Print this page
3752 want more verifiable dbuf user eviction
Submitted by: Justin Gibbs <justing@spectralogic.com>
Submitted by: Will Andrews <willa@spectralogic.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/dnode_sync.c
+++ new/usr/src/uts/common/fs/zfs/dnode_sync.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21
22 22 /*
23 23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 24 * Copyright (c) 2012 by Delphix. All rights reserved.
25 25 */
26 26
27 27 #include <sys/zfs_context.h>
28 28 #include <sys/dbuf.h>
29 29 #include <sys/dnode.h>
30 30 #include <sys/dmu.h>
31 31 #include <sys/dmu_tx.h>
32 32 #include <sys/dmu_objset.h>
33 33 #include <sys/dsl_dataset.h>
34 34 #include <sys/spa.h>
35 35
36 36 static void
37 37 dnode_increase_indirection(dnode_t *dn, dmu_tx_t *tx)
38 38 {
39 39 dmu_buf_impl_t *db;
40 40 int txgoff = tx->tx_txg & TXG_MASK;
41 41 int nblkptr = dn->dn_phys->dn_nblkptr;
42 42 int old_toplvl = dn->dn_phys->dn_nlevels - 1;
43 43 int new_level = dn->dn_next_nlevels[txgoff];
44 44 int i;
45 45
46 46 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
47 47
48 48 /* this dnode can't be paged out because it's dirty */
49 49 ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
50 50 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
51 51 ASSERT(new_level > 1 && dn->dn_phys->dn_nlevels > 0);
52 52
53 53 db = dbuf_hold_level(dn, dn->dn_phys->dn_nlevels, 0, FTAG);
54 54 ASSERT(db != NULL);
55 55
56 56 dn->dn_phys->dn_nlevels = new_level;
57 57 dprintf("os=%p obj=%llu, increase to %d\n", dn->dn_objset,
58 58 dn->dn_object, dn->dn_phys->dn_nlevels);
59 59
60 60 /* check for existing blkptrs in the dnode */
61 61 for (i = 0; i < nblkptr; i++)
62 62 if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[i]))
63 63 break;
64 64 if (i != nblkptr) {
65 65 /* transfer dnode's block pointers to new indirect block */
66 66 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED|DB_RF_HAVESTRUCT);
67 67 ASSERT(db->db.db_data);
68 68 ASSERT(arc_released(db->db_buf));
69 69 ASSERT3U(sizeof (blkptr_t) * nblkptr, <=, db->db.db_size);
70 70 bcopy(dn->dn_phys->dn_blkptr, db->db.db_data,
71 71 sizeof (blkptr_t) * nblkptr);
72 72 arc_buf_freeze(db->db_buf);
73 73 }
74 74
75 75 /* set dbuf's parent pointers to new indirect buf */
76 76 for (i = 0; i < nblkptr; i++) {
77 77 dmu_buf_impl_t *child = dbuf_find(dn, old_toplvl, i);
78 78
79 79 if (child == NULL)
80 80 continue;
81 81 #ifdef DEBUG
82 82 DB_DNODE_ENTER(child);
83 83 ASSERT3P(DB_DNODE(child), ==, dn);
84 84 DB_DNODE_EXIT(child);
85 85 #endif /* DEBUG */
86 86 if (child->db_parent && child->db_parent != dn->dn_dbuf) {
87 87 ASSERT(child->db_parent->db_level == db->db_level);
88 88 ASSERT(child->db_blkptr !=
89 89 &dn->dn_phys->dn_blkptr[child->db_blkid]);
90 90 mutex_exit(&child->db_mtx);
91 91 continue;
92 92 }
93 93 ASSERT(child->db_parent == NULL ||
94 94 child->db_parent == dn->dn_dbuf);
95 95
96 96 child->db_parent = db;
97 97 dbuf_add_ref(db, child);
