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 ASSERT0(ddl->ddl_object);
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 = 0;
180 if (dmu_object_info(ddt->ddt_os, ddt->ddt_log_active->ddl_object,
181 &doi) == 0)
182 nblocks += doi.doi_physical_blocks_512;
183 if (dmu_object_info(ddt->ddt_os, ddt->ddt_log_flushing->ddl_object,
184 &doi) == 0)
185 nblocks += doi.doi_physical_blocks_512;
186
187 ddt_object_t *ddo = &ddt->ddt_log_stats;
188 ddo->ddo_count =
189 avl_numnodes(&ddt->ddt_log_active->ddl_tree) +
190 avl_numnodes(&ddt->ddt_log_flushing->ddl_tree);
191 ddo->ddo_mspace = ddo->ddo_count * DDT_LOG_ENTRY_SIZE(ddt);
192 ddo->ddo_dspace = nblocks << 9;
193 }
194
195 void
ddt_log_begin(ddt_t * ddt,size_t nentries,dmu_tx_t * tx,ddt_log_update_t * dlu)196 ddt_log_begin(ddt_t *ddt, size_t nentries, dmu_tx_t *tx, ddt_log_update_t *dlu)
197 {
198 ASSERT3U(nentries, >, 0);
199 ASSERT0P(dlu->dlu_dbp);
200
201 if (ddt->ddt_log_active->ddl_object == 0)
202 ddt_log_create(ddt, tx);
203
204 /*
205 * We want to store as many entries as we can in a block, but never
206 * split an entry across block boundaries.
207 */
208 size_t reclen = P2ALIGN_TYPED(
209 sizeof (ddt_log_record_t) + sizeof (ddt_log_record_entry_t) +
210 DDT_PHYS_SIZE(ddt), sizeof (uint64_t), size_t);
211 ASSERT3U(reclen, <=, UINT16_MAX);
212 dlu->dlu_reclen = reclen;
213
214 VERIFY0(dnode_hold(ddt->ddt_os, ddt->ddt_log_active->ddl_object, FTAG,
215 &dlu->dlu_dn));
216 dnode_set_storage_type(dlu->dlu_dn, DMU_OT_DDT_ZAP);
217
218 uint64_t nblocks = howmany(nentries,
219 dlu->dlu_dn->dn_datablksz / dlu->dlu_reclen);
220 uint64_t offset = ddt->ddt_log_active->ddl_length;
221 uint64_t length = nblocks * dlu->dlu_dn->dn_datablksz;
222
223 VERIFY0(dmu_buf_hold_array_by_dnode(dlu->dlu_dn, offset, length,
224 B_FALSE, FTAG, &dlu->dlu_ndbp, &dlu->dlu_dbp,
225 DMU_READ_NO_PREFETCH));
226
227 dlu->dlu_tx = tx;
228 dlu->dlu_block = dlu->dlu_offset = 0;
229 }
230
231 static ddt_log_entry_t *
ddt_log_alloc_entry(ddt_t * ddt)232 ddt_log_alloc_entry(ddt_t *ddt)
233 {
234 ddt_log_entry_t *ddle;
235
236 if (ddt->ddt_flags & DDT_FLAG_FLAT) {
237 ddle = kmem_cache_alloc(ddt_log_entry_flat_cache, KM_SLEEP);
238 memset(ddle, 0, DDT_LOG_ENTRY_FLAT_SIZE);
239 } else {
240 ddle = kmem_cache_alloc(ddt_log_entry_trad_cache, KM_SLEEP);
241 memset(ddle, 0, DDT_LOG_ENTRY_TRAD_SIZE);
242 }
243
244 return (ddle);
245 }
246
247 static void
ddt_log_free_entry(ddt_t * ddt,ddt_log_entry_t * ddle)248 ddt_log_free_entry(ddt_t *ddt, ddt_log_entry_t *ddle)
249 {
250 kmem_cache_free(ddt->ddt_flags & DDT_FLAG_FLAT ?
