]>
Commit | Line | Data |
---|---|---|
959ef981 | 1 | // SPDX-License-Identifier: GPL-2.0+ |
9e0f480e DW |
2 | /* |
3 | * Copyright (C) 2016 Oracle. All Rights Reserved. | |
9e0f480e | 4 | * Author: Darrick J. Wong <darrick.wong@oracle.com> |
9e0f480e DW |
5 | */ |
6 | #include <libxfs.h> | |
7 | #include "btree.h" | |
8 | #include "err_protos.h" | |
9 | #include "libxlog.h" | |
10 | #include "incore.h" | |
11 | #include "globals.h" | |
12 | #include "dinode.h" | |
13 | #include "slab.h" | |
14 | #include "rmap.h" | |
a58400ed | 15 | #include "libfrog/bitmap.h" |
9e0f480e DW |
16 | |
17 | #undef RMAP_DEBUG | |
18 | ||
19 | #ifdef RMAP_DEBUG | |
20 | # define dbg_printf(f, a...) do {printf(f, ## a); fflush(stdout); } while (0) | |
21 | #else | |
22 | # define dbg_printf(f, a...) | |
23 | #endif | |
24 | ||
25 | /* per-AG rmap object anchor */ | |
26 | struct xfs_ag_rmap { | |
27 | struct xfs_slab *ar_rmaps; /* rmap observations, p4 */ | |
1102c155 | 28 | struct xfs_slab *ar_raw_rmaps; /* unmerged rmaps */ |
62cf990a DW |
29 | int ar_flcount; /* agfl entries from leftover */ |
30 | /* agbt allocations */ | |
b7f12e53 | 31 | struct xfs_rmap_irec ar_last_rmap; /* last rmap seen */ |
00f34bca | 32 | struct xfs_slab *ar_refcount_items; /* refcount items, p4-5 */ |
9e0f480e DW |
33 | }; |
34 | ||
35 | static struct xfs_ag_rmap *ag_rmaps; | |
11b9e510 | 36 | static bool rmapbt_suspect; |
80dbc783 | 37 | static bool refcbt_suspect; |
9e0f480e | 38 | |
197c2c6a | 39 | static inline int rmap_compare(const void *a, const void *b) |
9e0f480e | 40 | { |
197c2c6a | 41 | return libxfs_rmap_compare(a, b); |
9e0f480e DW |
42 | } |
43 | ||
44 | /* | |
45 | * Returns true if we must reconstruct either the reference count or reverse | |
46 | * mapping trees. | |
47 | */ | |
48 | bool | |
2d273771 | 49 | rmap_needs_work( |
9e0f480e DW |
50 | struct xfs_mount *mp) |
51 | { | |
00f34bca DW |
52 | return xfs_sb_version_hasreflink(&mp->m_sb) || |
53 | xfs_sb_version_hasrmapbt(&mp->m_sb); | |
9e0f480e DW |
54 | } |
55 | ||
56 | /* | |
57 | * Initialize per-AG reverse map data. | |
58 | */ | |
59 | void | |
2d273771 | 60 | rmaps_init( |
9e0f480e DW |
61 | struct xfs_mount *mp) |
62 | { | |
63 | xfs_agnumber_t i; | |
64 | int error; | |
65 | ||
2d273771 | 66 | if (!rmap_needs_work(mp)) |
9e0f480e DW |
67 | return; |
68 | ||
69 | ag_rmaps = calloc(mp->m_sb.sb_agcount, sizeof(struct xfs_ag_rmap)); | |
70 | if (!ag_rmaps) | |
71 | do_error(_("couldn't allocate per-AG reverse map roots\n")); | |
72 | ||
73 | for (i = 0; i < mp->m_sb.sb_agcount; i++) { | |
74 | error = init_slab(&ag_rmaps[i].ar_rmaps, | |
75 | sizeof(struct xfs_rmap_irec)); | |
76 | if (error) | |
77 | do_error( | |
78 | _("Insufficient memory while allocating reverse mapping slabs.")); | |
1102c155 DW |
79 | error = init_slab(&ag_rmaps[i].ar_raw_rmaps, |
80 | sizeof(struct xfs_rmap_irec)); | |
81 | if (error) | |
82 | do_error( | |
83 | _("Insufficient memory while allocating raw metadata reverse mapping slabs.")); | |
b7f12e53 | 84 | ag_rmaps[i].ar_last_rmap.rm_owner = XFS_RMAP_OWN_UNKNOWN; |
00f34bca DW |
85 | error = init_slab(&ag_rmaps[i].ar_refcount_items, |
86 | sizeof(struct xfs_refcount_irec)); | |
87 | if (error) | |
88 | do_error( | |
89 | _("Insufficient memory while allocating refcount item slabs.")); | |
9e0f480e DW |
90 | } |
91 | } | |
92 | ||
93 | /* | |
94 | * Free the per-AG reverse-mapping data. | |
95 | */ | |
96 | void | |
2d273771 | 97 | rmaps_free( |
9e0f480e DW |
98 | struct xfs_mount *mp) |
99 | { | |
100 | xfs_agnumber_t i; | |
101 | ||
2d273771 | 102 | if (!rmap_needs_work(mp)) |
9e0f480e DW |
103 | return; |
104 | ||
1102c155 | 105 | for (i = 0; i < mp->m_sb.sb_agcount; i++) { |
9e0f480e | 106 | free_slab(&ag_rmaps[i].ar_rmaps); |
1102c155 | 107 | free_slab(&ag_rmaps[i].ar_raw_rmaps); |
00f34bca | 108 | free_slab(&ag_rmaps[i].ar_refcount_items); |
1102c155 | 109 | } |
9e0f480e DW |
110 | free(ag_rmaps); |
111 | ag_rmaps = NULL; | |
112 | } | |
113 | ||
1102c155 DW |
114 | /* |
115 | * Decide if two reverse-mapping records can be merged. | |
116 | */ | |
117 | bool | |
2d273771 | 118 | rmaps_are_mergeable( |
1102c155 DW |
119 | struct xfs_rmap_irec *r1, |
120 | struct xfs_rmap_irec *r2) | |
121 | { | |
122 | if (r1->rm_owner != r2->rm_owner) | |
123 | return false; | |
124 | if (r1->rm_startblock + r1->rm_blockcount != r2->rm_startblock) | |
125 | return false; | |
126 | if ((unsigned long long)r1->rm_blockcount + r2->rm_blockcount > | |
127 | XFS_RMAP_LEN_MAX) | |
128 | return false; | |
129 | if (XFS_RMAP_NON_INODE_OWNER(r2->rm_owner)) | |
130 | return true; | |
131 | /* must be an inode owner below here */ | |
132 | if (r1->rm_flags != r2->rm_flags) | |
133 | return false; | |
134 | if (r1->rm_flags & XFS_RMAP_BMBT_BLOCK) | |
135 | return true; | |
136 | return r1->rm_offset + r1->rm_blockcount == r2->rm_offset; | |
137 | } | |
138 | ||
9e0f480e DW |
139 | /* |
140 | * Add an observation about a block mapping in an inode's data or attribute | |
141 | * fork for later btree reconstruction. | |
142 | */ | |
143 | int | |
2d273771 | 144 | rmap_add_rec( |
9e0f480e DW |
145 | struct xfs_mount *mp, |
146 | xfs_ino_t ino, | |
147 | int whichfork, | |
148 | struct xfs_bmbt_irec *irec) | |
149 | { | |
9e0f480e DW |
150 | struct xfs_rmap_irec rmap; |
151 | xfs_agnumber_t agno; | |
152 | xfs_agblock_t agbno; | |
b7f12e53 DW |
153 | struct xfs_rmap_irec *last_rmap; |
154 | int error = 0; | |
9e0f480e | 155 | |
2d273771 | 156 | if (!rmap_needs_work(mp)) |
9e0f480e DW |
157 | return 0; |
158 | ||
159 | agno = XFS_FSB_TO_AGNO(mp, irec->br_startblock); | |
160 | agbno = XFS_FSB_TO_AGBNO(mp, irec->br_startblock); | |
161 | ASSERT(agno != NULLAGNUMBER); | |
162 | ASSERT(agno < mp->m_sb.sb_agcount); | |
163 | ASSERT(agbno + irec->br_blockcount <= mp->m_sb.sb_agblocks); | |
164 | ASSERT(ino != NULLFSINO); | |
165 | ASSERT(whichfork == XFS_DATA_FORK || whichfork == XFS_ATTR_FORK); | |
166 | ||
9e0f480e DW |
167 | rmap.rm_owner = ino; |
168 | rmap.rm_offset = irec->br_startoff; | |
169 | rmap.rm_flags = 0; | |
170 | if (whichfork == XFS_ATTR_FORK) | |
171 | rmap.rm_flags |= XFS_RMAP_ATTR_FORK; | |
172 | rmap.rm_startblock = agbno; | |
173 | rmap.rm_blockcount = irec->br_blockcount; | |
174 | if (irec->br_state == XFS_EXT_UNWRITTEN) | |
175 | rmap.rm_flags |= XFS_RMAP_UNWRITTEN; | |
b7f12e53 DW |
176 | last_rmap = &ag_rmaps[agno].ar_last_rmap; |
177 | if (last_rmap->rm_owner == XFS_RMAP_OWN_UNKNOWN) | |
178 | *last_rmap = rmap; | |
2d273771 | 179 | else if (rmaps_are_mergeable(last_rmap, &rmap)) |
b7f12e53 DW |
180 | last_rmap->rm_blockcount += rmap.rm_blockcount; |
181 | else { | |
182 | error = slab_add(ag_rmaps[agno].ar_rmaps, last_rmap); | |
183 | if (error) | |
184 | return error; | |
185 | *last_rmap = rmap; | |
186 | } | |
187 | ||
188 | return error; | |
189 | } | |
