]>
Commit | Line | Data |
---|---|---|
9e0f480e DW |
1 | /* |
2 | * Copyright (C) 2016 Oracle. All Rights Reserved. | |
3 | * | |
4 | * Author: Darrick J. Wong <darrick.wong@oracle.com> | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU General Public License | |
8 | * as published by the Free Software Foundation; either version 2 | |
9 | * of the License, or (at your option) any later version. | |
10 | * | |
11 | * This program is distributed in the hope that it would be useful, | |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | * GNU General Public License for more details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * along with this program; if not, write the Free Software Foundation, | |
18 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. | |
19 | */ | |
20 | #include <libxfs.h> | |
21 | #include "btree.h" | |
22 | #include "err_protos.h" | |
23 | #include "libxlog.h" | |
24 | #include "incore.h" | |
25 | #include "globals.h" | |
26 | #include "dinode.h" | |
27 | #include "slab.h" | |
28 | #include "rmap.h" | |
29 | ||
30 | #undef RMAP_DEBUG | |
31 | ||
32 | #ifdef RMAP_DEBUG | |
33 | # define dbg_printf(f, a...) do {printf(f, ## a); fflush(stdout); } while (0) | |
34 | #else | |
35 | # define dbg_printf(f, a...) | |
36 | #endif | |
37 | ||
38 | /* per-AG rmap object anchor */ | |
39 | struct xfs_ag_rmap { | |
40 | struct xfs_slab *ar_rmaps; /* rmap observations, p4 */ | |
1102c155 | 41 | struct xfs_slab *ar_raw_rmaps; /* unmerged rmaps */ |
62cf990a DW |
42 | int ar_flcount; /* agfl entries from leftover */ |
43 | /* agbt allocations */ | |
b7f12e53 | 44 | struct xfs_rmap_irec ar_last_rmap; /* last rmap seen */ |
00f34bca | 45 | struct xfs_slab *ar_refcount_items; /* refcount items, p4-5 */ |
9e0f480e DW |
46 | }; |
47 | ||
48 | static struct xfs_ag_rmap *ag_rmaps; | |
11b9e510 | 49 | static bool rmapbt_suspect; |
80dbc783 | 50 | static bool refcbt_suspect; |
9e0f480e DW |
51 | |
52 | /* | |
53 | * Compare rmap observations for array sorting. | |
54 | */ | |
55 | static int | |
56 | rmap_compare( | |
57 | const void *a, | |
58 | const void *b) | |
59 | { | |
60 | const struct xfs_rmap_irec *pa; | |
61 | const struct xfs_rmap_irec *pb; | |
62 | __u64 oa; | |
63 | __u64 ob; | |
64 | ||
65 | pa = a; pb = b; | |
e2f60652 DW |
66 | oa = libxfs_rmap_irec_offset_pack(pa); |
67 | ob = libxfs_rmap_irec_offset_pack(pb); | |
9e0f480e DW |
68 | |
69 | if (pa->rm_startblock < pb->rm_startblock) | |
70 | return -1; | |
71 | else if (pa->rm_startblock > pb->rm_startblock) | |
72 | return 1; | |
73 | else if (pa->rm_owner < pb->rm_owner) | |
74 | return -1; | |
75 | else if (pa->rm_owner > pb->rm_owner) | |
76 | return 1; | |
77 | else if (oa < ob) | |
78 | return -1; | |
79 | else if (oa > ob) | |
80 | return 1; | |
81 | else | |
82 | return 0; | |
83 | } | |
84 | ||
85 | /* | |
86 | * Returns true if we must reconstruct either the reference count or reverse | |
87 | * mapping trees. | |
88 | */ | |
89 | bool | |
2d273771 | 90 | rmap_needs_work( |
9e0f480e DW |
91 | struct xfs_mount *mp) |
92 | { | |
00f34bca DW |
93 | return xfs_sb_version_hasreflink(&mp->m_sb) || |
94 | xfs_sb_version_hasrmapbt(&mp->m_sb); | |
9e0f480e DW |
95 | } |
96 | ||
97 | /* | |
98 | * Initialize per-AG reverse map data. | |
99 | */ | |
100 | void | |
2d273771 | 101 | rmaps_init( |
9e0f480e DW |
102 | struct xfs_mount *mp) |
103 | { | |
104 | xfs_agnumber_t i; | |
105 | int error; | |
106 | ||
2d273771 | 107 | if (!rmap_needs_work(mp)) |
9e0f480e DW |
108 | return; |
109 | ||
110 | ag_rmaps = calloc(mp->m_sb.sb_agcount, sizeof(struct xfs_ag_rmap)); | |
111 | if (!ag_rmaps) | |
112 | do_error(_("couldn't allocate per-AG reverse map roots\n")); | |
113 | ||
114 | for (i = 0; i < mp->m_sb.sb_agcount; i++) { | |
115 | error = init_slab(&ag_rmaps[i].ar_rmaps, | |
116 | sizeof(struct xfs_rmap_irec)); | |
117 | if (error) | |
118 | do_error( | |
119 | _("Insufficient memory while allocating reverse mapping slabs.")); | |
1102c155 DW |
120 | error = init_slab(&ag_rmaps[i].ar_raw_rmaps, |
121 | sizeof(struct xfs_rmap_irec)); | |
122 | if (error) | |
123 | do_error( | |
124 | _("Insufficient memory while allocating raw metadata reverse mapping slabs.")); | |
b7f12e53 | 125 | ag_rmaps[i].ar_last_rmap.rm_owner = XFS_RMAP_OWN_UNKNOWN; |
00f34bca DW |
126 | error = init_slab(&ag_rmaps[i].ar_refcount_items, |
127 | sizeof(struct xfs_refcount_irec)); | |
128 | if (error) | |
129 | do_error( | |
130 | _("Insufficient memory while allocating refcount item slabs.")); | |
9e0f480e DW |
131 | } |
132 | } | |
133 | ||
134 | /* | |
135 | * Free the per-AG reverse-mapping data. | |
136 | */ | |
137 | void | |
2d273771 | 138 | rmaps_free( |
9e0f480e DW |
139 | struct xfs_mount *mp) |
140 | { | |
141 | xfs_agnumber_t i; | |
142 | ||
2d273771 | 143 | if (!rmap_needs_work(mp)) |
9e0f480e DW |
144 | return; |
145 | ||
1102c155 | 146 | for (i = 0; i < mp->m_sb.sb_agcount; i++) { |
9e0f480e | 147 | free_slab(&ag_rmaps[i].ar_rmaps); |
1102c155 | 148 | free_slab(&ag_rmaps[i].ar_raw_rmaps); |
00f34bca | 149 | free_slab(&ag_rmaps[i].ar_refcount_items); |
1102c155 | 150 | } |
9e0f480e DW |
151 | free(ag_rmaps); |
152 | ag_rmaps = NULL; | |
153 | } | |
154 | ||
1102c155 DW |
155 | /* |
156 | * Decide if two reverse-mapping records can be merged. | |
157 | */ | |
158 | bool | |
2d273771 | 159 | rmaps_are_mergeable( |
1102c155 DW |
160 | struct xfs_rmap_irec *r1, |
161 | struct xfs_rmap_irec *r2) | |
162 | { | |
163 | if (r1->rm_owner != r2->rm_owner) | |
164 | return false; | |
165 | if (r1->rm_startblock + r1->rm_blockcount != r2->rm_startblock) | |
166 | return false; | |
167 | if ((unsigned long long)r1->rm_blockcount + r2->rm_blockcount > | |
168 | XFS_RMAP_LEN_MAX) | |
169 | return false; | |
170 | if (XFS_RMAP_NON_INODE_OWNER(r2->rm_owner)) | |
171 | return true; | |
172 | /* must be an inode owner below here */ | |
173 | if (r1->rm_flags != r2->rm_flags) | |
174 | return false; | |
175 | if (r1->rm_flags & XFS_RMAP_BMBT_BLOCK) | |
176 | return true; | |
177 | return r1->rm_offset + r1->rm_blockcount == r2->rm_offset; | |
178 | } | |
179 | ||
9e0f480e DW |
180 | /* |
181 | * Add an observation about a block mapping in an inode's data or attribute | |
182 | * fork for later btree reconstruction. | |
183 | */ | |
184 | int | |
2d273771 | 185 | rmap_add_rec( |
9e0f480e DW |
186 | struct xfs_mount *mp, |
187 | xfs_ino_t ino, | |
188 | int whichfork, | |
189 | struct xfs_bmbt_irec *irec) | |
190 | { | |
9e0f480e DW |
191 | struct xfs_rmap_irec rmap; |
192 | xfs_agnumber_t agno; | |
193 | xfs_agblock_t agbno; | |
b7f12e53 DW |
