_holds, txh)) {
900 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
901 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
902 match_object = TRUE;
903 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
904 int datablkshift = dn->dn_datablkshift ?
905 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
906 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
907 int shift = datablkshift + epbs * db->db_level;
908 uint64_t beginblk = shift >= 64 ? 0 :
909 (txh->txh_arg1 >> shift);
910 uint64_t endblk = shift >= 64 ? 0 :
911 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
912 uint64_t blkid = db->db_blkid;
913
914 /* XXX txh_arg2 better not be zero... */
915
916 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
917 txh->txh_type, beginblk, endblk);
918
919 switch (txh->txh_type) {
920 case THT_WRITE:
921 if (blkid >= beginblk && blkid <= endblk)
922 match_offset = TRUE;
923 /*
924 * We will let this hold work for the bonus
925 * or spill buffer so that we don't need to
926 * hold it when creating a new object.
927 */
928 if (blkid == DMU_BONUS_BLKID ||
929 blkid == DMU_SPILL_BLKID)
930 match_offset = TRUE;
931 /*
932 * They might have to increase nlevels,
933 * thus dirtying the new TLIBs. Or the
934 * might have to change the block size,
935 * thus dirying the new lvl=0 blk=0.
936 */
937 if (blkid == 0)
938 match_offset = TRUE;
939 break;
940 case THT_FREE:
941 /*
942 * We will dirty all the level 1 blocks in
943 * the free range and perhaps the first and
944 * last level 0 block.
945 */
946 if (blkid >= beginblk && (blkid <= endblk ||
947 txh->txh_arg2 == DMU_OBJECT_END))
948 match_offset = TRUE;
949 break;
950 case THT_SPILL:
951 if (blkid == DMU_SPILL_BLKID)
952 match_offset = TRUE;
953 break;
954 case THT_BONUS:
955 if (blkid == DMU_BONUS_BLKID)
956 match_offset = TRUE;
957 break;
958 case THT_ZAP:
959 match_offset = TRUE;
960 break;
961 case THT_NEWOBJECT:
962 match_object = TRUE;
963 break;
964 default:
965 cmn_err(CE_PANIC, "bad txh_type %d",
966 txh->txh_type);
967 }
968 }
969 if (match_object && match_offset) {
970 DB_DNODE_EXIT(db);
971 return;
972 }
973 }
974 DB_DNODE_EXIT(db);
975 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
976 (u_longlong_t)db->db.db_object, db->db_level,
977 (u_longlong_t)db->db_blkid);
978 }
979 #endif
980
981 /*
982 * If we can't do 10 iops, something is wrong. Let us go ahead
983 * and hit zfs_dirty_data_max.
984 */
985 hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */
986 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
987
988 /*
989 * We delay transactions when we've determined that the backend storage
990 * isn't able to accommodate the rate of incoming writes.
991 *
992 * If there is already a transaction waiting, we delay relative to when
993 * that transaction finishes waiting. This way the calculated min_time
994 * is independent of the number of threads concurrently executing
995 * transactions.
996 *
997 * If we are the only waiter, wait relative to when the transaction
998 * started, rather than the current time. This credits the transaction for
999 * "time already served", e.g. reading indirect blocks.
1000 *
1001 * The minimum time for a transaction to take is calculated as:
1002 * min_time = scale * (dirty - min) / (max - dirty)
1003 * min_time is then capped at zfs_delay_max_ns.
1004 *
1005 * The delay has two degrees of freedom that can be adjusted via tunables.
1006 * The percentage of dirty data at which we start to delay is defined by
1007 * zfs_delay_min_dirty_percent. This should typically be at or above
1008 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
1009 * delay after writing at full speed has failed to keep up with the incoming
1010 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
1011 * speaking, this variable determines the amount of delay at the midpoint of
1012 * the curve.
1013 *
1014 * delay
1015 * 10ms +-------------------------------------------------------------*+
1016 * | *|
1017 * 9ms + *+
1018 * | *|
1019 * 8ms + *+
1020 * | * |
1021 * 7ms + * +
1022 * | * |
1023 * 6ms + * +
1024 * | * |
1025 * 5ms + * +
1026 * | * |
1027 * 4ms + * +
1028 * | * |
1029 * 3ms + * +
1030 * | * |
1031 * 2ms + (midpoint) * +
1032 * | | ** |
1033 * 1ms + v *** +
1034 * | zfs_delay_scale ----------> ******** |
1035 * 0 +-------------------------------------*********----------------+
1036 * 0% <- zfs_dirty_data_max -> 100%
1037 *
1038 * Note that since the delay is added to the outstanding time remaining on the
1039 * most recent transaction, the delay is effectively the inverse of IOPS.