190 | ||
191 | /* Finish collecting inode data/attr fork rmaps. */ | |
192 | int | |
2d273771 | 193 | rmap_finish_collecting_fork_recs( |
b7f12e53 DW |
194 | struct xfs_mount *mp, |
195 | xfs_agnumber_t agno) | |
196 | { | |
2d273771 | 197 | if (!rmap_needs_work(mp) || |
b7f12e53 DW |
198 | ag_rmaps[agno].ar_last_rmap.rm_owner == XFS_RMAP_OWN_UNKNOWN) |
199 | return 0; | |
200 | return slab_add(ag_rmaps[agno].ar_rmaps, &ag_rmaps[agno].ar_last_rmap); | |
9e0f480e DW |
201 | } |
202 | ||
1102c155 DW |
203 | /* add a raw rmap; these will be merged later */ |
204 | static int | |
2d273771 | 205 | __rmap_add_raw_rec( |
1102c155 DW |
206 | struct xfs_mount *mp, |
207 | xfs_agnumber_t agno, | |
208 | xfs_agblock_t agbno, | |
209 | xfs_extlen_t len, | |
210 | uint64_t owner, | |
211 | bool is_attr, | |
212 | bool is_bmbt) | |
213 | { | |
214 | struct xfs_rmap_irec rmap; | |
215 | ||
216 | ASSERT(len != 0); | |
217 | rmap.rm_owner = owner; | |
218 | rmap.rm_offset = 0; | |
219 | rmap.rm_flags = 0; | |
220 | if (is_attr) | |
221 | rmap.rm_flags |= XFS_RMAP_ATTR_FORK; | |
222 | if (is_bmbt) | |
223 | rmap.rm_flags |= XFS_RMAP_BMBT_BLOCK; | |
224 | rmap.rm_startblock = agbno; | |
225 | rmap.rm_blockcount = len; | |
226 | return slab_add(ag_rmaps[agno].ar_raw_rmaps, &rmap); | |
227 | } | |
228 | ||
00efc33a DW |
229 | /* |
230 | * Add a reverse mapping for an inode fork's block mapping btree block. | |
231 | */ | |
232 | int | |
2d273771 | 233 | rmap_add_bmbt_rec( |
00efc33a DW |
234 | struct xfs_mount *mp, |
235 | xfs_ino_t ino, | |
236 | int whichfork, | |
237 | xfs_fsblock_t fsbno) | |
238 | { | |
239 | xfs_agnumber_t agno; | |
240 | xfs_agblock_t agbno; | |
241 | ||
2d273771 | 242 | if (!rmap_needs_work(mp)) |
00efc33a DW |
243 | return 0; |
244 | ||
245 | agno = XFS_FSB_TO_AGNO(mp, fsbno); | |
246 | agbno = XFS_FSB_TO_AGBNO(mp, fsbno); | |
247 | ASSERT(agno != NULLAGNUMBER); | |
248 | ASSERT(agno < mp->m_sb.sb_agcount); | |
249 | ASSERT(agbno + 1 <= mp->m_sb.sb_agblocks); | |
250 | ||
2d273771 | 251 | return __rmap_add_raw_rec(mp, agno, agbno, 1, ino, |
00efc33a DW |
252 | whichfork == XFS_ATTR_FORK, true); |
253 | } | |
254 | ||
1102c155 DW |
255 | /* |
256 | * Add a reverse mapping for a per-AG fixed metadata extent. | |
257 | */ | |
258 | int | |
2d273771 | 259 | rmap_add_ag_rec( |
1102c155 DW |
260 | struct xfs_mount *mp, |
261 | xfs_agnumber_t agno, | |
262 | xfs_agblock_t agbno, | |
263 | xfs_extlen_t len, | |
264 | uint64_t owner) | |
265 | { | |
2d273771 | 266 | if (!rmap_needs_work(mp)) |
1102c155 DW |
267 | return 0; |
268 | ||
269 | ASSERT(agno != NULLAGNUMBER); | |
270 | ASSERT(agno < mp->m_sb.sb_agcount); | |
271 | ASSERT(agbno + len <= mp->m_sb.sb_agblocks); | |
272 | ||
2d273771 | 273 | return __rmap_add_raw_rec(mp, agno, agbno, len, owner, false, false); |
1102c155 DW |
274 | } |
275 | ||
276 | /* | |
277 | * Merge adjacent raw rmaps and add them to the main rmap list. | |
278 | */ | |
279 | int | |
2d273771 | 280 | rmap_fold_raw_recs( |
1102c155 DW |
281 | struct xfs_mount *mp, |
282 | xfs_agnumber_t agno) | |
283 | { | |
284 | struct xfs_slab_cursor *cur = NULL; | |
285 | struct xfs_rmap_irec *prev, *rec; | |
286 | size_t old_sz; | |
138ce9ff | 287 | int error = 0; |
1102c155 DW |
288 | |
289 | old_sz = slab_count(ag_rmaps[agno].ar_rmaps); | |
290 | if (slab_count(ag_rmaps[agno].ar_raw_rmaps) == 0) | |
291 | goto no_raw; | |
292 | qsort_slab(ag_rmaps[agno].ar_raw_rmaps, rmap_compare); | |
293 | error = init_slab_cursor(ag_rmaps[agno].ar_raw_rmaps, rmap_compare, | |
294 | &cur); | |
295 | if (error) | |
296 | goto err; | |
297 | ||
298 | prev = pop_slab_cursor(cur); | |
299 | rec = pop_slab_cursor(cur); | |
138ce9ff | 300 | while (prev && rec) { |
2d273771 | 301 | if (rmaps_are_mergeable(prev, rec)) { |
1102c155 DW |
302 | prev->rm_blockcount += rec->rm_blockcount; |
303 | rec = pop_slab_cursor(cur); | |
304 | continue; | |
305 | } | |
306 | error = slab_add(ag_rmaps[agno].ar_rmaps, prev); | |
307 | if (error) | |
308 | goto err; | |
309 | prev = rec; | |
310 | rec = pop_slab_cursor(cur); | |
311 | } | |
312 | if (prev) { | |
313 | error = slab_add(ag_rmaps[agno].ar_rmaps, prev); | |
314 | if (error) | |
315 | goto err; | |
316 | } | |
317 | free_slab(&ag_rmaps[agno].ar_raw_rmaps); | |
318 | error = init_slab(&ag_rmaps[agno].ar_raw_rmaps, | |
319 | sizeof(struct xfs_rmap_irec)); | |
320 | if (error) | |
321 | do_error( | |
322 | _("Insufficient memory while allocating raw metadata reverse mapping slabs.")); | |
323 | no_raw: | |
324 | if (old_sz) | |
325 | qsort_slab(ag_rmaps[agno].ar_rmaps, rmap_compare); | |
326 | err: | |
327 | free_slab_cursor(&cur); | |
328 | return error; | |
329 | } | |
330 | ||
713b6817 DW |
331 | static int |
332 | find_first_zero_bit( | |
14f8b681 | 333 | uint64_t mask) |
713b6817 DW |
334 | { |
335 | int n; | |
336 | int b = 0; | |
337 | ||
338 | for (n = 0; n < sizeof(mask) * NBBY && (mask & 1); n++, mask >>= 1) | |
339 | b++; | |
340 | ||
341 | return b; | |
342 | } | |
343 | ||
344 | static int | |
345 | popcnt( | |
14f8b681 | 346 | uint64_t mask) |
713b6817 DW |
347 | { |
348 | int n; | |
349 | int b = 0; | |
350 | ||
351 | if (mask == 0) | |
352 | return 0; | |
353 | ||
354 | for (n = 0; n < sizeof(mask) * NBBY; n++, mask >>= 1) | |
355 | if (mask & 1) | |
356 | b++; | |
357 | ||
358 | return b; | |
359 | } | |
360 | ||
361 | /* | |
362 | * Add an allocation group's fixed metadata to the rmap list. This includes | |
363 | * sb/agi/agf/agfl headers, inode chunks, and the log. | |
364 | */ | |
365 | int | |
2d273771 | 366 | rmap_add_fixed_ag_rec( |
713b6817 DW |
367 | struct xfs_mount *mp, |
368 | xfs_agnumber_t agno) | |
369 | { | |
370 | xfs_fsblock_t fsbno; | |
371 | xfs_agblock_t agbno; | |
372 | ino_tree_node_t *ino_rec; | |
373 | xfs_agino_t agino; | |
374 | int error; | |
375 | int startidx; | |
376 | int nr; | |
377 | ||
2d273771 | 378 | if (!rmap_needs_work(mp)) |
713b6817 DW |
379 | return 0; |
380 | ||
381 | /* sb/agi/agf/agfl headers */ | |
2d273771 | 382 | error = rmap_add_ag_rec(mp, agno, 0, XFS_BNO_BLOCK(mp), |
713b6817 DW |
383 | XFS_RMAP_OWN_FS); |
384 | if (error) | |
385 | goto out; | |
386 | ||
387 | /* inodes */ | |
388 | ino_rec = findfirst_inode_rec(agno); | |
389 | for (; ino_rec != NULL; ino_rec = next_ino_rec(ino_rec)) { | |
390 | if (xfs_sb_version_hassparseinodes(&mp->m_sb)) { | |
391 | startidx = find_first_zero_bit(ino_rec->ir_sparse); | |
392 | nr = XFS_INODES_PER_CHUNK - popcnt(ino_rec->ir_sparse); | |
393 | } else { | |
394 | startidx = 0; | |
395 | nr = XFS_INODES_PER_CHUNK; | |
396 | } | |
397 | nr /= mp->m_sb.sb_inopblock; | |
398 | if (nr == 0) | |
399 | nr = 1; | |
400 | agino = ino_rec->ino_startnum + startidx; | |
401 | agbno = XFS_AGINO_TO_AGBNO(mp, agino); | |
402 | if (XFS_AGINO_TO_OFFSET(mp, agino) == 0) { | |