98 98 if (db->db.db_data)
99 99 child->db_blkptr = (blkptr_t *)db->db.db_data + i;
100 100 else
101 101 child->db_blkptr = NULL;
102 102 dprintf_dbuf_bp(child, child->db_blkptr,
103 103 "changed db_blkptr to new indirect %s", "");
104 104
105 105 mutex_exit(&child->db_mtx);
106 106 }
107 107
108 108 bzero(dn->dn_phys->dn_blkptr, sizeof (blkptr_t) * nblkptr);
109 109
110 110 dbuf_rele(db, FTAG);
111 111
112 112 rw_exit(&dn->dn_struct_rwlock);
113 113 }
114 114
115 115 static int
116 116 free_blocks(dnode_t *dn, blkptr_t *bp, int num, dmu_tx_t *tx)
117 117 {
118 118 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
119 119 uint64_t bytesfreed = 0;
120 120 int i, blocks_freed = 0;
121 121
122 122 dprintf("ds=%p obj=%llx num=%d\n", ds, dn->dn_object, num);
123 123
124 124 for (i = 0; i < num; i++, bp++) {
125 125 if (BP_IS_HOLE(bp))
126 126 continue;
127 127
128 128 bytesfreed += dsl_dataset_block_kill(ds, bp, tx, B_FALSE);
129 129 ASSERT3U(bytesfreed, <=, DN_USED_BYTES(dn->dn_phys));
130 130 bzero(bp, sizeof (blkptr_t));
131 131 blocks_freed += 1;
132 132 }
133 133 dnode_diduse_space(dn, -bytesfreed);
134 134 return (blocks_freed);
135 135 }
136 136
137 137 #ifdef ZFS_DEBUG
138 138 static void
139 139 free_verify(dmu_buf_impl_t *db, uint64_t start, uint64_t end, dmu_tx_t *tx)
140 140 {
141 141 int off, num;
142 142 int i, err, epbs;
143 143 uint64_t txg = tx->tx_txg;
144 144 dnode_t *dn;
145 145
146 146 DB_DNODE_ENTER(db);
147 147 dn = DB_DNODE(db);
148 148 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
149 149 off = start - (db->db_blkid * 1<<epbs);
150 150 num = end - start + 1;
151 151
152 152 ASSERT3U(off, >=, 0);
153 153 ASSERT3U(num, >=, 0);
154 154 ASSERT3U(db->db_level, >, 0);
155 155 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
156 156 ASSERT3U(off+num, <=, db->db.db_size >> SPA_BLKPTRSHIFT);
157 157 ASSERT(db->db_blkptr != NULL);
158 158
159 159 for (i = off; i < off+num; i++) {
160 160 uint64_t *buf;
161 161 dmu_buf_impl_t *child;
162 162 dbuf_dirty_record_t *dr;
163 163 int j;
164 164
165 165 ASSERT(db->db_level == 1);
166 166
167 167 rw_enter(&dn->dn_struct_rwlock, RW_READER);
168 168 err = dbuf_hold_impl(dn, db->db_level-1,
169 169 (db->db_blkid << epbs) + i, TRUE, FTAG, &child);
170 170 rw_exit(&dn->dn_struct_rwlock);
171 171 if (err == ENOENT)
172 172 continue;
173 173 ASSERT(err == 0);
174 174 ASSERT(child->db_level == 0);
175 175 dr = child->db_last_dirty;
176 176 while (dr && dr->dr_txg > txg)
177 177 dr = dr->dr_next;
178 178 ASSERT(dr == NULL || dr->dr_txg == txg);
179 179
180 180 /* data_old better be zeroed */
181 181 if (dr) {
182 182 buf = dr->dt.dl.dr_data->b_data;
183 183 for (j = 0; j < child->db.db_size >> 3; j++) {
184 184 if (buf[j] != 0) {
185 185 panic("freed data not zero: "
186 186 "child=%p i=%d off=%d num=%d\n",
187 187 (void *)child, i, off, num);
188 188 }
189 189 }
190 190 }
191 191
192 192 /*
193 193 * db_data better be zeroed unless it's dirty in a
194 194 * future txg.
