1eda14cbcSMatt Macy /*
2eda14cbcSMatt Macy * CDDL HEADER START
3eda14cbcSMatt Macy *
4eda14cbcSMatt Macy * The contents of this file are subject to the terms of the
5eda14cbcSMatt Macy * Common Development and Distribution License (the "License").
6eda14cbcSMatt Macy * You may not use this file except in compliance with the License.
7eda14cbcSMatt Macy *
8eda14cbcSMatt Macy * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9271171e0SMartin Matuska * or https://opensource.org/licenses/CDDL-1.0.
10eda14cbcSMatt Macy * See the License for the specific language governing permissions
11eda14cbcSMatt Macy * and limitations under the License.
12eda14cbcSMatt Macy *
13eda14cbcSMatt Macy * When distributing Covered Code, include this CDDL HEADER in each
14eda14cbcSMatt Macy * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15eda14cbcSMatt Macy * If applicable, add the following below this CDDL HEADER, with the
16eda14cbcSMatt Macy * fields enclosed by brackets "[]" replaced with your own identifying
17eda14cbcSMatt Macy * information: Portions Copyright [yyyy] [name of copyright owner]
18eda14cbcSMatt Macy *
19eda14cbcSMatt Macy * CDDL HEADER END
20eda14cbcSMatt Macy */
21eda14cbcSMatt Macy /*
22eda14cbcSMatt Macy * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23eda14cbcSMatt Macy * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24eda14cbcSMatt Macy * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
25eda14cbcSMatt Macy */
26eda14cbcSMatt Macy
27eda14cbcSMatt Macy #include <sys/dmu.h>
28eda14cbcSMatt Macy #include <sys/dmu_impl.h>
29eda14cbcSMatt Macy #include <sys/dbuf.h>
30eda14cbcSMatt Macy #include <sys/dmu_tx.h>
31eda14cbcSMatt Macy #include <sys/dmu_objset.h>
32eda14cbcSMatt Macy #include <sys/dsl_dataset.h>
33eda14cbcSMatt Macy #include <sys/dsl_dir.h>
34eda14cbcSMatt Macy #include <sys/dsl_pool.h>
35eda14cbcSMatt Macy #include <sys/zap_impl.h>
36eda14cbcSMatt Macy #include <sys/spa.h>
37eda14cbcSMatt Macy #include <sys/sa.h>
38eda14cbcSMatt Macy #include <sys/sa_impl.h>
39eda14cbcSMatt Macy #include <sys/zfs_context.h>
40eda14cbcSMatt Macy #include <sys/trace_zfs.h>
41eda14cbcSMatt Macy
42eda14cbcSMatt Macy typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
43eda14cbcSMatt Macy uint64_t arg1, uint64_t arg2);
44eda14cbcSMatt Macy
45eda14cbcSMatt Macy dmu_tx_stats_t dmu_tx_stats = {
46eda14cbcSMatt Macy { "dmu_tx_assigned", KSTAT_DATA_UINT64 },
47eda14cbcSMatt Macy { "dmu_tx_delay", KSTAT_DATA_UINT64 },
48eda14cbcSMatt Macy { "dmu_tx_error", KSTAT_DATA_UINT64 },
49eda14cbcSMatt Macy { "dmu_tx_suspended", KSTAT_DATA_UINT64 },
50eda14cbcSMatt Macy { "dmu_tx_group", KSTAT_DATA_UINT64 },
51eda14cbcSMatt Macy { "dmu_tx_memory_reserve", KSTAT_DATA_UINT64 },
52eda14cbcSMatt Macy { "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64 },
53eda14cbcSMatt Macy { "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64 },
54eda14cbcSMatt Macy { "dmu_tx_dirty_delay", KSTAT_DATA_UINT64 },
55eda14cbcSMatt Macy { "dmu_tx_dirty_over_max", KSTAT_DATA_UINT64 },
56eda14cbcSMatt Macy { "dmu_tx_dirty_frees_delay", KSTAT_DATA_UINT64 },
57e3aa18adSMartin Matuska { "dmu_tx_wrlog_delay", KSTAT_DATA_UINT64 },
58eda14cbcSMatt Macy { "dmu_tx_quota", KSTAT_DATA_UINT64 },
59eda14cbcSMatt Macy };
60eda14cbcSMatt Macy
61eda14cbcSMatt Macy static kstat_t *dmu_tx_ksp;
62eda14cbcSMatt Macy
63eda14cbcSMatt Macy dmu_tx_t *
dmu_tx_create_dd(dsl_dir_t * dd)64eda14cbcSMatt Macy dmu_tx_create_dd(dsl_dir_t *dd)
65eda14cbcSMatt Macy {
66eda14cbcSMatt Macy dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
67eda14cbcSMatt Macy tx->tx_dir = dd;
68eda14cbcSMatt Macy if (dd != NULL)
69eda14cbcSMatt Macy tx->tx_pool = dd->dd_pool;
70eda14cbcSMatt Macy list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
71eda14cbcSMatt Macy offsetof(dmu_tx_hold_t, txh_node));
72eda14cbcSMatt Macy list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
73eda14cbcSMatt Macy offsetof(dmu_tx_callback_t, dcb_node));
74eda14cbcSMatt Macy tx->tx_start = gethrtime();
75eda14cbcSMatt Macy return (tx);
76eda14cbcSMatt Macy }
77eda14cbcSMatt Macy
78eda14cbcSMatt Macy dmu_tx_t *
dmu_tx_create(objset_t * os)79eda14cbcSMatt Macy dmu_tx_create(objset_t *os)
80eda14cbcSMatt Macy {
81eda14cbcSMatt Macy dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
82eda14cbcSMatt Macy tx->tx_objset = os;
83eda14cbcSMatt Macy return (tx);
84eda14cbcSMatt Macy }
85eda14cbcSMatt Macy
86eda14cbcSMatt Macy dmu_tx_t *
dmu_tx_create_assigned(struct dsl_pool * dp,uint64_t txg)87eda14cbcSMatt Macy dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
88eda14cbcSMatt Macy {
89eda14cbcSMatt Macy dmu_tx_t *tx = dmu_tx_create_dd(NULL);
90eda14cbcSMatt Macy
91eda14cbcSMatt Macy TXG_VERIFY(dp->dp_spa, txg);
92eda14cbcSMatt Macy tx->tx_pool = dp;
93eda14cbcSMatt Macy tx->tx_txg = txg;
94eda14cbcSMatt Macy tx->tx_anyobj = TRUE;
95eda14cbcSMatt Macy
96eda14cbcSMatt Macy return (tx);
97eda14cbcSMatt Macy }
98eda14cbcSMatt Macy
99eda14cbcSMatt Macy int
dmu_tx_is_syncing(dmu_tx_t * tx)100eda14cbcSMatt Macy dmu_tx_is_syncing(dmu_tx_t *tx)
101eda14cbcSMatt Macy {
102eda14cbcSMatt Macy return (tx->tx_anyobj);
103eda14cbcSMatt Macy }
104eda14cbcSMatt Macy
105eda14cbcSMatt Macy int
dmu_tx_private_ok(dmu_tx_t * tx)106eda14cbcSMatt Macy dmu_tx_private_ok(dmu_tx_t *tx)
107eda14cbcSMatt Macy {
108eda14cbcSMatt Macy return (tx->tx_anyobj);
109eda14cbcSMatt Macy }
110eda14cbcSMatt Macy
111eda14cbcSMatt Macy static dmu_tx_hold_t *
dmu_tx_hold_dnode_impl(dmu_tx_t * tx,dnode_t * dn,enum dmu_tx_hold_type type,uint64_t arg1,uint64_t arg2)112eda14cbcSMatt Macy dmu_tx_hold_dnode_impl(dmu_tx_t *tx, dnode_t *dn, enum dmu_tx_hold_type type,
113eda14cbcSMatt Macy uint64_t arg1, uint64_t arg2)
114eda14cbcSMatt Macy {
115eda14cbcSMatt Macy dmu_tx_hold_t *txh;
116eda14cbcSMatt Macy
117eda14cbcSMatt Macy if (dn != NULL) {
118eda14cbcSMatt Macy (void) zfs_refcount_add(&dn->dn_holds, tx);
119eda14cbcSMatt Macy if (tx->tx_txg != 0) {
120eda14cbcSMatt Macy mutex_enter(&dn->dn_mtx);
121eda14cbcSMatt Macy /*
122eda14cbcSMatt Macy * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
123eda14cbcSMatt Macy * problem, but there's no way for it to happen (for
124eda14cbcSMatt Macy * now, at least).
125eda14cbcSMatt Macy */
126eda14cbcSMatt Macy ASSERT(dn->dn_assigned_txg == 0);
127eda14cbcSMatt Macy dn->dn_assigned_txg = tx->tx_txg;
128eda14cbcSMatt Macy (void) zfs_refcount_add(&dn->dn_tx_holds, tx);
129eda14cbcSMatt Macy mutex_exit(&dn->dn_mtx);
130eda14cbcSMatt Macy }
131eda14cbcSMatt Macy }
132eda14cbcSMatt Macy
133eda14cbcSMatt Macy txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
134eda14cbcSMatt Macy txh->txh_tx = tx;
135eda14cbcSMatt Macy txh->txh_dnode = dn;
136eda14cbcSMatt Macy zfs_refcount_create(&txh->txh_space_towrite);
137eda14cbcSMatt Macy zfs_refcount_create(&txh->txh_memory_tohold);
138eda14cbcSMatt Macy txh->txh_type = type;
139eda14cbcSMatt Macy txh->txh_arg1 = arg1;
140eda14cbcSMatt Macy txh->txh_arg2 = arg2;
141eda14cbcSMatt Macy list_insert_tail(&tx->tx_holds, txh);
142eda14cbcSMatt Macy
143eda14cbcSMatt Macy return (txh);
144eda14cbcSMatt Macy }
145eda14cbcSMatt Macy
146eda14cbcSMatt Macy static dmu_tx_hold_t *
dmu_tx_hold_object_impl(dmu_tx_t * tx,objset_t * os,uint64_t object,enum dmu_tx_hold_type type,uint64_t arg1,uint64_t arg2)147eda14cbcSMatt Macy dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
148eda14cbcSMatt Macy enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
149eda14cbcSMatt Macy {
150eda14cbcSMatt Macy dnode_t *dn = NULL;
151eda14cbcSMatt Macy dmu_tx_hold_t *txh;
152eda14cbcSMatt Macy int err;
153eda14cbcSMatt Macy
154eda14cbcSMatt Macy if (object != DMU_NEW_OBJECT) {
155eda14cbcSMatt Macy err = dnode_hold(os, object, FTAG, &dn);
156eda14cbcSMatt Macy if (err != 0) {
157eda14cbcSMatt Macy tx->tx_err = err;
158eda14cbcSMatt Macy return (NULL);
159eda14cbcSMatt Macy }
160eda14cbcSMatt Macy }
161eda14cbcSMatt Macy txh = dmu_tx_hold_dnode_impl(tx, dn, type, arg1, arg2);
162eda14cbcSMatt Macy if (dn != NULL)
163eda14cbcSMatt Macy dnode_rele(dn, FTAG);
164eda14cbcSMatt Macy return (txh);
165eda14cbcSMatt Macy }
166eda14cbcSMatt Macy
167eda14cbcSMatt Macy void
dmu_tx_add_new_object(dmu_tx_t * tx,dnode_t * dn)168eda14cbcSMatt Macy dmu_tx_add_new_object(dmu_tx_t *tx, dnode_t *dn)
169eda14cbcSMatt Macy {
170eda14cbcSMatt Macy /*
171eda14cbcSMatt Macy * If we're syncing, they can manipulate any object anyhow, and
172eda14cbcSMatt Macy * the hold on the dnode_t can cause problems.
173eda14cbcSMatt Macy */
174eda14cbcSMatt Macy if (!dmu_tx_is_syncing(tx))
175eda14cbcSMatt Macy (void) dmu_tx_hold_dnode_impl(tx, dn, THT_NEWOBJECT, 0, 0);
176eda14cbcSMatt Macy }
177eda14cbcSMatt Macy
178eda14cbcSMatt Macy /*
179eda14cbcSMatt Macy * This function reads specified data from disk. The specified data will
180eda14cbcSMatt Macy * be needed to perform the transaction -- i.e, it will be read after
181eda14cbcSMatt Macy * we do dmu_tx_assign(). There are two reasons that we read the data now
182eda14cbcSMatt Macy * (before dmu_tx_assign()):
183eda14cbcSMatt Macy *
184eda14cbcSMatt Macy * 1. Reading it now has potentially better performance. The transaction
185eda14cbcSMatt Macy * has not yet been assigned, so the TXG is not held open, and also the
186eda14cbcSMatt Macy * caller typically has less locks held when calling dmu_tx_hold_*() than
187eda14cbcSMatt Macy * after the transaction has been assigned. This reduces the lock (and txg)
188eda14cbcSMatt Macy * hold times, thus reducing lock contention.