1040 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
1041 * was chosen such that small changes in the amount of accumulated dirty data
1042 * in the first 3/4 of the curve yield relatively small differences in the
1043 * amount of delay.
1044 *
1045 * The effects can be easier to understand when the amount of delay is
1046 * represented on a log scale:
1047 *
1048 * delay
1049 * 100ms +-------------------------------------------------------------++
1050 * + +
1051 * | |
1052 * + *+
1053 * 10ms + *+
1054 * + ** +
1055 * | (midpoint) ** |
1056 * + | ** +
1057 * 1ms + v **** +
1058 * + zfs_delay_scale ----------> ***** +
1059 * | **** |
1060 * + **** +
1061 * 100us + ** +
1062 * + * +
1063 * | * |
1064 * + * +
1065 * 10us + * +
1066 * + +
1067 * | |
1068 * + +
1069 * +--------------------------------------------------------------+
1070 * 0% <- zfs_dirty_data_max -> 100%
1071 *
1072 * Note here that only as the amount of dirty data approaches its limit does
1073 * the delay start to increase rapidly. The goal of a properly tuned system
1074 * should be to keep the amount of dirty data out of that range by first
1075 * ensuring that the appropriate limits are set for the I/O scheduler to reach
1076 * optimal throughput on the backend storage, and then by changing the value
1077 * of zfs_delay_scale to increase the steepness of the curve.
1078 */
1079 static void
1080 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
1081 {
1082 dsl_pool_t *dp = tx->tx_pool;
1083 uint64_t delay_min_bytes =
1084 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
1085 hrtime_t wakeup, min_tx_time, now;
1086
1087 if (dirty <= delay_min_bytes)
1088 return;
1089
1090 /*
1091 * The caller has already waited until we are under the max.
1092 * We make them pass us the amount of dirty data so we don't
1093 * have to handle the case of it being >= the max, which could
1094 * cause a divide-by-zero if it's == the max.
1095 */
1096 ASSERT3U(dirty, <, zfs_dirty_data_max);
1097
1098 now = gethrtime();
1099 min_tx_time = zfs_delay_scale *
1100 (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
1101 min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
1102 if (now > tx->tx_start + min_tx_time)
1103 return;
1104
1105 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1106 uint64_t, min_tx_time);
1107
1108 mutex_enter(&dp->dp_lock);
1109 wakeup = MAX(tx->tx_start + min_tx_time,
1110 dp->dp_last_wakeup + min_tx_time);
1111 dp->dp_last_wakeup = wakeup;
1112 mutex_exit(&dp->dp_lock);
1113
1114 zfs_sleep_until(wakeup);
1115 }
1116
1117 static int
1118 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1119 {
1120 dmu_tx_hold_t *txh;
1121 spa_t *spa = tx->tx_pool->dp_spa;
1122 uint64_t memory, asize, fsize, usize;
1123 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1124
1125 ASSERT0(tx->tx_txg);
1126
1127 if (tx->tx_err) {
1128 DMU_TX_STAT_BUMP(dmu_tx_error);
1129 return (tx->tx_err);
1130 }
1131
1132 if (spa_suspended(spa)) {
1133 DMU_TX_STAT_BUMP(dmu_tx_suspended);
1134
1135 /*
1136 * If the user has indicated a blocking failure mode
1137 * then return ERESTART which will block in dmu_tx_wait().
1138 * Otherwise, return EIO so that an error can get
1139 * propagated back to the VOP calls.
1140 *
1141 * Note that we always honor the txg_how flag regardless
1142 * of the failuremode setting.
1143 */
1144 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1145 txg_how != TXG_WAIT)
1146 return (SET_ERROR(EIO));
1147
1148 return (SET_ERROR(ERESTART));
1149 }
1150
1151 if (!tx->tx_waited &&
1152 dsl_pool_need_dirty_delay(tx->tx_pool)) {
1153 tx->tx_wait_dirty = B_TRUE;
1154 DMU_TX_STAT_BUMP(dmu_tx_dirty_delay);
1155 return (ERESTART);
1156 }
1157
1158 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1159 tx->tx_needassign_txh = NULL;
1160
1161 /*
1162 * NB: No error returns are allowed after txg_hold_open, but
1163 * before processing the dnode holds, due to the
1164 * dmu_tx_unassign() logic.