195 195 */
196 196 mutex_enter(&child->db_mtx);
197 197 buf = child->db.db_data;
198 198 if (buf != NULL && child->db_state != DB_FILL &&
199 199 child->db_last_dirty == NULL) {
200 200 for (j = 0; j < child->db.db_size >> 3; j++) {
201 201 if (buf[j] != 0) {
202 202 panic("freed data not zero: "
203 203 "child=%p i=%d off=%d num=%d\n",
204 204 (void *)child, i, off, num);
205 205 }
206 206 }
207 207 }
208 208 mutex_exit(&child->db_mtx);
209 209
210 210 dbuf_rele(child, FTAG);
211 211 }
212 212 DB_DNODE_EXIT(db);
213 213 }
214 214 #endif
215 215
216 216 #define ALL -1
217 217
218 218 static int
219 219 free_children(dmu_buf_impl_t *db, uint64_t blkid, uint64_t nblks, int trunc,
220 220 dmu_tx_t *tx)
221 221 {
222 222 dnode_t *dn;
223 223 blkptr_t *bp;
224 224 dmu_buf_impl_t *subdb;
225 225 uint64_t start, end, dbstart, dbend, i;
226 226 int epbs, shift, err;
227 227 int all = TRUE;
228 228 int blocks_freed = 0;
229 229
230 230 /*
231 231 * There is a small possibility that this block will not be cached:
232 232 * 1 - if level > 1 and there are no children with level <= 1
233 233 * 2 - if we didn't get a dirty hold (because this block had just
234 234 * finished being written -- and so had no holds), and then this
235 235 * block got evicted before we got here.
236 236 */
237 237 if (db->db_state != DB_CACHED)
238 238 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
239 239
240 240 dbuf_release_bp(db);
241 241 bp = (blkptr_t *)db->db.db_data;
242 242
243 243 DB_DNODE_ENTER(db);
244 244 dn = DB_DNODE(db);
245 245 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
246 246 shift = (db->db_level - 1) * epbs;
247 247 dbstart = db->db_blkid << epbs;
248 248 start = blkid >> shift;
249 249 if (dbstart < start) {
250 250 bp += start - dbstart;
251 251 all = FALSE;
252 252 } else {
253 253 start = dbstart;
254 254 }
255 255 dbend = ((db->db_blkid + 1) << epbs) - 1;
256 256 end = (blkid + nblks - 1) >> shift;
257 257 if (dbend <= end)
258 258 end = dbend;
259 259 else if (all)
260 260 all = trunc;
261 261 ASSERT3U(start, <=, end);
262 262
263 263 if (db->db_level == 1) {
264 264 FREE_VERIFY(db, start, end, tx);
265 265 blocks_freed = free_blocks(dn, bp, end-start+1, tx);
266 266 arc_buf_freeze(db->db_buf);
267 267 ASSERT(all || blocks_freed == 0 || db->db_last_dirty);
268 268 DB_DNODE_EXIT(db);
269 269 return (all ? ALL : blocks_freed);
270 270 }
271 271
272 272 for (i = start; i <= end; i++, bp++) {
273 273 if (BP_IS_HOLE(bp))
274 274 continue;
275 275 rw_enter(&dn->dn_struct_rwlock, RW_READER);
276 276 err = dbuf_hold_impl(dn, db->db_level-1, i, TRUE, FTAG, &subdb);
277 277 ASSERT0(err);
278 278 rw_exit(&dn->dn_struct_rwlock);
279 279
280 280 if (free_children(subdb, blkid, nblks, trunc, tx) == ALL) {
281 281 ASSERT3P(subdb->db_blkptr, ==, bp);
282 282 blocks_freed += free_blocks(dn, bp, 1, tx);
283 283 } else {
284 284 all = FALSE;
285 285 }
286 286 dbuf_rele(subdb, FTAG);
287 287 }
288 288 DB_DNODE_EXIT(db);
289 289 arc_buf_freeze(db->db_buf);
290 290 #ifdef ZFS_DEBUG
291 291 bp -= (end-start)+1;
292 292 for (i = start; i <= end; i++, bp++) {
293 293 if (i == start && blkid != 0)
294 294 continue;
295 295 else if (i == end && !trunc)
296 296 continue;
297 297 ASSERT0(bp->blk_birth);
298 298 }
299 299 #endif
300 300 ASSERT(all || blocks_freed == 0 || db->db_last_dirty);
301 301 return (all ? ALL : blocks_freed);
302 302 }
303 303
304 304 /*
305 305 * free_range: Traverse the indicated range of the provided file
306 306 * and "free" all the blocks contained there.