2d273771 | 403 | error = rmap_add_ag_rec(mp, agno, agbno, nr, |
713b6817 DW |
404 | XFS_RMAP_OWN_INODES); |
405 | if (error) | |
406 | goto out; | |
407 | } | |
408 | } | |
409 | ||
410 | /* log */ | |
411 | fsbno = mp->m_sb.sb_logstart; | |
412 | if (fsbno && XFS_FSB_TO_AGNO(mp, fsbno) == agno) { | |
413 | agbno = XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart); | |
2d273771 | 414 | error = rmap_add_ag_rec(mp, agno, agbno, mp->m_sb.sb_logblocks, |
713b6817 DW |
415 | XFS_RMAP_OWN_LOG); |
416 | if (error) | |
417 | goto out; | |
418 | } | |
419 | out: | |
420 | return error; | |
421 | } | |
422 | ||
62cf990a DW |
423 | /* |
424 | * Copy the per-AG btree reverse-mapping data into the rmapbt. | |
425 | * | |
426 | * At rmapbt reconstruction time, the rmapbt will be populated _only_ with | |
427 | * rmaps for file extents, inode chunks, AG headers, and bmbt blocks. While | |
428 | * building the AG btrees we can record all the blocks allocated for each | |
429 | * btree, but we cannot resolve the conflict between the fact that one has to | |
430 | * finish allocating the space for the rmapbt before building the bnobt and the | |
431 | * fact that allocating blocks for the bnobt requires adding rmapbt entries. | |
432 | * Therefore we record in-core the rmaps for each btree and here use the | |
433 | * libxfs rmap functions to finish building the rmap btree. | |
434 | * | |
435 | * During AGF/AGFL reconstruction in phase 5, rmaps for the AG btrees are | |
436 | * recorded in memory. The rmapbt has not been set up yet, so we need to be | |
437 | * able to "expand" the AGFL without updating the rmapbt. After we've written | |
438 | * out the new AGF header the new rmapbt is available, so this function reads | |
439 | * each AGFL to generate rmap entries. These entries are merged with the AG | |
440 | * btree rmap entries, and then we use libxfs' rmap functions to add them to | |
441 | * the rmapbt, after which it is fully regenerated. | |
442 | */ | |
443 | int | |
2d273771 | 444 | rmap_store_ag_btree_rec( |
62cf990a DW |
445 | struct xfs_mount *mp, |
446 | xfs_agnumber_t agno) | |
447 | { | |
448 | struct xfs_slab_cursor *rm_cur; | |
449 | struct xfs_rmap_irec *rm_rec = NULL; | |
450 | struct xfs_buf *agbp = NULL; | |
451 | struct xfs_buf *agflbp = NULL; | |
452 | struct xfs_trans *tp; | |
62cf990a | 453 | __be32 *agfl_bno, *b; |
1cdc777d DW |
454 | struct xfs_ag_rmap *ag_rmap = &ag_rmaps[agno]; |
455 | struct bitmap *own_ag_bitmap = NULL; | |
62cf990a | 456 | int error = 0; |
62cf990a DW |
457 | |
458 | if (!xfs_sb_version_hasrmapbt(&mp->m_sb)) | |
459 | return 0; | |
460 | ||
461 | /* Release the ar_rmaps; they were put into the rmapbt during p5. */ | |
1cdc777d DW |
462 | free_slab(&ag_rmap->ar_rmaps); |
463 | error = init_slab(&ag_rmap->ar_rmaps, sizeof(struct xfs_rmap_irec)); | |
62cf990a DW |
464 | if (error) |
465 | goto err; | |
466 | ||
467 | /* Add the AGFL blocks to the rmap list */ | |
e2f60652 | 468 | error = -libxfs_trans_read_buf( |
62cf990a DW |
469 | mp, NULL, mp->m_ddev_targp, |
470 | XFS_AG_DADDR(mp, agno, XFS_AGFL_DADDR(mp)), | |
471 | XFS_FSS_TO_BB(mp, 1), 0, &agflbp, &xfs_agfl_buf_ops); | |
472 | if (error) | |
473 | goto err; | |
474 | ||
636f06d8 DW |
475 | /* |
476 | * Sometimes, the blocks at the beginning of the AGFL are there | |
477 | * because we overestimated how many blocks we needed to rebuild | |
478 | * the freespace btrees. ar_flcount records the number of | |
479 | * blocks in this situation. Since those blocks already have an | |
480 | * rmap, we only need to add rmap records for AGFL blocks past | |
481 | * that point in the AGFL because those blocks are a result of a | |
482 | * no-rmap no-shrink freelist fixup that we did earlier. | |
1cdc777d DW |
483 | * |
484 | * However, some blocks end up on the AGFL because the free space | |
485 | * btrees shed blocks as a result of allocating space to fix the | |
486 | * freelist. We already created in-core rmap records for the free | |
487 | * space btree blocks, so we must be careful not to create those | |
488 | * records again. Create a bitmap of already-recorded OWN_AG rmaps. | |
636f06d8 | 489 | */ |
1cdc777d DW |
490 | error = init_slab_cursor(ag_rmap->ar_raw_rmaps, rmap_compare, &rm_cur); |
491 | if (error) | |
492 | goto err; | |
93d69bc7 | 493 | error = -bitmap_alloc(&own_ag_bitmap); |
93ab49dd | 494 | if (error) |
1cdc777d | 495 | goto err_slab; |
1cdc777d DW |
496 | while ((rm_rec = pop_slab_cursor(rm_cur)) != NULL) { |
497 | if (rm_rec->rm_owner != XFS_RMAP_OWN_AG) | |
498 | continue; | |
93d69bc7 | 499 | error = -bitmap_set(own_ag_bitmap, rm_rec->rm_startblock, |
93ab49dd DW |
500 | rm_rec->rm_blockcount); |
501 | if (error) { | |
502 | /* | |
503 | * If this range is already set, then the incore rmap | |
504 | * records for the AG free space btrees overlap and | |
505 | * we're toast because that is not allowed. | |
506 | */ | |
507 | if (error == EEXIST) | |
508 | error = EFSCORRUPTED; | |
1cdc777d DW |
509 | goto err_slab; |
510 | } | |
511 | } | |
512 | free_slab_cursor(&rm_cur); | |
513 | ||
514 | /* Create rmaps for any AGFL blocks that aren't already rmapped. */ | |
62cf990a | 515 | agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp); |
1cdc777d | 516 | b = agfl_bno + ag_rmap->ar_flcount; |
8e1338fb ES |
517 | while (*b != cpu_to_be32(NULLAGBLOCK) && |
518 | b - agfl_bno < libxfs_agfl_size(mp)) { | |
1cdc777d DW |
519 | xfs_agblock_t agbno; |
520 | ||
521 | agbno = be32_to_cpu(*b); | |
522 | if (!bitmap_test(own_ag_bitmap, agbno, 1)) { | |
523 | error = rmap_add_ag_rec(mp, agno, agbno, 1, | |
524 | XFS_RMAP_OWN_AG); | |
525 | if (error) | |
526 | goto err; | |
527 | } | |
62cf990a DW |
528 | b++; |
529 | } | |
530 | libxfs_putbuf(agflbp); | |
531 | agflbp = NULL; | |
1cdc777d | 532 | bitmap_free(&own_ag_bitmap); |
62cf990a DW |
533 | |
534 | /* Merge all the raw rmaps into the main list */ | |
2d273771 | 535 | error = rmap_fold_raw_recs(mp, agno); |
62cf990a DW |
536 | if (error) |
537 | goto err; | |
538 | ||
539 | /* Create cursors to refcount structures */ | |
1cdc777d | 540 | error = init_slab_cursor(ag_rmap->ar_rmaps, rmap_compare, &rm_cur); |
62cf990a DW |
541 | if (error) |
542 | goto err; | |
543 | ||
544 | /* Insert rmaps into the btree one at a time */ | |
545 | rm_rec = pop_slab_cursor(rm_cur); | |
546 | while (rm_rec) { | |
007347e3 DW |
547 | struct xfs_owner_info oinfo = {}; |
548 | ||
225e4bb2 | 549 | error = -libxfs_trans_alloc_rollable(mp, 16, &tp); |
62cf990a DW |
550 | if (error) |
551 | goto err_slab; | |
552 | ||
e2f60652 | 553 | error = -libxfs_alloc_read_agf(mp, tp, agno, 0, &agbp); |
62cf990a DW |
554 | if (error) |
555 | goto err_trans; | |
556 | ||
557 | ASSERT(XFS_RMAP_NON_INODE_OWNER(rm_rec->rm_owner)); | |
007347e3 | 558 | oinfo.oi_owner = rm_rec->rm_owner; |
