1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3 * CDDL HEADER START
4 *
5 * The contents of this file are subject to the terms of the
6 * Common Development and Distribution License (the "License").
7 * You may not use this file except in compliance with the License.
8 *
9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10 * or https://opensource.org/licenses/CDDL-1.0.
11 * See the License for the specific language governing permissions
12 * and limitations under the License.
13 *
14 * When distributing Covered Code, include this CDDL HEADER in each
15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16 * If applicable, add the following below this CDDL HEADER, with the
17 * fields enclosed by brackets "[]" replaced with your own identifying
18 * information: Portions Copyright [yyyy] [name of copyright owner]
19 *
20 * CDDL HEADER END
21 */
22
23 /*
24 * Copyright (c) 2023, Klara Inc.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/spa.h>
29 #include <sys/ddt.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu.h>
32 #include <sys/ddt_impl.h>
33 #include <sys/dnode.h>
34 #include <sys/dbuf.h>
35 #include <sys/zap.h>
36 #include <sys/zio_checksum.h>
37
38 /*
39 * No more than this many txgs before swapping logs.
40 */
41 uint_t zfs_dedup_log_txg_max = 8;
42
43 /*
44 * Max memory for the log AVL trees. If zfs_dedup_log_mem_max is zero at module
45 * load, it will be set to zfs_dedup_log_mem_max_percent% of total memory.
46 */
47 uint64_t zfs_dedup_log_mem_max = 0;
48 uint_t zfs_dedup_log_mem_max_percent = 1;
49
50
51 static kmem_cache_t *ddt_log_entry_flat_cache;
52 static kmem_cache_t *ddt_log_entry_trad_cache;
53
54 #define DDT_LOG_ENTRY_FLAT_SIZE \
55 (sizeof (ddt_log_entry_t) + DDT_FLAT_PHYS_SIZE)
56 #define DDT_LOG_ENTRY_TRAD_SIZE \
57 (sizeof (ddt_log_entry_t) + DDT_TRAD_PHYS_SIZE)
58
59 #define DDT_LOG_ENTRY_SIZE(ddt) \
60 _DDT_PHYS_SWITCH(ddt, DDT_LOG_ENTRY_FLAT_SIZE, DDT_LOG_ENTRY_TRAD_SIZE)
61
62 void
ddt_log_init(void)63 ddt_log_init(void)
64 {
65 ddt_log_entry_flat_cache = kmem_cache_create("ddt_log_entry_flat_cache",
66 DDT_LOG_ENTRY_FLAT_SIZE, 0, NULL, NULL, NULL, NULL, NULL, 0);
67 ddt_log_entry_trad_cache = kmem_cache_create("ddt_log_entry_trad_cache",
68 DDT_LOG_ENTRY_TRAD_SIZE, 0, NULL, NULL, NULL, NULL, NULL, 0);
69
70 /*
71 * Max memory for log AVL entries. At least 1M, because we need
72 * something (that's ~3800 entries per tree). They can say 100% if they
73 * want; it just means they're at the mercy of the the txg flush limit.
74 */
75 if (zfs_dedup_log_mem_max == 0) {
76 zfs_dedup_log_mem_max_percent =
77 MIN(zfs_dedup_log_mem_max_percent, 100);
78 zfs_dedup_log_mem_max = (physmem * PAGESIZE) *
79 zfs_dedup_log_mem_max_percent / 100;
80 }
81 zfs_dedup_log_mem_max = MAX(zfs_dedup_log_mem_max, 1*1024*1024);
82 }
83
84 void
ddt_log_fini(void)85 ddt_log_fini(void)
86 {
87 kmem_cache_destroy(ddt_log_entry_trad_cache);
88 kmem_cache_destroy(ddt_log_entry_flat_cache);
89 }
90
91 static void
ddt_log_name(ddt_t * ddt,char * name,uint_t n)92 ddt_log_name(ddt_t *ddt, char *name, uint_t n)
93 {
94 snprintf(name, DDT_NAMELEN, DMU_POOL_DDT_LOG,
95 zio_checksum_table[ddt->ddt_checksum].ci_name, n);
96 }
97
98 static void
ddt_log_update_header(ddt_t * ddt,ddt_log_t * ddl,dmu_tx_t * tx)99 ddt_log_update_header(ddt_t *ddt, ddt_log_t *ddl, dmu_tx_t *tx)
100 {
101 dmu_buf_t *db;
102 VERIFY0(dmu_bonus_hold(ddt->ddt_os, ddl->ddl_object, FTAG, &db));
103 dmu_buf_will_dirty(db, tx);
104