307 307 */
308 308 static void
309 309 dnode_sync_free_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
310 310 {
311 311 blkptr_t *bp = dn->dn_phys->dn_blkptr;
312 312 dmu_buf_impl_t *db;
313 313 int trunc, start, end, shift, i, err;
314 314 int dnlevel = dn->dn_phys->dn_nlevels;
315 315
316 316 if (blkid > dn->dn_phys->dn_maxblkid)
317 317 return;
318 318
319 319 ASSERT(dn->dn_phys->dn_maxblkid < UINT64_MAX);
320 320 trunc = blkid + nblks > dn->dn_phys->dn_maxblkid;
321 321 if (trunc)
322 322 nblks = dn->dn_phys->dn_maxblkid - blkid + 1;
323 323
324 324 /* There are no indirect blocks in the object */
325 325 if (dnlevel == 1) {
326 326 if (blkid >= dn->dn_phys->dn_nblkptr) {
327 327 /* this range was never made persistent */
328 328 return;
329 329 }
330 330 ASSERT3U(blkid + nblks, <=, dn->dn_phys->dn_nblkptr);
331 331 (void) free_blocks(dn, bp + blkid, nblks, tx);
332 332 if (trunc) {
333 333 uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
334 334 (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
335 335 dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0);
336 336 ASSERT(off < dn->dn_phys->dn_maxblkid ||
337 337 dn->dn_phys->dn_maxblkid == 0 ||
338 338 dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
339 339 }
340 340 return;
341 341 }
342 342
343 343 shift = (dnlevel - 1) * (dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT);
344 344 start = blkid >> shift;
345 345 ASSERT(start < dn->dn_phys->dn_nblkptr);
346 346 end = (blkid + nblks - 1) >> shift;
347 347 bp += start;
348 348 for (i = start; i <= end; i++, bp++) {
349 349 if (BP_IS_HOLE(bp))
350 350 continue;
351 351 rw_enter(&dn->dn_struct_rwlock, RW_READER);
352 352 err = dbuf_hold_impl(dn, dnlevel-1, i, TRUE, FTAG, &db);
353 353 ASSERT0(err);
354 354 rw_exit(&dn->dn_struct_rwlock);
355 355
356 356 if (free_children(db, blkid, nblks, trunc, tx) == ALL) {
357 357 ASSERT3P(db->db_blkptr, ==, bp);
358 358 (void) free_blocks(dn, bp, 1, tx);
359 359 }
360 360 dbuf_rele(db, FTAG);
361 361 }
362 362 if (trunc) {
363 363 uint64_t off = (dn->dn_phys->dn_maxblkid + 1) *
364 364 (dn->dn_phys->dn_datablkszsec << SPA_MINBLOCKSHIFT);
365 365 dn->dn_phys->dn_maxblkid = (blkid ? blkid - 1 : 0);
366 366 ASSERT(off < dn->dn_phys->dn_maxblkid ||
367 367 dn->dn_phys->dn_maxblkid == 0 ||
368 368 dnode_next_offset(dn, 0, &off, 1, 1, 0) != 0);
369 369 }
|
↓ open down ↓ |
369 lines elided |
↑ open up ↑ |
370 370 }
371 371
372 372 /*
373 373 * Try to kick all the dnodes dbufs out of the cache...