1165 */
1166
1167 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1168 for (txh = list_head(&tx->tx_holds); txh;
1169 txh = list_next(&tx->tx_holds, txh)) {
1170 dnode_t *dn = txh->txh_dnode;
1171 if (dn != NULL) {
1172 mutex_enter(&dn->dn_mtx);
1173 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1174 mutex_exit(&dn->dn_mtx);
1175 tx->tx_needassign_txh = txh;
1176 DMU_TX_STAT_BUMP(dmu_tx_group);
1177 return (SET_ERROR(ERESTART));
1178 }
1179 if (dn->dn_assigned_txg == 0)
1180 dn->dn_assigned_txg = tx->tx_txg;
1181 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1182 (void) refcount_add(&dn->dn_tx_holds, tx);
1183 mutex_exit(&dn->dn_mtx);
1184 }
1185 towrite += txh->txh_space_towrite;
1186 tofree += txh->txh_space_tofree;
1187 tooverwrite += txh->txh_space_tooverwrite;
1188 tounref += txh->txh_space_tounref;
1189 tohold += txh->txh_memory_tohold;
1190 fudge += txh->txh_fudge;
1191 }
1192
1193 /*
1194 * If a snapshot has been taken since we made our estimates,
1195 * assume that we won't be able to free or overwrite anything.
1196 */
1197 if (tx->tx_objset &&
1198 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1199 tx->tx_lastsnap_txg) {
1200 towrite += tooverwrite;
1201 tooverwrite = tofree = 0;
1202 }
1203
1204 /* needed allocation: worst-case estimate of write space */
1205 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1206 /* freed space estimate: worst-case overwrite + free estimate */
1207 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1208 /* convert unrefd space to worst-case estimate */
1209 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1210 /* calculate memory footprint estimate */
1211 memory = towrite + tooverwrite + tohold;
1212
1213 #ifdef DEBUG_DMU_TX
1214 /*
1215 * Add in 'tohold' to account for our dirty holds on this memory
1216 * XXX - the "fudge" factor is to account for skipped blocks that
1217 * we missed because dnode_next_offset() misses in-core-only blocks.
1218 */
1219 tx->tx_space_towrite = asize +
1220 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1221 tx->tx_space_tofree = tofree;
1222 tx->tx_space_tooverwrite = tooverwrite;
1223 tx->tx_space_tounref = tounref;
1224 #endif
1225
1226 if (tx->tx_dir && asize != 0) {
1227 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1228 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1229 if (err)
1230 return (err);
1231 }
1232
1233 DMU_TX_STAT_BUMP(dmu_tx_assigned);
1234
1235 return (0);
1236 }
1237
1238 static void
1239 dmu_tx_unassign(dmu_tx_t *tx)
1240 {
1241 dmu_tx_hold_t *txh;
1242
1243 if (tx->tx_txg == 0)
1244 return;
1245
1246 txg_rele_to_quiesce(&tx->tx_txgh);
1247
1248 /*
1249 * Walk the transaction's hold list, removing the hold on the
1250 * associated dnode, and notifying waiters if the refcount drops to 0.
1251 */
1252 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1253 txh = list_next(&tx->tx_holds, txh)) {
1254 dnode_t *dn = txh->txh_dnode;
1255
1256 if (dn == NULL)
1257 continue;
1258 mutex_enter(&dn->dn_mtx);
1259 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1260
1261 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1262 dn->dn_assigned_txg = 0;
1263 cv_broadcast(&dn->dn_notxholds);
1264 }
1265 mutex_exit(&dn->dn_mtx);
1266 }
1267
1268 txg_rele_to_sync(&tx->tx_txgh);
1269
1270 tx->tx_lasttried_txg = tx->tx_txg;
1271 tx->tx_txg = 0;
1272 }
1273
1274 /*
1275 * Assign tx to a transaction group. txg_how can be one of:
1276 *
1277 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1278 * a new one. This should be used when you're not holding locks.
1279 * It will only fail if we're truly out of space (or over quota).
1280 *
1281 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1282 * blocking, returns immediately with ERESTART. This should be used
1283 * whenever you're holding locks. On an ERESTART error, the caller
1284 * should drop locks, do a dmu_tx_wait(tx), and try again.