105 ddt_log_header_t *hdr = (ddt_log_header_t *)db->db_data;
106 DLH_SET_VERSION(hdr, 1);
107 DLH_SET_FLAGS(hdr, ddl->ddl_flags);
108 hdr->dlh_length = ddl->ddl_length;
109 hdr->dlh_first_txg = ddl->ddl_first_txg;
110 hdr->dlh_checkpoint = ddl->ddl_checkpoint;
111
112 dmu_buf_rele(db, FTAG);
113 }
114
115 static void
ddt_log_create_one(ddt_t * ddt,ddt_log_t * ddl,uint_t n,dmu_tx_t * tx)116 ddt_log_create_one(ddt_t *ddt, ddt_log_t *ddl, uint_t n, dmu_tx_t *tx)
117 {
118 ASSERT3U(ddt->ddt_dir_object, >, 0);
119 ASSERT3U(ddl->ddl_object, ==, 0);
120
121 char name[DDT_NAMELEN];
122 ddt_log_name(ddt, name, n);
123
124 ddl->ddl_object = dmu_object_alloc(ddt->ddt_os,
125 DMU_OTN_UINT64_METADATA, SPA_OLD_MAXBLOCKSIZE,
126 DMU_OTN_UINT64_METADATA, sizeof (ddt_log_header_t), tx);
127 VERIFY0(zap_add(ddt->ddt_os, ddt->ddt_dir_object, name,
128 sizeof (uint64_t), 1, &ddl->ddl_object, tx));
129 ddl->ddl_length = 0;
130 ddl->ddl_first_txg = tx->tx_txg;
131 ddt_log_update_header(ddt, ddl, tx);
132 }
133
134 static void
ddt_log_create(ddt_t * ddt,dmu_tx_t * tx)135 ddt_log_create(ddt_t *ddt, dmu_tx_t *tx)
136 {
137 ddt_log_create_one(ddt, ddt->ddt_log_active, 0, tx);
138 ddt_log_create_one(ddt, ddt->ddt_log_flushing, 1, tx);
139 }
140
141 static void
ddt_log_destroy_one(ddt_t * ddt,ddt_log_t * ddl,uint_t n,dmu_tx_t * tx)142 ddt_log_destroy_one(ddt_t *ddt, ddt_log_t *ddl, uint_t n, dmu_tx_t *tx)
143 {
144 ASSERT3U(ddt->ddt_dir_object, >, 0);
145
146 if (ddl->ddl_object == 0)
147 return;
148
149 ASSERT0(ddl->ddl_length);
150
151 char name[DDT_NAMELEN];
152 ddt_log_name(ddt, name, n);
153
154 VERIFY0(zap_remove(ddt->ddt_os, ddt->ddt_dir_object, name, tx));
155 VERIFY0(dmu_object_free(ddt->ddt_os, ddl->ddl_object, tx));
156
157 ddl->ddl_object = 0;
158 }
159
160 void
ddt_log_destroy(ddt_t * ddt,dmu_tx_t * tx)161 ddt_log_destroy(ddt_t *ddt, dmu_tx_t *tx)
162 {
163 ddt_log_destroy_one(ddt, ddt->ddt_log_active, 0, tx);
164 ddt_log_destroy_one(ddt, ddt->ddt_log_flushing, 1, tx);
165 }
166
167 static void
ddt_log_update_stats(ddt_t * ddt)168 ddt_log_update_stats(ddt_t *ddt)
169 {
170 /*
171 * Log object stats. We count the number of live entries in the log
172 * tree, even if there are more than on disk, and even if the same
173 * entry is on both append and flush trees, because that's more what
174 * the user expects to see. This does mean the on-disk size is not
175 * really correlated with the number of entries, but I don't think
176 * that's reasonable to expect anyway.
177 */
178 dmu_object_info_t doi;
179 uint64_t nblocks;
180 dmu_object_info(ddt->ddt_os, ddt->ddt_log_active->ddl_object, &doi);
181 nblocks = doi.doi_physical_blocks_512;
182 dmu_object_info(ddt->ddt_os, ddt->ddt_log_flushing->ddl_object, &doi);
183 nblocks += doi.doi_physical_blocks_512;
184
185 ddt_object_t *ddo = &ddt->ddt_log_stats;
186 ddo->ddo_count =
187 avl_numnodes(&ddt->ddt_log_active->ddl_tree) +
188 avl_numnodes(&ddt->ddt_log_flushing->ddl_tree);
189 ddo->ddo_mspace = ddo->ddo_count * DDT_LOG_ENTRY_SIZE(ddt);
190 ddo->ddo_dspace = nblocks << 9;
191 }
192
193 void
ddt_log_begin(ddt_t * ddt,size_t nentries,dmu_tx_t * tx,ddt_log_update_t * dlu)194 ddt_log_begin(ddt_t *ddt, size_t nentries, dmu_tx_t *tx, ddt_log_update_t *dlu)
195 {
196 ASSERT3U(nentries, >, 0);
197 ASSERT3P(dlu->dlu_dbp, ==, NULL);
198
199 if (ddt->ddt_log_active->ddl_object == 0)
200 ddt_log_create(ddt, tx);
201
202 /*
203 * We want to store as many entries as we can in a block, but never
204 * split an entry across block boundaries.