251 ddt_log_entry_flat_cache : ddt_log_entry_trad_cache, ddle);
252 }
253
254 static void
ddt_log_update_entry(ddt_t * ddt,ddt_log_t * ddl,ddt_lightweight_entry_t * ddlwe)255 ddt_log_update_entry(ddt_t *ddt, ddt_log_t *ddl, ddt_lightweight_entry_t *ddlwe)
256 {
257 /* Create the log tree entry from a live or stored entry */
258 avl_index_t where;
259 ddt_log_entry_t *ddle =
260 avl_find(&ddl->ddl_tree, &ddlwe->ddlwe_key, &where);
261 if (ddle == NULL) {
262 ddle = ddt_log_alloc_entry(ddt);
263 ddle->ddle_key = ddlwe->ddlwe_key;
264 avl_insert(&ddl->ddl_tree, ddle, where);
265 }
266 ddle->ddle_type = ddlwe->ddlwe_type;
267 ddle->ddle_class = ddlwe->ddlwe_class;
268 memcpy(ddle->ddle_phys, &ddlwe->ddlwe_phys, DDT_PHYS_SIZE(ddt));
269 }
270
271 void
ddt_log_entry(ddt_t * ddt,ddt_lightweight_entry_t * ddlwe,ddt_log_update_t * dlu)272 ddt_log_entry(ddt_t *ddt, ddt_lightweight_entry_t *ddlwe, ddt_log_update_t *dlu)
273 {
274 ASSERT3U(dlu->dlu_dbp, !=, NULL);
275
276 ddt_log_update_entry(ddt, ddt->ddt_log_active, ddlwe);
277 ddt_histogram_add_entry(ddt, &ddt->ddt_log_histogram, ddlwe);
278
279 /* Get our block */
280 ASSERT3U(dlu->dlu_block, <, dlu->dlu_ndbp);
281 dmu_buf_t *db = dlu->dlu_dbp[dlu->dlu_block];
282
283 /*
284 * If this would take us past the end of the block, finish it and
285 * move to the next one.
286 */
287 if (db->db_size < (dlu->dlu_offset + dlu->dlu_reclen)) {
288 ASSERT3U(dlu->dlu_offset, >, 0);
289 dmu_buf_fill_done(db, dlu->dlu_tx, B_FALSE);
290 dlu->dlu_block++;
291 dlu->dlu_offset = 0;
292 ASSERT3U(dlu->dlu_block, <, dlu->dlu_ndbp);
293 db = dlu->dlu_dbp[dlu->dlu_block];
294 }
295
296 /*
297 * If this is the first time touching the block, inform the DMU that
298 * we will fill it, and zero it out.
299 */
300 if (dlu->dlu_offset == 0) {
301 dmu_buf_will_fill(db, dlu->dlu_tx, B_FALSE);
302 memset(db->db_data, 0, db->db_size);
303 }
304
305 /* Create the log record directly in the buffer */
306 ddt_log_record_t *dlr = (db->db_data + dlu->dlu_offset);
307 DLR_SET_TYPE(dlr, DLR_ENTRY);
308 DLR_SET_RECLEN(dlr, dlu->dlu_reclen);
309 DLR_SET_ENTRY_TYPE(dlr, ddlwe->ddlwe_type);
310 DLR_SET_ENTRY_CLASS(dlr, ddlwe->ddlwe_class);
311
312 ddt_log_record_entry_t *dlre =
313 (ddt_log_record_entry_t *)&dlr->dlr_payload;
314 dlre->dlre_key = ddlwe->ddlwe_key;
315 memcpy(dlre->dlre_phys, &ddlwe->ddlwe_phys, DDT_PHYS_SIZE(ddt));
316
317 /* Advance offset for next record. */
318 dlu->dlu_offset += dlu->dlu_reclen;
319 }
320
321 void
ddt_log_commit(ddt_t * ddt,ddt_log_update_t * dlu)322 ddt_log_commit(ddt_t *ddt, ddt_log_update_t *dlu)
323 {
324 ASSERT3U(dlu->dlu_dbp, !=, NULL);
325 ASSERT3U(dlu->dlu_block+1, ==, dlu->dlu_ndbp);
326 ASSERT3U(dlu->dlu_offset, >, 0);
327
328 /*
329 * Close out the last block. Whatever we haven't used will be zeroed,
330 * which matches DLR_INVALID, so we can detect this during load.