189eda14cbcSMatt Macy *
190eda14cbcSMatt Macy * 2. It is easier for callers (primarily the ZPL) to handle i/o errors
191eda14cbcSMatt Macy * that are detected before they start making changes to the DMU state
192eda14cbcSMatt Macy * (i.e. now). Once the transaction has been assigned, and some DMU
193eda14cbcSMatt Macy * state has been changed, it can be difficult to recover from an i/o
194eda14cbcSMatt Macy * error (e.g. to undo the changes already made in memory at the DMU
195eda14cbcSMatt Macy * layer). Typically code to do so does not exist in the caller -- it
196eda14cbcSMatt Macy * assumes that the data has already been cached and thus i/o errors are
197eda14cbcSMatt Macy * not possible.
198eda14cbcSMatt Macy *
199eda14cbcSMatt Macy * It has been observed that the i/o initiated here can be a performance
200eda14cbcSMatt Macy * problem, and it appears to be optional, because we don't look at the
201eda14cbcSMatt Macy * data which is read. However, removing this read would only serve to
202eda14cbcSMatt Macy * move the work elsewhere (after the dmu_tx_assign()), where it may
203eda14cbcSMatt Macy * have a greater impact on performance (in addition to the impact on
204eda14cbcSMatt Macy * fault tolerance noted above).
205eda14cbcSMatt Macy */
206eda14cbcSMatt Macy static int
dmu_tx_check_ioerr(zio_t * zio,dnode_t * dn,int level,uint64_t blkid)207eda14cbcSMatt Macy dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
208eda14cbcSMatt Macy {
209eda14cbcSMatt Macy int err;
210eda14cbcSMatt Macy dmu_buf_impl_t *db;
211eda14cbcSMatt Macy
212eda14cbcSMatt Macy rw_enter(&dn->dn_struct_rwlock, RW_READER);
21323cf27dbSMartin Matuska err = dbuf_hold_impl(dn, level, blkid, TRUE, FALSE, FTAG, &db);
214eda14cbcSMatt Macy rw_exit(&dn->dn_struct_rwlock);
21523cf27dbSMartin Matuska if (err == ENOENT)
21623cf27dbSMartin Matuska return (0);
21723cf27dbSMartin Matuska if (err != 0)
21823cf27dbSMartin Matuska return (err);
21915f0b8c3SMartin Matuska /*
22015f0b8c3SMartin Matuska * PARTIAL_FIRST allows caching for uncacheable blocks. It will
22115f0b8c3SMartin Matuska * be cleared after dmu_buf_will_dirty() call dbuf_read() again.
22215f0b8c3SMartin Matuska */
22315f0b8c3SMartin Matuska err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH |
22415f0b8c3SMartin Matuska (level == 0 ? DB_RF_PARTIAL_FIRST : 0));
225eda14cbcSMatt Macy dbuf_rele(db, FTAG);
226eda14cbcSMatt Macy return (err);
227eda14cbcSMatt Macy }
228eda14cbcSMatt Macy
229eda14cbcSMatt Macy static void
dmu_tx_count_write(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)230eda14cbcSMatt Macy dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
231eda14cbcSMatt Macy {
232eda14cbcSMatt Macy dnode_t *dn = txh->txh_dnode;
233eda14cbcSMatt Macy int err = 0;
234eda14cbcSMatt Macy
235eda14cbcSMatt Macy if (len == 0)
236eda14cbcSMatt Macy return;
237eda14cbcSMatt Macy
238eda14cbcSMatt Macy (void) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG);
239eda14cbcSMatt Macy
240eda14cbcSMatt Macy if (dn == NULL)
241eda14cbcSMatt Macy return;
242eda14cbcSMatt Macy
243eda14cbcSMatt Macy /*
244eda14cbcSMatt Macy * For i/o error checking, read the blocks that will be needed
245eda14cbcSMatt Macy * to perform the write: the first and last level-0 blocks (if
246eda14cbcSMatt Macy * they are not aligned, i.e. if they are partial-block writes),
247eda14cbcSMatt Macy * and all the level-1 blocks.
248eda14cbcSMatt Macy */
249eda14cbcSMatt Macy if (dn->dn_maxblkid == 0) {
250eda14cbcSMatt Macy if (off < dn->dn_datablksz &&
251eda14cbcSMatt Macy (off > 0 || len < dn->dn_datablksz)) {
252eda14cbcSMatt Macy err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
253eda14cbcSMatt Macy if (err != 0) {
254eda14cbcSMatt Macy txh->txh_tx->tx_err = err;
255eda14cbcSMatt Macy }
256eda14cbcSMatt Macy }
257eda14cbcSMatt Macy } else {
258eda14cbcSMatt Macy zio_t *zio = zio_root(dn->dn_objset->os_spa,
259eda14cbcSMatt Macy NULL, NULL, ZIO_FLAG_CANFAIL);
260eda14cbcSMatt Macy
261eda14cbcSMatt Macy /* first level-0 block */
262eda14cbcSMatt Macy uint64_t start = off >> dn->dn_datablkshift;
263eda14cbcSMatt Macy if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
264eda14cbcSMatt Macy err = dmu_tx_check_ioerr(zio, dn, 0, start);
265eda14cbcSMatt Macy if (err != 0) {
266eda14cbcSMatt Macy txh->txh_tx->tx_err = err;
267eda14cbcSMatt Macy }
268eda14cbcSMatt Macy }
269eda14cbcSMatt Macy
270eda14cbcSMatt Macy /* last level-0 block */
271eda14cbcSMatt Macy uint64_t end = (off + len - 1) >> dn->dn_datablkshift;
272eda14cbcSMatt Macy if (end != start && end <= dn->dn_maxblkid &&
273eda14cbcSMatt Macy P2PHASE(off + len, dn->dn_datablksz)) {
274eda14cbcSMatt Macy err = dmu_tx_check_ioerr(zio, dn, 0, end);
275eda14cbcSMatt Macy if (err != 0) {
276eda14cbcSMatt Macy txh->txh_tx->tx_err = err;
277eda14cbcSMatt Macy }
278eda14cbcSMatt Macy }
279eda14cbcSMatt Macy
280eda14cbcSMatt Macy /* level-1 blocks */
281eda14cbcSMatt Macy if (dn->dn_nlevels > 1) {
282eda14cbcSMatt Macy int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
283eda14cbcSMatt Macy for (uint64_t i = (start >> shft) + 1;
284eda14cbcSMatt Macy i < end >> shft; i++) {
285eda14cbcSMatt Macy err = dmu_tx_check_ioerr(zio, dn, 1, i);
286eda14cbcSMatt Macy if (err != 0) {
287eda14cbcSMatt Macy txh->txh_tx->tx_err = err;
288eda14cbcSMatt Macy }
289eda14cbcSMatt Macy }
290eda14cbcSMatt Macy }
291eda14cbcSMatt Macy
292eda14cbcSMatt Macy err = zio_wait(zio);
293eda14cbcSMatt Macy if (err != 0) {
294eda14cbcSMatt Macy txh->txh_tx->tx_err = err;
295eda14cbcSMatt Macy }
296eda14cbcSMatt Macy }
297eda14cbcSMatt Macy }
298eda14cbcSMatt Macy
299eda14cbcSMatt Macy static void
dmu_tx_count_append(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)300e639e0d2SMartin Matuska dmu_tx_count_append(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
301e639e0d2SMartin Matuska {
302e639e0d2SMartin Matuska dnode_t *dn = txh->txh_dnode;
303e639e0d2SMartin Matuska int err = 0;
304e639e0d2SMartin Matuska
305e639e0d2SMartin Matuska if (len == 0)
306e639e0d2SMartin Matuska return;
307e639e0d2SMartin Matuska
308e639e0d2SMartin Matuska (void) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG);
309e639e0d2SMartin Matuska
310e639e0d2SMartin Matuska if (dn == NULL)
311e639e0d2SMartin Matuska return;
312e639e0d2SMartin Matuska
313e639e0d2SMartin Matuska /*
314e639e0d2SMartin Matuska * For i/o error checking, read the blocks that will be needed
315e639e0d2SMartin Matuska * to perform the append; first level-0 block (if not aligned, i.e.
316e639e0d2SMartin Matuska * if they are partial-block writes), no additional blocks are read.
317e639e0d2SMartin Matuska */
318e639e0d2SMartin Matuska if (dn->dn_maxblkid == 0) {
319e639e0d2SMartin Matuska if (off < dn->dn_datablksz &&
320e639e0d2SMartin Matuska (off > 0 || len < dn->dn_datablksz)) {
321e639e0d2SMartin Matuska err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
322e639e0d2SMartin Matuska if (err != 0) {
323e639e0d2SMartin Matuska txh->txh_tx->tx_err = err;
324e639e0d2SMartin Matuska }
325e639e0d2SMartin Matuska }
326e639e0d2SMartin Matuska } else {
327e639e0d2SMartin Matuska zio_t *zio = zio_root(dn->dn_objset->os_spa,
328e639e0d2SMartin Matuska NULL, NULL, ZIO_FLAG_CANFAIL);
329e639e0d2SMartin Matuska
330e639e0d2SMartin Matuska /* first level-0 block */
331e639e0d2SMartin Matuska uint64_t start = off >> dn->dn_datablkshift;
332e639e0d2SMartin Matuska if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
333e639e0d2SMartin Matuska err = dmu_tx_check_ioerr(zio, dn, 0, start);
334e639e0d2SMartin Matuska if (err != 0) {
335e639e0d2SMartin Matuska txh->txh_tx->tx_err = err;
336e639e0d2SMartin Matuska }
337e639e0d2SMartin Matuska }
338e639e0d2SMartin Matuska
339e639e0d2SMartin Matuska err = zio_wait(zio);
340e639e0d2SMartin Matuska if (err != 0) {
341e639e0d2SMartin Matuska txh->txh_tx->tx_err = err;
342e639e0d2SMartin Matuska }
343e639e0d2SMartin Matuska }
344e639e0d2SMartin Matuska }
345e639e0d2SMartin Matuska
346e639e0d2SMartin Matuska static void
dmu_tx_count_dnode(dmu_tx_hold_t * txh)347eda14cbcSMatt Macy dmu_tx_count_dnode(dmu_tx_hold_t *txh)
348eda14cbcSMatt Macy {
349eda14cbcSMatt Macy (void) zfs_refcount_add_many(&txh->txh_space_towrite,
350eda14cbcSMatt Macy DNODE_MIN_SIZE, FTAG);
351eda14cbcSMatt Macy }
352eda14cbcSMatt Macy
353eda14cbcSMatt Macy void
dmu_tx_hold_write(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)354eda14cbcSMatt Macy dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
355eda14cbcSMatt Macy {
356eda14cbcSMatt Macy dmu_tx_hold_t *txh;
357eda14cbcSMatt Macy
358eda14cbcSMatt Macy ASSERT0(tx->tx_txg);
359eda14cbcSMatt Macy ASSERT3U(len, <=, DMU_MAX_ACCESS);
360eda14cbcSMatt Macy ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
361eda14cbcSMatt Macy
362eda14cbcSMatt Macy txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
363eda14cbcSMatt Macy object, THT_WRITE, off, len);
364eda14cbcSMatt Macy if (txh != NULL) {
365eda14cbcSMatt Macy dmu_tx_count_write(txh, off, len);
366eda14cbcSMatt Macy dmu_tx_count_dnode(txh);
367eda14cbcSMatt Macy }
368eda14cbcSMatt Macy }
369eda14cbcSMatt Macy
370eda14cbcSMatt Macy void
dmu_tx_hold_write_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,int len)371eda14cbcSMatt Macy dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
372eda14cbcSMatt Macy {
373eda14cbcSMatt Macy dmu_tx_hold_t *txh;
374eda14cbcSMatt Macy
375eda14cbcSMatt Macy ASSERT0(tx->tx_txg);
376eda14cbcSMatt Macy ASSERT3U(len, <=, DMU_MAX_ACCESS);
377eda14cbcSMatt Macy ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
378eda14cbcSMatt Macy
379eda14cbcSMatt Macy txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len);
380eda14cbcSMatt Macy if (txh != NULL) {
381eda14cbcSMatt Macy dmu_tx_count_write(txh, off, len);
382eda14cbcSMatt Macy dmu_tx_count_dnode(txh);
383eda14cbcSMatt Macy }
384eda14cbcSMatt Macy }
385eda14cbcSMatt Macy
386eda14cbcSMatt Macy /*
387e639e0d2SMartin Matuska * Should be used when appending to an object and the exact offset is unknown.