e2f60652 | 559 | error = -libxfs_rmap_alloc(tp, agbp, agno, rm_rec->rm_startblock, |
62cf990a DW |
560 | rm_rec->rm_blockcount, &oinfo); |
561 | if (error) | |
562 | goto err_trans; | |
563 | ||
564 | error = -libxfs_trans_commit(tp); | |
565 | if (error) | |
566 | goto err_slab; | |
567 | ||
568 | fix_freelist(mp, agno, false); | |
569 | ||
570 | rm_rec = pop_slab_cursor(rm_cur); | |
571 | } | |
572 | ||
573 | free_slab_cursor(&rm_cur); | |
574 | return 0; | |
575 | ||
576 | err_trans: | |
577 | libxfs_trans_cancel(tp); | |
578 | err_slab: | |
579 | free_slab_cursor(&rm_cur); | |
580 | err: | |
581 | if (agflbp) | |
582 | libxfs_putbuf(agflbp); | |
1cdc777d DW |
583 | if (own_ag_bitmap) |
584 | bitmap_free(&own_ag_bitmap); | |
62cf990a DW |
585 | return error; |
586 | } | |
587 | ||
9e0f480e DW |
588 | #ifdef RMAP_DEBUG |
589 | static void | |
2d273771 | 590 | rmap_dump( |
9e0f480e DW |
591 | const char *msg, |
592 | xfs_agnumber_t agno, | |
593 | struct xfs_rmap_irec *rmap) | |
594 | { | |
595 | printf("%s: %p agno=%u pblk=%llu own=%lld lblk=%llu len=%u flags=0x%x\n", | |
596 | msg, rmap, | |
597 | (unsigned int)agno, | |
598 | (unsigned long long)rmap->rm_startblock, | |
599 | (unsigned long long)rmap->rm_owner, | |
600 | (unsigned long long)rmap->rm_offset, | |
601 | (unsigned int)rmap->rm_blockcount, | |
602 | (unsigned int)rmap->rm_flags); | |
603 | } | |
604 | #else | |
2d273771 | 605 | # define rmap_dump(m, a, r) |
9e0f480e | 606 | #endif |
11b9e510 | 607 | |
00f34bca DW |
608 | /* |
609 | * Rebuilding the Reference Count & Reverse Mapping Btrees | |
610 | * | |
611 | * The reference count (refcnt) and reverse mapping (rmap) btrees are | |
612 | * rebuilt during phase 5, like all other AG btrees. Therefore, reverse | |
613 | * mappings must be processed into reference counts at the end of phase | |
614 | * 4, and the rmaps must be recorded during phase 4. There is a need to | |
615 | * access the rmaps in physical block order, but no particular need for | |
616 | * random access, so the slab.c code provides a big logical array | |
617 | * (consisting of smaller slabs) and some inorder iterator functions. | |
618 | * | |
619 | * Once we've recorded all the reverse mappings, we're ready to | |
620 | * translate the rmaps into refcount entries. Imagine the rmap entries | |
621 | * as rectangles representing extents of physical blocks, and that the | |
622 | * rectangles can be laid down to allow them to overlap each other; then | |
623 | * we know that we must emit a refcnt btree entry wherever the amount of | |
624 | * overlap changes, i.e. the emission stimulus is level-triggered: | |
625 | * | |
626 | * - --- | |
627 | * -- ----- ---- --- ------ | |
628 | * -- ---- ----------- ---- --------- | |
629 | * -------------------------------- ----------- | |
630 | * ^ ^ ^^ ^^ ^ ^^ ^^^ ^^^^ ^ ^^ ^ ^ ^ | |
631 | * 2 1 23 21 3 43 234 2123 1 01 2 3 0 | |
632 | * | |
633 | * For our purposes, a rmap is a tuple (startblock, len, fileoff, owner). | |
634 | * | |
635 | * Note that in the actual refcnt btree we don't store the refcount < 2 | |
636 | * cases because the bnobt tells us which blocks are free; single-use | |
637 | * blocks aren't recorded in the bnobt or the refcntbt. If the rmapbt | |
638 | * supports storing multiple entries covering a given block we could | |
639 | * theoretically dispense with the refcntbt and simply count rmaps, but | |
640 | * that's inefficient in the (hot) write path, so we'll take the cost of | |
641 | * the extra tree to save time. Also there's no guarantee that rmap | |
642 | * will be enabled. | |
643 | * | |
644 | * Given an array of rmaps sorted by physical block number, a starting | |
645 | * physical block (sp), a bag to hold rmaps that cover sp, and the next | |
646 | * physical block where the level changes (np), we can reconstruct the | |
647 | * refcount btree as follows: | |
648 | * | |
649 | * While there are still unprocessed rmaps in the array, | |
650 | * - Set sp to the physical block (pblk) of the next unprocessed rmap. | |
651 | * - Add to the bag all rmaps in the array where startblock == sp. | |
652 | * - Set np to the physical block where the bag size will change. This | |
653 | * is the minimum of (the pblk of the next unprocessed rmap) and | |
654 | * (startblock + len of each rmap in the bag). | |
655 | * - Record the bag size as old_bag_size. | |
656 | * | |
657 | * - While the bag isn't empty, | |
658 | * - Remove from the bag all rmaps where startblock + len == np. | |
659 | * - Add to the bag all rmaps in the array where startblock == np. | |
660 | * - If the bag size isn't old_bag_size, store the refcount entry | |
661 | * (sp, np - sp, bag_size) in the refcnt btree. | |
662 | * - If the bag is empty, break out of the inner loop. | |
663 | * - Set old_bag_size to the bag size | |
664 | * - Set sp = np. | |
665 | * - Set np to the physical block where the bag size will change. | |
666 | * This is the minimum of (the pblk of the next unprocessed rmap) | |
667 | * and (startblock + len of each rmap in the bag). | |
668 | * | |
669 | * An implementation detail is that because this processing happens | |
670 | * during phase 4, the refcount entries are stored in an array so that | |
671 | * phase 5 can load them into the refcount btree. The rmaps can be | |
672 | * loaded directly into the rmap btree during phase 5 as well. | |
673 | */ | |
674 | ||
ca8d7d6a DW |
675 | /* |
676 | * Mark all inodes in the reverse-mapping observation stack as requiring the | |
677 | * reflink inode flag, if the stack depth is greater than 1. | |
678 | */ | |
679 | static void | |
680 | mark_inode_rl( | |
681 | struct xfs_mount *mp, | |
682 | struct xfs_bag *rmaps) | |
683 | { | |
684 | xfs_agnumber_t iagno; | |
685 | struct xfs_rmap_irec *rmap; | |
686 | struct ino_tree_node *irec; | |
687 | int off; | |
688 | size_t idx; | |
689 | xfs_agino_t ino; | |
690 | ||
691 | if (bag_count(rmaps) < 2) | |
692 | return; | |
693 | ||
694 | /* Reflink flag accounting */ | |
695 | foreach_bag_ptr(rmaps, idx, rmap) { | |
696 | ASSERT(!XFS_RMAP_NON_INODE_OWNER(rmap->rm_owner)); | |
697 | iagno = XFS_INO_TO_AGNO(mp, rmap->rm_owner); | |
698 | ino = XFS_INO_TO_AGINO(mp, rmap->rm_owner); | |
699 | pthread_mutex_lock(&ag_locks[iagno].lock); | |
700 | irec = find_inode_rec(mp, iagno, ino); | |
701 | off = get_inode_offset(mp, rmap->rm_owner, irec); | |
702 | /* lock here because we might go outside this ag */ | |
703 | set_inode_is_rl(irec, off); | |
704 | pthread_mutex_unlock(&ag_locks[iagno].lock); | |
705 | } | |
706 | } | |
707 | ||
00f34bca DW |
708 | /* |
709 | * Emit a refcount object for refcntbt reconstruction during phase 5. | |
710 | */ | |
711 | #define REFCOUNT_CLAMP(nr) ((nr) > MAXREFCOUNT ? MAXREFCOUNT : (nr)) | |
712 | static void | |
713 | refcount_emit( | |
714 | struct xfs_mount *mp, | |
715 | xfs_agnumber_t agno, | |