331 */
332 dmu_buf_fill_done(dlu->dlu_dbp[dlu->dlu_block], dlu->dlu_tx, B_FALSE);
333
334 dmu_buf_rele_array(dlu->dlu_dbp, dlu->dlu_ndbp, FTAG);
335
336 ddt->ddt_log_active->ddl_length +=
337 dlu->dlu_ndbp * (uint64_t)dlu->dlu_dn->dn_datablksz;
338 dnode_rele(dlu->dlu_dn, FTAG);
339
340 ddt_log_update_header(ddt, ddt->ddt_log_active, dlu->dlu_tx);
341
342 memset(dlu, 0, sizeof (ddt_log_update_t));
343
344 ddt_log_update_stats(ddt);
345 }
346
347 boolean_t
ddt_log_take_first(ddt_t * ddt,ddt_log_t * ddl,ddt_lightweight_entry_t * ddlwe)348 ddt_log_take_first(ddt_t *ddt, ddt_log_t *ddl, ddt_lightweight_entry_t *ddlwe)
349 {
350 ddt_log_entry_t *ddle = avl_first(&ddl->ddl_tree);
351 if (ddle == NULL)
352 return (B_FALSE);
353
354 DDT_LOG_ENTRY_TO_LIGHTWEIGHT(ddt, ddle, ddlwe);
355
356 ddt_histogram_sub_entry(ddt, &ddt->ddt_log_histogram, ddlwe);
357
358 avl_remove(&ddl->ddl_tree, ddle);
359 ddt_log_free_entry(ddt, ddle);
360
361 return (B_TRUE);
362 }
363
364 boolean_t
ddt_log_remove_key(ddt_t * ddt,ddt_log_t * ddl,const ddt_key_t * ddk)365 ddt_log_remove_key(ddt_t *ddt, ddt_log_t *ddl, const ddt_key_t *ddk)
366 {
367 ddt_log_entry_t *ddle = avl_find(&ddl->ddl_tree, ddk, NULL);
368 if (ddle == NULL)
369 return (B_FALSE);
370
371 ddt_lightweight_entry_t ddlwe;
372 DDT_LOG_ENTRY_TO_LIGHTWEIGHT(ddt, ddle, &ddlwe);
373 ddt_histogram_sub_entry(ddt, &ddt->ddt_log_histogram, &ddlwe);
374
375 avl_remove(&ddl->ddl_tree, ddle);
376 ddt_log_free_entry(ddt, ddle);
377
378 return (B_TRUE);
379 }
380
381 boolean_t
ddt_log_find_key(ddt_t * ddt,const ddt_key_t * ddk,ddt_lightweight_entry_t * ddlwe)382 ddt_log_find_key(ddt_t *ddt, const ddt_key_t *ddk,
383 ddt_lightweight_entry_t *ddlwe)
384 {
385 ddt_log_entry_t *ddle =
386 avl_find(&ddt->ddt_log_active->ddl_tree, ddk, NULL);
387 if (!ddle)
388 ddle = avl_find(&ddt->ddt_log_flushing->ddl_tree, ddk, NULL);
389 if (!ddle)
390 return (B_FALSE);
391 if (ddlwe)
392 DDT_LOG_ENTRY_TO_LIGHTWEIGHT(ddt, ddle, ddlwe);
393 return (B_TRUE);
394 }
395
396 void
ddt_log_checkpoint(ddt_t * ddt,ddt_lightweight_entry_t * ddlwe,dmu_tx_t * tx)397 ddt_log_checkpoint(ddt_t *ddt, ddt_lightweight_entry_t *ddlwe, dmu_tx_t *tx)
398 {
399 ddt_log_t *ddl = ddt->ddt_log_flushing;
400
401 ASSERT3U(ddl->ddl_object, !=, 0);
402
403 #ifdef ZFS_DEBUG
404 /*
405 * There should not be any entries on the log tree before the given
406 * checkpoint. Assert that this is the case.