205 */
206 size_t reclen = P2ALIGN_TYPED(
207 sizeof (ddt_log_record_t) + sizeof (ddt_log_record_entry_t) +
208 DDT_PHYS_SIZE(ddt), sizeof (uint64_t), size_t);
209 ASSERT3U(reclen, <=, UINT16_MAX);
210 dlu->dlu_reclen = reclen;
211
212 VERIFY0(dnode_hold(ddt->ddt_os, ddt->ddt_log_active->ddl_object, FTAG,
213 &dlu->dlu_dn));
214 dnode_set_storage_type(dlu->dlu_dn, DMU_OT_DDT_ZAP);
215
216 uint64_t nblocks = howmany(nentries,
217 dlu->dlu_dn->dn_datablksz / dlu->dlu_reclen);
218 uint64_t offset = ddt->ddt_log_active->ddl_length;
219 uint64_t length = nblocks * dlu->dlu_dn->dn_datablksz;
220
221 VERIFY0(dmu_buf_hold_array_by_dnode(dlu->dlu_dn, offset, length,
222 B_FALSE, FTAG, &dlu->dlu_ndbp, &dlu->dlu_dbp,
223 DMU_READ_NO_PREFETCH));
224
225 dlu->dlu_tx = tx;
226 dlu->dlu_block = dlu->dlu_offset = 0;
227 }
228
229 static ddt_log_entry_t *
ddt_log_alloc_entry(ddt_t * ddt)230 ddt_log_alloc_entry(ddt_t *ddt)
231 {
232 ddt_log_entry_t *ddle;
233
234 if (ddt->ddt_flags & DDT_FLAG_FLAT) {
235 ddle = kmem_cache_alloc(ddt_log_entry_flat_cache, KM_SLEEP);
236 memset(ddle, 0, DDT_LOG_ENTRY_FLAT_SIZE);
237 } else {
238 ddle = kmem_cache_alloc(ddt_log_entry_trad_cache, KM_SLEEP);
239 memset(ddle, 0, DDT_LOG_ENTRY_TRAD_SIZE);
240 }
241
242 return (ddle);
243 }
244
245 static void
ddt_log_update_entry(ddt_t * ddt,ddt_log_t * ddl,ddt_lightweight_entry_t * ddlwe)246 ddt_log_update_entry(ddt_t *ddt, ddt_log_t *ddl, ddt_lightweight_entry_t *ddlwe)
247 {
248 /* Create the log tree entry from a live or stored entry */
249 avl_index_t where;
250 ddt_log_entry_t *ddle =
251 avl_find(&ddl->ddl_tree, &ddlwe->ddlwe_key, &where);
252 if (ddle == NULL) {
253 ddle = ddt_log_alloc_entry(ddt);
254 ddle->ddle_key = ddlwe->ddlwe_key;
255 avl_insert(&ddl->ddl_tree, ddle, where);
256 }
257 ddle->ddle_type = ddlwe->ddlwe_type;
258 ddle->ddle_class = ddlwe->ddlwe_class;
259 memcpy(ddle->ddle_phys, &ddlwe->ddlwe_phys, DDT_PHYS_SIZE(ddt));
260 }
261
262 void
ddt_log_entry(ddt_t * ddt,ddt_lightweight_entry_t * ddlwe,ddt_log_update_t * dlu)263 ddt_log_entry(ddt_t *ddt, ddt_lightweight_entry_t *ddlwe, ddt_log_update_t *dlu)
264 {
265 ASSERT3U(dlu->dlu_dbp, !=, NULL);
266
267 ddt_log_update_entry(ddt, ddt->ddt_log_active, ddlwe);
268 ddt_histogram_add_entry(ddt, &ddt->ddt_log_histogram, ddlwe);
269
270 /* Get our block */
271 ASSERT3U(dlu->dlu_block, <, dlu->dlu_ndbp);
272 dmu_buf_t *db = dlu->dlu_dbp[dlu->dlu_block];
273
274 /*
275 * If this would take us past the end of the block, finish it and
276 * move to the next one.