194 | struct xfs_rmap_irec *last_rmap; |
195 | int error = 0; | |
9e0f480e | 196 | |
2d273771 | 197 | if (!rmap_needs_work(mp)) |
9e0f480e DW |
198 | return 0; |
199 | ||
200 | agno = XFS_FSB_TO_AGNO(mp, irec->br_startblock); | |
201 | agbno = XFS_FSB_TO_AGBNO(mp, irec->br_startblock); | |
202 | ASSERT(agno != NULLAGNUMBER); | |
203 | ASSERT(agno < mp->m_sb.sb_agcount); | |
204 | ASSERT(agbno + irec->br_blockcount <= mp->m_sb.sb_agblocks); | |
205 | ASSERT(ino != NULLFSINO); | |
206 | ASSERT(whichfork == XFS_DATA_FORK || whichfork == XFS_ATTR_FORK); | |
207 | ||
9e0f480e DW |
208 | rmap.rm_owner = ino; |
209 | rmap.rm_offset = irec->br_startoff; | |
210 | rmap.rm_flags = 0; | |
211 | if (whichfork == XFS_ATTR_FORK) | |
212 | rmap.rm_flags |= XFS_RMAP_ATTR_FORK; | |
213 | rmap.rm_startblock = agbno; | |
214 | rmap.rm_blockcount = irec->br_blockcount; | |
215 | if (irec->br_state == XFS_EXT_UNWRITTEN) | |
216 | rmap.rm_flags |= XFS_RMAP_UNWRITTEN; | |
b7f12e53 DW |
217 | last_rmap = &ag_rmaps[agno].ar_last_rmap; |
218 | if (last_rmap->rm_owner == XFS_RMAP_OWN_UNKNOWN) | |
219 | *last_rmap = rmap; | |
2d273771 | 220 | else if (rmaps_are_mergeable(last_rmap, &rmap)) |
b7f12e53 DW |
221 | last_rmap->rm_blockcount += rmap.rm_blockcount; |
222 | else { | |
223 | error = slab_add(ag_rmaps[agno].ar_rmaps, last_rmap); | |
224 | if (error) | |
225 | return error; | |
226 | *last_rmap = rmap; | |
227 | } | |
228 | ||
229 | return error; | |
230 | } | |
231 | ||
232 | /* Finish collecting inode data/attr fork rmaps. */ | |
233 | int | |
2d273771 | 234 | rmap_finish_collecting_fork_recs( |
b7f12e53 DW |
235 | struct xfs_mount *mp, |
236 | xfs_agnumber_t agno) | |
237 | { | |
2d273771 | 238 | if (!rmap_needs_work(mp) || |
b7f12e53 DW |
239 | ag_rmaps[agno].ar_last_rmap.rm_owner == XFS_RMAP_OWN_UNKNOWN) |
240 | return 0; | |
241 | return slab_add(ag_rmaps[agno].ar_rmaps, &ag_rmaps[agno].ar_last_rmap); | |
9e0f480e DW |
242 | } |
243 | ||
1102c155 DW |
244 | /* add a raw rmap; these will be merged later */ |
245 | static int | |
2d273771 | 246 | __rmap_add_raw_rec( |
1102c155 DW |
247 | struct xfs_mount *mp, |
248 | xfs_agnumber_t agno, | |
249 | xfs_agblock_t agbno, | |
250 | xfs_extlen_t len, | |
251 | uint64_t owner, | |
252 | bool is_attr, | |
253 | bool is_bmbt) | |
254 | { | |
255 | struct xfs_rmap_irec rmap; | |
256 | ||
257 | ASSERT(len != 0); | |
258 | rmap.rm_owner = owner; | |
259 | rmap.rm_offset = 0; | |
260 | rmap.rm_flags = 0; | |
261 | if (is_attr) | |
262 | rmap.rm_flags |= XFS_RMAP_ATTR_FORK; | |
263 | if (is_bmbt) | |
264 | rmap.rm_flags |= XFS_RMAP_BMBT_BLOCK; | |
265 | rmap.rm_startblock = agbno; | |
266 | rmap.rm_blockcount = len; | |
267 | return slab_add(ag_rmaps[agno].ar_raw_rmaps, &rmap); | |
268 | } | |
269 | ||
00efc33a DW |
270 | /* |
271 | * Add a reverse mapping for an inode fork's block mapping btree block. | |
272 | */ | |
273 | int | |
2d273771 | 274 | rmap_add_bmbt_rec( |
00efc33a DW |
275 | struct xfs_mount *mp, |
276 | xfs_ino_t ino, | |
277 | int whichfork, | |
278 | xfs_fsblock_t fsbno) | |
279 | { | |
280 | xfs_agnumber_t agno; | |
281 | xfs_agblock_t agbno; | |
282 | ||
2d273771 | 283 | if (!rmap_needs_work(mp)) |
00efc33a DW |
284 | return 0; |
285 | ||
286 | agno = XFS_FSB_TO_AGNO(mp, fsbno); | |
287 | agbno = XFS_FSB_TO_AGBNO(mp, fsbno); | |
288 | ASSERT(agno != NULLAGNUMBER); | |
289 | ASSERT(agno < mp->m_sb.sb_agcount); | |
290 | ASSERT(agbno + 1 <= mp->m_sb.sb_agblocks); | |
291 | ||
2d273771 | 292 | return __rmap_add_raw_rec(mp, agno, agbno, 1, ino, |
00efc33a DW |
293 | whichfork == XFS_ATTR_FORK, true); |
294 | } | |
295 | ||
1102c155 DW |
296 | /* |
297 | * Add a reverse mapping for a per-AG fixed metadata extent. | |
298 | */ | |
299 | int | |
2d273771 | 300 | rmap_add_ag_rec( |
1102c155 DW |
301 | struct xfs_mount *mp, |
302 | xfs_agnumber_t agno, | |
303 | xfs_agblock_t agbno, | |
304 | xfs_extlen_t len, | |
305 | uint64_t owner) | |
306 | { | |
2d273771 | 307 | if (!rmap_needs_work(mp)) |
1102c155 DW |
308 | return 0; |
309 | ||
310 | ASSERT(agno != NULLAGNUMBER); | |
311 | ASSERT(agno < mp->m_sb.sb_agcount); | |
312 | ASSERT(agbno + len <= mp->m_sb.sb_agblocks); | |
313 | ||
2d273771 | 314 | return __rmap_add_raw_rec(mp, agno, agbno, len, owner, false, false); |
1102c155 DW |
315 | } |
316 | ||
317 | /* | |
318 | * Merge adjacent raw rmaps and add them to the main rmap list. | |
319 | */ | |
320 | int | |
2d273771 | 321 | rmap_fold_raw_recs( |
1102c155 DW |
322 | struct xfs_mount *mp, |
323 | xfs_agnumber_t agno) | |
324 | { | |
325 | struct xfs_slab_cursor *cur = NULL; | |
326 | struct xfs_rmap_irec *prev, *rec; | |
327 | size_t old_sz; | |
138ce9ff | 328 | int error = 0; |
1102c155 DW |
329 | |
330 | old_sz = slab_count(ag_rmaps[agno].ar_rmaps); | |
331 | if (slab_count(ag_rmaps[agno].ar_raw_rmaps) == 0) | |
332 | goto no_raw; | |
333 | qsort_slab(ag_rmaps[agno].ar_raw_rmaps, rmap_compare); | |
334 | error = init_slab_cursor(ag_rmaps[agno].ar_raw_rmaps, rmap_compare, | |
335 | &cur); | |
336 | if (error) | |
337 | goto err; | |
338 | ||
339 | prev = pop_slab_cursor(cur); | |
340 | rec = pop_slab_cursor(cur); | |
138ce9ff | 341 | while (prev && rec) { |
2d273771 | 342 | if (rmaps_are_mergeable(prev, rec)) { |
1102c155 DW |
343 | prev->rm_blockcount += rec->rm_blockcount; |
344 | rec = pop_slab_cursor(cur); | |
345 | continue; | |
346 | } | |
347 | error = slab_add(ag_rmaps[agno].ar_rmaps, prev); | |
348 | if (error) | |
349 | goto err; | |
350 | prev = rec; | |
351 | rec = pop_slab_cursor(cur); | |
352 | } | |
353 | if (prev) { | |
354 | error = slab_add(ag_rmaps[agno].ar_rmaps, prev); | |
355 | if (error) | |
356 | goto err; | |
357 | } | |
358 | free_slab(&ag_rmaps[agno].ar_raw_rmaps); | |
359 | error = init_slab(&ag_rmaps[agno].ar_raw_rmaps, | |
360 | sizeof(struct xfs_rmap_irec)); | |
361 | if (error) | |
362 | do_error( | |
363 | _("Insufficient memory while allocating raw metadata reverse mapping slabs.")); | |
364 | no_raw: | |
365 | if (old_sz) | |
366 | qsort_slab(ag_rmaps[agno].ar_rmaps, rmap_compare); | |
367 | err: | |
368 | free_slab_cursor(&cur); | |
369 | return error; | |
370 | } | |
371 | ||
713b6817 DW |
372 | static int |
373 | find_first_zero_bit( | |
374 | __uint64_t mask) | |
375 | { | |
376 | int n; | |
377 | int b = 0; | |
378 | ||
379 | for (n = 0; n < sizeof(mask) * NBBY && (mask & 1); n++, mask >>= 1) | |
380 | b++; | |
381 | ||
382 | return b; | |
383 | } | |
384 | ||
385 | static int | |
386 | popcnt( | |
387 | __uint64_t mask) | |
388 | { | |
389 | int n; | |
390 | int b = 0; | |
391 | ||
392 | if (mask == 0) | |