407 */
408 ddt_log_entry_t *ddle = avl_first(&ddl->ddl_tree);
409 if (ddle != NULL)
410 VERIFY3U(ddt_key_compare(&ddle->ddle_key, &ddlwe->ddlwe_key),
411 >, 0);
412 #endif
413
414 ddl->ddl_flags |= DDL_FLAG_CHECKPOINT;
415 ddl->ddl_checkpoint = ddlwe->ddlwe_key;
416 ddt_log_update_header(ddt, ddl, tx);
417
418 ddt_log_update_stats(ddt);
419 }
420
421 void
ddt_log_truncate(ddt_t * ddt,dmu_tx_t * tx)422 ddt_log_truncate(ddt_t *ddt, dmu_tx_t *tx)
423 {
424 ddt_log_t *ddl = ddt->ddt_log_flushing;
425
426 if (ddl->ddl_object == 0)
427 return;
428
429 ASSERT(avl_is_empty(&ddl->ddl_tree));
430
431 /* Eject the entire object */
432 dmu_free_range(ddt->ddt_os, ddl->ddl_object, 0, DMU_OBJECT_END, tx);
433
434 ddl->ddl_length = 0;
435 ddl->ddl_flags &= ~DDL_FLAG_CHECKPOINT;
436 memset(&ddl->ddl_checkpoint, 0, sizeof (ddt_key_t));
437 ddt_log_update_header(ddt, ddl, tx);
438
439 ddt_log_update_stats(ddt);
440 }
441
442 boolean_t
ddt_log_swap(ddt_t * ddt,dmu_tx_t * tx)443 ddt_log_swap(ddt_t *ddt, dmu_tx_t *tx)
444 {
445 /* Swap the logs. The old flushing one must be empty */
446 VERIFY(avl_is_empty(&ddt->ddt_log_flushing->ddl_tree));
447
448 /*
449 * If there are still blocks on the flushing log, truncate it first.
450 * This can happen if there were entries on the flushing log that were
451 * removed in memory via ddt_lookup(); their vestigal remains are
452 * on disk.
453 */
454 if (ddt->ddt_log_flushing->ddl_length > 0)
455 ddt_log_truncate(ddt, tx);
456
457 /*
458 * Swap policy. We swap the logs (and so begin flushing) when the
459 * active tree grows too large, or when we haven't swapped it in
460 * some amount of time, or if something has requested the logs be
461 * flushed ASAP (see ddt_walk_init()).
462 */
463
464 /*
465 * The log tree is too large if the memory usage of its entries is over
466 * half of the memory limit. This effectively gives each log tree half
467 * the available memory.
468 */
469 const boolean_t too_large =
470 (avl_numnodes(&ddt->ddt_log_active->ddl_tree) *
471 DDT_LOG_ENTRY_SIZE(ddt)) >= (zfs_dedup_log_mem_max >> 1);
472
473 const boolean_t too_old =
474 tx->tx_txg >=
475 (ddt->ddt_log_active->ddl_first_txg +
476 MAX(1, zfs_dedup_log_txg_max));
477
478 const boolean_t force =
479 ddt->ddt_log_active->ddl_first_txg <= ddt->ddt_flush_force_txg;
480
481 if (!(too_large || too_old || force))
482 return (B_FALSE);
483
484 ddt_log_t *swap = ddt->ddt_log_active;
485 ddt->ddt_log_active = ddt->ddt_log_flushing;
486 ddt->ddt_log_flushing = swap;
487
488 ASSERT(ddt->ddt_log_active->ddl_flags & DDL_FLAG_FLUSHING);
489 ddt->ddt_log_active->ddl_flags &=
490 ~(DDL_FLAG_FLUSHING | DDL_FLAG_CHECKPOINT);
491
492 ASSERT(!(ddt->ddt_log_flushing->ddl_flags & DDL_FLAG_FLUSHING));
493 ddt->ddt_log_flushing->ddl_flags |= DDL_FLAG_FLUSHING;