388e639e0d2SMartin Matuska * The write must occur at or beyond the specified offset. Only the L0 block
389e639e0d2SMartin Matuska * at provided offset will be prefetched.
390e639e0d2SMartin Matuska */
391e639e0d2SMartin Matuska void
dmu_tx_hold_append(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)392e639e0d2SMartin Matuska dmu_tx_hold_append(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
393e639e0d2SMartin Matuska {
394e639e0d2SMartin Matuska dmu_tx_hold_t *txh;
395e639e0d2SMartin Matuska
396e639e0d2SMartin Matuska ASSERT0(tx->tx_txg);
397e639e0d2SMartin Matuska ASSERT3U(len, <=, DMU_MAX_ACCESS);
398e639e0d2SMartin Matuska
399e639e0d2SMartin Matuska txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
400e639e0d2SMartin Matuska object, THT_APPEND, off, DMU_OBJECT_END);
401e639e0d2SMartin Matuska if (txh != NULL) {
402e639e0d2SMartin Matuska dmu_tx_count_append(txh, off, len);
403e639e0d2SMartin Matuska dmu_tx_count_dnode(txh);
404e639e0d2SMartin Matuska }
405e639e0d2SMartin Matuska }
406e639e0d2SMartin Matuska
407e639e0d2SMartin Matuska void
dmu_tx_hold_append_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,int len)408e639e0d2SMartin Matuska dmu_tx_hold_append_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
409e639e0d2SMartin Matuska {
410e639e0d2SMartin Matuska dmu_tx_hold_t *txh;
411e639e0d2SMartin Matuska
412e639e0d2SMartin Matuska ASSERT0(tx->tx_txg);
413e639e0d2SMartin Matuska ASSERT3U(len, <=, DMU_MAX_ACCESS);
414e639e0d2SMartin Matuska
415e639e0d2SMartin Matuska txh = dmu_tx_hold_dnode_impl(tx, dn, THT_APPEND, off, DMU_OBJECT_END);
416e639e0d2SMartin Matuska if (txh != NULL) {
417e639e0d2SMartin Matuska dmu_tx_count_append(txh, off, len);
418e639e0d2SMartin Matuska dmu_tx_count_dnode(txh);
419e639e0d2SMartin Matuska }
420e639e0d2SMartin Matuska }
421e639e0d2SMartin Matuska
422e639e0d2SMartin Matuska /*
423eda14cbcSMatt Macy * This function marks the transaction as being a "net free". The end
424eda14cbcSMatt Macy * result is that refquotas will be disabled for this transaction, and
425eda14cbcSMatt Macy * this transaction will be able to use half of the pool space overhead
426eda14cbcSMatt Macy * (see dsl_pool_adjustedsize()). Therefore this function should only
427eda14cbcSMatt Macy * be called for transactions that we expect will not cause a net increase
428eda14cbcSMatt Macy * in the amount of space used (but it's OK if that is occasionally not true).
429eda14cbcSMatt Macy */
430eda14cbcSMatt Macy void
dmu_tx_mark_netfree(dmu_tx_t * tx)431eda14cbcSMatt Macy dmu_tx_mark_netfree(dmu_tx_t *tx)
432eda14cbcSMatt Macy {
433eda14cbcSMatt Macy tx->tx_netfree = B_TRUE;
434eda14cbcSMatt Macy }
435eda14cbcSMatt Macy
436eda14cbcSMatt Macy static void
dmu_tx_count_free(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)4372a58b312SMartin Matuska dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
438eda14cbcSMatt Macy {
439eda14cbcSMatt Macy dmu_tx_t *tx = txh->txh_tx;
440eda14cbcSMatt Macy dnode_t *dn = txh->txh_dnode;
441eda14cbcSMatt Macy int err;
442eda14cbcSMatt Macy
443eda14cbcSMatt Macy ASSERT(tx->tx_txg == 0);
444eda14cbcSMatt Macy
445eda14cbcSMatt Macy if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz)
446eda14cbcSMatt Macy return;
447eda14cbcSMatt Macy if (len == DMU_OBJECT_END)
448eda14cbcSMatt Macy len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off;
449eda14cbcSMatt Macy
450eda14cbcSMatt Macy /*
451eda14cbcSMatt Macy * For i/o error checking, we read the first and last level-0
452eda14cbcSMatt Macy * blocks if they are not aligned, and all the level-1 blocks.
453eda14cbcSMatt Macy *
454eda14cbcSMatt Macy * Note: dbuf_free_range() assumes that we have not instantiated
455eda14cbcSMatt Macy * any level-0 dbufs that will be completely freed. Therefore we must
456eda14cbcSMatt Macy * exercise care to not read or count the first and last blocks
457eda14cbcSMatt Macy * if they are blocksize-aligned.
458eda14cbcSMatt Macy */
459eda14cbcSMatt Macy if (dn->dn_datablkshift == 0) {
460eda14cbcSMatt Macy if (off != 0 || len < dn->dn_datablksz)
461eda14cbcSMatt Macy dmu_tx_count_write(txh, 0, dn->dn_datablksz);
462eda14cbcSMatt Macy } else {
463eda14cbcSMatt Macy /* first block will be modified if it is not aligned */
464eda14cbcSMatt Macy if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
465eda14cbcSMatt Macy dmu_tx_count_write(txh, off, 1);
466eda14cbcSMatt Macy /* last block will be modified if it is not aligned */
467eda14cbcSMatt Macy if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
468eda14cbcSMatt Macy dmu_tx_count_write(txh, off + len, 1);
469eda14cbcSMatt Macy }
470eda14cbcSMatt Macy
471eda14cbcSMatt Macy /*
472eda14cbcSMatt Macy * Check level-1 blocks.
473eda14cbcSMatt Macy */
474eda14cbcSMatt Macy if (dn->dn_nlevels > 1) {
475eda14cbcSMatt Macy int shift = dn->dn_datablkshift + dn->dn_indblkshift -
476eda14cbcSMatt Macy SPA_BLKPTRSHIFT;
477eda14cbcSMatt Macy uint64_t start = off >> shift;
478eda14cbcSMatt Macy uint64_t end = (off + len) >> shift;
479eda14cbcSMatt Macy
480eda14cbcSMatt Macy ASSERT(dn->dn_indblkshift != 0);
481eda14cbcSMatt Macy
482eda14cbcSMatt Macy /*
483eda14cbcSMatt Macy * dnode_reallocate() can result in an object with indirect
484eda14cbcSMatt Macy * blocks having an odd data block size. In this case,
485eda14cbcSMatt Macy * just check the single block.
486eda14cbcSMatt Macy */
487eda14cbcSMatt Macy if (dn->dn_datablkshift == 0)
488eda14cbcSMatt Macy start = end = 0;
489eda14cbcSMatt Macy
490eda14cbcSMatt Macy zio_t *zio = zio_root(tx->tx_pool->dp_spa,
491eda14cbcSMatt Macy NULL, NULL, ZIO_FLAG_CANFAIL);
492eda14cbcSMatt Macy for (uint64_t i = start; i <= end; i++) {
493eda14cbcSMatt Macy uint64_t ibyte = i << shift;
494eda14cbcSMatt Macy err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
495eda14cbcSMatt Macy i = ibyte >> shift;
496eda14cbcSMatt Macy if (err == ESRCH || i > end)
497eda14cbcSMatt Macy break;
498eda14cbcSMatt Macy if (err != 0) {
499eda14cbcSMatt Macy tx->tx_err = err;
500eda14cbcSMatt Macy (void) zio_wait(zio);
501eda14cbcSMatt Macy return;
502eda14cbcSMatt Macy }
503eda14cbcSMatt Macy
504eda14cbcSMatt Macy (void) zfs_refcount_add_many(&txh->txh_memory_tohold,
505eda14cbcSMatt Macy 1 << dn->dn_indblkshift, FTAG);
506eda14cbcSMatt Macy
507eda14cbcSMatt Macy err = dmu_tx_check_ioerr(zio, dn, 1, i);
508eda14cbcSMatt Macy if (err != 0) {
509eda14cbcSMatt Macy tx->tx_err = err;
510eda14cbcSMatt Macy (void) zio_wait(zio);
511eda14cbcSMatt Macy return;
512eda14cbcSMatt Macy }
513eda14cbcSMatt Macy }
514eda14cbcSMatt Macy err = zio_wait(zio);
515eda14cbcSMatt Macy if (err != 0) {
516eda14cbcSMatt Macy tx->tx_err = err;
517eda14cbcSMatt Macy return;
518eda14cbcSMatt Macy }
519eda14cbcSMatt Macy }
520eda14cbcSMatt Macy }
521eda14cbcSMatt Macy
522eda14cbcSMatt Macy void
dmu_tx_hold_free(dmu_tx_t * tx,uint64_t object,uint64_t off,uint64_t len)523eda14cbcSMatt Macy dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
524eda14cbcSMatt Macy {
525eda14cbcSMatt Macy dmu_tx_hold_t *txh;
526eda14cbcSMatt Macy
527eda14cbcSMatt Macy txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
528eda14cbcSMatt Macy object, THT_FREE, off, len);
5292a58b312SMartin Matuska if (txh != NULL) {
5302a58b312SMartin Matuska dmu_tx_count_dnode(txh);
5312a58b312SMartin Matuska dmu_tx_count_free(txh, off, len);
5322a58b312SMartin Matuska }
533eda14cbcSMatt Macy }
534eda14cbcSMatt Macy
535eda14cbcSMatt Macy void
dmu_tx_hold_free_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,uint64_t len)536eda14cbcSMatt Macy dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len)
537eda14cbcSMatt Macy {
538eda14cbcSMatt Macy dmu_tx_hold_t *txh;
539eda14cbcSMatt Macy
540eda14cbcSMatt Macy txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len);
5412a58b312SMartin Matuska if (txh != NULL) {
5422a58b312SMartin Matuska dmu_tx_count_dnode(txh);
5432a58b312SMartin Matuska dmu_tx_count_free(txh, off, len);
5442a58b312SMartin Matuska }
5452a58b312SMartin Matuska }
5462a58b312SMartin Matuska
5472a58b312SMartin Matuska static void
dmu_tx_count_clone(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)5482a58b312SMartin Matuska dmu_tx_count_clone(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
5492a58b312SMartin Matuska {
5502a58b312SMartin Matuska
5512a58b312SMartin Matuska /*
5522a58b312SMartin Matuska * Reuse dmu_tx_count_free(), it does exactly what we need for clone.
5532a58b312SMartin Matuska */
5542a58b312SMartin Matuska dmu_tx_count_free(txh, off, len);
5552a58b312SMartin Matuska }
5562a58b312SMartin Matuska
5572a58b312SMartin Matuska void
dmu_tx_hold_clone_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,int len)5582a58b312SMartin Matuska dmu_tx_hold_clone_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
5592a58b312SMartin Matuska {
5602a58b312SMartin Matuska dmu_tx_hold_t *txh;
5612a58b312SMartin Matuska
5622a58b312SMartin Matuska ASSERT0(tx->tx_txg);
5632a58b312SMartin Matuska ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
5642a58b312SMartin Matuska
5652a58b312SMartin Matuska txh = dmu_tx_hold_dnode_impl(tx, dn, THT_CLONE, off, len);
5662a58b312SMartin Matuska if (txh != NULL) {
5672a58b312SMartin Matuska dmu_tx_count_dnode(txh);
5682a58b312SMartin Matuska dmu_tx_count_clone(txh, off, len);
5692a58b312SMartin Matuska }
570eda14cbcSMatt Macy }
571eda14cbcSMatt Macy
572eda14cbcSMatt Macy static void
dmu_tx_hold_zap_impl(dmu_tx_hold_t * txh,const char * name)573eda14cbcSMatt Macy dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name)
574eda14cbcSMatt Macy {
575eda14cbcSMatt Macy dmu_tx_t *tx = txh->txh_tx;
576eda14cbcSMatt Macy dnode_t *dn = txh->txh_dnode;
577eda14cbcSMatt Macy int err;
57815f0b8c3SMartin Matuska extern int zap_micro_max_size;
579eda14cbcSMatt Macy
580eda14cbcSMatt Macy ASSERT(tx->tx_txg == 0);
581eda14cbcSMatt Macy
582eda14cbcSMatt Macy dmu_tx_count_dnode(txh);
583eda14cbcSMatt Macy
584eda14cbcSMatt Macy /*
585eda14cbcSMatt Macy * Modifying a almost-full microzap is around the worst case (128KB)
586eda14cbcSMatt Macy *
587eda14cbcSMatt Macy * If it is a fat zap, the worst case would be 7*16KB=112KB:
588eda14cbcSMatt Macy * - 3 blocks overwritten: target leaf, ptrtbl block, header block
589eda14cbcSMatt Macy * - 4 new blocks written if adding:
590eda14cbcSMatt Macy * - 2 blocks for possibly split leaves,
591eda14cbcSMatt Macy * - 2 grown ptrtbl blocks
592eda14cbcSMatt Macy */
593eda14cbcSMatt Macy (void) zfs_refcount_add_many(&txh->txh_space_towrite,
59415f0b8c3SMartin Matuska zap_micro_max_size, FTAG);
595eda14cbcSMatt Macy
596eda14cbcSMatt Macy if (dn == NULL)
597eda14cbcSMatt Macy return;
598eda14cbcSMatt Macy
599eda14cbcSMatt Macy ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
600eda14cbcSMatt Macy
601eda14cbcSMatt Macy if (dn->dn_maxblkid == 0 || name == NULL) {
602eda14cbcSMatt Macy /*
603eda14cbcSMatt Macy * This is a microzap (only one block), or we don't know
604eda14cbcSMatt Macy * the name. Check the first block for i/o errors.