393 | return 0; | |
394 | ||
395 | for (n = 0; n < sizeof(mask) * NBBY; n++, mask >>= 1) | |
396 | if (mask & 1) | |
397 | b++; | |
398 | ||
399 | return b; | |
400 | } | |
401 | ||
402 | /* | |
403 | * Add an allocation group's fixed metadata to the rmap list. This includes | |
404 | * sb/agi/agf/agfl headers, inode chunks, and the log. | |
405 | */ | |
406 | int | |
2d273771 | 407 | rmap_add_fixed_ag_rec( |
713b6817 DW |
408 | struct xfs_mount *mp, |
409 | xfs_agnumber_t agno) | |
410 | { | |
411 | xfs_fsblock_t fsbno; | |
412 | xfs_agblock_t agbno; | |
413 | ino_tree_node_t *ino_rec; | |
414 | xfs_agino_t agino; | |
415 | int error; | |
416 | int startidx; | |
417 | int nr; | |
418 | ||
2d273771 | 419 | if (!rmap_needs_work(mp)) |
713b6817 DW |
420 | return 0; |
421 | ||
422 | /* sb/agi/agf/agfl headers */ | |
2d273771 | 423 | error = rmap_add_ag_rec(mp, agno, 0, XFS_BNO_BLOCK(mp), |
713b6817 DW |
424 | XFS_RMAP_OWN_FS); |
425 | if (error) | |
426 | goto out; | |
427 | ||
428 | /* inodes */ | |
429 | ino_rec = findfirst_inode_rec(agno); | |
430 | for (; ino_rec != NULL; ino_rec = next_ino_rec(ino_rec)) { | |
431 | if (xfs_sb_version_hassparseinodes(&mp->m_sb)) { | |
432 | startidx = find_first_zero_bit(ino_rec->ir_sparse); | |
433 | nr = XFS_INODES_PER_CHUNK - popcnt(ino_rec->ir_sparse); | |
434 | } else { | |
435 | startidx = 0; | |
436 | nr = XFS_INODES_PER_CHUNK; | |
437 | } | |
438 | nr /= mp->m_sb.sb_inopblock; | |
439 | if (nr == 0) | |
440 | nr = 1; | |
441 | agino = ino_rec->ino_startnum + startidx; | |
442 | agbno = XFS_AGINO_TO_AGBNO(mp, agino); | |
443 | if (XFS_AGINO_TO_OFFSET(mp, agino) == 0) { | |
2d273771 | 444 | error = rmap_add_ag_rec(mp, agno, agbno, nr, |
713b6817 DW |
445 | XFS_RMAP_OWN_INODES); |
446 | if (error) | |
447 | goto out; | |
448 | } | |
449 | } | |
450 | ||
451 | /* log */ | |
452 | fsbno = mp->m_sb.sb_logstart; | |
453 | if (fsbno && XFS_FSB_TO_AGNO(mp, fsbno) == agno) { | |
454 | agbno = XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart); | |
2d273771 | 455 | error = rmap_add_ag_rec(mp, agno, agbno, mp->m_sb.sb_logblocks, |
713b6817 DW |
456 | XFS_RMAP_OWN_LOG); |
457 | if (error) | |
458 | goto out; | |
459 | } | |
460 | out: | |
461 | return error; | |
462 | } | |
463 | ||
62cf990a DW |
464 | /* |
465 | * Copy the per-AG btree reverse-mapping data into the rmapbt. | |
466 | * | |
467 | * At rmapbt reconstruction time, the rmapbt will be populated _only_ with | |
468 | * rmaps for file extents, inode chunks, AG headers, and bmbt blocks. While | |
469 | * building the AG btrees we can record all the blocks allocated for each | |
470 | * btree, but we cannot resolve the conflict between the fact that one has to | |
471 | * finish allocating the space for the rmapbt before building the bnobt and the | |
472 | * fact that allocating blocks for the bnobt requires adding rmapbt entries. | |
473 | * Therefore we record in-core the rmaps for each btree and here use the | |
474 | * libxfs rmap functions to finish building the rmap btree. | |
475 | * | |
476 | * During AGF/AGFL reconstruction in phase 5, rmaps for the AG btrees are | |
477 | * recorded in memory. The rmapbt has not been set up yet, so we need to be | |
478 | * able to "expand" the AGFL without updating the rmapbt. After we've written | |
479 | * out the new AGF header the new rmapbt is available, so this function reads | |
480 | * each AGFL to generate rmap entries. These entries are merged with the AG | |
481 | * btree rmap entries, and then we use libxfs' rmap functions to add them to | |
482 | * the rmapbt, after which it is fully regenerated. | |
483 | */ | |
484 | int | |
2d273771 | 485 | rmap_store_ag_btree_rec( |
62cf990a DW |
486 | struct xfs_mount *mp, |
487 | xfs_agnumber_t agno) | |
488 | { | |
489 | struct xfs_slab_cursor *rm_cur; | |
490 | struct xfs_rmap_irec *rm_rec = NULL; | |
491 | struct xfs_buf *agbp = NULL; | |
492 | struct xfs_buf *agflbp = NULL; | |
493 | struct xfs_trans *tp; | |
494 | struct xfs_trans_res tres = {0}; | |
495 | __be32 *agfl_bno, *b; | |
496 | int error = 0; | |
497 | struct xfs_owner_info oinfo; | |
498 | ||
499 | if (!xfs_sb_version_hasrmapbt(&mp->m_sb)) | |
500 | return 0; | |
501 | ||
502 | /* Release the ar_rmaps; they were put into the rmapbt during p5. */ | |
503 | free_slab(&ag_rmaps[agno].ar_rmaps); | |
504 | error = init_slab(&ag_rmaps[agno].ar_rmaps, | |
505 | sizeof(struct xfs_rmap_irec)); | |
506 | if (error) | |
507 | goto err; | |
508 | ||
509 | /* Add the AGFL blocks to the rmap list */ | |
e2f60652 | 510 | error = -libxfs_trans_read_buf( |
62cf990a DW |
511 | mp, NULL, mp->m_ddev_targp, |
512 | XFS_AG_DADDR(mp, agno, XFS_AGFL_DADDR(mp)), | |
513 | XFS_FSS_TO_BB(mp, 1), 0, &agflbp, &xfs_agfl_buf_ops); | |
514 | if (error) | |
515 | goto err; | |
516 | ||
636f06d8 DW |
517 | /* |
518 | * Sometimes, the blocks at the beginning of the AGFL are there | |
519 | * because we overestimated how many blocks we needed to rebuild | |
520 | * the freespace btrees. ar_flcount records the number of | |
521 | * blocks in this situation. Since those blocks already have an | |
522 | * rmap, we only need to add rmap records for AGFL blocks past | |
523 | * that point in the AGFL because those blocks are a result of a | |
524 | * no-rmap no-shrink freelist fixup that we did earlier. | |
525 | */ | |
62cf990a | 526 | agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); |
636f06d8 DW |
527 | b = agfl_bno + ag_rmaps[agno].ar_flcount; |
528 | while (*b != NULLAGBLOCK && b - agfl_bno < XFS_AGFL_SIZE(mp)) { | |
2d273771 | 529 | error = rmap_add_ag_rec(mp, agno, be32_to_cpu(*b), 1, |
62cf990a DW |
530 | XFS_RMAP_OWN_AG); |
531 | if (error) | |
532 | goto err; | |
533 | b++; | |
534 | } | |
535 | libxfs_putbuf(agflbp); | |
536 | agflbp = NULL; | |
537 | ||
538 | /* Merge all the raw rmaps into the main list */ | |
2d273771 | 539 | error = rmap_fold_raw_recs(mp, agno); |
62cf990a DW |
540 | if (error) |
541 | goto err; | |
542 | ||
543 | /* Create cursors to refcount structures */ | |
544 | error = init_slab_cursor(ag_rmaps[agno].ar_rmaps, rmap_compare, | |
545 | &rm_cur); | |
546 | if (error) | |
547 | goto err; | |
548 | ||
549 | /* Insert rmaps into the btree one at a time */ | |
550 | rm_rec = pop_slab_cursor(rm_cur); | |
551 | while (rm_rec) { | |
552 | error = -libxfs_trans_alloc(mp, &tres, 16, 0, 0, &tp); | |
553 | if (error) | |
554 | goto err_slab; | |
555 | ||
e2f60652 | 556 | error = -libxfs_alloc_read_agf(mp, tp, agno, 0, &agbp); |
62cf990a DW |
557 | if (error) |
558 | goto err_trans; | |
559 | ||