494
495 ddt->ddt_log_active->ddl_first_txg = tx->tx_txg;
496
497 ddt_log_update_header(ddt, ddt->ddt_log_active, tx);
498 ddt_log_update_header(ddt, ddt->ddt_log_flushing, tx);
499
500 ddt_log_update_stats(ddt);
501
502 return (B_TRUE);
503 }
504
505 static inline void
ddt_log_load_entry(ddt_t * ddt,ddt_log_t * ddl,ddt_log_record_t * dlr,const ddt_key_t * checkpoint)506 ddt_log_load_entry(ddt_t *ddt, ddt_log_t *ddl, ddt_log_record_t *dlr,
507 const ddt_key_t *checkpoint)
508 {
509 ASSERT3U(DLR_GET_TYPE(dlr), ==, DLR_ENTRY);
510
511 ddt_log_record_entry_t *dlre =
512 (ddt_log_record_entry_t *)dlr->dlr_payload;
513 if (checkpoint != NULL &&
514 ddt_key_compare(&dlre->dlre_key, checkpoint) <= 0) {
515 /* Skip pre-checkpoint entries; they're already flushed. */
516 return;
517 }
518
519 ddt_lightweight_entry_t ddlwe;
520 ddlwe.ddlwe_type = DLR_GET_ENTRY_TYPE(dlr);
521 ddlwe.ddlwe_class = DLR_GET_ENTRY_CLASS(dlr);
522
523 ddlwe.ddlwe_key = dlre->dlre_key;
524 memcpy(&ddlwe.ddlwe_phys, dlre->dlre_phys, DDT_PHYS_SIZE(ddt));
525
526 ddt_log_update_entry(ddt, ddl, &ddlwe);
527 }
528
529 static void
ddt_log_empty(ddt_t * ddt,ddt_log_t * ddl)530 ddt_log_empty(ddt_t *ddt, ddt_log_t *ddl)
531 {
532 void *cookie = NULL;
533 ddt_log_entry_t *ddle;
534 IMPLY(ddt->ddt_version == UINT64_MAX, avl_is_empty(&ddl->ddl_tree));
535 while ((ddle =
536 avl_destroy_nodes(&ddl->ddl_tree, &cookie)) != NULL) {
537 ddt_log_free_entry(ddt, ddle);
538 }
539 ASSERT(avl_is_empty(&ddl->ddl_tree));
540 }
541
542 static int
ddt_log_load_one(ddt_t * ddt,uint_t n)543 ddt_log_load_one(ddt_t *ddt, uint_t n)
544 {
545 ASSERT3U(n, <, 2);
546
547 ddt_log_t *ddl = &ddt->ddt_log[n];
548
549 char name[DDT_NAMELEN];
550 ddt_log_name(ddt, name, n);
551
552 uint64_t obj;
553 int err = zap_lookup(ddt->ddt_os, ddt->ddt_dir_object, name,
554 sizeof (uint64_t), 1, &obj);
555 if (err == ENOENT)
556 return (0);
557 if (err != 0)
558 return (err);
559
560 dnode_t *dn;
561 err = dnode_hold(ddt->ddt_os, obj, FTAG, &dn);
562 if (err != 0)
563 return (err);
564
565 ddt_log_header_t hdr;
566 dmu_buf_t *db;
567 err = dmu_bonus_hold_by_dnode(dn, FTAG, &db, DMU_READ_NO_PREFETCH);
568 if (err != 0) {
569 dnode_rele(dn, FTAG);
570 return (err);
571 }
572 memcpy(&hdr, db->db_data, sizeof (ddt_log_header_t));
573 dmu_buf_rele(db, FTAG);
574
575 if (DLH_GET_VERSION(&hdr) != 1) {
576 dnode_rele(dn, FTAG);
577 zfs_dbgmsg("ddt_log_load: spa=%s ddt_log=%s "
578 "unknown version=%llu", spa_name(ddt->ddt_spa), name,
579 (u_longlong_t)DLH_GET_VERSION(&hdr));
580 return (SET_ERROR(EINVAL));
581 }
582
583 ddt_key_t *checkpoint = NULL;
584 if (DLH_GET_FLAGS(&hdr) & DDL_FLAG_CHECKPOINT) {
585 /*
586 * If the log has a checkpoint, then we can ignore any entries
587 * that have already been flushed.