716 | xfs_agblock_t agbno, | |
717 | xfs_extlen_t len, | |
718 | size_t nr_rmaps) | |
719 | { | |
720 | struct xfs_refcount_irec rlrec; | |
721 | int error; | |
722 | struct xfs_slab *rlslab; | |
723 | ||
724 | rlslab = ag_rmaps[agno].ar_refcount_items; | |
725 | ASSERT(nr_rmaps > 0); | |
726 | ||
727 | dbg_printf("REFL: agno=%u pblk=%u, len=%u -> refcount=%zu\n", | |
728 | agno, agbno, len, nr_rmaps); | |
729 | rlrec.rc_startblock = agbno; | |
730 | rlrec.rc_blockcount = len; | |
731 | rlrec.rc_refcount = REFCOUNT_CLAMP(nr_rmaps); | |
732 | error = slab_add(rlslab, &rlrec); | |
733 | if (error) | |
734 | do_error( | |
735 | _("Insufficient memory while recreating refcount tree.")); | |
736 | } | |
737 | #undef REFCOUNT_CLAMP | |
738 | ||
739 | /* | |
740 | * Transform a pile of physical block mapping observations into refcount data | |
741 | * for eventual rebuilding of the btrees. | |
742 | */ | |
743 | #define RMAP_END(r) ((r)->rm_startblock + (r)->rm_blockcount) | |
744 | int | |
745 | compute_refcounts( | |
746 | struct xfs_mount *mp, | |
747 | xfs_agnumber_t agno) | |
748 | { | |
749 | struct xfs_bag *stack_top = NULL; | |
750 | struct xfs_slab *rmaps; | |
751 | struct xfs_slab_cursor *rmaps_cur; | |
752 | struct xfs_rmap_irec *array_cur; | |
753 | struct xfs_rmap_irec *rmap; | |
754 | xfs_agblock_t sbno; /* first bno of this rmap set */ | |
755 | xfs_agblock_t cbno; /* first bno of this refcount set */ | |
756 | xfs_agblock_t nbno; /* next bno where rmap set changes */ | |
757 | size_t n, idx; | |
758 | size_t old_stack_nr; | |
759 | int error; | |
760 | ||
761 | if (!xfs_sb_version_hasreflink(&mp->m_sb)) | |
762 | return 0; | |
763 | ||
764 | rmaps = ag_rmaps[agno].ar_rmaps; | |
765 | ||
766 | error = init_slab_cursor(rmaps, rmap_compare, &rmaps_cur); | |
767 | if (error) | |
768 | return error; | |
769 | ||
770 | error = init_bag(&stack_top); | |
771 | if (error) | |
772 | goto err; | |
773 | ||
774 | /* While there are rmaps to be processed... */ | |
775 | n = 0; | |
776 | while (n < slab_count(rmaps)) { | |
777 | array_cur = peek_slab_cursor(rmaps_cur); | |
778 | sbno = cbno = array_cur->rm_startblock; | |
779 | /* Push all rmaps with pblk == sbno onto the stack */ | |
780 | for (; | |
781 | array_cur && array_cur->rm_startblock == sbno; | |
782 | array_cur = peek_slab_cursor(rmaps_cur)) { | |
783 | advance_slab_cursor(rmaps_cur); n++; | |
784 | rmap_dump("push0", agno, array_cur); | |
785 | error = bag_add(stack_top, array_cur); | |
786 | if (error) | |
787 | goto err; | |
788 | } | |
ca8d7d6a | 789 | mark_inode_rl(mp, stack_top); |
00f34bca DW |
790 | |
791 | /* Set nbno to the bno of the next refcount change */ | |
ff14f594 | 792 | if (n < slab_count(rmaps) && array_cur) |
00f34bca DW |
793 | nbno = array_cur->rm_startblock; |
794 | else | |
795 | nbno = NULLAGBLOCK; | |
796 | foreach_bag_ptr(stack_top, idx, rmap) { | |
797 | nbno = min(nbno, RMAP_END(rmap)); | |
798 | } | |
799 | ||
800 | /* Emit reverse mappings, if needed */ | |
801 | ASSERT(nbno > sbno); | |
802 | old_stack_nr = bag_count(stack_top); | |
803 | ||
804 | /* While stack isn't empty... */ | |
805 | while (bag_count(stack_top)) { | |
806 | /* Pop all rmaps that end at nbno */ | |
807 | foreach_bag_ptr_reverse(stack_top, idx, rmap) { | |
808 | if (RMAP_END(rmap) != nbno) | |
809 | continue; | |
810 | rmap_dump("pop", agno, rmap); | |
811 | error = bag_remove(stack_top, idx); | |
812 | if (error) | |
813 | goto err; | |
814 | } | |
815 | ||
816 | /* Push array items that start at nbno */ | |
817 | for (; | |
818 | array_cur && array_cur->rm_startblock == nbno; | |
819 | array_cur = peek_slab_cursor(rmaps_cur)) { | |
820 | advance_slab_cursor(rmaps_cur); n++; | |
821 | rmap_dump("push1", agno, array_cur); | |
822 | error = bag_add(stack_top, array_cur); | |
823 | if (error) | |
824 | goto err; | |
825 | } | |
ca8d7d6a | 826 | mark_inode_rl(mp, stack_top); |
00f34bca DW |
827 | |
828 | /* Emit refcount if necessary */ | |
829 | ASSERT(nbno > cbno); | |
830 | if (bag_count(stack_top) != old_stack_nr) { | |
831 | if (old_stack_nr > 1) { | |
832 | refcount_emit(mp, agno, cbno, | |
833 | nbno - cbno, | |
834 | old_stack_nr); | |
835 | } | |
836 | cbno = nbno; | |
837 | } | |
838 | ||
839 | /* Stack empty, go find the next rmap */ | |
840 | if (bag_count(stack_top) == 0) | |
841 | break; | |
842 | old_stack_nr = bag_count(stack_top); | |
843 | sbno = nbno; | |
844 | ||
845 | /* Set nbno to the bno of the next refcount change */ | |
846 | if (n < slab_count(rmaps)) | |
847 | nbno = array_cur->rm_startblock; | |
848 | else | |
849 | nbno = NULLAGBLOCK; | |
850 | foreach_bag_ptr(stack_top, idx, rmap) { | |
851 | nbno = min(nbno, RMAP_END(rmap)); | |
852 | } | |
853 | ||
854 | /* Emit reverse mappings, if needed */ | |
855 | ASSERT(nbno > sbno); | |
856 | } | |
857 | } | |
858 | err: | |
859 | free_bag(&stack_top); | |
860 | free_slab_cursor(&rmaps_cur); | |
861 | ||
862 | return error; | |
863 | } | |
864 | #undef RMAP_END | |
865 | ||
11b9e510 DW |
866 | /* |
867 | * Return the number of rmap objects for an AG. | |
868 | */ | |
869 | size_t | |
870 | rmap_record_count( | |
871 | struct xfs_mount *mp, | |
872 | xfs_agnumber_t agno) | |
873 | { | |
874 | return slab_count(ag_rmaps[agno].ar_rmaps); | |
875 | } | |
876 | ||
877 | /* | |
878 | * Return a slab cursor that will return rmap objects in order. | |
879 | */ | |
880 | int | |
2d273771 | 881 | rmap_init_cursor( |
11b9e510 DW |
882 | xfs_agnumber_t agno, |
883 | struct xfs_slab_cursor **cur) | |
884 | { | |
885 | return init_slab_cursor(ag_rmaps[agno].ar_rmaps, rmap_compare, cur); | |
886 | } | |
887 | ||
888 | /* | |
889 | * Disable the refcount btree check. | |
890 | */ | |
891 | void | |
892 | rmap_avoid_check(void) | |
893 | { | |
894 | rmapbt_suspect = true; | |
895 | } | |
896 | ||
897 | /* Look for an rmap in the rmapbt that matches a given rmap. */ | |
898 | static int | |
2d273771 | 899 | rmap_lookup( |
11b9e510 DW |
900 | struct xfs_btree_cur *bt_cur, |
901 | struct xfs_rmap_irec *rm_rec, | |
902 | struct xfs_rmap_irec *tmp, | |
903 | int *have) | |
904 | { | |
905 | int error; | |
906 | ||
907 | /* Use the regular btree retrieval routine. */ | |
908 | error = -libxfs_rmap_lookup_le(bt_cur, rm_rec->rm_startblock, | |
909 | rm_rec->rm_blockcount, | |
910 | rm_rec->rm_owner, rm_rec->rm_offset, | |
911 | rm_rec->rm_flags, have); | |
912 | if (error) | |
913 | return error; | |
914 | if (*have == 0) | |
915 | return error; | |
916 | return -libxfs_rmap_get_rec(bt_cur, tmp, have); | |
917 | } | |
918 | ||
7ba02033 DW |
919 | /* Look for an rmap in the rmapbt that matches a given rmap. */ |
920 | static int | |
921 | rmap_lookup_overlapped( | |
922 | struct xfs_btree_cur *bt_cur, | |
923 | struct xfs_rmap_irec *rm_rec, | |
924 | struct xfs_rmap_irec *tmp, | |
925 | int *have) | |