1285 *
1286 * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1287 * has already been called on behalf of this operation (though
1288 * most likely on a different tx).
1289 */
1290 int
1291 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1292 {
1293 int err;
1294
1295 ASSERT(tx->tx_txg == 0);
1296 ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1297 txg_how == TXG_WAITED);
1298 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1299
1300 if (txg_how == TXG_WAITED)
1301 tx->tx_waited = B_TRUE;
1302
1303 /* If we might wait, we must not hold the config lock. */
1304 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1305
1306 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1307 dmu_tx_unassign(tx);
1308
1309 if (err != ERESTART || txg_how != TXG_WAIT)
1310 return (err);
1311
1312 dmu_tx_wait(tx);
1313 }
1314
1315 txg_rele_to_quiesce(&tx->tx_txgh);
1316
1317 return (0);
1318 }
1319
1320 void
1321 dmu_tx_wait(dmu_tx_t *tx)
1322 {
1323 spa_t *spa = tx->tx_pool->dp_spa;
1324 dsl_pool_t *dp = tx->tx_pool;
1325 hrtime_t before;
1326
1327 ASSERT(tx->tx_txg == 0);
1328 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1329
1330 before = gethrtime();
1331
1332 if (tx->tx_wait_dirty) {
1333 uint64_t dirty;
1334
1335 /*
1336 * dmu_tx_try_assign() has determined that we need to wait
1337 * because we've consumed much or all of the dirty buffer
1338 * space.
1339 */
1340 mutex_enter(&dp->dp_lock);
1341 if (dp->dp_dirty_total >= zfs_dirty_data_max)
1342 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
1343 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1344 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1345 dirty = dp->dp_dirty_total;
1346 mutex_exit(&dp->dp_lock);
1347
1348 dmu_tx_delay(tx, dirty);
1349
1350 tx->tx_wait_dirty = B_FALSE;
1351
1352 /*
1353 * Note: setting tx_waited only has effect if the caller
1354 * used TX_WAIT. Otherwise they are going to destroy
1355 * this tx and try again. The common case, zfs_write(),
1356 * uses TX_WAIT.
1357 */
1358 tx->tx_waited = B_TRUE;
1359 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1360 /*
1361 * If the pool is suspended we need to wait until it
1362 * is resumed. Note that it's possible that the pool
1363 * has become active after this thread has tried to
1364 * obtain a tx. If that's the case then tx_lasttried_txg
1365 * would not have been set.
1366 */
1367 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1368 } else if (tx->tx_needassign_txh) {
1369 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1370
1371 mutex_enter(&dn->dn_mtx);
1372 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1373 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1374 mutex_exit(&dn->dn_mtx);
1375 tx->tx_needassign_txh = NULL;
1376 } else {
1377 /*
1378 * A dnode is assigned to the quiescing txg. Wait for its
1379 * transaction to complete.
1380 */
1381 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1382 }
1383
1384 spa_tx_assign_add_nsecs(spa, gethrtime() - before);
1385 }
1386
1387 void
1388 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1389 {
1390 #ifdef DEBUG_DMU_TX
1391 if (tx->tx_dir == NULL || delta == 0)
1392 return;
1393
1394 if (delta > 0) {
1395 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1396 tx->tx_space_towrite);
1397 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1398 } else {
1399 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1400 }
1401 #endif
1402 }
1403
1404 void
1405 dmu_tx_commit(dmu_tx_t *tx)
1406 {
1407 dmu_tx_hold_t *txh;
1408
1409 ASSERT(tx->tx_txg != 0);
1410
1411 /*
1412 * Go through the transaction's hold list and remove holds on
1413 * associated dnodes, notifying waiters if no holds remain.