588 */
589 ASSERT(DLH_GET_FLAGS(&hdr) & DDL_FLAG_FLUSHING);
590 checkpoint = &hdr.dlh_checkpoint;
591 }
592
593 if (hdr.dlh_length > 0) {
594 dmu_prefetch_by_dnode(dn, 0, 0, hdr.dlh_length,
595 ZIO_PRIORITY_SYNC_READ);
596
597 for (uint64_t offset = 0; offset < hdr.dlh_length;
598 offset += dn->dn_datablksz) {
599 err = dmu_buf_hold_by_dnode(dn, offset, FTAG, &db,
600 DMU_READ_PREFETCH);
601 if (err != 0) {
602 dnode_rele(dn, FTAG);
603 ddt_log_empty(ddt, ddl);
604 return (err);
605 }
606
607 uint64_t boffset = 0;
608 while (boffset < db->db_size) {
609 ddt_log_record_t *dlr =
610 (ddt_log_record_t *)(db->db_data + boffset);
611
612 /* Partially-filled block, skip the rest */
613 if (DLR_GET_TYPE(dlr) == DLR_INVALID)
614 break;
615
616 switch (DLR_GET_TYPE(dlr)) {
617 case DLR_ENTRY:
618 ddt_log_load_entry(ddt, ddl, dlr,
619 checkpoint);
620 break;
621
622 default:
623 dmu_buf_rele(db, FTAG);
624 dnode_rele(dn, FTAG);
625 ddt_log_empty(ddt, ddl);
626 return (SET_ERROR(EINVAL));
627 }
628
629 boffset += DLR_GET_RECLEN(dlr);
630 }
631
632 dmu_buf_rele(db, FTAG);
633 }
634 }
635
636 dnode_rele(dn, FTAG);
637
638 ddl->ddl_object = obj;
639 ddl->ddl_flags = DLH_GET_FLAGS(&hdr);
640 ddl->ddl_length = hdr.dlh_length;
641 ddl->ddl_first_txg = hdr.dlh_first_txg;
642
643 if (ddl->ddl_flags & DDL_FLAG_FLUSHING)
644 ddt->ddt_log_flushing = ddl;
645 else
646 ddt->ddt_log_active = ddl;
647
648 return (0);
649 }
650
651 int
ddt_log_load(ddt_t * ddt)652 ddt_log_load(ddt_t *ddt)
653 {
654 int err;
655
656 if (spa_load_state(ddt->ddt_spa) == SPA_LOAD_TRYIMPORT) {
657 /*
658 * The DDT is going to be freed again in a moment, so there's
659 * no point loading the log; it'll just slow down import.
660 */
661 return (0);
662 }
663
664 ASSERT0(ddt->ddt_log[0].ddl_object);
665 ASSERT0(ddt->ddt_log[1].ddl_object);
666 if (ddt->ddt_dir_object == 0) {
667 /*
668 * If we're configured but the containing dir doesn't exist
669 * yet, then the log object can't possibly exist either.
670 */
671 ASSERT3U(ddt->ddt_version, !=, UINT64_MAX);
672 return (SET_ERROR(ENOENT));
673 }
674
675 if ((err = ddt_log_load_one(ddt, 0)) != 0)
676 return (err);
677 if ((err = ddt_log_load_one(ddt, 1)) != 0)
678 return (err);
679
680 VERIFY3P(ddt->ddt_log_active, !=, ddt->ddt_log_flushing);
681 VERIFY(!(ddt->ddt_log_active->ddl_flags & DDL_FLAG_FLUSHING));
682 VERIFY(!(ddt->ddt_log_active->ddl_flags & DDL_FLAG_CHECKPOINT));
683 VERIFY(ddt->ddt_log_flushing->ddl_flags & DDL_FLAG_FLUSHING);
684
685 /*
686 * We have two finalisation tasks:
687 *
688 * - rebuild the histogram. We do this at the end rather than while
689 * we're loading so we don't need to uncount and recount entries that
690 * appear multiple times in the log.
691 *
692 * - remove entries from the flushing tree that are on both trees. This
693 * happens when ddt_lookup() rehydrates an entry from the flushing
694 * tree, as ddt_log_take_key() removes the entry from the in-memory
695 * tree but doesn't remove it from disk.