560 | ASSERT(XFS_RMAP_NON_INODE_OWNER(rm_rec->rm_owner)); | |
561 | libxfs_rmap_ag_owner(&oinfo, rm_rec->rm_owner); | |
e2f60652 | 562 | error = -libxfs_rmap_alloc(tp, agbp, agno, rm_rec->rm_startblock, |
62cf990a DW |
563 | rm_rec->rm_blockcount, &oinfo); |
564 | if (error) | |
565 | goto err_trans; | |
566 | ||
567 | error = -libxfs_trans_commit(tp); | |
568 | if (error) | |
569 | goto err_slab; | |
570 | ||
571 | fix_freelist(mp, agno, false); | |
572 | ||
573 | rm_rec = pop_slab_cursor(rm_cur); | |
574 | } | |
575 | ||
576 | free_slab_cursor(&rm_cur); | |
577 | return 0; | |
578 | ||
579 | err_trans: | |
580 | libxfs_trans_cancel(tp); | |
581 | err_slab: | |
582 | free_slab_cursor(&rm_cur); | |
583 | err: | |
584 | if (agflbp) | |
585 | libxfs_putbuf(agflbp); | |
62cf990a DW |
586 | return error; |
587 | } | |
588 | ||
9e0f480e DW |
589 | #ifdef RMAP_DEBUG |
590 | static void | |
2d273771 | 591 | rmap_dump( |
9e0f480e DW |
592 | const char *msg, |
593 | xfs_agnumber_t agno, | |
594 | struct xfs_rmap_irec *rmap) | |
595 | { | |
596 | printf("%s: %p agno=%u pblk=%llu own=%lld lblk=%llu len=%u flags=0x%x\n", | |
597 | msg, rmap, | |
598 | (unsigned int)agno, | |
599 | (unsigned long long)rmap->rm_startblock, | |
600 | (unsigned long long)rmap->rm_owner, | |
601 | (unsigned long long)rmap->rm_offset, | |
602 | (unsigned int)rmap->rm_blockcount, | |
603 | (unsigned int)rmap->rm_flags); | |
604 | } | |
605 | #else | |
2d273771 | 606 | # define rmap_dump(m, a, r) |
9e0f480e | 607 | #endif |
11b9e510 | 608 | |
00f34bca DW |
609 | /* |
610 | * Rebuilding the Reference Count & Reverse Mapping Btrees | |
611 | * | |
612 | * The reference count (refcnt) and reverse mapping (rmap) btrees are | |
613 | * rebuilt during phase 5, like all other AG btrees. Therefore, reverse | |
614 | * mappings must be processed into reference counts at the end of phase | |
615 | * 4, and the rmaps must be recorded during phase 4. There is a need to | |
616 | * access the rmaps in physical block order, but no particular need for | |
617 | * random access, so the slab.c code provides a big logical array | |
618 | * (consisting of smaller slabs) and some inorder iterator functions. | |
619 | * | |
620 | * Once we've recorded all the reverse mappings, we're ready to | |
621 | * translate the rmaps into refcount entries. Imagine the rmap entries | |
622 | * as rectangles representing extents of physical blocks, and that the | |
623 | * rectangles can be laid down to allow them to overlap each other; then | |
624 | * we know that we must emit a refcnt btree entry wherever the amount of | |
625 | * overlap changes, i.e. the emission stimulus is level-triggered: | |
626 | * | |
627 | * - --- | |
628 | * -- ----- ---- --- ------ | |
629 | * -- ---- ----------- ---- --------- | |
630 | * -------------------------------- ----------- | |
631 | * ^ ^ ^^ ^^ ^ ^^ ^^^ ^^^^ ^ ^^ ^ ^ ^ | |
632 | * 2 1 23 21 3 43 234 2123 1 01 2 3 0 | |
633 | * | |
634 | * For our purposes, a rmap is a tuple (startblock, len, fileoff, owner). | |
635 | * | |
636 | * Note that in the actual refcnt btree we don't store the refcount < 2 | |
637 | * cases because the bnobt tells us which blocks are free; single-use | |
638 | * blocks aren't recorded in the bnobt or the refcntbt. If the rmapbt | |
639 | * supports storing multiple entries covering a given block we could | |
640 | * theoretically dispense with the refcntbt and simply count rmaps, but | |
641 | * that's inefficient in the (hot) write path, so we'll take the cost of | |
642 | * the extra tree to save time. Also there's no guarantee that rmap | |
643 | * will be enabled. | |
644 | * | |
645 | * Given an array of rmaps sorted by physical block number, a starting | |
646 | * physical block (sp), a bag to hold rmaps that cover sp, and the next | |
647 | * physical block where the level changes (np), we can reconstruct the | |
648 | * refcount btree as follows: | |
649 | * | |
650 | * While there are still unprocessed rmaps in the array, | |
651 | * - Set sp to the physical block (pblk) of the next unprocessed rmap. | |
652 | * - Add to the bag all rmaps in the array where startblock == sp. | |
653 | * - Set np to the physical block where the bag size will change. This | |
654 | * is the minimum of (the pblk of the next unprocessed rmap) and | |
655 | * (startblock + len of each rmap in the bag). | |
656 | * - Record the bag size as old_bag_size. | |
657 | * | |
658 | * - While the bag isn't empty, | |
659 | * - Remove from the bag all rmaps where startblock + len == np. | |
660 | * - Add to the bag all rmaps in the array where startblock == np. | |
661 | * - If the bag size isn't old_bag_size, store the refcount entry | |
662 | * (sp, np - sp, bag_size) in the refcnt btree. | |
663 | * - If the bag is empty, break out of the inner loop. | |
664 | * - Set old_bag_size to the bag size | |
665 | * - Set sp = np. | |
666 | * - Set np to the physical block where the bag size will change. | |
667 | * This is the minimum of (the pblk of the next unprocessed rmap) | |
668 | * and (startblock + len of each rmap in the bag). | |
669 | * | |
670 | * An implementation detail is that because this processing happens | |
671 | * during phase 4, the refcount entries are stored in an array so that | |
672 | * phase 5 can load them into the refcount btree. The rmaps can be | |
673 | * loaded directly into the rmap btree during phase 5 as well. | |
674 | */ | |
675 | ||
ca8d7d6a DW |
676 | /* |
677 | * Mark all inodes in the reverse-mapping observation stack as requiring the | |
678 | * reflink inode flag, if the stack depth is greater than 1. | |
679 | */ | |
680 | static void | |
681 | mark_inode_rl( | |
682 | struct xfs_mount *mp, | |
683 | struct xfs_bag *rmaps) | |
684 | { | |
685 | xfs_agnumber_t iagno; | |
686 | struct xfs_rmap_irec *rmap; | |
687 | struct ino_tree_node *irec; | |
688 | int off; | |
689 | size_t idx; | |
690 | xfs_agino_t ino; | |
691 | ||
692 | if (bag_count(rmaps) < 2) | |
693 | return; | |
694 | ||
695 | /* Reflink flag accounting */ | |
696 | foreach_bag_ptr(rmaps, idx, rmap) { | |
697 | ASSERT(!XFS_RMAP_NON_INODE_OWNER(rmap->rm_owner)); | |
698 | iagno = XFS_INO_TO_AGNO(mp, rmap->rm_owner); | |
699 | ino = XFS_INO_TO_AGINO(mp, rmap->rm_owner); | |
700 | pthread_mutex_lock(&ag_locks[iagno].lock); | |
701 | irec = find_inode_rec(mp, iagno, ino); | |
702 | off = get_inode_offset(mp, rmap->rm_owner, irec); | |
703 | /* lock here because we might go outside this ag */ | |
704 | set_inode_is_rl(irec, off); | |
705 | pthread_mutex_unlock(&ag_locks[iagno].lock); | |
706 | } | |
707 | } | |
708 | ||
00f34bca DW |