374 374 */
375 375 void
376 376 dnode_evict_dbufs(dnode_t *dn)
377 377 {
378 378 int progress;
379 379 int pass = 0;
380 + list_t evict_list;
381 +
382 + dmu_buf_create_user_evict_list(&evict_list);
380 383
381 384 do {
382 385 dmu_buf_impl_t *db, marker;
383 386 int evicting = FALSE;
384 387
385 388 progress = FALSE;
386 389 mutex_enter(&dn->dn_dbufs_mtx);
387 390 list_insert_tail(&dn->dn_dbufs, &marker);
388 391 db = list_head(&dn->dn_dbufs);
389 392 for (; db != ▮ db = list_head(&dn->dn_dbufs)) {
390 393 list_remove(&dn->dn_dbufs, db);
391 394 list_insert_tail(&dn->dn_dbufs, db);
392 395 #ifdef DEBUG
393 396 DB_DNODE_ENTER(db);
394 397 ASSERT3P(DB_DNODE(db), ==, dn);
|
↓ open down ↓ |
5 lines elided |
↑ open up ↑ |
395 398 DB_DNODE_EXIT(db);
396 399 #endif /* DEBUG */
397 400
398 401 mutex_enter(&db->db_mtx);
399 402 if (db->db_state == DB_EVICTING) {
400 403 progress = TRUE;
401 404 evicting = TRUE;
402 405 mutex_exit(&db->db_mtx);
403 406 } else if (refcount_is_zero(&db->db_holds)) {
404 407 progress = TRUE;
405 - dbuf_clear(db); /* exits db_mtx for us */
408 + dbuf_clear(db, &evict_list); /* exits db_mtx */
406 409 } else {
407 410 mutex_exit(&db->db_mtx);
408 411 }
409 -
412 + ASSERT(MUTEX_NOT_HELD(&db->db_mtx));
413 + dmu_buf_process_user_evicts(&evict_list);
410 414 }
411 415 list_remove(&dn->dn_dbufs, &marker);
412 416 /*
413 417 * NB: we need to drop dn_dbufs_mtx between passes so
414 418 * that any DB_EVICTING dbufs can make progress.
415 419 * Ideally, we would have some cv we could wait on, but
416 420 * since we don't, just wait a bit to give the other
417 421 * thread a chance to run.
418 422 */
419 423 mutex_exit(&dn->dn_dbufs_mtx);
420 424 if (evicting)
421 425 delay(1);
422 426 pass++;
423 427 ASSERT(pass < 100); /* sanity check */
424 428 } while (progress);
425 429
426 430 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
427 431 if (dn->dn_bonus && refcount_is_zero(&dn->dn_bonus->db_holds)) {
428 432 mutex_enter(&dn->dn_bonus->db_mtx);
429 - dbuf_evict(dn->dn_bonus);
433 + dbuf_evict(dn->dn_bonus, &evict_list);
430 434 dn->dn_bonus = NULL;
431 435 }
432 436 rw_exit(&dn->dn_struct_rwlock);
437 + dmu_buf_destroy_user_evict_list(&evict_list);
433 438 }
434 439
435 440 static void
436 441 dnode_undirty_dbufs(list_t *list)
437 442 {
438 443 dbuf_dirty_record_t *dr;
439 444
440 445 while (dr = list_head(list)) {
441 446 dmu_buf_impl_t *db = dr->dr_dbuf;
442 447 uint64_t txg = dr->dr_txg;
443 448
444 449 if (db->db_level != 0)
445 450 dnode_undirty_dbufs(&dr->dt.di.dr_children);
446 451
447 452 mutex_enter(&db->db_mtx);
448 453 /* XXX - use dbuf_undirty()? */
449 454 list_remove(list, dr);
450 455 ASSERT(db->db_last_dirty == dr);
451 456 db->db_last_dirty = NULL;
452 457 db->db_dirtycnt -= 1;
453 458 if (db->db_level == 0) {
454 459 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
455 460 dr->dt.dl.dr_data == db->db_buf);
456 461 dbuf_unoverride(dr);
457 462 }
458 463 kmem_free(dr, sizeof (dbuf_dirty_record_t));
459 464 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
460 465 }
461 466 }
462 467
463 468 static void
464 469 dnode_sync_free(dnode_t *dn, dmu_tx_t *tx)
465 470 {
466 471 int txgoff = tx->tx_txg & TXG_MASK;
467 472
468 473 ASSERT(dmu_tx_is_syncing(tx));
469 474
470 475 /*
471 476 * Our contents should have been freed in dnode_sync() by the
472 477 * free range record inserted by the caller of dnode_free().