277 */
278 if (db->db_size < (dlu->dlu_offset + dlu->dlu_reclen)) {
279 ASSERT3U(dlu->dlu_offset, >, 0);
280 dmu_buf_fill_done(db, dlu->dlu_tx, B_FALSE);
281 dlu->dlu_block++;
282 dlu->dlu_offset = 0;
283 ASSERT3U(dlu->dlu_block, <, dlu->dlu_ndbp);
284 db = dlu->dlu_dbp[dlu->dlu_block];
285 }
286
287 /*
288 * If this is the first time touching the block, inform the DMU that
289 * we will fill it, and zero it out.
290 */
291 if (dlu->dlu_offset == 0) {
292 dmu_buf_will_fill(db, dlu->dlu_tx, B_FALSE);
293 memset(db->db_data, 0, db->db_size);
294 }
295
296 /* Create the log record directly in the buffer */
297 ddt_log_record_t *dlr = (db->db_data + dlu->dlu_offset);
298 DLR_SET_TYPE(dlr, DLR_ENTRY);
299 DLR_SET_RECLEN(dlr, dlu->dlu_reclen);
300 DLR_SET_ENTRY_TYPE(dlr, ddlwe->ddlwe_type);
301 DLR_SET_ENTRY_CLASS(dlr, ddlwe->ddlwe_class);
302
303 ddt_log_record_entry_t *dlre =
304 (ddt_log_record_entry_t *)&dlr->dlr_payload;
305 dlre->dlre_key = ddlwe->ddlwe_key;
306 memcpy(dlre->dlre_phys, &ddlwe->ddlwe_phys, DDT_PHYS_SIZE(ddt));
307
308 /* Advance offset for next record. */
309 dlu->dlu_offset += dlu->dlu_reclen;
310 }
311
312 void
ddt_log_commit(ddt_t * ddt,ddt_log_update_t * dlu)313 ddt_log_commit(ddt_t *ddt, ddt_log_update_t *dlu)
314 {
315 ASSERT3U(dlu->dlu_dbp, !=, NULL);
316 ASSERT3U(dlu->dlu_block+1, ==, dlu->dlu_ndbp);
317 ASSERT3U(dlu->dlu_offset, >, 0);
318
319 /*
320 * Close out the last block. Whatever we haven't used will be zeroed,
321 * which matches DLR_INVALID, so we can detect this during load.
322 */
323 dmu_buf_fill_done(dlu->dlu_dbp[dlu->dlu_block], dlu->dlu_tx, B_FALSE);
324
325 dmu_buf_rele_array(dlu->dlu_dbp, dlu->dlu_ndbp, FTAG);
326
327 ddt->ddt_log_active->ddl_length +=
328 dlu->dlu_ndbp * (uint64_t)dlu->dlu_dn->dn_datablksz;
329 dnode_rele(dlu->dlu_dn, FTAG);
330
331 ddt_log_update_header(ddt, ddt->ddt_log_active, dlu->dlu_tx);
332
333 memset(dlu, 0, sizeof (ddt_log_update_t));
334
335 ddt_log_update_stats(ddt);
336 }
337
338 boolean_t
ddt_log_take_first(ddt_t * ddt,ddt_log_t * ddl,ddt_lightweight_entry_t * ddlwe)339 ddt_log_take_first(ddt_t *ddt, ddt_log_t *ddl, ddt_lightweight_entry_t *ddlwe)
340 {
341 ddt_log_entry_t *ddle = avl_first(&ddl->ddl_tree);
342 if (ddle == NULL)
343 return (B_FALSE);
344
345 DDT_LOG_ENTRY_TO_LIGHTWEIGHT(ddt, ddle, ddlwe);
346
347 ddt_histogram_sub_entry(ddt, &ddt->ddt_log_histogram, ddlwe);
348
349 avl_remove(&ddl->ddl_tree, ddle);
350 kmem_cache_free(ddt->ddt_flags & DDT_FLAG_FLAT ?
351 ddt_log_entry_flat_cache : ddt_log_entry_trad_cache, ddle);
352
353 return (B_TRUE);
354 }
355
356 boolean_t
ddt_log_remove_key(ddt_t * ddt,ddt_log_t * ddl,const ddt_key_t * ddk)357 ddt_log_remove_key(ddt_t *ddt, ddt_log_t *ddl, const ddt_key_t *ddk)
358 {
359 ddt_log_entry_t *ddle = avl_find(&ddl->ddl_tree, ddk, NULL);
360 if (ddle == NULL)
361 return (B_FALSE);
362
363 ddt_lightweight_entry_t ddlwe;
364 DDT_LOG_ENTRY_TO_LIGHTWEIGHT(ddt, ddle, &ddlwe);
365 ddt_histogram_sub_entry(ddt, &ddt->ddt_log_histogram, &ddlwe);
366
367 avl_remove(&ddl->ddl_tree, ddle);
368 kmem_cache_free(ddt->ddt_flags & DDT_FLAG_FLAT ?