709 | /* |
710 | * Emit a refcount object for refcntbt reconstruction during phase 5. | |
711 | */ | |
712 | #define REFCOUNT_CLAMP(nr) ((nr) > MAXREFCOUNT ? MAXREFCOUNT : (nr)) | |
713 | static void | |
714 | refcount_emit( | |
715 | struct xfs_mount *mp, | |
716 | xfs_agnumber_t agno, | |
717 | xfs_agblock_t agbno, | |
718 | xfs_extlen_t len, | |
719 | size_t nr_rmaps) | |
720 | { | |
721 | struct xfs_refcount_irec rlrec; | |
722 | int error; | |
723 | struct xfs_slab *rlslab; | |
724 | ||
725 | rlslab = ag_rmaps[agno].ar_refcount_items; | |
726 | ASSERT(nr_rmaps > 0); | |
727 | ||
728 | dbg_printf("REFL: agno=%u pblk=%u, len=%u -> refcount=%zu\n", | |
729 | agno, agbno, len, nr_rmaps); | |
730 | rlrec.rc_startblock = agbno; | |
731 | rlrec.rc_blockcount = len; | |
732 | rlrec.rc_refcount = REFCOUNT_CLAMP(nr_rmaps); | |
733 | error = slab_add(rlslab, &rlrec); | |
734 | if (error) | |
735 | do_error( | |
736 | _("Insufficient memory while recreating refcount tree.")); | |
737 | } | |
738 | #undef REFCOUNT_CLAMP | |
739 | ||
740 | /* | |
741 | * Transform a pile of physical block mapping observations into refcount data | |
742 | * for eventual rebuilding of the btrees. | |
743 | */ | |
744 | #define RMAP_END(r) ((r)->rm_startblock + (r)->rm_blockcount) | |
745 | int | |
746 | compute_refcounts( | |
747 | struct xfs_mount *mp, | |
748 | xfs_agnumber_t agno) | |
749 | { | |
750 | struct xfs_bag *stack_top = NULL; | |
751 | struct xfs_slab *rmaps; | |
752 | struct xfs_slab_cursor *rmaps_cur; | |
753 | struct xfs_rmap_irec *array_cur; | |
754 | struct xfs_rmap_irec *rmap; | |
755 | xfs_agblock_t sbno; /* first bno of this rmap set */ | |
756 | xfs_agblock_t cbno; /* first bno of this refcount set */ | |
757 | xfs_agblock_t nbno; /* next bno where rmap set changes */ | |
758 | size_t n, idx; | |
759 | size_t old_stack_nr; | |
760 | int error; | |
761 | ||
762 | if (!xfs_sb_version_hasreflink(&mp->m_sb)) | |
763 | return 0; | |
764 | ||
765 | rmaps = ag_rmaps[agno].ar_rmaps; | |
766 | ||
767 | error = init_slab_cursor(rmaps, rmap_compare, &rmaps_cur); | |
768 | if (error) | |
769 | return error; | |
770 | ||
771 | error = init_bag(&stack_top); | |
772 | if (error) | |
773 | goto err; | |
774 | ||
775 | /* While there are rmaps to be processed... */ | |
776 | n = 0; | |
777 | while (n < slab_count(rmaps)) { | |
778 | array_cur = peek_slab_cursor(rmaps_cur); | |
779 | sbno = cbno = array_cur->rm_startblock; | |
780 | /* Push all rmaps with pblk == sbno onto the stack */ | |
781 | for (; | |
782 | array_cur && array_cur->rm_startblock == sbno; | |
783 | array_cur = peek_slab_cursor(rmaps_cur)) { | |
784 | advance_slab_cursor(rmaps_cur); n++; | |
785 | rmap_dump("push0", agno, array_cur); | |
786 | error = bag_add(stack_top, array_cur); | |
787 | if (error) | |
788 | goto err; | |
789 | } | |
ca8d7d6a | 790 | mark_inode_rl(mp, stack_top); |
00f34bca DW |
791 | |
792 | /* Set nbno to the bno of the next refcount change */ | |
793 | if (n < slab_count(rmaps)) | |
794 | nbno = array_cur->rm_startblock; | |
795 | else | |
796 | nbno = NULLAGBLOCK; | |
797 | foreach_bag_ptr(stack_top, idx, rmap) { | |
798 | nbno = min(nbno, RMAP_END(rmap)); | |
799 | } | |
800 | ||
801 | /* Emit reverse mappings, if needed */ | |
802 | ASSERT(nbno > sbno); | |
803 | old_stack_nr = bag_count(stack_top); | |
804 | ||
805 | /* While stack isn't empty... */ | |
806 | while (bag_count(stack_top)) { | |
807 | /* Pop all rmaps that end at nbno */ | |
808 | foreach_bag_ptr_reverse(stack_top, idx, rmap) { | |
809 | if (RMAP_END(rmap) != nbno) | |
810 | continue; | |
811 | rmap_dump("pop", agno, rmap); | |
812 | error = bag_remove(stack_top, idx); | |
813 | if (error) | |
814 | goto err; | |
815 | } | |
816 | ||
817 | /* Push array items that start at nbno */ | |
818 | for (; | |
819 | array_cur && array_cur->rm_startblock == nbno; | |
820 | array_cur = peek_slab_cursor(rmaps_cur)) { | |
821 | advance_slab_cursor(rmaps_cur); n++; | |
822 | rmap_dump("push1", agno, array_cur); | |
823 | error = bag_add(stack_top, array_cur); | |
824 | if (error) | |
825 | goto err; | |
826 | } | |
ca8d7d6a | 827 | mark_inode_rl(mp, stack_top); |
00f34bca DW |
828 | |
829 | /* Emit refcount if necessary */ | |
830 | ASSERT(nbno > cbno); | |
831 | if (bag_count(stack_top) != old_stack_nr) { | |
832 | if (old_stack_nr > 1) { | |
833 | refcount_emit(mp, agno, cbno, | |
834 | nbno - cbno, | |
835 | old_stack_nr); | |
836 | } | |
837 | cbno = nbno; | |
838 | } | |
839 | ||
840 | /* Stack empty, go find the next rmap */ | |
841 | if (bag_count(stack_top) == 0) | |
842 | break; | |
843 | old_stack_nr = bag_count(stack_top); | |
844 | sbno = nbno; | |
845 | ||
846 | /* Set nbno to the bno of the next refcount change */ | |
847 | if (n < slab_count(rmaps)) | |
848 | nbno = array_cur->rm_startblock; | |
849 | else | |
850 | nbno = NULLAGBLOCK; | |
851 | foreach_bag_ptr(stack_top, idx, rmap) { | |
852 | nbno = min(nbno, RMAP_END(rmap)); | |
853 | } | |
854 | ||
855 | /* Emit reverse mappings, if needed */ | |
856 | ASSERT(nbno > sbno); | |
857 | } | |
858 | } | |
859 | err: | |
860 | free_bag(&stack_top); | |
861 | free_slab_cursor(&rmaps_cur); | |
862 | ||
863 | return error; | |
864 | } | |
865 | #undef RMAP_END | |
866 | ||
11b9e510 DW |
867 | /* |
868 | * Return the number of rmap objects for an AG. | |
869 | */ | |
870 | size_t | |
871 | rmap_record_count( | |
872 | struct xfs_mount *mp, | |
873 | xfs_agnumber_t agno) | |
874 | { | |
875 | return slab_count(ag_rmaps[agno].ar_rmaps); | |
876 | } | |
877 | ||
878 | /* | |
879 | * Return a slab cursor that will return rmap objects in order. | |
880 | */ | |
881 | int | |
2d273771 | 882 | rmap_init_cursor( |
11b9e510 DW |
883 | xfs_agnumber_t agno, |
884 | struct xfs_slab_cursor **cur) | |
885 | { | |
886 | return init_slab_cursor(ag_rmaps[agno].ar_rmaps, rmap_compare, cur); | |
887 | } | |
888 | ||
889 | /* | |
890 | * Disable the refcount btree check. | |
891 | */ | |
892 | void | |
893 | rmap_avoid_check(void) | |
894 | { | |
895 | rmapbt_suspect = true; | |
896 | } | |
897 | ||
898 | /* Look for an rmap in the rmapbt that matches a given rmap. */ | |
899 | static int | |
2d273771 | 900 | rmap_lookup( |
11b9e510 DW |
901 | struct xfs_btree_cur *bt_cur, |
902 | struct xfs_rmap_irec *rm_rec, | |
903 | struct xfs_rmap_irec *tmp, | |
904 | int *have) | |
905 | { | |
906 | int error; | |
907 | ||
908 | /* Use the regular btree retrieval routine. */ | |
909 | error = -libxfs_rmap_lookup_le(bt_cur, rm_rec->rm_startblock, | |
910 | rm_rec->rm_blockcount, | |
911 | rm_rec->rm_owner, rm_rec->rm_offset, | |
