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