369 ddt_log_entry_flat_cache : ddt_log_entry_trad_cache, ddle);
370
371 return (B_TRUE);
372 }
373
374 boolean_t
ddt_log_find_key(ddt_t * ddt,const ddt_key_t * ddk,ddt_lightweight_entry_t * ddlwe)375 ddt_log_find_key(ddt_t *ddt, const ddt_key_t *ddk,
376 ddt_lightweight_entry_t *ddlwe)
377 {
378 ddt_log_entry_t *ddle =
379 avl_find(&ddt->ddt_log_active->ddl_tree, ddk, NULL);
380 if (!ddle)
381 ddle = avl_find(&ddt->ddt_log_flushing->ddl_tree, ddk, NULL);
382 if (!ddle)
383 return (B_FALSE);
384 if (ddlwe)
385 DDT_LOG_ENTRY_TO_LIGHTWEIGHT(ddt, ddle, ddlwe);
386 return (B_TRUE);
387 }
388
389 void
ddt_log_checkpoint(ddt_t * ddt,ddt_lightweight_entry_t * ddlwe,dmu_tx_t * tx)390 ddt_log_checkpoint(ddt_t *ddt, ddt_lightweight_entry_t *ddlwe, dmu_tx_t *tx)
391 {
392 ddt_log_t *ddl = ddt->ddt_log_flushing;
393
394 ASSERT3U(ddl->ddl_object, !=, 0);
395
396 #ifdef ZFS_DEBUG
397 /*
398 * There should not be any entries on the log tree before the given
399 * checkpoint. Assert that this is the case.
400 */
401 ddt_log_entry_t *ddle = avl_first(&ddl->ddl_tree);
402 if (ddle != NULL)
403 VERIFY3U(ddt_key_compare(&ddle->ddle_key, &ddlwe->ddlwe_key),
404 >, 0);
405 #endif
406
407 ddl->ddl_flags |= DDL_FLAG_CHECKPOINT;
408 ddl->ddl_checkpoint = ddlwe->ddlwe_key;
409 ddt_log_update_header(ddt, ddl, tx);
410
411 ddt_log_update_stats(ddt);
412 }
413
414 void
ddt_log_truncate(ddt_t * ddt,dmu_tx_t * tx)415 ddt_log_truncate(ddt_t *ddt, dmu_tx_t *tx)
416 {
417 ddt_log_t *ddl = ddt->ddt_log_flushing;
418
419 if (ddl->ddl_object == 0)
420 return;
421
422 ASSERT(avl_is_empty(&ddl->ddl_tree));
423
424 /* Eject the entire object */
425 dmu_free_range(ddt->ddt_os, ddl->ddl_object, 0, DMU_OBJECT_END, tx);
426
427 ddl->ddl_length = 0;
428 ddl->ddl_flags &= ~DDL_FLAG_CHECKPOINT;
429 memset(&ddl->ddl_checkpoint, 0, sizeof (ddt_key_t));
430 ddt_log_update_header(ddt, ddl, tx);
431
432 ddt_log_update_stats(ddt);
433 }
434
435 boolean_t
ddt_log_swap(ddt_t * ddt,dmu_tx_t * tx)436 ddt_log_swap(ddt_t *ddt, dmu_tx_t *tx)
437 {
438 /* Swap the logs. The old flushing one must be empty */
439 VERIFY(avl_is_empty(&ddt->ddt_log_flushing->ddl_tree));
440
441 /*
442 * If there are still blocks on the flushing log, truncate it first.
443 * This can happen if there were entries on the flushing log that were
444 * removed in memory via ddt_lookup(); their vestigal remains are
445 * on disk.
446 */
447 if (ddt->ddt_log_flushing->ddl_length > 0)
448 ddt_log_truncate(ddt, tx);
449
450 /*
451 * Swap policy. We swap the logs (and so begin flushing) when the
452 * active tree grows too large, or when we haven't swapped it in
453 * some amount of time, or if something has requested the logs be
454 * flushed ASAP (see ddt_walk_init()).
455 */
456
457 /*
458 * The log tree is too large if the memory usage of its entries is over
459 * half of the memory limit. This effectively gives each log tree half
460 * the available memory.