605eda14cbcSMatt Macy */
606eda14cbcSMatt Macy err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
607eda14cbcSMatt Macy if (err != 0) {
608eda14cbcSMatt Macy tx->tx_err = err;
609eda14cbcSMatt Macy }
610eda14cbcSMatt Macy } else {
611eda14cbcSMatt Macy /*
612eda14cbcSMatt Macy * Access the name so that we'll check for i/o errors to
613eda14cbcSMatt Macy * the leaf blocks, etc. We ignore ENOENT, as this name
614eda14cbcSMatt Macy * may not yet exist.
615eda14cbcSMatt Macy */
616eda14cbcSMatt Macy err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
617eda14cbcSMatt Macy if (err == EIO || err == ECKSUM || err == ENXIO) {
618eda14cbcSMatt Macy tx->tx_err = err;
619eda14cbcSMatt Macy }
620eda14cbcSMatt Macy }
621eda14cbcSMatt Macy }
622eda14cbcSMatt Macy
623eda14cbcSMatt Macy void
dmu_tx_hold_zap(dmu_tx_t * tx,uint64_t object,int add,const char * name)624eda14cbcSMatt Macy dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
625eda14cbcSMatt Macy {
626eda14cbcSMatt Macy dmu_tx_hold_t *txh;
627eda14cbcSMatt Macy
628eda14cbcSMatt Macy ASSERT0(tx->tx_txg);
629eda14cbcSMatt Macy
630eda14cbcSMatt Macy txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
631eda14cbcSMatt Macy object, THT_ZAP, add, (uintptr_t)name);
632eda14cbcSMatt Macy if (txh != NULL)
633eda14cbcSMatt Macy dmu_tx_hold_zap_impl(txh, name);
634eda14cbcSMatt Macy }
635eda14cbcSMatt Macy
636eda14cbcSMatt Macy void
dmu_tx_hold_zap_by_dnode(dmu_tx_t * tx,dnode_t * dn,int add,const char * name)637eda14cbcSMatt Macy dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name)
638eda14cbcSMatt Macy {
639eda14cbcSMatt Macy dmu_tx_hold_t *txh;
640eda14cbcSMatt Macy
641eda14cbcSMatt Macy ASSERT0(tx->tx_txg);
642eda14cbcSMatt Macy ASSERT(dn != NULL);
643eda14cbcSMatt Macy
644eda14cbcSMatt Macy txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name);
645eda14cbcSMatt Macy if (txh != NULL)
646eda14cbcSMatt Macy dmu_tx_hold_zap_impl(txh, name);
647eda14cbcSMatt Macy }
648eda14cbcSMatt Macy
649eda14cbcSMatt Macy void
dmu_tx_hold_bonus(dmu_tx_t * tx,uint64_t object)650eda14cbcSMatt Macy dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
651eda14cbcSMatt Macy {
652eda14cbcSMatt Macy dmu_tx_hold_t *txh;
653eda14cbcSMatt Macy
654eda14cbcSMatt Macy ASSERT(tx->tx_txg == 0);
655eda14cbcSMatt Macy
656eda14cbcSMatt Macy txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
657eda14cbcSMatt Macy object, THT_BONUS, 0, 0);
658eda14cbcSMatt Macy if (txh)
659eda14cbcSMatt Macy dmu_tx_count_dnode(txh);
660eda14cbcSMatt Macy }
661eda14cbcSMatt Macy
662eda14cbcSMatt Macy void
dmu_tx_hold_bonus_by_dnode(dmu_tx_t * tx,dnode_t * dn)663eda14cbcSMatt Macy dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn)
664eda14cbcSMatt Macy {
665eda14cbcSMatt Macy dmu_tx_hold_t *txh;
666eda14cbcSMatt Macy
667eda14cbcSMatt Macy ASSERT0(tx->tx_txg);
668eda14cbcSMatt Macy
669eda14cbcSMatt Macy txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0);
670eda14cbcSMatt Macy if (txh)
671eda14cbcSMatt Macy dmu_tx_count_dnode(txh);
672eda14cbcSMatt Macy }
673eda14cbcSMatt Macy
674eda14cbcSMatt Macy void
dmu_tx_hold_space(dmu_tx_t * tx,uint64_t space)675eda14cbcSMatt Macy dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
676eda14cbcSMatt Macy {
677eda14cbcSMatt Macy dmu_tx_hold_t *txh;
678eda14cbcSMatt Macy
679eda14cbcSMatt Macy ASSERT(tx->tx_txg == 0);
680eda14cbcSMatt Macy
681eda14cbcSMatt Macy txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
682eda14cbcSMatt Macy DMU_NEW_OBJECT, THT_SPACE, space, 0);
683eda14cbcSMatt Macy if (txh) {
684eda14cbcSMatt Macy (void) zfs_refcount_add_many(
685eda14cbcSMatt Macy &txh->txh_space_towrite, space, FTAG);
686eda14cbcSMatt Macy }
687eda14cbcSMatt Macy }
688eda14cbcSMatt Macy
689eda14cbcSMatt Macy #ifdef ZFS_DEBUG
690eda14cbcSMatt Macy void
dmu_tx_dirty_buf(dmu_tx_t * tx,dmu_buf_impl_t * db)691eda14cbcSMatt Macy dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
692eda14cbcSMatt Macy {
693eda14cbcSMatt Macy boolean_t match_object = B_FALSE;
694eda14cbcSMatt Macy boolean_t match_offset = B_FALSE;
695eda14cbcSMatt Macy
696eda14cbcSMatt Macy DB_DNODE_ENTER(db);
697eda14cbcSMatt Macy dnode_t *dn = DB_DNODE(db);
698eda14cbcSMatt Macy ASSERT(tx->tx_txg != 0);
699eda14cbcSMatt Macy ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
700eda14cbcSMatt Macy ASSERT3U(dn->dn_object, ==, db->db.db_object);
701eda14cbcSMatt Macy
702eda14cbcSMatt Macy if (tx->tx_anyobj) {
703eda14cbcSMatt Macy DB_DNODE_EXIT(db);
704eda14cbcSMatt Macy return;
705eda14cbcSMatt Macy }
706eda14cbcSMatt Macy
707eda14cbcSMatt Macy /* XXX No checking on the meta dnode for now */
708eda14cbcSMatt Macy if (db->db.db_object == DMU_META_DNODE_OBJECT) {
709eda14cbcSMatt Macy DB_DNODE_EXIT(db);
710eda14cbcSMatt Macy return;
711eda14cbcSMatt Macy }
712eda14cbcSMatt Macy
713eda14cbcSMatt Macy for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
714eda14cbcSMatt Macy txh = list_next(&tx->tx_holds, txh)) {
715eda14cbcSMatt Macy ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
716eda14cbcSMatt Macy if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
717eda14cbcSMatt Macy match_object = TRUE;
718eda14cbcSMatt Macy if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
719eda14cbcSMatt Macy int datablkshift = dn->dn_datablkshift ?
720eda14cbcSMatt Macy dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
721eda14cbcSMatt Macy int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
722eda14cbcSMatt Macy int shift = datablkshift + epbs * db->db_level;
723eda14cbcSMatt Macy uint64_t beginblk = shift >= 64 ? 0 :
724eda14cbcSMatt Macy (txh->txh_arg1 >> shift);
725eda14cbcSMatt Macy uint64_t endblk = shift >= 64 ? 0 :
726eda14cbcSMatt Macy ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
727eda14cbcSMatt Macy uint64_t blkid = db->db_blkid;
728eda14cbcSMatt Macy
729eda14cbcSMatt Macy /* XXX txh_arg2 better not be zero... */
730eda14cbcSMatt Macy
731eda14cbcSMatt Macy dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
73233b8c039SMartin Matuska txh->txh_type, (u_longlong_t)beginblk,
73333b8c039SMartin Matuska (u_longlong_t)endblk);
734eda14cbcSMatt Macy
735eda14cbcSMatt Macy switch (txh->txh_type) {
736eda14cbcSMatt Macy case THT_WRITE:
737eda14cbcSMatt Macy if (blkid >= beginblk && blkid <= endblk)
738eda14cbcSMatt Macy match_offset = TRUE;
739eda14cbcSMatt Macy /*
740eda14cbcSMatt Macy * We will let this hold work for the bonus
741eda14cbcSMatt Macy * or spill buffer so that we don't need to
742eda14cbcSMatt Macy * hold it when creating a new object.
743eda14cbcSMatt Macy */
744eda14cbcSMatt Macy if (blkid == DMU_BONUS_BLKID ||
745eda14cbcSMatt Macy blkid == DMU_SPILL_BLKID)
746eda14cbcSMatt Macy match_offset = TRUE;
747eda14cbcSMatt Macy /*
748eda14cbcSMatt Macy * They might have to increase nlevels,
749eda14cbcSMatt Macy * thus dirtying the new TLIBs. Or the
750eda14cbcSMatt Macy * might have to change the block size,
751eda14cbcSMatt Macy * thus dirying the new lvl=0 blk=0.
752eda14cbcSMatt Macy */
753eda14cbcSMatt Macy if (blkid == 0)
754eda14cbcSMatt Macy match_offset = TRUE;
755eda14cbcSMatt Macy break;
756e639e0d2SMartin Matuska case THT_APPEND:
757e639e0d2SMartin Matuska if (blkid >= beginblk && (blkid <= endblk ||
758e639e0d2SMartin Matuska txh->txh_arg2 == DMU_OBJECT_END))
759e639e0d2SMartin Matuska match_offset = TRUE;
760e639e0d2SMartin Matuska
761e639e0d2SMartin Matuska /*
762e639e0d2SMartin Matuska * THT_WRITE used for bonus and spill blocks.
763e639e0d2SMartin Matuska */
764e639e0d2SMartin Matuska ASSERT(blkid != DMU_BONUS_BLKID &&
765e639e0d2SMartin Matuska blkid != DMU_SPILL_BLKID);
766e639e0d2SMartin Matuska
767e639e0d2SMartin Matuska /*
768e639e0d2SMartin Matuska * They might have to increase nlevels,
769e639e0d2SMartin Matuska * thus dirtying the new TLIBs. Or the
770e639e0d2SMartin Matuska * might have to change the block size,
771e639e0d2SMartin Matuska * thus dirying the new lvl=0 blk=0.
772e639e0d2SMartin Matuska */
773e639e0d2SMartin Matuska if (blkid == 0)
774e639e0d2SMartin Matuska match_offset = TRUE;
775e639e0d2SMartin Matuska break;
776eda14cbcSMatt Macy case THT_FREE:
777eda14cbcSMatt Macy /*
778eda14cbcSMatt Macy * We will dirty all the level 1 blocks in
779eda14cbcSMatt Macy * the free range and perhaps the first and
780eda14cbcSMatt Macy * last level 0 block.