696 */
697
698 /*
699 * We don't technically need a config lock here, since there shouldn't
700 * be pool config changes during DDT load. dva_get_dsize_sync() via
701 * ddt_stat_generate() is expecting it though, and it won't hurt
702 * anything, so we take it.
703 */
704 spa_config_enter(ddt->ddt_spa, SCL_STATE, FTAG, RW_READER);
705
706 avl_tree_t *al = &ddt->ddt_log_active->ddl_tree;
707 avl_tree_t *fl = &ddt->ddt_log_flushing->ddl_tree;
708 ddt_log_entry_t *ae = avl_first(al);
709 ddt_log_entry_t *fe = avl_first(fl);
710 while (ae != NULL || fe != NULL) {
711 ddt_log_entry_t *ddle;
712 if (ae == NULL) {
713 /* active exhausted, take flushing */
714 ddle = fe;
715 fe = AVL_NEXT(fl, fe);
716 } else if (fe == NULL) {
717 /* flushing exuhausted, take active */
718 ddle = ae;
719 ae = AVL_NEXT(al, ae);
720 } else {
721 /* compare active and flushing */
722 int c = ddt_key_compare(&ae->ddle_key, &fe->ddle_key);
723 if (c < 0) {
724 /* active behind, take and advance */
725 ddle = ae;
726 ae = AVL_NEXT(al, ae);
727 } else if (c > 0) {
728 /* flushing behind, take and advance */
729 ddle = fe;
730 fe = AVL_NEXT(fl, fe);
731 } else {
732 /* match. remove from flushing, take active */
733 ddle = fe;
734 fe = AVL_NEXT(fl, fe);
735 avl_remove(fl, ddle);
736 ddt_log_free_entry(ddt, ddle);
737 ddle = ae;
738 ae = AVL_NEXT(al, ae);
739 }
740 }
741
742 ddt_lightweight_entry_t ddlwe;
743 DDT_LOG_ENTRY_TO_LIGHTWEIGHT(ddt, ddle, &ddlwe);
744 ddt_histogram_add_entry(ddt, &ddt->ddt_log_histogram, &ddlwe);
745 }
746
747 spa_config_exit(ddt->ddt_spa, SCL_STATE, FTAG);
748
749 ddt_log_update_stats(ddt);
750
751 return (0);
752 }
753
754 void
ddt_log_alloc(ddt_t * ddt)755 ddt_log_alloc(ddt_t *ddt)
756 {
757 ASSERT0P(ddt->ddt_log_active);
758 ASSERT0P(ddt->ddt_log_flushing);
759
760 avl_create(&ddt->ddt_log[0].ddl_tree, ddt_key_compare,
761 sizeof (ddt_log_entry_t), offsetof(ddt_log_entry_t, ddle_node));
762 avl_create(&ddt->ddt_log[1].ddl_tree, ddt_key_compare,
763 sizeof (ddt_log_entry_t), offsetof(ddt_log_entry_t, ddle_node));
764 ddt->ddt_log_active = &ddt->ddt_log[0];
765 ddt->ddt_log_flushing = &ddt->ddt_log[1];
766 ddt->ddt_log_flushing->ddl_flags |= DDL_FLAG_FLUSHING;
767 }
768
769 void
ddt_log_free(ddt_t * ddt)770 ddt_log_free(ddt_t *ddt)
771 {
772 ddt_log_empty(ddt, &ddt->ddt_log[0]);
773 ddt_log_empty(ddt, &ddt->ddt_log[1]);
774 avl_destroy(&ddt->ddt_log[0].ddl_tree);
775 avl_destroy(&ddt->ddt_log[1].ddl_tree);
776 }
777
778 ZFS_MODULE_PARAM(zfs_dedup, zfs_dedup_, log_txg_max, UINT, ZMOD_RW,
779 "Max transactions before starting to flush dedup logs");
780
781 ZFS_MODULE_PARAM(zfs_dedup, zfs_dedup_, log_mem_max, U64, ZMOD_RD,
782 "Max memory for dedup logs");
783
784 ZFS_MODULE_PARAM(zfs_dedup, zfs_dedup_, log_mem_max_percent, UINT, ZMOD_RD,
785 "Max memory for dedup logs, as % of total memory");
786