461 */
462 const boolean_t too_large =
463 (avl_numnodes(&ddt->ddt_log_active->ddl_tree) *
464 DDT_LOG_ENTRY_SIZE(ddt)) >= (zfs_dedup_log_mem_max >> 1);
465
466 const boolean_t too_old =
467 tx->tx_txg >=
468 (ddt->ddt_log_active->ddl_first_txg +
469 MAX(1, zfs_dedup_log_txg_max));
470
471 const boolean_t force =
472 ddt->ddt_log_active->ddl_first_txg <= ddt->ddt_flush_force_txg;
473
474 if (!(too_large || too_old || force))
475 return (B_FALSE);
476
477 ddt_log_t *swap = ddt->ddt_log_active;
478 ddt->ddt_log_active = ddt->ddt_log_flushing;
479 ddt->ddt_log_flushing = swap;
480
481 ASSERT(ddt->ddt_log_active->ddl_flags & DDL_FLAG_FLUSHING);
482 ddt->ddt_log_active->ddl_flags &=
483 ~(DDL_FLAG_FLUSHING | DDL_FLAG_CHECKPOINT);
484
485 ASSERT(!(ddt->ddt_log_flushing->ddl_flags & DDL_FLAG_FLUSHING));
486 ddt->ddt_log_flushing->ddl_flags |= DDL_FLAG_FLUSHING;
487
488 ddt->ddt_log_active->ddl_first_txg = tx->tx_txg;
489
490 ddt_log_update_header(ddt, ddt->ddt_log_active, tx);
491 ddt_log_update_header(ddt, ddt->ddt_log_flushing, tx);
492
493 ddt_log_update_stats(ddt);
494
495 return (B_TRUE);
496 }
497
498 static inline void
ddt_log_load_entry(ddt_t * ddt,ddt_log_t * ddl,ddt_log_record_t * dlr,const ddt_key_t * checkpoint)499 ddt_log_load_entry(ddt_t *ddt, ddt_log_t *ddl, ddt_log_record_t *dlr,
500 const ddt_key_t *checkpoint)
501 {
502 ASSERT3U(DLR_GET_TYPE(dlr), ==, DLR_ENTRY);
503
504 ddt_log_record_entry_t *dlre =
505 (ddt_log_record_entry_t *)dlr->dlr_payload;
506 if (checkpoint != NULL &&
507 ddt_key_compare(&dlre->dlre_key, checkpoint) <= 0) {
508 /* Skip pre-checkpoint entries; they're already flushed. */
509 return;
510 }
511
512 ddt_lightweight_entry_t ddlwe;
513 ddlwe.ddlwe_type = DLR_GET_ENTRY_TYPE(dlr);
514 ddlwe.ddlwe_class = DLR_GET_ENTRY_CLASS(dlr);
515
516 ddlwe.ddlwe_key = dlre->dlre_key;
517 memcpy(&ddlwe.ddlwe_phys, dlre->dlre_phys, DDT_PHYS_SIZE(ddt));
518
519 ddt_log_update_entry(ddt, ddl, &ddlwe);
520 }
521
522 static void
ddt_log_empty(ddt_t * ddt,ddt_log_t * ddl)523 ddt_log_empty(ddt_t *ddt, ddt_log_t *ddl)
524 {
525 void *cookie = NULL;
526 ddt_log_entry_t *ddle;
527 IMPLY(ddt->ddt_version == UINT64_MAX, avl_is_empty(&ddl->ddl_tree));
528 while ((ddle =
529 avl_destroy_nodes(&ddl->ddl_tree, &cookie)) != NULL) {
530 kmem_cache_free(ddt->ddt_flags & DDT_FLAG_FLAT ?
531 ddt_log_entry_flat_cache : ddt_log_entry_trad_cache, ddle);
532 }
533 ASSERT(avl_is_empty(&ddl->ddl_tree));
534 }
535
536 static int
ddt_log_load_one(ddt_t * ddt,uint_t n)537 ddt_log_load_one(ddt_t *ddt, uint_t n)
538 {
539 ASSERT3U(n, <, 2);
540
541 ddt_log_t *ddl = &ddt->ddt_log[n];
542
543 char name[DDT_NAMELEN];
544 ddt_log_name(ddt, name, n);
545
546 uint64_t obj;
547 int err = zap_lookup(ddt->ddt_os, ddt->ddt_dir_object, name,
548 sizeof (uint64_t), 1, &obj);
549 if (err == ENOENT)
550 return (0);
551 if (err != 0)
552 return (err);
553
554 dnode_t *dn;
555 err = dnode_hold(ddt->ddt_os, obj, FTAG, &dn);
556 if (err != 0)
557 return (err);
558
559 ddt_log_header_t hdr;
560 dmu_buf_t *db;
561 err = dmu_bonus_hold_by_dnode(dn, FTAG, &db, DMU_READ_NO_PREFETCH);
562 if (err != 0) {
563 dnode_rele(dn, FTAG);
564 return (err);
565 }
566 memcpy(&hdr, db->db_data, sizeof (ddt_log_header_t));
567 dmu_buf_rele(db, FTAG);
568
569 if (DLH_GET_VERSION(&hdr) != 1) {
570 dnode_rele(dn, FTAG);
571 zfs_dbgmsg("ddt_log_load: spa=%s ddt_log=%s "
572 "unknown version=%llu", spa_name(ddt->ddt_spa), name,
573 (u_longlong_t)DLH_GET_VERSION(&hdr));
574 return (SET_ERROR(EINVAL));
575 }
576
577 ddt_key_t *checkpoint = NULL;
578 if (DLH_GET_FLAGS(&hdr) & DDL_FLAG_CHECKPOINT) {
579 /*
580 * If the log has a checkpoint, then we can ignore any entries
581 * that have already been flushed.