781eda14cbcSMatt Macy */
782eda14cbcSMatt Macy if (blkid >= beginblk && (blkid <= endblk ||
783eda14cbcSMatt Macy txh->txh_arg2 == DMU_OBJECT_END))
784eda14cbcSMatt Macy match_offset = TRUE;
785eda14cbcSMatt Macy break;
786eda14cbcSMatt Macy case THT_SPILL:
787eda14cbcSMatt Macy if (blkid == DMU_SPILL_BLKID)
788eda14cbcSMatt Macy match_offset = TRUE;
789eda14cbcSMatt Macy break;
790eda14cbcSMatt Macy case THT_BONUS:
791eda14cbcSMatt Macy if (blkid == DMU_BONUS_BLKID)
792eda14cbcSMatt Macy match_offset = TRUE;
793eda14cbcSMatt Macy break;
794eda14cbcSMatt Macy case THT_ZAP:
795eda14cbcSMatt Macy match_offset = TRUE;
796eda14cbcSMatt Macy break;
797eda14cbcSMatt Macy case THT_NEWOBJECT:
798eda14cbcSMatt Macy match_object = TRUE;
799eda14cbcSMatt Macy break;
8002a58b312SMartin Matuska case THT_CLONE:
8012a58b312SMartin Matuska if (blkid >= beginblk && blkid <= endblk)
8022a58b312SMartin Matuska match_offset = TRUE;
8032a58b312SMartin Matuska break;
804eda14cbcSMatt Macy default:
805eda14cbcSMatt Macy cmn_err(CE_PANIC, "bad txh_type %d",
806eda14cbcSMatt Macy txh->txh_type);
807eda14cbcSMatt Macy }
808eda14cbcSMatt Macy }
809eda14cbcSMatt Macy if (match_object && match_offset) {
810eda14cbcSMatt Macy DB_DNODE_EXIT(db);
811eda14cbcSMatt Macy return;
812eda14cbcSMatt Macy }
813eda14cbcSMatt Macy }
814eda14cbcSMatt Macy DB_DNODE_EXIT(db);
815eda14cbcSMatt Macy panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
816eda14cbcSMatt Macy (u_longlong_t)db->db.db_object, db->db_level,
817eda14cbcSMatt Macy (u_longlong_t)db->db_blkid);
818eda14cbcSMatt Macy }
819eda14cbcSMatt Macy #endif
820eda14cbcSMatt Macy
821eda14cbcSMatt Macy /*
822eda14cbcSMatt Macy * If we can't do 10 iops, something is wrong. Let us go ahead
823eda14cbcSMatt Macy * and hit zfs_dirty_data_max.
824eda14cbcSMatt Macy */
825e92ffd9bSMartin Matuska static const hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */
826eda14cbcSMatt Macy
827eda14cbcSMatt Macy /*
828eda14cbcSMatt Macy * We delay transactions when we've determined that the backend storage
829eda14cbcSMatt Macy * isn't able to accommodate the rate of incoming writes.
830eda14cbcSMatt Macy *
831eda14cbcSMatt Macy * If there is already a transaction waiting, we delay relative to when
832eda14cbcSMatt Macy * that transaction finishes waiting. This way the calculated min_time
833eda14cbcSMatt Macy * is independent of the number of threads concurrently executing
834eda14cbcSMatt Macy * transactions.
835eda14cbcSMatt Macy *
836eda14cbcSMatt Macy * If we are the only waiter, wait relative to when the transaction
837eda14cbcSMatt Macy * started, rather than the current time. This credits the transaction for
838eda14cbcSMatt Macy * "time already served", e.g. reading indirect blocks.
839eda14cbcSMatt Macy *
840eda14cbcSMatt Macy * The minimum time for a transaction to take is calculated as:
841eda14cbcSMatt Macy * min_time = scale * (dirty - min) / (max - dirty)
842eda14cbcSMatt Macy * min_time is then capped at zfs_delay_max_ns.
843eda14cbcSMatt Macy *
844eda14cbcSMatt Macy * The delay has two degrees of freedom that can be adjusted via tunables.
845eda14cbcSMatt Macy * The percentage of dirty data at which we start to delay is defined by
846eda14cbcSMatt Macy * zfs_delay_min_dirty_percent. This should typically be at or above
847eda14cbcSMatt Macy * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
848eda14cbcSMatt Macy * delay after writing at full speed has failed to keep up with the incoming
849eda14cbcSMatt Macy * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
850eda14cbcSMatt Macy * speaking, this variable determines the amount of delay at the midpoint of
851eda14cbcSMatt Macy * the curve.
852eda14cbcSMatt Macy *
853eda14cbcSMatt Macy * delay
854eda14cbcSMatt Macy * 10ms +-------------------------------------------------------------*+
855eda14cbcSMatt Macy * | *|
856eda14cbcSMatt Macy * 9ms + *+
857eda14cbcSMatt Macy * | *|
858eda14cbcSMatt Macy * 8ms + *+
859eda14cbcSMatt Macy * | * |
860eda14cbcSMatt Macy * 7ms + * +
861eda14cbcSMatt Macy * | * |
862eda14cbcSMatt Macy * 6ms + * +
863eda14cbcSMatt Macy * | * |
864eda14cbcSMatt Macy * 5ms + * +
865eda14cbcSMatt Macy * | * |
866eda14cbcSMatt Macy * 4ms + * +
867eda14cbcSMatt Macy * | * |
868eda14cbcSMatt Macy * 3ms + * +
869eda14cbcSMatt Macy * | * |
870eda14cbcSMatt Macy * 2ms + (midpoint) * +
871eda14cbcSMatt Macy * | | ** |
872eda14cbcSMatt Macy * 1ms + v *** +
873eda14cbcSMatt Macy * | zfs_delay_scale ----------> ******** |
874eda14cbcSMatt Macy * 0 +-------------------------------------*********----------------+
875eda14cbcSMatt Macy * 0% <- zfs_dirty_data_max -> 100%
876eda14cbcSMatt Macy *
877eda14cbcSMatt Macy * Note that since the delay is added to the outstanding time remaining on the
878eda14cbcSMatt Macy * most recent transaction, the delay is effectively the inverse of IOPS.
879eda14cbcSMatt Macy * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
880eda14cbcSMatt Macy * was chosen such that small changes in the amount of accumulated dirty data
881eda14cbcSMatt Macy * in the first 3/4 of the curve yield relatively small differences in the
882eda14cbcSMatt Macy * amount of delay.
883eda14cbcSMatt Macy *
884eda14cbcSMatt Macy * The effects can be easier to understand when the amount of delay is
885eda14cbcSMatt Macy * represented on a log scale:
886eda14cbcSMatt Macy *
887eda14cbcSMatt Macy * delay
888eda14cbcSMatt Macy * 100ms +-------------------------------------------------------------++
889eda14cbcSMatt Macy * + +
890eda14cbcSMatt Macy * | |
891eda14cbcSMatt Macy * + *+
892eda14cbcSMatt Macy * 10ms + *+
893eda14cbcSMatt Macy * + ** +
894eda14cbcSMatt Macy * | (midpoint) ** |
895eda14cbcSMatt Macy * + | ** +
896eda14cbcSMatt Macy * 1ms + v **** +
897eda14cbcSMatt Macy * + zfs_delay_scale ----------> ***** +
898eda14cbcSMatt Macy * | **** |
899eda14cbcSMatt Macy * + **** +
900eda14cbcSMatt Macy * 100us + ** +
901eda14cbcSMatt Macy * + * +
902eda14cbcSMatt Macy * | * |
903eda14cbcSMatt Macy * + * +
904eda14cbcSMatt Macy * 10us + * +
905eda14cbcSMatt Macy * + +
906eda14cbcSMatt Macy * | |
907eda14cbcSMatt Macy * + +
908eda14cbcSMatt Macy * +--------------------------------------------------------------+
909eda14cbcSMatt Macy * 0% <- zfs_dirty_data_max -> 100%
910eda14cbcSMatt Macy *
911eda14cbcSMatt Macy * Note here that only as the amount of dirty data approaches its limit does
912eda14cbcSMatt Macy * the delay start to increase rapidly. The goal of a properly tuned system
913eda14cbcSMatt Macy * should be to keep the amount of dirty data out of that range by first
914eda14cbcSMatt Macy * ensuring that the appropriate limits are set for the I/O scheduler to reach
915eda14cbcSMatt Macy * optimal throughput on the backend storage, and then by changing the value
916eda14cbcSMatt Macy * of zfs_delay_scale to increase the steepness of the curve.
917eda14cbcSMatt Macy */
918eda14cbcSMatt Macy static void
dmu_tx_delay(dmu_tx_t * tx,uint64_t dirty)919eda14cbcSMatt Macy dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
920eda14cbcSMatt Macy {
921eda14cbcSMatt Macy dsl_pool_t *dp = tx->tx_pool;
922e3aa18adSMartin Matuska uint64_t delay_min_bytes, wrlog;
923e3aa18adSMartin Matuska hrtime_t wakeup, tx_time = 0, now;
924e3aa18adSMartin Matuska
925e3aa18adSMartin Matuska /* Calculate minimum transaction time for the dirty data amount. */
926e3aa18adSMartin Matuska delay_min_bytes =
927eda14cbcSMatt Macy zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
928e3aa18adSMartin Matuska if (dirty > delay_min_bytes) {
929eda14cbcSMatt Macy /*
930eda14cbcSMatt Macy * The caller has already waited until we are under the max.
931eda14cbcSMatt Macy * We make them pass us the amount of dirty data so we don't
932e3aa18adSMartin Matuska * have to handle the case of it being >= the max, which
933e3aa18adSMartin Matuska * could cause a divide-by-zero if it's == the max.
934eda14cbcSMatt Macy */
935eda14cbcSMatt Macy ASSERT3U(dirty, <, zfs_dirty_data_max);
936eda14cbcSMatt Macy
937e3aa18adSMartin Matuska tx_time = zfs_delay_scale * (dirty - delay_min_bytes) /
938e3aa18adSMartin Matuska (zfs_dirty_data_max - dirty);
939e3aa18adSMartin Matuska }
940e3aa18adSMartin Matuska
941e3aa18adSMartin Matuska /* Calculate minimum transaction time for the TX_WRITE log size. */
942e3aa18adSMartin Matuska wrlog = aggsum_upper_bound(&dp->dp_wrlog_total);
943e3aa18adSMartin Matuska delay_min_bytes =
944e3aa18adSMartin Matuska zfs_wrlog_data_max * zfs_delay_min_dirty_percent / 100;
945e3aa18adSMartin Matuska if (wrlog >= zfs_wrlog_data_max) {
946e3aa18adSMartin Matuska tx_time = zfs_delay_max_ns;
947e3aa18adSMartin Matuska } else if (wrlog > delay_min_bytes) {
948e3aa18adSMartin Matuska tx_time = MAX(zfs_delay_scale * (wrlog - delay_min_bytes) /
949e3aa18adSMartin Matuska (zfs_wrlog_data_max - wrlog), tx_time);
950e3aa18adSMartin Matuska }
951e3aa18adSMartin Matuska
952e3aa18adSMartin Matuska if (tx_time == 0)
953e3aa18adSMartin Matuska return;
954e3aa18adSMartin Matuska
955e3aa18adSMartin Matuska tx_time = MIN(tx_time, zfs_delay_max_ns);
956eda14cbcSMatt Macy now = gethrtime();
957e3aa18adSMartin Matuska if (now > tx->tx_start + tx_time)
958eda14cbcSMatt Macy return;
959eda14cbcSMatt Macy
960eda14cbcSMatt Macy DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
961e3aa18adSMartin Matuska uint64_t, tx_time);
962eda14cbcSMatt Macy
963eda14cbcSMatt Macy mutex_enter(&dp->dp_lock);
964e3aa18adSMartin Matuska wakeup = MAX(tx->tx_start + tx_time, dp->dp_last_wakeup + tx_time);
965eda14cbcSMatt Macy dp->dp_last_wakeup = wakeup;
966eda14cbcSMatt Macy mutex_exit(&dp->dp_lock);
967eda14cbcSMatt Macy
968eda14cbcSMatt Macy zfs_sleep_until(wakeup);
969eda14cbcSMatt Macy }
970eda14cbcSMatt Macy
971eda14cbcSMatt Macy /*
972eda14cbcSMatt Macy * This routine attempts to assign the transaction to a transaction group.
973eda14cbcSMatt Macy * To do so, we must determine if there is sufficient free space on disk.
974eda14cbcSMatt Macy *
975eda14cbcSMatt Macy * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
976eda14cbcSMatt Macy * on it), then it is assumed that there is sufficient free space,
977eda14cbcSMatt Macy * unless there's insufficient slop space in the pool (see the comment
978eda14cbcSMatt Macy * above spa_slop_shift in spa_misc.c).
979eda14cbcSMatt Macy *
980eda14cbcSMatt Macy * If it is not a "netfree" transaction, then if the data already on disk
981eda14cbcSMatt Macy * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
982eda14cbcSMatt Macy * ENOSPC. Otherwise, if the current rough estimate of pending changes,
983eda14cbcSMatt Macy * plus the rough estimate of this transaction's changes, may exceed the
984eda14cbcSMatt Macy * allowed usage, then this will fail with ERESTART, which will cause the
985eda14cbcSMatt Macy * caller to wait for the pending changes to be written to disk (by waiting
986eda14cbcSMatt Macy * for the next TXG to open), and then check the space usage again.
987eda14cbcSMatt Macy *
988eda14cbcSMatt Macy * The rough estimate of pending changes is comprised of the sum of:
989eda14cbcSMatt Macy *
990eda14cbcSMatt Macy * - this transaction's holds' txh_space_towrite
991eda14cbcSMatt Macy *
992eda14cbcSMatt Macy * - dd_tempreserved[], which is the sum of in-flight transactions'
993eda14cbcSMatt Macy * holds' txh_space_towrite (i.e. those transactions that have called
994eda14cbcSMatt Macy * dmu_tx_assign() but not yet called dmu_tx_commit()).