926 | { | |
927 | /* Have to use our fancy version for overlapped */ | |
928 | return -libxfs_rmap_lookup_le_range(bt_cur, rm_rec->rm_startblock, | |
929 | rm_rec->rm_owner, rm_rec->rm_offset, | |
930 | rm_rec->rm_flags, tmp, have); | |
931 | } | |
932 | ||
11b9e510 DW |
933 | /* Does the btree rmap cover the observed rmap? */ |
934 | #define NEXTP(x) ((x)->rm_startblock + (x)->rm_blockcount) | |
935 | #define NEXTL(x) ((x)->rm_offset + (x)->rm_blockcount) | |
936 | static bool | |
2d273771 | 937 | rmap_is_good( |
11b9e510 DW |
938 | struct xfs_rmap_irec *observed, |
939 | struct xfs_rmap_irec *btree) | |
940 | { | |
941 | /* Can't have mismatches in the flags or the owner. */ | |
942 | if (btree->rm_flags != observed->rm_flags || | |
943 | btree->rm_owner != observed->rm_owner) | |
944 | return false; | |
945 | ||
946 | /* | |
947 | * Btree record can't physically start after the observed | |
948 | * record, nor can it end before the observed record. | |
949 | */ | |
950 | if (btree->rm_startblock > observed->rm_startblock || | |
951 | NEXTP(btree) < NEXTP(observed)) | |
952 | return false; | |
953 | ||
954 | /* If this is metadata or bmbt, we're done. */ | |
955 | if (XFS_RMAP_NON_INODE_OWNER(observed->rm_owner) || | |
956 | (observed->rm_flags & XFS_RMAP_BMBT_BLOCK)) | |
957 | return true; | |
958 | /* | |
959 | * Btree record can't logically start after the observed | |
960 | * record, nor can it end before the observed record. | |
961 | */ | |
962 | if (btree->rm_offset > observed->rm_offset || | |
963 | NEXTL(btree) < NEXTL(observed)) | |
964 | return false; | |
965 | ||
966 | return true; | |
967 | } | |
968 | #undef NEXTP | |
969 | #undef NEXTL | |
970 | ||
971 | /* | |
972 | * Compare the observed reverse mappings against what's in the ag btree. | |
973 | */ | |
974 | int | |
2d273771 | 975 | rmaps_verify_btree( |
11b9e510 DW |
976 | struct xfs_mount *mp, |
977 | xfs_agnumber_t agno) | |
978 | { | |
979 | struct xfs_slab_cursor *rm_cur; | |
980 | struct xfs_btree_cur *bt_cur = NULL; | |
981 | int error; | |
982 | int have; | |
983 | struct xfs_buf *agbp = NULL; | |
984 | struct xfs_rmap_irec *rm_rec; | |
985 | struct xfs_rmap_irec tmp; | |
986 | struct xfs_perag *pag; /* per allocation group data */ | |
987 | ||
988 | if (!xfs_sb_version_hasrmapbt(&mp->m_sb)) | |
989 | return 0; | |
990 | if (rmapbt_suspect) { | |
991 | if (no_modify && agno == 0) | |
992 | do_warn(_("would rebuild corrupt rmap btrees.\n")); | |
993 | return 0; | |
994 | } | |
995 | ||
996 | /* Create cursors to refcount structures */ | |
2d273771 | 997 | error = rmap_init_cursor(agno, &rm_cur); |
11b9e510 DW |
998 | if (error) |
999 | return error; | |
1000 | ||
1001 | error = -libxfs_alloc_read_agf(mp, NULL, agno, 0, &agbp); | |
1002 | if (error) | |
1003 | goto err; | |
1004 | ||
1005 | /* Leave the per-ag data "uninitialized" since we rewrite it later */ | |
e2f60652 | 1006 | pag = libxfs_perag_get(mp, agno); |
11b9e510 | 1007 | pag->pagf_init = 0; |
e2f60652 | 1008 | libxfs_perag_put(pag); |
11b9e510 DW |
1009 | |
1010 | bt_cur = libxfs_rmapbt_init_cursor(mp, NULL, agbp, agno); | |
1011 | if (!bt_cur) { | |
1012 | error = -ENOMEM; | |
1013 | goto err; | |
1014 | } | |
1015 | ||
1016 | rm_rec = pop_slab_cursor(rm_cur); | |
1017 | while (rm_rec) { | |
2d273771 | 1018 | error = rmap_lookup(bt_cur, rm_rec, &tmp, &have); |
11b9e510 DW |
1019 | if (error) |
1020 | goto err; | |
7ba02033 DW |
1021 | /* |
1022 | * Using the range query is expensive, so only do it if | |
1023 | * the regular lookup doesn't find anything or if it doesn't | |
1024 | * match the observed rmap. | |
1025 | */ | |
1026 | if (xfs_sb_version_hasreflink(&bt_cur->bc_mp->m_sb) && | |
1027 | (!have || !rmap_is_good(rm_rec, &tmp))) { | |
1028 | error = rmap_lookup_overlapped(bt_cur, rm_rec, | |
1029 | &tmp, &have); | |
1030 | if (error) | |
1031 | goto err; | |
1032 | } | |
11b9e510 DW |
1033 | if (!have) { |
1034 | do_warn( | |
1035 | _("Missing reverse-mapping record for (%u/%u) %slen %u owner %"PRId64" \ | |
1036 | %s%soff %"PRIu64"\n"), | |
1037 | agno, rm_rec->rm_startblock, | |
1038 | (rm_rec->rm_flags & XFS_RMAP_UNWRITTEN) ? | |
1039 | _("unwritten ") : "", | |
1040 | rm_rec->rm_blockcount, | |
1041 | rm_rec->rm_owner, | |
1042 | (rm_rec->rm_flags & XFS_RMAP_ATTR_FORK) ? | |
1043 | _("attr ") : "", | |
1044 | (rm_rec->rm_flags & XFS_RMAP_BMBT_BLOCK) ? | |
1045 | _("bmbt ") : "", | |
1046 | rm_rec->rm_offset); | |
1047 | goto next_loop; | |
1048 | } | |
1049 | ||
1050 | /* Compare each refcount observation against the btree's */ | |
2d273771 | 1051 | if (!rmap_is_good(rm_rec, &tmp)) { |
11b9e510 DW |
1052 | do_warn( |
1053 | _("Incorrect reverse-mapping: saw (%u/%u) %slen %u owner %"PRId64" %s%soff \ | |
1054 | %"PRIu64"; should be (%u/%u) %slen %u owner %"PRId64" %s%soff %"PRIu64"\n"), | |
1055 | agno, tmp.rm_startblock, | |
1056 | (tmp.rm_flags & XFS_RMAP_UNWRITTEN) ? | |
1057 | _("unwritten ") : "", | |
1058 | tmp.rm_blockcount, | |
1059 | tmp.rm_owner, | |
1060 | (tmp.rm_flags & XFS_RMAP_ATTR_FORK) ? | |
1061 | _("attr ") : "", | |
1062 | (tmp.rm_flags & XFS_RMAP_BMBT_BLOCK) ? | |
1063 | _("bmbt ") : "", | |
1064 | tmp.rm_offset, | |
1065 | agno, rm_rec->rm_startblock, | |
1066 | (rm_rec->rm_flags & XFS_RMAP_UNWRITTEN) ? | |
1067 | _("unwritten ") : "", | |
1068 | rm_rec->rm_blockcount, | |
1069 | rm_rec->rm_owner, | |
1070 | (rm_rec->rm_flags & XFS_RMAP_ATTR_FORK) ? | |
1071 | _("attr ") : "", | |
1072 | (rm_rec->rm_flags & XFS_RMAP_BMBT_BLOCK) ? | |
1073 | _("bmbt ") : "", | |
1074 | rm_rec->rm_offset); | |
1075 | goto next_loop; | |
1076 | } | |
1077 | next_loop: | |
1078 | rm_rec = pop_slab_cursor(rm_cur); | |
1079 | } | |
1080 | ||
1081 | err: | |
1082 | if (bt_cur) | |
1083 | libxfs_btree_del_cursor(bt_cur, XFS_BTREE_NOERROR); | |
1084 | if (agbp) | |
1085 | libxfs_putbuf(agbp); | |
1086 | free_slab_cursor(&rm_cur); | |
1087 | return 0; | |
1088 | } | |
1089 | ||
1090 | /* | |
1091 | * Compare the key fields of two rmap records -- positive if key1 > key2, | |
1092 | * negative if key1 < key2, and zero if equal. | |
1093 | */ | |
14f8b681 | 1094 | int64_t |
11b9e510 DW |
1095 | rmap_diffkeys( |
1096 | struct xfs_rmap_irec *kp1, | |
1097 | struct xfs_rmap_irec *kp2) | |
1098 | { | |
1099 | __u64 oa; | |
1100 | __u64 ob; | |
14f8b681 | 1101 | int64_t d; |
11b9e510 DW |
1102 | struct xfs_rmap_irec tmp; |
1103 | ||
1104 | tmp = *kp1; | |
1105 | tmp.rm_flags &= ~XFS_RMAP_REC_FLAGS; | |
e2f60652 | 1106 | oa = libxfs_rmap_irec_offset_pack(&tmp); |
11b9e510 DW |
1107 | tmp = *kp2; |
1108 | tmp.rm_flags &= ~XFS_RMAP_REC_FLAGS; | |
e2f60652 | 1109 | ob = libxfs_rmap_irec_offset_pack(&tmp); |
11b9e510 | 1110 | |
14f8b681 | 1111 | d = (int64_t)kp1->rm_startblock - kp2->rm_startblock; |
11b9e510 DW |
1112 | if (d) |
1113 | return d; | |
1114 | ||
1115 | if (kp1->rm_owner > kp2->rm_owner) | |