582 */
583 ASSERT(DLH_GET_FLAGS(&hdr) & DDL_FLAG_FLUSHING);
584 checkpoint = &hdr.dlh_checkpoint;
585 }
586
587 if (hdr.dlh_length > 0) {
588 dmu_prefetch_by_dnode(dn, 0, 0, hdr.dlh_length,
589 ZIO_PRIORITY_SYNC_READ);
590
591 for (uint64_t offset = 0; offset < hdr.dlh_length;
592 offset += dn->dn_datablksz) {
593 err = dmu_buf_hold_by_dnode(dn, offset, FTAG, &db,
594 DMU_READ_PREFETCH);
595 if (err != 0) {
596 dnode_rele(dn, FTAG);
597 ddt_log_empty(ddt, ddl);
598 return (err);
599 }
600
601 uint64_t boffset = 0;
602 while (boffset < db->db_size) {
603 ddt_log_record_t *dlr =
604 (ddt_log_record_t *)(db->db_data + boffset);
605
606 /* Partially-filled block, skip the rest */
607 if (DLR_GET_TYPE(dlr) == DLR_INVALID)
608 break;
609
610 switch (DLR_GET_TYPE(dlr)) {
611 case DLR_ENTRY:
612 ddt_log_load_entry(ddt, ddl, dlr,
613 checkpoint);
614 break;
615
616 default:
617 dmu_buf_rele(db, FTAG);
618 dnode_rele(dn, FTAG);
619 ddt_log_empty(ddt, ddl);
620 return (SET_ERROR(EINVAL));
621 }
622
623 boffset += DLR_GET_RECLEN(dlr);
624 }
625
626 dmu_buf_rele(db, FTAG);
627 }
628 }
629
630 dnode_rele(dn, FTAG);
631
632 ddl->ddl_object = obj;
633 ddl->ddl_flags = DLH_GET_FLAGS(&hdr);
634 ddl->ddl_length = hdr.dlh_length;
635 ddl->ddl_first_txg = hdr.dlh_first_txg;
636
637 if (ddl->ddl_flags & DDL_FLAG_FLUSHING)
638 ddt->ddt_log_flushing = ddl;
639 else
640 ddt->ddt_log_active = ddl;
641
642 return (0);
643 }
644
645 int
ddt_log_load(ddt_t * ddt)646 ddt_log_load(ddt_t *ddt)
647 {
648 int err;
649
650 if (spa_load_state(ddt->ddt_spa) == SPA_LOAD_TRYIMPORT) {
651 /*
652 * The DDT is going to be freed again in a moment, so there's
653 * no point loading the log; it'll just slow down import.
654 */
655 return (0);
656 }
657
658 ASSERT0(ddt->ddt_log[0].ddl_object);
659 ASSERT0(ddt->ddt_log[1].ddl_object);
660 if (ddt->ddt_dir_object == 0) {
661 /*
662 * If we're configured but the containing dir doesn't exist
663 * yet, then the log object can't possibly exist either.
664 */
665 ASSERT3U(ddt->ddt_version, !=, UINT64_MAX);
666 return (SET_ERROR(ENOENT));
667 }
668
669 if ((err = ddt_log_load_one(ddt, 0)) != 0)
670 return (err);
671 if ((err = ddt_log_load_one(ddt, 1)) != 0)
672 return (err);
673
674 VERIFY3P(ddt->ddt_log_active, !=, ddt->ddt_log_flushing);
675 VERIFY(!(ddt->ddt_log_active->ddl_flags & DDL_FLAG_FLUSHING));
676 VERIFY(!(ddt->ddt_log_active->ddl_flags & DDL_FLAG_CHECKPOINT));
677 VERIFY(ddt->ddt_log_flushing->ddl_flags & DDL_FLAG_FLUSHING);
678
679 /*
680 * We have two finalisation tasks:
681 *
682 * - rebuild the histogram. We do this at the end rather than while
683 * we're loading so we don't need to uncount and recount entries that
684 * appear multiple times in the log.