995eda14cbcSMatt Macy *
996eda14cbcSMatt Macy * - dd_space_towrite[], which is the amount of dirtied dbufs.
997eda14cbcSMatt Macy *
998eda14cbcSMatt Macy * Note that all of these values are inflated by spa_get_worst_case_asize(),
999eda14cbcSMatt Macy * which means that we may get ERESTART well before we are actually in danger
1000eda14cbcSMatt Macy * of running out of space, but this also mitigates any small inaccuracies
1001eda14cbcSMatt Macy * in the rough estimate (e.g. txh_space_towrite doesn't take into account
1002eda14cbcSMatt Macy * indirect blocks, and dd_space_towrite[] doesn't take into account changes
1003eda14cbcSMatt Macy * to the MOS).
1004eda14cbcSMatt Macy *
1005eda14cbcSMatt Macy * Note that due to this algorithm, it is possible to exceed the allowed
1006eda14cbcSMatt Macy * usage by one transaction. Also, as we approach the allowed usage,
1007eda14cbcSMatt Macy * we will allow a very limited amount of changes into each TXG, thus
1008eda14cbcSMatt Macy * decreasing performance.
1009eda14cbcSMatt Macy */
1010eda14cbcSMatt Macy static int
dmu_tx_try_assign(dmu_tx_t * tx,uint64_t txg_how)1011eda14cbcSMatt Macy dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
1012eda14cbcSMatt Macy {
1013eda14cbcSMatt Macy spa_t *spa = tx->tx_pool->dp_spa;
1014eda14cbcSMatt Macy
1015eda14cbcSMatt Macy ASSERT0(tx->tx_txg);
1016eda14cbcSMatt Macy
1017eda14cbcSMatt Macy if (tx->tx_err) {
1018eda14cbcSMatt Macy DMU_TX_STAT_BUMP(dmu_tx_error);
1019eda14cbcSMatt Macy return (tx->tx_err);
1020eda14cbcSMatt Macy }
1021eda14cbcSMatt Macy
1022eda14cbcSMatt Macy if (spa_suspended(spa)) {
1023eda14cbcSMatt Macy DMU_TX_STAT_BUMP(dmu_tx_suspended);
1024eda14cbcSMatt Macy
1025eda14cbcSMatt Macy /*
1026eda14cbcSMatt Macy * If the user has indicated a blocking failure mode
1027eda14cbcSMatt Macy * then return ERESTART which will block in dmu_tx_wait().
1028eda14cbcSMatt Macy * Otherwise, return EIO so that an error can get
1029eda14cbcSMatt Macy * propagated back to the VOP calls.
1030eda14cbcSMatt Macy *
1031eda14cbcSMatt Macy * Note that we always honor the txg_how flag regardless
1032eda14cbcSMatt Macy * of the failuremode setting.
1033eda14cbcSMatt Macy */
1034eda14cbcSMatt Macy if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1035eda14cbcSMatt Macy !(txg_how & TXG_WAIT))
1036eda14cbcSMatt Macy return (SET_ERROR(EIO));
1037eda14cbcSMatt Macy
1038eda14cbcSMatt Macy return (SET_ERROR(ERESTART));
1039eda14cbcSMatt Macy }
1040eda14cbcSMatt Macy
1041eda14cbcSMatt Macy if (!tx->tx_dirty_delayed &&
1042e3aa18adSMartin Matuska dsl_pool_need_wrlog_delay(tx->tx_pool)) {
1043e3aa18adSMartin Matuska tx->tx_wait_dirty = B_TRUE;
1044e3aa18adSMartin Matuska DMU_TX_STAT_BUMP(dmu_tx_wrlog_delay);
10453f9d360cSMartin Matuska return (SET_ERROR(ERESTART));
10463f9d360cSMartin Matuska }
10473f9d360cSMartin Matuska
10483f9d360cSMartin Matuska if (!tx->tx_dirty_delayed &&
1049eda14cbcSMatt Macy dsl_pool_need_dirty_delay(tx->tx_pool)) {
1050eda14cbcSMatt Macy tx->tx_wait_dirty = B_TRUE;
1051eda14cbcSMatt Macy DMU_TX_STAT_BUMP(dmu_tx_dirty_delay);
1052eda14cbcSMatt Macy return (SET_ERROR(ERESTART));
1053eda14cbcSMatt Macy }
1054eda14cbcSMatt Macy
1055eda14cbcSMatt Macy tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1056eda14cbcSMatt Macy tx->tx_needassign_txh = NULL;
1057eda14cbcSMatt Macy
1058eda14cbcSMatt Macy /*
1059eda14cbcSMatt Macy * NB: No error returns are allowed after txg_hold_open, but
1060eda14cbcSMatt Macy * before processing the dnode holds, due to the
1061eda14cbcSMatt Macy * dmu_tx_unassign() logic.
1062eda14cbcSMatt Macy */
1063eda14cbcSMatt Macy
1064eda14cbcSMatt Macy uint64_t towrite = 0;
1065eda14cbcSMatt Macy uint64_t tohold = 0;
1066eda14cbcSMatt Macy for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1067eda14cbcSMatt Macy txh = list_next(&tx->tx_holds, txh)) {
1068eda14cbcSMatt Macy dnode_t *dn = txh->txh_dnode;
1069eda14cbcSMatt Macy if (dn != NULL) {
1070eda14cbcSMatt Macy /*
1071eda14cbcSMatt Macy * This thread can't hold the dn_struct_rwlock
1072eda14cbcSMatt Macy * while assigning the tx, because this can lead to
1073eda14cbcSMatt Macy * deadlock. Specifically, if this dnode is already
1074eda14cbcSMatt Macy * assigned to an earlier txg, this thread may need
1075eda14cbcSMatt Macy * to wait for that txg to sync (the ERESTART case
1076eda14cbcSMatt Macy * below). The other thread that has assigned this
1077eda14cbcSMatt Macy * dnode to an earlier txg prevents this txg from
1078eda14cbcSMatt Macy * syncing until its tx can complete (calling
1079eda14cbcSMatt Macy * dmu_tx_commit()), but it may need to acquire the
1080eda14cbcSMatt Macy * dn_struct_rwlock to do so (e.g. via
1081eda14cbcSMatt Macy * dmu_buf_hold*()).
1082eda14cbcSMatt Macy *
1083eda14cbcSMatt Macy * Note that this thread can't hold the lock for
1084eda14cbcSMatt Macy * read either, but the rwlock doesn't record
1085eda14cbcSMatt Macy * enough information to make that assertion.
1086eda14cbcSMatt Macy */
1087eda14cbcSMatt Macy ASSERT(!RW_WRITE_HELD(&dn->dn_struct_rwlock));
1088eda14cbcSMatt Macy
1089eda14cbcSMatt Macy mutex_enter(&dn->dn_mtx);
1090eda14cbcSMatt Macy if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1091eda14cbcSMatt Macy mutex_exit(&dn->dn_mtx);
1092eda14cbcSMatt Macy tx->tx_needassign_txh = txh;
1093eda14cbcSMatt Macy DMU_TX_STAT_BUMP(dmu_tx_group);
1094eda14cbcSMatt Macy return (SET_ERROR(ERESTART));
1095eda14cbcSMatt Macy }
1096eda14cbcSMatt Macy if (dn->dn_assigned_txg == 0)
1097eda14cbcSMatt Macy dn->dn_assigned_txg = tx->tx_txg;
1098eda14cbcSMatt Macy ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1099eda14cbcSMatt Macy (void) zfs_refcount_add(&dn->dn_tx_holds, tx);
1100eda14cbcSMatt Macy mutex_exit(&dn->dn_mtx);
1101eda14cbcSMatt Macy }
1102eda14cbcSMatt Macy towrite += zfs_refcount_count(&txh->txh_space_towrite);
1103eda14cbcSMatt Macy tohold += zfs_refcount_count(&txh->txh_memory_tohold);
1104eda14cbcSMatt Macy }
1105eda14cbcSMatt Macy
1106eda14cbcSMatt Macy /* needed allocation: worst-case estimate of write space */
1107eda14cbcSMatt Macy uint64_t asize = spa_get_worst_case_asize(tx->tx_pool->dp_spa, towrite);
1108eda14cbcSMatt Macy /* calculate memory footprint estimate */
1109eda14cbcSMatt Macy uint64_t memory = towrite + tohold;
1110eda14cbcSMatt Macy
1111eda14cbcSMatt Macy if (tx->tx_dir != NULL && asize != 0) {
1112eda14cbcSMatt Macy int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1113eda14cbcSMatt Macy asize, tx->tx_netfree, &tx->tx_tempreserve_cookie, tx);
1114eda14cbcSMatt Macy if (err != 0)
1115eda14cbcSMatt Macy return (err);
1116eda14cbcSMatt Macy }
1117eda14cbcSMatt Macy
1118eda14cbcSMatt Macy DMU_TX_STAT_BUMP(dmu_tx_assigned);
1119eda14cbcSMatt Macy
1120eda14cbcSMatt Macy return (0);
1121eda14cbcSMatt Macy }
1122eda14cbcSMatt Macy
1123eda14cbcSMatt Macy static void
dmu_tx_unassign(dmu_tx_t * tx)1124eda14cbcSMatt Macy dmu_tx_unassign(dmu_tx_t *tx)
1125eda14cbcSMatt Macy {
1126eda14cbcSMatt Macy if (tx->tx_txg == 0)
1127eda14cbcSMatt Macy return;
1128eda14cbcSMatt Macy
1129eda14cbcSMatt Macy txg_rele_to_quiesce(&tx->tx_txgh);
1130eda14cbcSMatt Macy
1131eda14cbcSMatt Macy /*
1132eda14cbcSMatt Macy * Walk the transaction's hold list, removing the hold on the
1133eda14cbcSMatt Macy * associated dnode, and notifying waiters if the refcount drops to 0.
1134eda14cbcSMatt Macy */
1135eda14cbcSMatt Macy for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds);
1136eda14cbcSMatt Macy txh && txh != tx->tx_needassign_txh;
1137eda14cbcSMatt Macy txh = list_next(&tx->tx_holds, txh)) {
1138eda14cbcSMatt Macy dnode_t *dn = txh->txh_dnode;
1139eda14cbcSMatt Macy
1140eda14cbcSMatt Macy if (dn == NULL)
1141eda14cbcSMatt Macy continue;
1142eda14cbcSMatt Macy mutex_enter(&dn->dn_mtx);
1143eda14cbcSMatt Macy ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1144eda14cbcSMatt Macy
1145eda14cbcSMatt Macy if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1146eda14cbcSMatt Macy dn->dn_assigned_txg = 0;
1147eda14cbcSMatt Macy cv_broadcast(&dn->dn_notxholds);
1148eda14cbcSMatt Macy }
1149eda14cbcSMatt Macy mutex_exit(&dn->dn_mtx);
1150eda14cbcSMatt Macy }
1151eda14cbcSMatt Macy
1152eda14cbcSMatt Macy txg_rele_to_sync(&tx->tx_txgh);
1153eda14cbcSMatt Macy
1154eda14cbcSMatt Macy tx->tx_lasttried_txg = tx->tx_txg;
1155eda14cbcSMatt Macy tx->tx_txg = 0;
1156eda14cbcSMatt Macy }
1157eda14cbcSMatt Macy
1158eda14cbcSMatt Macy /*
1159eda14cbcSMatt Macy * Assign tx to a transaction group; txg_how is a bitmask:
1160eda14cbcSMatt Macy *
1161eda14cbcSMatt Macy * If TXG_WAIT is set and the currently open txg is full, this function
1162eda14cbcSMatt Macy * will wait until there's a new txg. This should be used when no locks
1163eda14cbcSMatt Macy * are being held. With this bit set, this function will only fail if
1164eda14cbcSMatt Macy * we're truly out of space (or over quota).
1165eda14cbcSMatt Macy *
1166eda14cbcSMatt Macy * If TXG_WAIT is *not* set and we can't assign into the currently open
1167eda14cbcSMatt Macy * txg without blocking, this function will return immediately with
1168eda14cbcSMatt Macy * ERESTART. This should be used whenever locks are being held. On an
1169eda14cbcSMatt Macy * ERESTART error, the caller should drop all locks, call dmu_tx_wait(),
1170eda14cbcSMatt Macy * and try again.