912 | rm_rec->rm_flags, have); | |
913 | if (error) | |
914 | return error; | |
915 | if (*have == 0) | |
916 | return error; | |
917 | return -libxfs_rmap_get_rec(bt_cur, tmp, have); | |
918 | } | |
919 | ||
7ba02033 DW |
920 | /* Look for an rmap in the rmapbt that matches a given rmap. */ |
921 | static int | |
922 | rmap_lookup_overlapped( | |
923 | struct xfs_btree_cur *bt_cur, | |
924 | struct xfs_rmap_irec *rm_rec, | |
925 | struct xfs_rmap_irec *tmp, | |
926 | int *have) | |
927 | { | |
928 | /* Have to use our fancy version for overlapped */ | |
929 | return -libxfs_rmap_lookup_le_range(bt_cur, rm_rec->rm_startblock, | |
930 | rm_rec->rm_owner, rm_rec->rm_offset, | |
931 | rm_rec->rm_flags, tmp, have); | |
932 | } | |
933 | ||
11b9e510 DW |
934 | /* Does the btree rmap cover the observed rmap? */ |
935 | #define NEXTP(x) ((x)->rm_startblock + (x)->rm_blockcount) | |
936 | #define NEXTL(x) ((x)->rm_offset + (x)->rm_blockcount) | |
937 | static bool | |
2d273771 | 938 | rmap_is_good( |
11b9e510 DW |
939 | struct xfs_rmap_irec *observed, |
940 | struct xfs_rmap_irec *btree) | |
941 | { | |
942 | /* Can't have mismatches in the flags or the owner. */ | |
943 | if (btree->rm_flags != observed->rm_flags || | |
944 | btree->rm_owner != observed->rm_owner) | |
945 | return false; | |
946 | ||
947 | /* | |
948 | * Btree record can't physically start after the observed | |
949 | * record, nor can it end before the observed record. | |
950 | */ | |
951 | if (btree->rm_startblock > observed->rm_startblock || | |
952 | NEXTP(btree) < NEXTP(observed)) | |
953 | return false; | |
954 | ||
955 | /* If this is metadata or bmbt, we're done. */ | |
956 | if (XFS_RMAP_NON_INODE_OWNER(observed->rm_owner) || | |
957 | (observed->rm_flags & XFS_RMAP_BMBT_BLOCK)) | |
958 | return true; | |
959 | /* | |
960 | * Btree record can't logically start after the observed | |
961 | * record, nor can it end before the observed record. | |
962 | */ | |
963 | if (btree->rm_offset > observed->rm_offset || | |
964 | NEXTL(btree) < NEXTL(observed)) | |
965 | return false; | |
966 | ||
967 | return true; | |
968 | } | |
969 | #undef NEXTP | |
970 | #undef NEXTL | |
971 | ||
972 | /* | |
973 | * Compare the observed reverse mappings against what's in the ag btree. | |
974 | */ | |
975 | int | |
2d273771 | 976 | rmaps_verify_btree( |
11b9e510 DW |
977 | struct xfs_mount *mp, |
978 | xfs_agnumber_t agno) | |
979 | { | |
980 | struct xfs_slab_cursor *rm_cur; | |
981 | struct xfs_btree_cur *bt_cur = NULL; | |
982 | int error; | |
983 | int have; | |
984 | struct xfs_buf *agbp = NULL; | |
985 | struct xfs_rmap_irec *rm_rec; | |
986 | struct xfs_rmap_irec tmp; | |
987 | struct xfs_perag *pag; /* per allocation group data */ | |
988 | ||
989 | if (!xfs_sb_version_hasrmapbt(&mp->m_sb)) | |
990 | return 0; | |
991 | if (rmapbt_suspect) { | |
992 | if (no_modify && agno == 0) | |
993 | do_warn(_("would rebuild corrupt rmap btrees.\n")); | |
994 | return 0; | |
995 | } | |
996 | ||
997 | /* Create cursors to refcount structures */ | |
2d273771 | 998 | error = rmap_init_cursor(agno, &rm_cur); |
11b9e510 DW |
999 | if (error) |
1000 | return error; | |
1001 | ||
1002 | error = -libxfs_alloc_read_agf(mp, NULL, agno, 0, &agbp); | |
1003 | if (error) | |
1004 | goto err; | |
1005 | ||
1006 | /* Leave the per-ag data "uninitialized" since we rewrite it later */ | |
e2f60652 | 1007 | pag = libxfs_perag_get(mp, agno); |
11b9e510 | 1008 | pag->pagf_init = 0; |
e2f60652 | 1009 | libxfs_perag_put(pag); |
11b9e510 DW |
1010 | |
1011 | bt_cur = libxfs_rmapbt_init_cursor(mp, NULL, agbp, agno); | |
1012 | if (!bt_cur) { | |
1013 | error = -ENOMEM; | |
1014 | goto err; | |
1015 | } | |
1016 | ||
1017 | rm_rec = pop_slab_cursor(rm_cur); | |
1018 | while (rm_rec) { | |
2d273771 | 1019 | error = rmap_lookup(bt_cur, rm_rec, &tmp, &have); |
11b9e510 DW |
1020 | if (error) |
1021 | goto err; | |
7ba02033 DW |
1022 | /* |
1023 | * Using the range query is expensive, so only do it if | |
1024 | * the regular lookup doesn't find anything or if it doesn't | |
1025 | * match the observed rmap. | |
1026 | */ | |
1027 | if (xfs_sb_version_hasreflink(&bt_cur->bc_mp->m_sb) && | |
1028 | (!have || !rmap_is_good(rm_rec, &tmp))) { | |
1029 | error = rmap_lookup_overlapped(bt_cur, rm_rec, | |
1030 | &tmp, &have); | |
1031 | if (error) | |
1032 | goto err; | |
1033 | } | |
11b9e510 DW |
1034 | if (!have) { |
1035 | do_warn( | |
1036 | _("Missing reverse-mapping record for (%u/%u) %slen %u owner %"PRId64" \ | |
1037 | %s%soff %"PRIu64"\n"), | |
1038 | agno, rm_rec->rm_startblock, | |
1039 | (rm_rec->rm_flags & XFS_RMAP_UNWRITTEN) ? | |
1040 | _("unwritten ") : "", | |
1041 | rm_rec->rm_blockcount, | |
1042 | rm_rec->rm_owner, | |
1043 | (rm_rec->rm_flags & XFS_RMAP_ATTR_FORK) ? | |
1044 | _("attr ") : "", | |
1045 | (rm_rec->rm_flags & XFS_RMAP_BMBT_BLOCK) ? | |
1046 | _("bmbt ") : "", | |
1047 | rm_rec->rm_offset); | |
1048 | goto next_loop; | |
1049 | } | |
1050 | ||
1051 | /* Compare each refcount observation against the btree's */ | |
2d273771 | 1052 | if (!rmap_is_good(rm_rec, &tmp)) { |
11b9e510 DW |
1053 | do_warn( |
1054 | _("Incorrect reverse-mapping: saw (%u/%u) %slen %u owner %"PRId64" %s%soff \ | |
1055 | %"PRIu64"; should be (%u/%u) %slen %u owner %"PRId64" %s%soff %"PRIu64"\n"), | |
1056 | agno, tmp.rm_startblock, | |
1057 | (tmp.rm_flags & XFS_RMAP_UNWRITTEN) ? | |
1058 | _("unwritten ") : "", | |
1059 | tmp.rm_blockcount, | |
1060 | tmp.rm_owner, | |
1061 | (tmp.rm_flags & XFS_RMAP_ATTR_FORK) ? | |
1062 | _("attr ") : "", | |
1063 | (tmp.rm_flags & XFS_RMAP_BMBT_BLOCK) ? | |
1064 | _("bmbt ") : "", | |
1065 | tmp.rm_offset, | |
1066 | agno, rm_rec->rm_startblock, | |
1067 | (rm_rec->rm_flags & XFS_RMAP_UNWRITTEN) ? | |
1068 | _("unwritten ") : "", | |
1069 | rm_rec->rm_blockcount, | |
1070 | rm_rec->rm_owner, | |
1071 | (rm_rec->rm_flags & XFS_RMAP_ATTR_FORK) ? | |
1072 | _("attr ") : "", | |
1073 | (rm_rec->rm_flags & XFS_RMAP_BMBT_BLOCK) ? | |
1074 | _("bmbt ") : "", | |
1075 | rm_rec->rm_offset); | |
1076 | goto next_loop; | |
1077 | } | |
1078 | next_loop: | |
1079 | rm_rec = pop_slab_cursor(rm_cur); | |
1080 | } | |
1081 | ||
1082 | err: | |
1083 | if (bt_cur) | |
1084 | libxfs_btree_del_cursor(bt_cur, XFS_BTREE_NOERROR); | |
1085 | if (agbp) | |
1086 | libxfs_putbuf(agbp); | |
1087 | free_slab_cursor(&rm_cur); | |
1088 | return 0; | |
1089 | } | |
1090 | ||
1091 | /* | |
1092 | * Compare the key fields of two rmap records -- positive if key1 > key2, | |
1093 | * negative if key1 < key2, and zero if equal. | |
1094 | */ | |