1116 | return 1; | |
1117 | else if (kp2->rm_owner > kp1->rm_owner) | |
1118 | return -1; | |
1119 | ||
1120 | if (oa > ob) | |
1121 | return 1; | |
1122 | else if (ob > oa) | |
1123 | return -1; | |
1124 | return 0; | |
1125 | } | |
1126 | ||
1127 | /* Compute the high key of an rmap record. */ | |
1128 | void | |
1129 | rmap_high_key_from_rec( | |
1130 | struct xfs_rmap_irec *rec, | |
1131 | struct xfs_rmap_irec *key) | |
1132 | { | |
1133 | int adj; | |
1134 | ||
1135 | adj = rec->rm_blockcount - 1; | |
1136 | ||
1137 | key->rm_startblock = rec->rm_startblock + adj; | |
1138 | key->rm_owner = rec->rm_owner; | |
1139 | key->rm_offset = rec->rm_offset; | |
1140 | key->rm_flags = rec->rm_flags & XFS_RMAP_KEY_FLAGS; | |
1141 | if (XFS_RMAP_NON_INODE_OWNER(rec->rm_owner) || | |
1142 | (rec->rm_flags & XFS_RMAP_BMBT_BLOCK)) | |
1143 | return; | |
1144 | key->rm_offset += adj; | |
1145 | } | |
62cf990a | 1146 | |
7e174ec7 DW |
1147 | /* |
1148 | * Record that an inode had the reflink flag set when repair started. The | |
1149 | * inode reflink flag will be adjusted as necessary. | |
1150 | */ | |
1151 | void | |
1152 | record_inode_reflink_flag( | |
1153 | struct xfs_mount *mp, | |
1154 | struct xfs_dinode *dino, | |
1155 | xfs_agnumber_t agno, | |
1156 | xfs_agino_t ino, | |
1157 | xfs_ino_t lino) | |
1158 | { | |
1159 | struct ino_tree_node *irec; | |
1160 | int off; | |
1161 | ||
1162 | ASSERT(XFS_AGINO_TO_INO(mp, agno, ino) == be64_to_cpu(dino->di_ino)); | |
1163 | if (!(be64_to_cpu(dino->di_flags2) & XFS_DIFLAG2_REFLINK)) | |
1164 | return; | |
1165 | irec = find_inode_rec(mp, agno, ino); | |
1166 | off = get_inode_offset(mp, lino, irec); | |
1167 | ASSERT(!inode_was_rl(irec, off)); | |
1168 | set_inode_was_rl(irec, off); | |
1169 | dbg_printf("set was_rl lino=%llu was=0x%llx\n", | |
1170 | (unsigned long long)lino, (unsigned long long)irec->ino_was_rl); | |
1171 | } | |
1172 | ||
904a5020 DW |
1173 | /* |
1174 | * Inform the user that we're clearing the reflink flag on an inode that | |
1175 | * doesn't actually share any blocks. This is an optimization (the kernel | |
1176 | * skips refcount checks for non-reflink files) and not a corruption repair, | |
1177 | * so we don't need to log every time we clear a flag unless verbose mode is | |
1178 | * enabled. | |
1179 | */ | |
1180 | static void | |
1181 | warn_clearing_reflink( | |
1182 | xfs_ino_t ino) | |
1183 | { | |
1184 | static bool warned = false; | |
1185 | static pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER; | |
1186 | ||
1187 | if (verbose) { | |
1188 | do_warn(_("clearing reflink flag on inode %"PRIu64"\n"), ino); | |
1189 | return; | |
1190 | } | |
1191 | ||
1192 | if (warned) | |
1193 | return; | |
1194 | ||
1195 | pthread_mutex_lock(&lock); | |
1196 | if (!warned) { | |
1197 | do_warn(_("clearing reflink flag on inodes when possible\n")); | |
1198 | warned = true; | |
1199 | } | |
1200 | pthread_mutex_unlock(&lock); | |
1201 | } | |
1202 | ||
ca8d7d6a DW |
1203 | /* |
1204 | * Fix an inode's reflink flag. | |
1205 | */ | |
1206 | static int | |
1207 | fix_inode_reflink_flag( | |
1208 | struct xfs_mount *mp, | |
1209 | xfs_agnumber_t agno, | |
1210 | xfs_agino_t agino, | |
1211 | bool set) | |
1212 | { | |
1213 | struct xfs_dinode *dino; | |
1214 | struct xfs_buf *buf; | |
1215 | ||
1216 | if (set) | |
1217 | do_warn( | |
1218 | _("setting reflink flag on inode %"PRIu64"\n"), | |
1219 | XFS_AGINO_TO_INO(mp, agno, agino)); | |
1220 | else if (!no_modify) /* && !set */ | |
904a5020 | 1221 | warn_clearing_reflink(XFS_AGINO_TO_INO(mp, agno, agino)); |
ca8d7d6a DW |
1222 | if (no_modify) |
1223 | return 0; | |
1224 | ||
1225 | buf = get_agino_buf(mp, agno, agino, &dino); | |
1226 | if (!buf) | |
1227 | return 1; | |
1228 | ASSERT(XFS_AGINO_TO_INO(mp, agno, agino) == be64_to_cpu(dino->di_ino)); | |
1229 | if (set) | |
1230 | dino->di_flags2 |= cpu_to_be64(XFS_DIFLAG2_REFLINK); | |
1231 | else | |
1232 | dino->di_flags2 &= cpu_to_be64(~XFS_DIFLAG2_REFLINK); | |
1233 | libxfs_dinode_calc_crc(mp, dino); | |
1234 | libxfs_writebuf(buf, 0); | |
1235 | ||
1236 | return 0; | |
1237 | } | |
1238 | ||
1239 | /* | |
1240 | * Fix discrepancies between the state of the inode reflink flag and our | |
1241 | * observations as to whether or not the inode really needs it. | |
1242 | */ | |
1243 | int | |
1244 | fix_inode_reflink_flags( | |
1245 | struct xfs_mount *mp, | |
1246 | xfs_agnumber_t agno) | |
1247 | { | |
1248 | struct ino_tree_node *irec; | |
1249 | int bit; | |
14f8b681 DW |
1250 | uint64_t was; |
1251 | uint64_t is; | |
1252 | uint64_t diff; | |
1253 | uint64_t mask; | |
ca8d7d6a DW |
1254 | int error = 0; |
1255 | xfs_agino_t agino; | |
1256 | ||
1257 | /* | |
1258 | * Update the reflink flag for any inode where there's a discrepancy | |
1259 | * between the inode flag and whether or not we found any reflinked | |
1260 | * extents. | |
1261 | */ | |
1262 | for (irec = findfirst_inode_rec(agno); | |
1263 | irec != NULL; | |
1264 | irec = next_ino_rec(irec)) { | |
1265 | ASSERT((irec->ino_was_rl & irec->ir_free) == 0); | |
1266 | ASSERT((irec->ino_is_rl & irec->ir_free) == 0); | |
1267 | was = irec->ino_was_rl; | |
1268 | is = irec->ino_is_rl; | |
1269 | if (was == is) | |
1270 | continue; | |
1271 | diff = was ^ is; | |
1272 | dbg_printf("mismatch ino=%llu was=0x%lx is=0x%lx dif=0x%lx\n", | |
1273 | (unsigned long long)XFS_AGINO_TO_INO(mp, agno, | |
1274 | irec->ino_startnum), | |
1275 | was, is, diff); | |
1276 | ||
1277 | for (bit = 0, mask = 1; bit < 64; bit++, mask <<= 1) { | |
1278 | agino = bit + irec->ino_startnum; | |
1279 | if (!(diff & mask)) | |
1280 | continue; | |
1281 | else if (was & mask) | |
1282 | error = fix_inode_reflink_flag(mp, agno, agino, | |
1283 | false); | |
1284 | else if (is & mask) | |
1285 | error = fix_inode_reflink_flag(mp, agno, agino, | |
1286 | true); | |
1287 | else | |
1288 | ASSERT(0); | |
1289 | if (error) | |
1290 | do_error( | |
1291 | _("Unable to fix reflink flag on inode %"PRIu64".\n"), | |
1292 | XFS_AGINO_TO_INO(mp, agno, agino)); | |
1293 | } | |
1294 | } | |
1295 | ||
1296 | return error; | |
1297 | } | |
1298 | ||
80dbc783 DW |
1299 | /* |
1300 | * Return the number of refcount objects for an AG. | |
1301 | */ | |
1302 | size_t | |
1303 | refcount_record_count( | |
1304 | struct xfs_mount *mp, | |
1305 | xfs_agnumber_t agno) | |
1306 | { | |
1307 | return slab_count(ag_rmaps[agno].ar_refcount_items); | |
1308 | } | |
1309 | ||
1310 | /* | |
1311 | * Return a slab cursor that will return refcount objects in order. | |
1312 | */ | |
1313 | int | |
1314 | init_refcount_cursor( | |
1315 | xfs_agnumber_t agno, | |
1316 | struct xfs_slab_cursor **cur) | |
1317 | { | |
1318 | return init_slab_cursor(ag_rmaps[agno].ar_refcount_items, NULL, cur); | |
1319 | } | |
1320 | ||
1321 | /* | |