1414 */
1415 while ((txh = list_head(&tx->tx_holds))) {
1416 dnode_t *dn = txh->txh_dnode;
1417
1418 list_remove(&tx->tx_holds, txh);
1419 kmem_free(txh, sizeof (dmu_tx_hold_t));
1420 if (dn == NULL)
1421 continue;
1422 mutex_enter(&dn->dn_mtx);
1423 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1424
1425 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1426 dn->dn_assigned_txg = 0;
1427 cv_broadcast(&dn->dn_notxholds);
1428 }
1429 mutex_exit(&dn->dn_mtx);
1430 dnode_rele(dn, tx);
1431 }
1432
1433 if (tx->tx_tempreserve_cookie)
1434 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1435
1436 if (!list_is_empty(&tx->tx_callbacks))
1437 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1438
1439 if (tx->tx_anyobj == FALSE)
1440 txg_rele_to_sync(&tx->tx_txgh);
1441
1442 list_destroy(&tx->tx_callbacks);
1443 list_destroy(&tx->tx_holds);
1444 #ifdef DEBUG_DMU_TX
1445 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1446 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1447 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1448 refcount_destroy_many(&tx->tx_space_written,
1449 refcount_count(&tx->tx_space_written));
1450 refcount_destroy_many(&tx->tx_space_freed,
1451 refcount_count(&tx->tx_space_freed));
1452 #endif
1453 kmem_free(tx, sizeof (dmu_tx_t));
1454 }
1455
1456 void
1457 dmu_tx_abort(dmu_tx_t *tx)
1458 {
1459 dmu_tx_hold_t *txh;
1460
1461 ASSERT(tx->tx_txg == 0);
1462
1463 while ((txh = list_head(&tx->tx_holds))) {
1464 dnode_t *dn = txh->txh_dnode;
1465
1466 list_remove(&tx->tx_holds, txh);
1467 kmem_free(txh, sizeof (dmu_tx_hold_t));
1468 if (dn != NULL)
1469 dnode_rele(dn, tx);
1470 }
1471
1472 /*
1473 * Call any registered callbacks with an error code.
1474 */
1475 if (!list_is_empty(&tx->tx_callbacks))
1476 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1477
1478 list_destroy(&tx->tx_callbacks);
1479 list_destroy(&tx->tx_holds);
1480 #ifdef DEBUG_DMU_TX
1481 refcount_destroy_many(&tx->tx_space_written,
1482 refcount_count(&tx->tx_space_written));
1483 refcount_destroy_many(&tx->tx_space_freed,
1484 refcount_count(&tx->tx_space_freed));
1485 #endif
1486 kmem_free(tx, sizeof (dmu_tx_t));
1487 }
1488
1489 uint64_t
1490 dmu_tx_get_txg(dmu_tx_t *tx)
1491 {
1492 ASSERT(tx->tx_txg != 0);
1493 return (tx->tx_txg);
1494 }
1495
1496 dsl_pool_t *
1497 dmu_tx_pool(dmu_tx_t *tx)
1498 {
1499 ASSERT(tx->tx_pool != NULL);
1500 return (tx->tx_pool);
1501 }
1502
1503 void
1504 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1505 {
1506 dmu_tx_callback_t *dcb;
1507
1508 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1509
1510 dcb->dcb_func = func;
1511 dcb->dcb_data = data;
1512
1513 list_insert_tail(&tx->tx_callbacks, dcb);
1514 }
1515
1516 /*
1517 * Call all the commit callbacks on a list, with a given error code.
1518 */
1519 void
1520 dmu_tx_do_callbacks(list_t *cb_list, int error)
1521 {
1522 dmu_tx_callback_t *dcb;
1523
1524 while ((dcb = list_head(cb_list))) {
1525 list_remove(cb_list, dcb);
1526 dcb->dcb_func(dcb->dcb_data, error);
1527 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1528 }
1529 }
1530
1531 /*
1532 * Interface to hold a bunch of attributes.
1533 * used for creating new files.
1534 * attrsize is the total size of all attributes
1535 * to be added during object creation
1536 *
1537 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1538 */
1539
1540 /*
1541 * hold necessary attribute name for attribute registration.
1542 * should be a very rare case where this is needed. If it does
1543 * happen it would only happen on the first write to the file system.