1171eda14cbcSMatt Macy *
1172eda14cbcSMatt Macy * If TXG_NOTHROTTLE is set, this indicates that this tx should not be
1173eda14cbcSMatt Macy * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for
1174eda14cbcSMatt Macy * details on the throttle). This is used by the VFS operations, after
1175eda14cbcSMatt Macy * they have already called dmu_tx_wait() (though most likely on a
1176eda14cbcSMatt Macy * different tx).
1177184c1b94SMartin Matuska *
1178184c1b94SMartin Matuska * It is guaranteed that subsequent successful calls to dmu_tx_assign()
1179184c1b94SMartin Matuska * will assign the tx to monotonically increasing txgs. Of course this is
1180184c1b94SMartin Matuska * not strong monotonicity, because the same txg can be returned multiple
1181184c1b94SMartin Matuska * times in a row. This guarantee holds both for subsequent calls from
1182184c1b94SMartin Matuska * one thread and for multiple threads. For example, it is impossible to
1183184c1b94SMartin Matuska * observe the following sequence of events:
1184184c1b94SMartin Matuska *
1185184c1b94SMartin Matuska * Thread 1 Thread 2
1186184c1b94SMartin Matuska *
1187184c1b94SMartin Matuska * dmu_tx_assign(T1, ...)
1188184c1b94SMartin Matuska * 1 <- dmu_tx_get_txg(T1)
1189184c1b94SMartin Matuska * dmu_tx_assign(T2, ...)
1190184c1b94SMartin Matuska * 2 <- dmu_tx_get_txg(T2)
1191184c1b94SMartin Matuska * dmu_tx_assign(T3, ...)
1192184c1b94SMartin Matuska * 1 <- dmu_tx_get_txg(T3)
1193eda14cbcSMatt Macy */
1194eda14cbcSMatt Macy int
dmu_tx_assign(dmu_tx_t * tx,uint64_t txg_how)1195eda14cbcSMatt Macy dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1196eda14cbcSMatt Macy {
1197eda14cbcSMatt Macy int err;
1198eda14cbcSMatt Macy
1199eda14cbcSMatt Macy ASSERT(tx->tx_txg == 0);
1200eda14cbcSMatt Macy ASSERT0(txg_how & ~(TXG_WAIT | TXG_NOTHROTTLE));
1201eda14cbcSMatt Macy ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1202eda14cbcSMatt Macy
1203eda14cbcSMatt Macy /* If we might wait, we must not hold the config lock. */
1204eda14cbcSMatt Macy IMPLY((txg_how & TXG_WAIT), !dsl_pool_config_held(tx->tx_pool));
1205eda14cbcSMatt Macy
1206eda14cbcSMatt Macy if ((txg_how & TXG_NOTHROTTLE))
1207eda14cbcSMatt Macy tx->tx_dirty_delayed = B_TRUE;
1208eda14cbcSMatt Macy
1209eda14cbcSMatt Macy while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1210eda14cbcSMatt Macy dmu_tx_unassign(tx);
1211eda14cbcSMatt Macy
1212eda14cbcSMatt Macy if (err != ERESTART || !(txg_how & TXG_WAIT))
1213eda14cbcSMatt Macy return (err);
1214eda14cbcSMatt Macy
1215eda14cbcSMatt Macy dmu_tx_wait(tx);
1216eda14cbcSMatt Macy }
1217eda14cbcSMatt Macy
1218eda14cbcSMatt Macy txg_rele_to_quiesce(&tx->tx_txgh);
1219eda14cbcSMatt Macy
1220eda14cbcSMatt Macy return (0);
1221eda14cbcSMatt Macy }
1222eda14cbcSMatt Macy
1223eda14cbcSMatt Macy void
dmu_tx_wait(dmu_tx_t * tx)1224eda14cbcSMatt Macy dmu_tx_wait(dmu_tx_t *tx)
1225eda14cbcSMatt Macy {
1226eda14cbcSMatt Macy spa_t *spa = tx->tx_pool->dp_spa;
1227eda14cbcSMatt Macy dsl_pool_t *dp = tx->tx_pool;
1228eda14cbcSMatt Macy hrtime_t before;
1229eda14cbcSMatt Macy
1230eda14cbcSMatt Macy ASSERT(tx->tx_txg == 0);
1231eda14cbcSMatt Macy ASSERT(!dsl_pool_config_held(tx->tx_pool));
1232eda14cbcSMatt Macy
1233eda14cbcSMatt Macy before = gethrtime();
1234eda14cbcSMatt Macy
1235eda14cbcSMatt Macy if (tx->tx_wait_dirty) {
1236eda14cbcSMatt Macy uint64_t dirty;
1237eda14cbcSMatt Macy
1238eda14cbcSMatt Macy /*
1239eda14cbcSMatt Macy * dmu_tx_try_assign() has determined that we need to wait
1240eda14cbcSMatt Macy * because we've consumed much or all of the dirty buffer
1241eda14cbcSMatt Macy * space.
1242eda14cbcSMatt Macy */
1243eda14cbcSMatt Macy mutex_enter(&dp->dp_lock);
1244eda14cbcSMatt Macy if (dp->dp_dirty_total >= zfs_dirty_data_max)
1245eda14cbcSMatt Macy DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
1246eda14cbcSMatt Macy while (dp->dp_dirty_total >= zfs_dirty_data_max)
1247eda14cbcSMatt Macy cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1248eda14cbcSMatt Macy dirty = dp->dp_dirty_total;
1249eda14cbcSMatt Macy mutex_exit(&dp->dp_lock);
1250eda14cbcSMatt Macy
1251eda14cbcSMatt Macy dmu_tx_delay(tx, dirty);
1252eda14cbcSMatt Macy
1253eda14cbcSMatt Macy tx->tx_wait_dirty = B_FALSE;
1254eda14cbcSMatt Macy
1255eda14cbcSMatt Macy /*
1256eda14cbcSMatt Macy * Note: setting tx_dirty_delayed only has effect if the
1257eda14cbcSMatt Macy * caller used TX_WAIT. Otherwise they are going to
1258eda14cbcSMatt Macy * destroy this tx and try again. The common case,
1259eda14cbcSMatt Macy * zfs_write(), uses TX_WAIT.
1260eda14cbcSMatt Macy */
1261eda14cbcSMatt Macy tx->tx_dirty_delayed = B_TRUE;
1262eda14cbcSMatt Macy } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1263eda14cbcSMatt Macy /*
1264eda14cbcSMatt Macy * If the pool is suspended we need to wait until it
1265eda14cbcSMatt Macy * is resumed. Note that it's possible that the pool
1266eda14cbcSMatt Macy * has become active after this thread has tried to
1267eda14cbcSMatt Macy * obtain a tx. If that's the case then tx_lasttried_txg
1268eda14cbcSMatt Macy * would not have been set.
1269eda14cbcSMatt Macy */
1270eda14cbcSMatt Macy txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1271eda14cbcSMatt Macy } else if (tx->tx_needassign_txh) {
1272eda14cbcSMatt Macy dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1273eda14cbcSMatt Macy
1274eda14cbcSMatt Macy mutex_enter(&dn->dn_mtx);
1275eda14cbcSMatt Macy while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1276eda14cbcSMatt Macy cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1277eda14cbcSMatt Macy mutex_exit(&dn->dn_mtx);
1278eda14cbcSMatt Macy tx->tx_needassign_txh = NULL;
1279eda14cbcSMatt Macy } else {
1280eda14cbcSMatt Macy /*
1281eda14cbcSMatt Macy * If we have a lot of dirty data just wait until we sync
1282eda14cbcSMatt Macy * out a TXG at which point we'll hopefully have synced
1283eda14cbcSMatt Macy * a portion of the changes.
1284eda14cbcSMatt Macy */
1285eda14cbcSMatt Macy txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1286eda14cbcSMatt Macy }
1287eda14cbcSMatt Macy
1288eda14cbcSMatt Macy spa_tx_assign_add_nsecs(spa, gethrtime() - before);
1289eda14cbcSMatt Macy }
1290eda14cbcSMatt Macy
1291eda14cbcSMatt Macy static void
dmu_tx_destroy(dmu_tx_t * tx)1292eda14cbcSMatt Macy dmu_tx_destroy(dmu_tx_t *tx)
1293eda14cbcSMatt Macy {
1294eda14cbcSMatt Macy dmu_tx_hold_t *txh;
1295eda14cbcSMatt Macy
1296eda14cbcSMatt Macy while ((txh = list_head(&tx->tx_holds)) != NULL) {
1297eda14cbcSMatt Macy dnode_t *dn = txh->txh_dnode;
1298eda14cbcSMatt Macy
1299eda14cbcSMatt Macy list_remove(&tx->tx_holds, txh);
1300eda14cbcSMatt Macy zfs_refcount_destroy_many(&txh->txh_space_towrite,
1301eda14cbcSMatt Macy zfs_refcount_count(&txh->txh_space_towrite));
1302eda14cbcSMatt Macy zfs_refcount_destroy_many(&txh->txh_memory_tohold,
1303eda14cbcSMatt Macy zfs_refcount_count(&txh->txh_memory_tohold));
1304eda14cbcSMatt Macy kmem_free(txh, sizeof (dmu_tx_hold_t));
1305eda14cbcSMatt Macy if (dn != NULL)
1306eda14cbcSMatt Macy dnode_rele(dn, tx);
1307eda14cbcSMatt Macy }
1308eda14cbcSMatt Macy
1309eda14cbcSMatt Macy list_destroy(&tx->tx_callbacks);
1310eda14cbcSMatt Macy list_destroy(&tx->tx_holds);
1311eda14cbcSMatt Macy kmem_free(tx, sizeof (dmu_tx_t));
1312eda14cbcSMatt Macy }
1313eda14cbcSMatt Macy
1314eda14cbcSMatt Macy void
dmu_tx_commit(dmu_tx_t * tx)1315eda14cbcSMatt Macy dmu_tx_commit(dmu_tx_t *tx)
1316eda14cbcSMatt Macy {
1317eda14cbcSMatt Macy ASSERT(tx->tx_txg != 0);
1318eda14cbcSMatt Macy
1319eda14cbcSMatt Macy /*
1320eda14cbcSMatt Macy * Go through the transaction's hold list and remove holds on
1321eda14cbcSMatt Macy * associated dnodes, notifying waiters if no holds remain.
1322eda14cbcSMatt Macy */
1323eda14cbcSMatt Macy for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1324eda14cbcSMatt Macy txh = list_next(&tx->tx_holds, txh)) {
1325eda14cbcSMatt Macy dnode_t *dn = txh->txh_dnode;
1326eda14cbcSMatt Macy
1327eda14cbcSMatt Macy if (dn == NULL)
1328eda14cbcSMatt Macy continue;
1329eda14cbcSMatt Macy
1330eda14cbcSMatt Macy mutex_enter(&dn->dn_mtx);
1331eda14cbcSMatt Macy ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1332eda14cbcSMatt Macy
1333eda14cbcSMatt Macy if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1334eda14cbcSMatt Macy dn->dn_assigned_txg = 0;
1335eda14cbcSMatt Macy cv_broadcast(&dn->dn_notxholds);
1336eda14cbcSMatt Macy }
1337eda14cbcSMatt Macy mutex_exit(&dn->dn_mtx);
1338eda14cbcSMatt Macy }
1339eda14cbcSMatt Macy
1340eda14cbcSMatt Macy if (tx->tx_tempreserve_cookie)
1341eda14cbcSMatt Macy dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1342eda14cbcSMatt Macy
1343eda14cbcSMatt Macy if (!list_is_empty(&tx->tx_callbacks))
1344eda14cbcSMatt Macy txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1345eda14cbcSMatt Macy
1346eda14cbcSMatt Macy if (tx->tx_anyobj == FALSE)
1347eda14cbcSMatt Macy txg_rele_to_sync(&tx->tx_txgh);
1348eda14cbcSMatt Macy
1349eda14cbcSMatt Macy dmu_tx_destroy(tx);
1350eda14cbcSMatt Macy }
1351eda14cbcSMatt Macy
1352eda14cbcSMatt Macy void
dmu_tx_abort(dmu_tx_t * tx)1353eda14cbcSMatt Macy dmu_tx_abort(dmu_tx_t *tx)
1354eda14cbcSMatt Macy {
1355eda14cbcSMatt Macy ASSERT(tx->tx_txg == 0);
1356eda14cbcSMatt Macy
1357eda14cbcSMatt Macy /*
1358eda14cbcSMatt Macy * Call any registered callbacks with an error code.