473 478 */
474 479 ASSERT0(DN_USED_BYTES(dn->dn_phys));
475 480 ASSERT(BP_IS_HOLE(dn->dn_phys->dn_blkptr));
476 481
477 482 dnode_undirty_dbufs(&dn->dn_dirty_records[txgoff]);
478 483 dnode_evict_dbufs(dn);
479 484 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
480 485 ASSERT3P(dn->dn_bonus, ==, NULL);
481 486
482 487 /*
483 488 * XXX - It would be nice to assert this, but we may still
484 489 * have residual holds from async evictions from the arc...
485 490 *
486 491 * zfs_obj_to_path() also depends on this being
487 492 * commented out.
488 493 *
489 494 * ASSERT3U(refcount_count(&dn->dn_holds), ==, 1);
490 495 */
491 496
492 497 /* Undirty next bits */
493 498 dn->dn_next_nlevels[txgoff] = 0;
494 499 dn->dn_next_indblkshift[txgoff] = 0;
495 500 dn->dn_next_blksz[txgoff] = 0;
496 501
497 502 /* ASSERT(blkptrs are zero); */
498 503 ASSERT(dn->dn_phys->dn_type != DMU_OT_NONE);
499 504 ASSERT(dn->dn_type != DMU_OT_NONE);
500 505
501 506 ASSERT(dn->dn_free_txg > 0);
502 507 if (dn->dn_allocated_txg != dn->dn_free_txg)
503 508 dbuf_will_dirty(dn->dn_dbuf, tx);
504 509 bzero(dn->dn_phys, sizeof (dnode_phys_t));
505 510
506 511 mutex_enter(&dn->dn_mtx);
507 512 dn->dn_type = DMU_OT_NONE;
508 513 dn->dn_maxblkid = 0;
509 514 dn->dn_allocated_txg = 0;
510 515 dn->dn_free_txg = 0;
511 516 dn->dn_have_spill = B_FALSE;
512 517 mutex_exit(&dn->dn_mtx);
513 518
514 519 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
515 520
516 521 dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
517 522 /*
518 523 * Now that we've released our hold, the dnode may
519 524 * be evicted, so we musn't access it.
520 525 */
521 526 }
522 527
523 528 /*
524 529 * Write out the dnode's dirty buffers.