1544 */
1545 static void
1546 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1547 {
1548 int i;
1549
1550 if (!sa->sa_need_attr_registration)
1551 return;
1552
1553 for (i = 0; i != sa->sa_num_attrs; i++) {
1554 if (!sa->sa_attr_table[i].sa_registered) {
1555 if (sa->sa_reg_attr_obj)
1556 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1557 B_TRUE, sa->sa_attr_table[i].sa_name);
1558 else
1559 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1560 B_TRUE, sa->sa_attr_table[i].sa_name);
1561 }
1562 }
1563 }
1564
1565
1566 void
1567 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1568 {
1569 dnode_t *dn;
1570 dmu_tx_hold_t *txh;
1571
1572 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1573 THT_SPILL, 0, 0);
1574 if (txh == NULL)
1575 return;
1576
1577 dn = txh->txh_dnode;
1578
1579 if (dn == NULL)
1580 return;
1581
1582 /* If blkptr doesn't exist then add space to towrite */
1583 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1584 txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1585 } else {
1586 blkptr_t *bp;
1587
1588 bp = DN_SPILL_BLKPTR(dn->dn_phys);
1589 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1590 bp, bp->blk_birth))
1591 txh->txh_space_tooverwrite += SPA_OLD_MAXBLOCKSIZE;
1592 else
1593 txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
1594 if (!BP_IS_HOLE(bp))
1595 txh->txh_space_tounref += SPA_OLD_MAXBLOCKSIZE;
1596 }
1597 }
1598
1599 void
1600 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1601 {
1602 sa_os_t *sa = tx->tx_objset->os_sa;
1603
1604 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1605
1606 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1607 return;
1608
1609 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1610 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1611 else {
1612 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1613 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1614 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1615 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1616 }
1617
1618 dmu_tx_sa_registration_hold(sa, tx);
1619
1620 if (attrsize <= DN_OLD_MAX_BONUSLEN && !sa->sa_force_spill)
1621 return;
1622
1623 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1624 THT_SPILL, 0, 0);
1625 }
1626
1627 /*
1628 * Hold SA attribute
1629 *
1630 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1631 *
1632 * variable_size is the total size of all variable sized attributes
1633 * passed to this function. It is not the total size of all
1634 * variable size attributes that *may* exist on this object.
1635 */
1636 void
1637 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1638 {
1639 uint64_t object;
1640 sa_os_t *sa = tx->tx_objset->os_sa;
1641
1642 ASSERT(hdl != NULL);
1643
1644 object = sa_handle_object(hdl);
1645
1646 dmu_tx_hold_bonus(tx, object);
1647
1648 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1649 return;
1650
1651 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1652 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1653 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1654 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1655 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1656 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1657 }
1658
1659 dmu_tx_sa_registration_hold(sa, tx);
1660
1661 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1662 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1663
1664 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1665 ASSERT(tx->tx_txg == 0);
1666 dmu_tx_hold_spill(tx, object);
1667 } else {
1668 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1669 dnode_t *dn;
1670
1671 DB_DNODE_ENTER(db);
1672 dn = DB_DNODE(db);
1673 if (dn->dn_have_spill) {
1674 ASSERT(tx->tx_txg == 0);
1675 dmu_tx_hold_spill(tx, object);
1676 }
1677 DB_DNODE_EXIT(db);
1678 }
1679 }
1680
1681 void
1682 dmu_tx_init(void)
1683 {
1684 dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc",
1685 KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t),
1686 KSTAT_FLAG_VIRTUAL);
1687
1688 if (dmu_tx_ksp != NULL) {
1689 dmu_tx_ksp->ks_data = &dmu_tx_stats;
1690 kstat_install(dmu_tx_ksp);
1691 }
1692 }
1693
1694 void
1695 dmu_tx_fini(void)
1696 {
1697 if (dmu_tx_ksp != NULL) {
1698 kstat_delete(dmu_tx_ksp);
1699 dmu_tx_ksp = NULL;
1700 }
1701 }
1702
1703 #if defined(_KERNEL) && defined(HAVE_SPL)
1704 EXPORT_SYMBOL(dmu_tx_create);
1705 EXPORT_SYMBOL(dmu_tx_hold_write);
1706 EXPORT_SYMBOL(dmu_tx_hold_free);
1707 EXPORT_SYMBOL(dmu_tx_hold_zap);
1708 EXPORT_SYMBOL(dmu_tx_hold_bonus);
1709 EXPORT_SYMBOL(dmu_tx_abort);
1710 EXPORT_SYMBOL(dmu_tx_assign);
1711 EXPORT_SYMBOL(dmu_tx_wait);
1712 EXPORT_SYMBOL(dmu_tx_commit);
1713 EXPORT_SYMBOL(dmu_tx_get_txg);
1714 EXPORT_SYMBOL(dmu_tx_callback_register);
1715 EXPORT_SYMBOL(dmu_tx_do_callbacks);
1716 EXPORT_SYMBOL(dmu_tx_hold_spill);
1717 EXPORT_SYMBOL(dmu_tx_hold_sa_create);
1718 EXPORT_SYMBOL(dmu_tx_hold_sa);
1719 #endif
ZFS On Linux mirror for PVE