1359eda14cbcSMatt Macy */
1360eda14cbcSMatt Macy if (!list_is_empty(&tx->tx_callbacks))
1361eda14cbcSMatt Macy dmu_tx_do_callbacks(&tx->tx_callbacks, SET_ERROR(ECANCELED));
1362eda14cbcSMatt Macy
1363eda14cbcSMatt Macy dmu_tx_destroy(tx);
1364eda14cbcSMatt Macy }
1365eda14cbcSMatt Macy
1366eda14cbcSMatt Macy uint64_t
dmu_tx_get_txg(dmu_tx_t * tx)1367eda14cbcSMatt Macy dmu_tx_get_txg(dmu_tx_t *tx)
1368eda14cbcSMatt Macy {
1369eda14cbcSMatt Macy ASSERT(tx->tx_txg != 0);
1370eda14cbcSMatt Macy return (tx->tx_txg);
1371eda14cbcSMatt Macy }
1372eda14cbcSMatt Macy
1373eda14cbcSMatt Macy dsl_pool_t *
dmu_tx_pool(dmu_tx_t * tx)1374eda14cbcSMatt Macy dmu_tx_pool(dmu_tx_t *tx)
1375eda14cbcSMatt Macy {
1376eda14cbcSMatt Macy ASSERT(tx->tx_pool != NULL);
1377eda14cbcSMatt Macy return (tx->tx_pool);
1378eda14cbcSMatt Macy }
1379eda14cbcSMatt Macy
1380eda14cbcSMatt Macy void
dmu_tx_callback_register(dmu_tx_t * tx,dmu_tx_callback_func_t * func,void * data)1381eda14cbcSMatt Macy dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1382eda14cbcSMatt Macy {
1383eda14cbcSMatt Macy dmu_tx_callback_t *dcb;
1384eda14cbcSMatt Macy
1385eda14cbcSMatt Macy dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1386eda14cbcSMatt Macy
1387eda14cbcSMatt Macy dcb->dcb_func = func;
1388eda14cbcSMatt Macy dcb->dcb_data = data;
1389eda14cbcSMatt Macy
1390eda14cbcSMatt Macy list_insert_tail(&tx->tx_callbacks, dcb);
1391eda14cbcSMatt Macy }
1392eda14cbcSMatt Macy
1393eda14cbcSMatt Macy /*
1394eda14cbcSMatt Macy * Call all the commit callbacks on a list, with a given error code.
1395eda14cbcSMatt Macy */
1396eda14cbcSMatt Macy void
dmu_tx_do_callbacks(list_t * cb_list,int error)1397eda14cbcSMatt Macy dmu_tx_do_callbacks(list_t *cb_list, int error)
1398eda14cbcSMatt Macy {
1399eda14cbcSMatt Macy dmu_tx_callback_t *dcb;
1400eda14cbcSMatt Macy
14014e8d558cSMartin Matuska while ((dcb = list_remove_tail(cb_list)) != NULL) {
1402eda14cbcSMatt Macy dcb->dcb_func(dcb->dcb_data, error);
1403eda14cbcSMatt Macy kmem_free(dcb, sizeof (dmu_tx_callback_t));
1404eda14cbcSMatt Macy }
1405eda14cbcSMatt Macy }
1406eda14cbcSMatt Macy
1407eda14cbcSMatt Macy /*
1408eda14cbcSMatt Macy * Interface to hold a bunch of attributes.
1409eda14cbcSMatt Macy * used for creating new files.
1410eda14cbcSMatt Macy * attrsize is the total size of all attributes
1411eda14cbcSMatt Macy * to be added during object creation
1412eda14cbcSMatt Macy *
1413eda14cbcSMatt Macy * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1414eda14cbcSMatt Macy */
1415eda14cbcSMatt Macy
1416eda14cbcSMatt Macy /*
1417eda14cbcSMatt Macy * hold necessary attribute name for attribute registration.
1418eda14cbcSMatt Macy * should be a very rare case where this is needed. If it does
1419eda14cbcSMatt Macy * happen it would only happen on the first write to the file system.
1420eda14cbcSMatt Macy */
1421eda14cbcSMatt Macy static void
dmu_tx_sa_registration_hold(sa_os_t * sa,dmu_tx_t * tx)1422eda14cbcSMatt Macy dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1423eda14cbcSMatt Macy {
1424eda14cbcSMatt Macy if (!sa->sa_need_attr_registration)
1425eda14cbcSMatt Macy return;
1426eda14cbcSMatt Macy
1427eda14cbcSMatt Macy for (int i = 0; i != sa->sa_num_attrs; i++) {
1428eda14cbcSMatt Macy if (!sa->sa_attr_table[i].sa_registered) {
1429eda14cbcSMatt Macy if (sa->sa_reg_attr_obj)
1430eda14cbcSMatt Macy dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1431eda14cbcSMatt Macy B_TRUE, sa->sa_attr_table[i].sa_name);
1432eda14cbcSMatt Macy else
1433eda14cbcSMatt Macy dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1434eda14cbcSMatt Macy B_TRUE, sa->sa_attr_table[i].sa_name);
1435eda14cbcSMatt Macy }
1436eda14cbcSMatt Macy }
1437eda14cbcSMatt Macy }
1438eda14cbcSMatt Macy
1439eda14cbcSMatt Macy void
dmu_tx_hold_spill(dmu_tx_t * tx,uint64_t object)1440eda14cbcSMatt Macy dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1441eda14cbcSMatt Macy {
1442eda14cbcSMatt Macy dmu_tx_hold_t *txh;
1443eda14cbcSMatt Macy
1444eda14cbcSMatt Macy txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1445eda14cbcSMatt Macy THT_SPILL, 0, 0);
1446eda14cbcSMatt Macy if (txh != NULL)
1447eda14cbcSMatt Macy (void) zfs_refcount_add_many(&txh->txh_space_towrite,
1448eda14cbcSMatt Macy SPA_OLD_MAXBLOCKSIZE, FTAG);
1449eda14cbcSMatt Macy }
1450eda14cbcSMatt Macy
1451eda14cbcSMatt Macy void
dmu_tx_hold_sa_create(dmu_tx_t * tx,int attrsize)1452eda14cbcSMatt Macy dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1453eda14cbcSMatt Macy {
1454eda14cbcSMatt Macy sa_os_t *sa = tx->tx_objset->os_sa;
1455eda14cbcSMatt Macy
1456eda14cbcSMatt Macy dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1457eda14cbcSMatt Macy
1458eda14cbcSMatt Macy if (tx->tx_objset->os_sa->sa_master_obj == 0)
1459eda14cbcSMatt Macy return;
1460eda14cbcSMatt Macy
1461eda14cbcSMatt Macy if (tx->tx_objset->os_sa->sa_layout_attr_obj) {
1462eda14cbcSMatt Macy dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1463eda14cbcSMatt Macy } else {
1464eda14cbcSMatt Macy dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1465eda14cbcSMatt Macy dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1466eda14cbcSMatt Macy dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1467eda14cbcSMatt Macy dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1468eda14cbcSMatt Macy }
1469eda14cbcSMatt Macy
1470eda14cbcSMatt Macy dmu_tx_sa_registration_hold(sa, tx);
1471eda14cbcSMatt Macy
1472eda14cbcSMatt Macy if (attrsize <= DN_OLD_MAX_BONUSLEN && !sa->sa_force_spill)
1473eda14cbcSMatt Macy return;
1474eda14cbcSMatt Macy
1475eda14cbcSMatt Macy (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1476eda14cbcSMatt Macy THT_SPILL, 0, 0);
1477eda14cbcSMatt Macy }
1478eda14cbcSMatt Macy
1479eda14cbcSMatt Macy /*
1480eda14cbcSMatt Macy * Hold SA attribute
1481eda14cbcSMatt Macy *
1482eda14cbcSMatt Macy * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1483eda14cbcSMatt Macy *
1484eda14cbcSMatt Macy * variable_size is the total size of all variable sized attributes
1485eda14cbcSMatt Macy * passed to this function. It is not the total size of all
1486eda14cbcSMatt Macy * variable size attributes that *may* exist on this object.
1487eda14cbcSMatt Macy */
1488eda14cbcSMatt Macy void
dmu_tx_hold_sa(dmu_tx_t * tx,sa_handle_t * hdl,boolean_t may_grow)1489eda14cbcSMatt Macy dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1490eda14cbcSMatt Macy {
1491eda14cbcSMatt Macy uint64_t object;
1492eda14cbcSMatt Macy sa_os_t *sa = tx->tx_objset->os_sa;
1493eda14cbcSMatt Macy
1494eda14cbcSMatt Macy ASSERT(hdl != NULL);
1495eda14cbcSMatt Macy
1496eda14cbcSMatt Macy object = sa_handle_object(hdl);
1497eda14cbcSMatt Macy
1498eda14cbcSMatt Macy dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1499eda14cbcSMatt Macy DB_DNODE_ENTER(db);
1500eda14cbcSMatt Macy dmu_tx_hold_bonus_by_dnode(tx, DB_DNODE(db));
1501eda14cbcSMatt Macy DB_DNODE_EXIT(db);
1502eda14cbcSMatt Macy
1503eda14cbcSMatt Macy if (tx->tx_objset->os_sa->sa_master_obj == 0)
1504eda14cbcSMatt Macy return;
1505eda14cbcSMatt Macy
1506eda14cbcSMatt Macy if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1507eda14cbcSMatt Macy tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1508eda14cbcSMatt Macy dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1509eda14cbcSMatt Macy dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1510eda14cbcSMatt Macy dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1511eda14cbcSMatt Macy dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1512eda14cbcSMatt Macy }
1513eda14cbcSMatt Macy
1514eda14cbcSMatt Macy dmu_tx_sa_registration_hold(sa, tx);
1515eda14cbcSMatt Macy
1516eda14cbcSMatt Macy if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1517eda14cbcSMatt Macy dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1518eda14cbcSMatt Macy
1519eda14cbcSMatt Macy if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1520eda14cbcSMatt Macy ASSERT(tx->tx_txg == 0);
1521eda14cbcSMatt Macy dmu_tx_hold_spill(tx, object);
1522eda14cbcSMatt Macy } else {
1523eda14cbcSMatt Macy dnode_t *dn;
1524eda14cbcSMatt Macy
1525eda14cbcSMatt Macy DB_DNODE_ENTER(db);
1526eda14cbcSMatt Macy dn = DB_DNODE(db);
1527eda14cbcSMatt Macy if (dn->dn_have_spill) {
1528eda14cbcSMatt Macy ASSERT(tx->tx_txg == 0);
1529eda14cbcSMatt Macy dmu_tx_hold_spill(tx, object);
1530eda14cbcSMatt Macy }
1531eda14cbcSMatt Macy DB_DNODE_EXIT(db);
1532eda14cbcSMatt Macy }
1533eda14cbcSMatt Macy }
1534eda14cbcSMatt Macy
1535eda14cbcSMatt Macy void
dmu_tx_init(void)1536eda14cbcSMatt Macy dmu_tx_init(void)
1537eda14cbcSMatt Macy {
1538eda14cbcSMatt Macy dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc",
1539eda14cbcSMatt Macy KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t),
1540eda14cbcSMatt Macy KSTAT_FLAG_VIRTUAL);
1541eda14cbcSMatt Macy
1542eda14cbcSMatt Macy if (dmu_tx_ksp != NULL) {
1543eda14cbcSMatt Macy dmu_tx_ksp->ks_data = &dmu_tx_stats;
1544eda14cbcSMatt Macy kstat_install(dmu_tx_ksp);
1545eda14cbcSMatt Macy }
1546eda14cbcSMatt Macy }
1547eda14cbcSMatt Macy
1548eda14cbcSMatt Macy void
dmu_tx_fini(void)1549eda14cbcSMatt Macy dmu_tx_fini(void)
1550eda14cbcSMatt Macy {
1551eda14cbcSMatt Macy if (dmu_tx_ksp != NULL) {
1552eda14cbcSMatt Macy kstat_delete(dmu_tx_ksp);
1553eda14cbcSMatt Macy dmu_tx_ksp = NULL;
1554eda14cbcSMatt Macy }
1555eda14cbcSMatt Macy }
1556eda14cbcSMatt Macy
1557eda14cbcSMatt Macy #if defined(_KERNEL)
1558eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_create);
1559eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_write);
1560eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_write_by_dnode);
1561e639e0d2SMartin Matuska EXPORT_SYMBOL(dmu_tx_hold_append);
1562e639e0d2SMartin Matuska EXPORT_SYMBOL(dmu_tx_hold_append_by_dnode);
1563eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_free);
1564eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_free_by_dnode);
1565eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_zap);
1566eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_zap_by_dnode);
1567eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_bonus);
1568eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_bonus_by_dnode);
1569eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_abort);
1570eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_assign);
1571eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_wait);
1572eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_commit);
1573eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_mark_netfree);
1574eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_get_txg);
1575eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_callback_register);
1576eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_do_callbacks);
1577eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_spill);
1578eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_sa_create);
1579eda14cbcSMatt Macy EXPORT_SYMBOL(dmu_tx_hold_sa);
1580eda14cbcSMatt Macy #endif
1581