525 530 */
526 531 void
527 532 dnode_sync(dnode_t *dn, dmu_tx_t *tx)
528 533 {
529 534 free_range_t *rp;
530 535 dnode_phys_t *dnp = dn->dn_phys;
531 536 int txgoff = tx->tx_txg & TXG_MASK;
532 537 list_t *list = &dn->dn_dirty_records[txgoff];
533 538 static const dnode_phys_t zerodn = { 0 };
534 539 boolean_t kill_spill = B_FALSE;
535 540
536 541 ASSERT(dmu_tx_is_syncing(tx));
537 542 ASSERT(dnp->dn_type != DMU_OT_NONE || dn->dn_allocated_txg);
538 543 ASSERT(dnp->dn_type != DMU_OT_NONE ||
539 544 bcmp(dnp, &zerodn, DNODE_SIZE) == 0);
540 545 DNODE_VERIFY(dn);
541 546
542 547 ASSERT(dn->dn_dbuf == NULL || arc_released(dn->dn_dbuf->db_buf));
543 548
544 549 if (dmu_objset_userused_enabled(dn->dn_objset) &&
545 550 !DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
546 551 mutex_enter(&dn->dn_mtx);
547 552 dn->dn_oldused = DN_USED_BYTES(dn->dn_phys);
548 553 dn->dn_oldflags = dn->dn_phys->dn_flags;
549 554 dn->dn_phys->dn_flags |= DNODE_FLAG_USERUSED_ACCOUNTED;
550 555 mutex_exit(&dn->dn_mtx);
551 556 dmu_objset_userquota_get_ids(dn, B_FALSE, tx);
552 557 } else {
553 558 /* Once we account for it, we should always account for it. */
554 559 ASSERT(!(dn->dn_phys->dn_flags &
555 560 DNODE_FLAG_USERUSED_ACCOUNTED));
556 561 }
557 562
558 563 mutex_enter(&dn->dn_mtx);
559 564 if (dn->dn_allocated_txg == tx->tx_txg) {
560 565 /* The dnode is newly allocated or reallocated */
561 566 if (dnp->dn_type == DMU_OT_NONE) {
562 567 /* this is a first alloc, not a realloc */
563 568 dnp->dn_nlevels = 1;
564 569 dnp->dn_nblkptr = dn->dn_nblkptr;
565 570 }
566 571
567 572 dnp->dn_type = dn->dn_type;
568 573 dnp->dn_bonustype = dn->dn_bonustype;
569 574 dnp->dn_bonuslen = dn->dn_bonuslen;
570 575 }
571 576
572 577 ASSERT(dnp->dn_nlevels > 1 ||
573 578 BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
574 579 BP_GET_LSIZE(&dnp->dn_blkptr[0]) ==
575 580 dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
576 581
577 582 if (dn->dn_next_blksz[txgoff]) {
578 583 ASSERT(P2PHASE(dn->dn_next_blksz[txgoff],
579 584 SPA_MINBLOCKSIZE) == 0);
580 585 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[0]) ||
581 586 dn->dn_maxblkid == 0 || list_head(list) != NULL ||
582 587 avl_last(&dn->dn_ranges[txgoff]) ||
583 588 dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT ==
584 589 dnp->dn_datablkszsec);
585 590 dnp->dn_datablkszsec =
586 591 dn->dn_next_blksz[txgoff] >> SPA_MINBLOCKSHIFT;
587 592 dn->dn_next_blksz[txgoff] = 0;
588 593 }
589 594
590 595 if (dn->dn_next_bonuslen[txgoff]) {
591 596 if (dn->dn_next_bonuslen[txgoff] == DN_ZERO_BONUSLEN)
592 597 dnp->dn_bonuslen = 0;
593 598 else
594 599 dnp->dn_bonuslen = dn->dn_next_bonuslen[txgoff];
595 600 ASSERT(dnp->dn_bonuslen <= DN_MAX_BONUSLEN);
596 601 dn->dn_next_bonuslen[txgoff] = 0;
597 602 }
598 603
599 604 if (dn->dn_next_bonustype[txgoff]) {
600 605 ASSERT(DMU_OT_IS_VALID(dn->dn_next_bonustype[txgoff]));
601 606 dnp->dn_bonustype = dn->dn_next_bonustype[txgoff];
602 607 dn->dn_next_bonustype[txgoff] = 0;
603 608 }
604 609
605 610 /*
606 611 * We will either remove a spill block when a file is being removed
607 612 * or we have been asked to remove it.
608 613 */
609 614 if (dn->dn_rm_spillblk[txgoff] ||
610 615 ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) &&
611 616 dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg)) {
612 617 if ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
613 618 kill_spill = B_TRUE;
614 619 dn->dn_rm_spillblk[txgoff] = 0;
615 620 }
616 621
617 622 if (dn->dn_next_indblkshift[txgoff]) {
618 623 ASSERT(dnp->dn_nlevels == 1);
619 624 dnp->dn_indblkshift = dn->dn_next_indblkshift[txgoff];
620 625 dn->dn_next_indblkshift[txgoff] = 0;
621 626 }
622 627
623 628 /*
624 629 * Just take the live (open-context) values for checksum and compress.