1095 | __int64_t | |
1096 | rmap_diffkeys( | |
1097 | struct xfs_rmap_irec *kp1, | |
1098 | struct xfs_rmap_irec *kp2) | |
1099 | { | |
1100 | __u64 oa; | |
1101 | __u64 ob; | |
1102 | __int64_t d; | |
1103 | struct xfs_rmap_irec tmp; | |
1104 | ||
1105 | tmp = *kp1; | |
1106 | tmp.rm_flags &= ~XFS_RMAP_REC_FLAGS; | |
e2f60652 | 1107 | oa = libxfs_rmap_irec_offset_pack(&tmp); |
11b9e510 DW |
1108 | tmp = *kp2; |
1109 | tmp.rm_flags &= ~XFS_RMAP_REC_FLAGS; | |
e2f60652 | 1110 | ob = libxfs_rmap_irec_offset_pack(&tmp); |
11b9e510 DW |
1111 | |
1112 | d = (__int64_t)kp1->rm_startblock - kp2->rm_startblock; | |
1113 | if (d) | |
1114 | return d; | |
1115 | ||
1116 | if (kp1->rm_owner > kp2->rm_owner) | |
1117 | return 1; | |
1118 | else if (kp2->rm_owner > kp1->rm_owner) | |
1119 | return -1; | |
1120 | ||
1121 | if (oa > ob) | |
1122 | return 1; | |
1123 | else if (ob > oa) | |
1124 | return -1; | |
1125 | return 0; | |
1126 | } | |
1127 | ||
1128 | /* Compute the high key of an rmap record. */ | |
1129 | void | |
1130 | rmap_high_key_from_rec( | |
1131 | struct xfs_rmap_irec *rec, | |
1132 | struct xfs_rmap_irec *key) | |
1133 | { | |
1134 | int adj; | |
1135 | ||
1136 | adj = rec->rm_blockcount - 1; | |
1137 | ||
1138 | key->rm_startblock = rec->rm_startblock + adj; | |
1139 | key->rm_owner = rec->rm_owner; | |
1140 | key->rm_offset = rec->rm_offset; | |
1141 | key->rm_flags = rec->rm_flags & XFS_RMAP_KEY_FLAGS; | |
1142 | if (XFS_RMAP_NON_INODE_OWNER(rec->rm_owner) || | |
1143 | (rec->rm_flags & XFS_RMAP_BMBT_BLOCK)) | |
1144 | return; | |
1145 | key->rm_offset += adj; | |
1146 | } | |
62cf990a | 1147 | |
7e174ec7 DW |
1148 | /* |
1149 | * Record that an inode had the reflink flag set when repair started. The | |
1150 | * inode reflink flag will be adjusted as necessary. | |
1151 | */ | |
1152 | void | |
1153 | record_inode_reflink_flag( | |
1154 | struct xfs_mount *mp, | |
1155 | struct xfs_dinode *dino, | |
1156 | xfs_agnumber_t agno, | |
1157 | xfs_agino_t ino, | |
1158 | xfs_ino_t lino) | |
1159 | { | |
1160 | struct ino_tree_node *irec; | |
1161 | int off; | |
1162 | ||
1163 | ASSERT(XFS_AGINO_TO_INO(mp, agno, ino) == be64_to_cpu(dino->di_ino)); | |
1164 | if (!(be64_to_cpu(dino->di_flags2) & XFS_DIFLAG2_REFLINK)) | |
1165 | return; | |
1166 | irec = find_inode_rec(mp, agno, ino); | |
1167 | off = get_inode_offset(mp, lino, irec); | |
1168 | ASSERT(!inode_was_rl(irec, off)); | |
1169 | set_inode_was_rl(irec, off); | |
1170 | dbg_printf("set was_rl lino=%llu was=0x%llx\n", | |
1171 | (unsigned long long)lino, (unsigned long long)irec->ino_was_rl); | |
1172 | } | |
1173 | ||
ca8d7d6a DW |
1174 | /* |
1175 | * Fix an inode's reflink flag. | |
1176 | */ | |
1177 | static int | |
1178 | fix_inode_reflink_flag( | |
1179 | struct xfs_mount *mp, | |
1180 | xfs_agnumber_t agno, | |
1181 | xfs_agino_t agino, | |
1182 | bool set) | |
1183 | { | |
1184 | struct xfs_dinode *dino; | |
1185 | struct xfs_buf *buf; | |
1186 | ||
1187 | if (set) | |
1188 | do_warn( | |
1189 | _("setting reflink flag on inode %"PRIu64"\n"), | |
1190 | XFS_AGINO_TO_INO(mp, agno, agino)); | |
1191 | else if (!no_modify) /* && !set */ | |
1192 | do_warn( | |
1193 | _("clearing reflink flag on inode %"PRIu64"\n"), | |
1194 | XFS_AGINO_TO_INO(mp, agno, agino)); | |
1195 | if (no_modify) | |
1196 | return 0; | |
1197 | ||
1198 | buf = get_agino_buf(mp, agno, agino, &dino); | |
1199 | if (!buf) | |
1200 | return 1; | |
1201 | ASSERT(XFS_AGINO_TO_INO(mp, agno, agino) == be64_to_cpu(dino->di_ino)); | |
1202 | if (set) | |
1203 | dino->di_flags2 |= cpu_to_be64(XFS_DIFLAG2_REFLINK); | |
1204 | else | |
1205 | dino->di_flags2 &= cpu_to_be64(~XFS_DIFLAG2_REFLINK); | |
1206 | libxfs_dinode_calc_crc(mp, dino); | |
1207 | libxfs_writebuf(buf, 0); | |
1208 | ||
1209 | return 0; | |
1210 | } | |
1211 | ||
1212 | /* | |
1213 | * Fix discrepancies between the state of the inode reflink flag and our | |
1214 | * observations as to whether or not the inode really needs it. | |
1215 | */ | |
1216 | int | |
1217 | fix_inode_reflink_flags( | |
1218 | struct xfs_mount *mp, | |
1219 | xfs_agnumber_t agno) | |
1220 | { | |
1221 | struct ino_tree_node *irec; | |
1222 | int bit; | |
1223 | __uint64_t was; | |
1224 | __uint64_t is; | |
1225 | __uint64_t diff; | |
1226 | __uint64_t mask; | |
1227 | int error = 0; | |
1228 | xfs_agino_t agino; | |
1229 | ||
1230 | /* | |
1231 | * Update the reflink flag for any inode where there's a discrepancy | |
1232 | * between the inode flag and whether or not we found any reflinked | |
1233 | * extents. | |
1234 | */ | |
1235 | for (irec = findfirst_inode_rec(agno); | |
1236 | irec != NULL; | |
1237 | irec = next_ino_rec(irec)) { | |
1238 | ASSERT((irec->ino_was_rl & irec->ir_free) == 0); | |
1239 | ASSERT((irec->ino_is_rl & irec->ir_free) == 0); | |
1240 | was = irec->ino_was_rl; | |
1241 | is = irec->ino_is_rl; | |
1242 | if (was == is) | |
1243 | continue; | |
1244 | diff = was ^ is; | |
1245 | dbg_printf("mismatch ino=%llu was=0x%lx is=0x%lx dif=0x%lx\n", | |
1246 | (unsigned long long)XFS_AGINO_TO_INO(mp, agno, | |
1247 | irec->ino_startnum), | |
1248 | was, is, diff); | |
1249 | ||
1250 | for (bit = 0, mask = 1; bit < 64; bit++, mask <<= 1) { | |
1251 | agino = bit + irec->ino_startnum; | |
1252 | if (!(diff & mask)) | |
1253 | continue; | |
1254 | else if (was & mask) | |
1255 | error = fix_inode_reflink_flag(mp, agno, agino, | |
1256 | false); | |
1257 | else if (is & mask) | |
1258 | error = fix_inode_reflink_flag(mp, agno, agino, | |
1259 | true); | |
1260 | else | |
1261 | ASSERT(0); | |
1262 | if (error) | |
1263 | do_error( | |
1264 | _("Unable to fix reflink flag on inode %"PRIu64".\n"), | |
1265 | XFS_AGINO_TO_INO(mp, agno, agino)); | |
1266 | } | |
1267 | } | |
1268 | ||
1269 | return error; | |
1270 | } | |
1271 | ||
80dbc783 DW |
1272 | /* |
1273 | * Return the number of refcount objects for an AG. | |
1274 | */ | |
1275 | size_t | |
1276 | refcount_record_count( | |
1277 | struct xfs_mount *mp, | |
1278 | xfs_agnumber_t agno) | |
1279 | { | |
1280 | return slab_count(ag_rmaps[agno].ar_refcount_items); | |
1281 | } | |
1282 | ||
1283 | /* | |
1284 | * Return a slab cursor that will return refcount objects in order. | |
1285 | */ | |
1286 | int | |
1287 | init_refcount_cursor( | |
1288 | xfs_agnumber_t agno, | |
1289 | struct xfs_slab_cursor **cur) | |
1290 | { | |
1291 | return init_slab_cursor(ag_rmaps[agno].ar_refcount_items, NULL, cur); | |
1292 | } | |
1293 | ||
1294 | /* | |
1295 | * Disable the refcount btree check. | |
1296 | */ | |
1297 | void | |
1298 | refcount_avoid_check(void) | |