685 *
686 * - remove entries from the flushing tree that are on both trees. This
687 * happens when ddt_lookup() rehydrates an entry from the flushing
688 * tree, as ddt_log_take_key() removes the entry from the in-memory
689 * tree but doesn't remove it from disk.
690 */
691
692 /*
693 * We don't technically need a config lock here, since there shouldn't
694 * be pool config changes during DDT load. dva_get_dsize_sync() via
695 * ddt_stat_generate() is expecting it though, and it won't hurt
696 * anything, so we take it.
697 */
698 spa_config_enter(ddt->ddt_spa, SCL_STATE, FTAG, RW_READER);
699
700 avl_tree_t *al = &ddt->ddt_log_active->ddl_tree;
701 avl_tree_t *fl = &ddt->ddt_log_flushing->ddl_tree;
702 ddt_log_entry_t *ae = avl_first(al);
703 ddt_log_entry_t *fe = avl_first(fl);
704 while (ae != NULL || fe != NULL) {
705 ddt_log_entry_t *ddle;
706 if (ae == NULL) {
707 /* active exhausted, take flushing */
708 ddle = fe;
709 fe = AVL_NEXT(fl, fe);
710 } else if (fe == NULL) {
711 /* flushing exuhausted, take active */
712 ddle = ae;
713 ae = AVL_NEXT(al, ae);
714 } else {
715 /* compare active and flushing */
716 int c = ddt_key_compare(&ae->ddle_key, &fe->ddle_key);
717 if (c < 0) {
718 /* active behind, take and advance */
719 ddle = ae;
720 ae = AVL_NEXT(al, ae);
721 } else if (c > 0) {
722 /* flushing behind, take and advance */
723 ddle = fe;
724 fe = AVL_NEXT(fl, fe);
725 } else {
726 /* match. remove from flushing, take active */
727 ddle = fe;
728 fe = AVL_NEXT(fl, fe);
729 avl_remove(fl, ddle);
730
731 ddle = ae;
732 ae = AVL_NEXT(al, ae);
733 }
734 }
735
736 ddt_lightweight_entry_t ddlwe;
737 DDT_LOG_ENTRY_TO_LIGHTWEIGHT(ddt, ddle, &ddlwe);
738 ddt_histogram_add_entry(ddt, &ddt->ddt_log_histogram, &ddlwe);
739 }
740
741 spa_config_exit(ddt->ddt_spa, SCL_STATE, FTAG);
742
743 ddt_log_update_stats(ddt);
744
745 return (0);
746 }
747
748 void
ddt_log_alloc(ddt_t * ddt)749 ddt_log_alloc(ddt_t *ddt)
750 {
751 ASSERT3P(ddt->ddt_log_active, ==, NULL);
752 ASSERT3P(ddt->ddt_log_flushing, ==, NULL);
753
754 avl_create(&ddt->ddt_log[0].ddl_tree, ddt_key_compare,
755 sizeof (ddt_log_entry_t), offsetof(ddt_log_entry_t, ddle_node));
756 avl_create(&ddt->ddt_log[1].ddl_tree, ddt_key_compare,
757 sizeof (ddt_log_entry_t), offsetof(ddt_log_entry_t, ddle_node));
758 ddt->ddt_log_active = &ddt->ddt_log[0];
759 ddt->ddt_log_flushing = &ddt->ddt_log[1];
760 ddt->ddt_log_flushing->ddl_flags |= DDL_FLAG_FLUSHING;
761 }
762
763 void
ddt_log_free(ddt_t * ddt)764 ddt_log_free(ddt_t *ddt)
765 {
766 ddt_log_empty(ddt, &ddt->ddt_log[0]);
767 ddt_log_empty(ddt, &ddt->ddt_log[1]);
768 avl_destroy(&ddt->ddt_log[0].ddl_tree);
769 avl_destroy(&ddt->ddt_log[1].ddl_tree);
770 }
771
772 ZFS_MODULE_PARAM(zfs_dedup, zfs_dedup_, log_txg_max, UINT, ZMOD_RW,
773 "Max transactions before starting to flush dedup logs");
774
775 ZFS_MODULE_PARAM(zfs_dedup, zfs_dedup_, log_mem_max, U64, ZMOD_RD,
776 "Max memory for dedup logs");
777
778 ZFS_MODULE_PARAM(zfs_dedup, zfs_dedup_, log_mem_max_percent, UINT, ZMOD_RD,
779 "Max memory for dedup logs, as % of total memory");
780