1322 | * Disable the refcount btree check. | |
1323 | */ | |
1324 | void | |
1325 | refcount_avoid_check(void) | |
1326 | { | |
1327 | refcbt_suspect = true; | |
1328 | } | |
1329 | ||
1330 | /* | |
1331 | * Compare the observed reference counts against what's in the ag btree. | |
1332 | */ | |
1333 | int | |
1334 | check_refcounts( | |
1335 | struct xfs_mount *mp, | |
1336 | xfs_agnumber_t agno) | |
1337 | { | |
1338 | struct xfs_slab_cursor *rl_cur; | |
1339 | struct xfs_btree_cur *bt_cur = NULL; | |
1340 | int error; | |
1341 | int have; | |
1342 | int i; | |
1343 | struct xfs_buf *agbp = NULL; | |
1344 | struct xfs_refcount_irec *rl_rec; | |
1345 | struct xfs_refcount_irec tmp; | |
1346 | struct xfs_perag *pag; /* per allocation group data */ | |
1347 | ||
1348 | if (!xfs_sb_version_hasreflink(&mp->m_sb)) | |
1349 | return 0; | |
1350 | if (refcbt_suspect) { | |
1351 | if (no_modify && agno == 0) | |
1352 | do_warn(_("would rebuild corrupt refcount btrees.\n")); | |
1353 | return 0; | |
1354 | } | |
1355 | ||
1356 | /* Create cursors to refcount structures */ | |
1357 | error = init_refcount_cursor(agno, &rl_cur); | |
1358 | if (error) | |
1359 | return error; | |
1360 | ||
1361 | error = -libxfs_alloc_read_agf(mp, NULL, agno, 0, &agbp); | |
1362 | if (error) | |
1363 | goto err; | |
1364 | ||
1365 | /* Leave the per-ag data "uninitialized" since we rewrite it later */ | |
1366 | pag = libxfs_perag_get(mp, agno); | |
1367 | pag->pagf_init = 0; | |
1368 | libxfs_perag_put(pag); | |
1369 | ||
5ff5ced0 | 1370 | bt_cur = libxfs_refcountbt_init_cursor(mp, NULL, agbp, agno); |
80dbc783 DW |
1371 | if (!bt_cur) { |
1372 | error = -ENOMEM; | |
1373 | goto err; | |
1374 | } | |
1375 | ||
1376 | rl_rec = pop_slab_cursor(rl_cur); | |
1377 | while (rl_rec) { | |
1378 | /* Look for a refcount record in the btree */ | |
1379 | error = -libxfs_refcount_lookup_le(bt_cur, | |
1380 | rl_rec->rc_startblock, &have); | |
1381 | if (error) | |
1382 | goto err; | |
1383 | if (!have) { | |
1384 | do_warn( | |
1385 | _("Missing reference count record for (%u/%u) len %u count %u\n"), | |
1386 | agno, rl_rec->rc_startblock, | |
1387 | rl_rec->rc_blockcount, rl_rec->rc_refcount); | |
1388 | goto next_loop; | |
1389 | } | |
1390 | ||
1391 | error = -libxfs_refcount_get_rec(bt_cur, &tmp, &i); | |
1392 | if (error) | |
1393 | goto err; | |
1394 | if (!i) { | |
1395 | do_warn( | |
1396 | _("Missing reference count record for (%u/%u) len %u count %u\n"), | |
1397 | agno, rl_rec->rc_startblock, | |
1398 | rl_rec->rc_blockcount, rl_rec->rc_refcount); | |
1399 | goto next_loop; | |
1400 | } | |
1401 | ||
1402 | /* Compare each refcount observation against the btree's */ | |
1403 | if (tmp.rc_startblock != rl_rec->rc_startblock || | |
1404 | tmp.rc_blockcount < rl_rec->rc_blockcount || | |
1405 | tmp.rc_refcount < rl_rec->rc_refcount) | |
1406 | do_warn( | |
1407 | _("Incorrect reference count: saw (%u/%u) len %u nlinks %u; should be (%u/%u) len %u nlinks %u\n"), | |
1408 | agno, tmp.rc_startblock, tmp.rc_blockcount, | |
1409 | tmp.rc_refcount, agno, rl_rec->rc_startblock, | |
1410 | rl_rec->rc_blockcount, rl_rec->rc_refcount); | |
1411 | next_loop: | |
1412 | rl_rec = pop_slab_cursor(rl_cur); | |
1413 | } | |
1414 | ||
1415 | err: | |
1416 | if (bt_cur) | |
8743fab4 DW |
1417 | libxfs_btree_del_cursor(bt_cur, error ? XFS_BTREE_ERROR : |
1418 | XFS_BTREE_NOERROR); | |
80dbc783 DW |
1419 | if (agbp) |
1420 | libxfs_putbuf(agbp); | |
1421 | free_slab_cursor(&rl_cur); | |
1422 | return 0; | |
1423 | } | |
1424 | ||
62cf990a DW |
1425 | /* |
1426 | * Regenerate the AGFL so that we don't run out of it while rebuilding the | |
1427 | * rmap btree. If skip_rmapbt is true, don't update the rmapbt (most probably | |
1428 | * because we're updating the rmapbt). | |
1429 | */ | |
1430 | void | |
1431 | fix_freelist( | |
1432 | struct xfs_mount *mp, | |
1433 | xfs_agnumber_t agno, | |
1434 | bool skip_rmapbt) | |
1435 | { | |
1436 | xfs_alloc_arg_t args; | |
1437 | xfs_trans_t *tp; | |
62cf990a DW |
1438 | int flags; |
1439 | int error; | |
1440 | ||
1441 | memset(&args, 0, sizeof(args)); | |
1442 | args.mp = mp; | |
1443 | args.agno = agno; | |
1444 | args.alignment = 1; | |
e2f60652 | 1445 | args.pag = libxfs_perag_get(mp, agno); |
f9c9fd94 | 1446 | error = -libxfs_trans_alloc_rollable(mp, 0, &tp); |
62cf990a DW |
1447 | if (error) |
1448 | do_error(_("failed to fix AGFL on AG %d, error %d\n"), | |
1449 | agno, error); | |
1450 | args.tp = tp; | |
1451 | ||
1452 | /* | |
1453 | * Prior to rmapbt, all we had to do to fix the freelist is "expand" | |
1454 | * the fresh AGFL header from empty to full. That hasn't changed. For | |
1455 | * rmapbt, however, things change a bit. | |
1456 | * | |
1457 | * When we're stuffing the rmapbt with the AG btree rmaps the tree can | |
1458 | * expand, so we need to keep the AGFL well-stocked for the expansion. | |
1459 | * However, this expansion can cause the bnobt/cntbt to shrink, which | |
1460 | * can make the AGFL eligible for shrinking. Shrinking involves | |
1461 | * freeing rmapbt entries, but since we haven't finished loading the | |
1462 | * rmapbt with the btree rmaps it's possible for the remove operation | |
1463 | * to fail. The AGFL block is large enough at this point to absorb any | |
1464 | * blocks freed from the bnobt/cntbt, so we can disable shrinking. | |
1465 | * | |
1466 | * During the initial AGFL regeneration during AGF generation in phase5 | |
1467 | * we must also disable rmapbt modifications because the AGF that | |
1468 | * libxfs reads does not yet point to the new rmapbt. These initial | |
1469 | * AGFL entries are added just prior to adding the AG btree block rmaps | |
1470 | * to the rmapbt. It's ok to pass NOSHRINK here too, since the AGFL is | |
1471 | * empty and cannot shrink. | |
1472 | */ | |
1473 | flags = XFS_ALLOC_FLAG_NOSHRINK; | |
1474 | if (skip_rmapbt) | |
1475 | flags |= XFS_ALLOC_FLAG_NORMAP; | |
e2f60652 DW |
1476 | error = -libxfs_alloc_fix_freelist(&args, flags); |
1477 | libxfs_perag_put(args.pag); | |
62cf990a DW |
1478 | if (error) { |
1479 | do_error(_("failed to fix AGFL on AG %d, error %d\n"), | |
1480 | agno, error); | |
1481 | } | |
f2279d8d DW |
1482 | error = -libxfs_trans_commit(tp); |
1483 | if (error) | |
1484 | do_error(_("%s: commit failed, error %d\n"), __func__, error); | |
62cf990a DW |
1485 | } |
1486 | ||
1487 | /* | |
1488 | * Remember how many AGFL entries came from excess AG btree allocations and | |
1489 | * therefore already have rmap entries. | |
1490 | */ | |
1491 | void | |
1492 | rmap_store_agflcount( | |
1493 | struct xfs_mount *mp, | |
1494 | xfs_agnumber_t agno, | |
1495 | int count) | |
1496 | { | |
2d273771 | 1497 | if (!rmap_needs_work(mp)) |
62cf990a DW |
1498 | return; |
1499 | ||
1500 | ag_rmaps[agno].ar_flcount = count; | |
1501 | } |