1299 | { | |
1300 | refcbt_suspect = true; | |
1301 | } | |
1302 | ||
1303 | /* | |
1304 | * Compare the observed reference counts against what's in the ag btree. | |
1305 | */ | |
1306 | int | |
1307 | check_refcounts( | |
1308 | struct xfs_mount *mp, | |
1309 | xfs_agnumber_t agno) | |
1310 | { | |
1311 | struct xfs_slab_cursor *rl_cur; | |
1312 | struct xfs_btree_cur *bt_cur = NULL; | |
1313 | int error; | |
1314 | int have; | |
1315 | int i; | |
1316 | struct xfs_buf *agbp = NULL; | |
1317 | struct xfs_refcount_irec *rl_rec; | |
1318 | struct xfs_refcount_irec tmp; | |
1319 | struct xfs_perag *pag; /* per allocation group data */ | |
1320 | ||
1321 | if (!xfs_sb_version_hasreflink(&mp->m_sb)) | |
1322 | return 0; | |
1323 | if (refcbt_suspect) { | |
1324 | if (no_modify && agno == 0) | |
1325 | do_warn(_("would rebuild corrupt refcount btrees.\n")); | |
1326 | return 0; | |
1327 | } | |
1328 | ||
1329 | /* Create cursors to refcount structures */ | |
1330 | error = init_refcount_cursor(agno, &rl_cur); | |
1331 | if (error) | |
1332 | return error; | |
1333 | ||
1334 | error = -libxfs_alloc_read_agf(mp, NULL, agno, 0, &agbp); | |
1335 | if (error) | |
1336 | goto err; | |
1337 | ||
1338 | /* Leave the per-ag data "uninitialized" since we rewrite it later */ | |
1339 | pag = libxfs_perag_get(mp, agno); | |
1340 | pag->pagf_init = 0; | |
1341 | libxfs_perag_put(pag); | |
1342 | ||
1343 | bt_cur = libxfs_refcountbt_init_cursor(mp, NULL, agbp, agno, NULL); | |
1344 | if (!bt_cur) { | |
1345 | error = -ENOMEM; | |
1346 | goto err; | |
1347 | } | |
1348 | ||
1349 | rl_rec = pop_slab_cursor(rl_cur); | |
1350 | while (rl_rec) { | |
1351 | /* Look for a refcount record in the btree */ | |
1352 | error = -libxfs_refcount_lookup_le(bt_cur, | |
1353 | rl_rec->rc_startblock, &have); | |
1354 | if (error) | |
1355 | goto err; | |
1356 | if (!have) { | |
1357 | do_warn( | |
1358 | _("Missing reference count record for (%u/%u) len %u count %u\n"), | |
1359 | agno, rl_rec->rc_startblock, | |
1360 | rl_rec->rc_blockcount, rl_rec->rc_refcount); | |
1361 | goto next_loop; | |
1362 | } | |
1363 | ||
1364 | error = -libxfs_refcount_get_rec(bt_cur, &tmp, &i); | |
1365 | if (error) | |
1366 | goto err; | |
1367 | if (!i) { | |
1368 | do_warn( | |
1369 | _("Missing reference count record for (%u/%u) len %u count %u\n"), | |
1370 | agno, rl_rec->rc_startblock, | |
1371 | rl_rec->rc_blockcount, rl_rec->rc_refcount); | |
1372 | goto next_loop; | |
1373 | } | |
1374 | ||
1375 | /* Compare each refcount observation against the btree's */ | |
1376 | if (tmp.rc_startblock != rl_rec->rc_startblock || | |
1377 | tmp.rc_blockcount < rl_rec->rc_blockcount || | |
1378 | tmp.rc_refcount < rl_rec->rc_refcount) | |
1379 | do_warn( | |
1380 | _("Incorrect reference count: saw (%u/%u) len %u nlinks %u; should be (%u/%u) len %u nlinks %u\n"), | |
1381 | agno, tmp.rc_startblock, tmp.rc_blockcount, | |
1382 | tmp.rc_refcount, agno, rl_rec->rc_startblock, | |
1383 | rl_rec->rc_blockcount, rl_rec->rc_refcount); | |
1384 | next_loop: | |
1385 | rl_rec = pop_slab_cursor(rl_cur); | |
1386 | } | |
1387 | ||
1388 | err: | |
1389 | if (bt_cur) | |
1390 | libxfs_btree_del_cursor(bt_cur, XFS_BTREE_NOERROR); | |
1391 | if (agbp) | |
1392 | libxfs_putbuf(agbp); | |
1393 | free_slab_cursor(&rl_cur); | |
1394 | return 0; | |
1395 | } | |
1396 | ||
62cf990a DW |
1397 | /* |
1398 | * Regenerate the AGFL so that we don't run out of it while rebuilding the | |
1399 | * rmap btree. If skip_rmapbt is true, don't update the rmapbt (most probably | |
1400 | * because we're updating the rmapbt). | |
1401 | */ | |
1402 | void | |
1403 | fix_freelist( | |
1404 | struct xfs_mount *mp, | |
1405 | xfs_agnumber_t agno, | |
1406 | bool skip_rmapbt) | |
1407 | { | |
1408 | xfs_alloc_arg_t args; | |
1409 | xfs_trans_t *tp; | |
1410 | struct xfs_trans_res tres = {0}; | |
1411 | int flags; | |
1412 | int error; | |
1413 | ||
1414 | memset(&args, 0, sizeof(args)); | |
1415 | args.mp = mp; | |
1416 | args.agno = agno; | |
1417 | args.alignment = 1; | |
e2f60652 | 1418 | args.pag = libxfs_perag_get(mp, agno); |
62cf990a DW |
1419 | error = -libxfs_trans_alloc(mp, &tres, |
1420 | libxfs_alloc_min_freelist(mp, args.pag), 0, 0, &tp); | |
1421 | if (error) | |
1422 | do_error(_("failed to fix AGFL on AG %d, error %d\n"), | |
1423 | agno, error); | |
1424 | args.tp = tp; | |
1425 | ||
1426 | /* | |
1427 | * Prior to rmapbt, all we had to do to fix the freelist is "expand" | |
1428 | * the fresh AGFL header from empty to full. That hasn't changed. For | |
1429 | * rmapbt, however, things change a bit. | |
1430 | * | |
1431 | * When we're stuffing the rmapbt with the AG btree rmaps the tree can | |
1432 | * expand, so we need to keep the AGFL well-stocked for the expansion. | |
1433 | * However, this expansion can cause the bnobt/cntbt to shrink, which | |
1434 | * can make the AGFL eligible for shrinking. Shrinking involves | |
1435 | * freeing rmapbt entries, but since we haven't finished loading the | |
1436 | * rmapbt with the btree rmaps it's possible for the remove operation | |
1437 | * to fail. The AGFL block is large enough at this point to absorb any | |
1438 | * blocks freed from the bnobt/cntbt, so we can disable shrinking. | |
1439 | * | |
1440 | * During the initial AGFL regeneration during AGF generation in phase5 | |
1441 | * we must also disable rmapbt modifications because the AGF that | |
1442 | * libxfs reads does not yet point to the new rmapbt. These initial | |
1443 | * AGFL entries are added just prior to adding the AG btree block rmaps | |
1444 | * to the rmapbt. It's ok to pass NOSHRINK here too, since the AGFL is | |
1445 | * empty and cannot shrink. | |
1446 | */ | |
1447 | flags = XFS_ALLOC_FLAG_NOSHRINK; | |
1448 | if (skip_rmapbt) | |
1449 | flags |= XFS_ALLOC_FLAG_NORMAP; | |
e2f60652 DW |
1450 | error = -libxfs_alloc_fix_freelist(&args, flags); |
1451 | libxfs_perag_put(args.pag); | |
62cf990a DW |
1452 | if (error) { |
1453 | do_error(_("failed to fix AGFL on AG %d, error %d\n"), | |
1454 | agno, error); | |
1455 | } | |
1456 | libxfs_trans_commit(tp); | |
1457 | } | |
1458 | ||
1459 | /* | |
1460 | * Remember how many AGFL entries came from excess AG btree allocations and | |
1461 | * therefore already have rmap entries. | |
1462 | */ | |
1463 | void | |
1464 | rmap_store_agflcount( | |
1465 | struct xfs_mount *mp, | |
1466 | xfs_agnumber_t agno, | |
1467 | int count) | |
1468 | { | |
2d273771 | 1469 | if (!rmap_needs_work(mp)) |
62cf990a DW |
1470 | return; |
1471 | ||
1472 | ag_rmaps[agno].ar_flcount = count; | |
1473 | } |