625 630 * Strictly speaking it's a future leak, but nothing bad happens if we
626 631 * start using the new checksum or compress algorithm a little early.
627 632 */
628 633 dnp->dn_checksum = dn->dn_checksum;
629 634 dnp->dn_compress = dn->dn_compress;
630 635
631 636 mutex_exit(&dn->dn_mtx);
632 637
633 638 if (kill_spill) {
634 639 (void) free_blocks(dn, &dn->dn_phys->dn_spill, 1, tx);
635 640 mutex_enter(&dn->dn_mtx);
636 641 dnp->dn_flags &= ~DNODE_FLAG_SPILL_BLKPTR;
637 642 mutex_exit(&dn->dn_mtx);
638 643 }
639 644
640 645 /* process all the "freed" ranges in the file */
641 646 while (rp = avl_last(&dn->dn_ranges[txgoff])) {
642 647 dnode_sync_free_range(dn, rp->fr_blkid, rp->fr_nblks, tx);
643 648 /* grab the mutex so we don't race with dnode_block_freed() */
644 649 mutex_enter(&dn->dn_mtx);
645 650 avl_remove(&dn->dn_ranges[txgoff], rp);
646 651 mutex_exit(&dn->dn_mtx);
647 652 kmem_free(rp, sizeof (free_range_t));
648 653 }
649 654
650 655 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= tx->tx_txg) {
651 656 dnode_sync_free(dn, tx);
652 657 return;
653 658 }
654 659
655 660 if (dn->dn_next_nblkptr[txgoff]) {
656 661 /* this should only happen on a realloc */
657 662 ASSERT(dn->dn_allocated_txg == tx->tx_txg);
658 663 if (dn->dn_next_nblkptr[txgoff] > dnp->dn_nblkptr) {
659 664 /* zero the new blkptrs we are gaining */
660 665 bzero(dnp->dn_blkptr + dnp->dn_nblkptr,
661 666 sizeof (blkptr_t) *
662 667 (dn->dn_next_nblkptr[txgoff] - dnp->dn_nblkptr));
663 668 #ifdef ZFS_DEBUG
664 669 } else {
665 670 int i;
666 671 ASSERT(dn->dn_next_nblkptr[txgoff] < dnp->dn_nblkptr);
667 672 /* the blkptrs we are losing better be unallocated */
668 673 for (i = dn->dn_next_nblkptr[txgoff];
669 674 i < dnp->dn_nblkptr; i++)
670 675 ASSERT(BP_IS_HOLE(&dnp->dn_blkptr[i]));
671 676 #endif
672 677 }
673 678 mutex_enter(&dn->dn_mtx);
674 679 dnp->dn_nblkptr = dn->dn_next_nblkptr[txgoff];
675 680 dn->dn_next_nblkptr[txgoff] = 0;
676 681 mutex_exit(&dn->dn_mtx);
677 682 }
678 683
679 684 if (dn->dn_next_nlevels[txgoff]) {
680 685 dnode_increase_indirection(dn, tx);
681 686 dn->dn_next_nlevels[txgoff] = 0;
682 687 }
683 688
684 689 dbuf_sync_list(list, tx);
685 690
686 691 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
687 692 ASSERT3P(list_head(list), ==, NULL);
688 693 dnode_rele(dn, (void *)(uintptr_t)tx->tx_txg);
689 694 }
690 695
691 696 /*
692 697 * Although we have dropped our reference to the dnode, it
693 698 * can't be evicted until its written, and we haven't yet
694 699 * initiated the IO for the dnode's dbuf.
695 700 */
696 701 }
|
↓ open down ↓ |
254 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX