xref: /linux/fs/xfs/libxfs/xfs_ag.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4  * Copyright (c) 2018 Red Hat, Inc.
5  * All rights reserved.
6  */
7 
8 #include "xfs.h"
9 #include "xfs_fs.h"
10 #include "xfs_shared.h"
11 #include "xfs_format.h"
12 #include "xfs_trans_resv.h"
13 #include "xfs_bit.h"
14 #include "xfs_sb.h"
15 #include "xfs_mount.h"
16 #include "xfs_btree.h"
17 #include "xfs_alloc_btree.h"
18 #include "xfs_rmap_btree.h"
19 #include "xfs_alloc.h"
20 #include "xfs_ialloc.h"
21 #include "xfs_rmap.h"
22 #include "xfs_ag.h"
23 #include "xfs_ag_resv.h"
24 #include "xfs_health.h"
25 #include "xfs_error.h"
26 #include "xfs_bmap.h"
27 #include "xfs_defer.h"
28 #include "xfs_log_format.h"
29 #include "xfs_trans.h"
30 #include "xfs_trace.h"
31 #include "xfs_inode.h"
32 #include "xfs_icache.h"
33 #include "xfs_group.h"
34 
35 /*
36  * xfs_initialize_perag_data
37  *
38  * Read in each per-ag structure so we can count up the number of
39  * allocated inodes, free inodes and used filesystem blocks as this
40  * information is no longer persistent in the superblock. Once we have
41  * this information, write it into the in-core superblock structure.
42  */
43 int
44 xfs_initialize_perag_data(
45 	struct xfs_mount	*mp,
46 	xfs_agnumber_t		agcount)
47 {
48 	xfs_agnumber_t		index;
49 	struct xfs_perag	*pag;
50 	struct xfs_sb		*sbp = &mp->m_sb;
51 	uint64_t		ifree = 0;
52 	uint64_t		ialloc = 0;
53 	uint64_t		bfree = 0;
54 	uint64_t		bfreelst = 0;
55 	uint64_t		btree = 0;
56 	uint64_t		fdblocks;
57 	int			error = 0;
58 
59 	for (index = 0; index < agcount; index++) {
60 		/*
61 		 * Read the AGF and AGI buffers to populate the per-ag
62 		 * structures for us.
63 		 */
64 		pag = xfs_perag_get(mp, index);
65 		error = xfs_alloc_read_agf(pag, NULL, 0, NULL);
66 		if (!error)
67 			error = xfs_ialloc_read_agi(pag, NULL, 0, NULL);
68 		if (error) {
69 			xfs_perag_put(pag);
70 			return error;
71 		}
72 
73 		ifree += pag->pagi_freecount;
74 		ialloc += pag->pagi_count;
75 		bfree += pag->pagf_freeblks;
76 		bfreelst += pag->pagf_flcount;
77 		btree += pag->pagf_btreeblks;
78 		xfs_perag_put(pag);
79 	}
80 	fdblocks = bfree + bfreelst + btree;
81 
82 	/*
83 	 * If the new summary counts are obviously incorrect, fail the
84 	 * mount operation because that implies the AGFs are also corrupt.
85 	 * Clear FS_COUNTERS so that we don't unmount with a dirty log, which
86 	 * will prevent xfs_repair from fixing anything.
87 	 */
88 	if (fdblocks > sbp->sb_dblocks || ifree > ialloc) {
89 		xfs_alert(mp, "AGF corruption. Please run xfs_repair.");
90 		xfs_fs_mark_sick(mp, XFS_SICK_FS_COUNTERS);
91 		error = -EFSCORRUPTED;
92 		goto out;
93 	}
94 
95 	/* Overwrite incore superblock counters with just-read data */
96 	spin_lock(&mp->m_sb_lock);
97 	sbp->sb_ifree = ifree;
98 	sbp->sb_icount = ialloc;
99 	sbp->sb_fdblocks = fdblocks;
100 	spin_unlock(&mp->m_sb_lock);
101 
102 	xfs_reinit_percpu_counters(mp);
103 out:
104 	xfs_fs_mark_healthy(mp, XFS_SICK_FS_COUNTERS);
105 	return error;
106 }
107 
108 static void
109 xfs_perag_uninit(
110 	struct xfs_group	*xg)
111 {
112 #ifdef __KERNEL__
113 	struct xfs_perag	*pag = to_perag(xg);
114 
115 	cancel_delayed_work_sync(&pag->pag_blockgc_work);
116 	xfs_buf_cache_destroy(&pag->pag_bcache);
117 #endif
118 }
119 
120 /*
121  * Free up the per-ag resources  within the specified AG range.
122  */
123 void
124 xfs_free_perag_range(
125 	struct xfs_mount	*mp,
126 	xfs_agnumber_t		first_agno,
127 	xfs_agnumber_t		end_agno)
128 
129 {
130 	xfs_agnumber_t		agno;
131 
132 	for (agno = first_agno; agno < end_agno; agno++)
133 		xfs_group_free(mp, agno, XG_TYPE_AG, xfs_perag_uninit);
134 }
135 
136 /* Find the size of the AG, in blocks. */
137 static xfs_agblock_t
138 __xfs_ag_block_count(
139 	struct xfs_mount	*mp,
140 	xfs_agnumber_t		agno,
141 	xfs_agnumber_t		agcount,
142 	xfs_rfsblock_t		dblocks)
143 {
144 	ASSERT(agno < agcount);
145 
146 	if (agno < agcount - 1)
147 		return mp->m_sb.sb_agblocks;
148 	return dblocks - (agno * mp->m_sb.sb_agblocks);
149 }
150 
151 xfs_agblock_t
152 xfs_ag_block_count(
153 	struct xfs_mount	*mp,
154 	xfs_agnumber_t		agno)
155 {
156 	return __xfs_ag_block_count(mp, agno, mp->m_sb.sb_agcount,
157 			mp->m_sb.sb_dblocks);
158 }
159 
160 /* Calculate the first and last possible inode number in an AG. */
161 static void
162 __xfs_agino_range(
163 	struct xfs_mount	*mp,
164 	xfs_agblock_t		eoag,
165 	xfs_agino_t		*first,
166 	xfs_agino_t		*last)
167 {
168 	xfs_agblock_t		bno;
169 
170 	/*
171 	 * Calculate the first inode, which will be in the first
172 	 * cluster-aligned block after the AGFL.
173 	 */
174 	bno = round_up(XFS_AGFL_BLOCK(mp) + 1, M_IGEO(mp)->cluster_align);
175 	*first = XFS_AGB_TO_AGINO(mp, bno);
176 
177 	/*
178 	 * Calculate the last inode, which will be at the end of the
179 	 * last (aligned) cluster that can be allocated in the AG.
180 	 */
181 	bno = round_down(eoag, M_IGEO(mp)->cluster_align);
182 	*last = XFS_AGB_TO_AGINO(mp, bno) - 1;
183 }
184 
185 void
186 xfs_agino_range(
187 	struct xfs_mount	*mp,
188 	xfs_agnumber_t		agno,
189 	xfs_agino_t		*first,
190 	xfs_agino_t		*last)
191 {
192 	return __xfs_agino_range(mp, xfs_ag_block_count(mp, agno), first, last);
193 }
194 
195 /*
196  * Update the perag of the previous tail AG if it has been changed during
197  * recovery (i.e. recovery of a growfs).
198  */
199 int
200 xfs_update_last_ag_size(
201 	struct xfs_mount	*mp,
202 	xfs_agnumber_t		prev_agcount)
203 {
204 	struct xfs_perag	*pag = xfs_perag_grab(mp, prev_agcount - 1);
205 
206 	if (!pag)
207 		return -EFSCORRUPTED;
208 	pag_group(pag)->xg_block_count = __xfs_ag_block_count(mp,
209 			prev_agcount - 1, mp->m_sb.sb_agcount,
210 			mp->m_sb.sb_dblocks);
211 	__xfs_agino_range(mp, pag_group(pag)->xg_block_count, &pag->agino_min,
212 			&pag->agino_max);
213 	xfs_perag_rele(pag);
214 	return 0;
215 }
216 
217 static int
218 xfs_perag_alloc(
219 	struct xfs_mount	*mp,
220 	xfs_agnumber_t		index,
221 	xfs_agnumber_t		agcount,
222 	xfs_rfsblock_t		dblocks)
223 {
224 	struct xfs_perag	*pag;
225 	int			error;
226 
227 	pag = kzalloc(sizeof(*pag), GFP_KERNEL);
228 	if (!pag)
229 		return -ENOMEM;
230 
231 #ifdef __KERNEL__
232 	/* Place kernel structure only init below this point. */
233 	spin_lock_init(&pag->pag_ici_lock);
234 	INIT_DELAYED_WORK(&pag->pag_blockgc_work, xfs_blockgc_worker);
235 	INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
236 #endif /* __KERNEL__ */
237 
238 	error = xfs_buf_cache_init(&pag->pag_bcache);
239 	if (error)
240 		goto out_free_perag;
241 
242 	/*
243 	 * Pre-calculated geometry
244 	 */
245 	pag_group(pag)->xg_block_count = __xfs_ag_block_count(mp, index, agcount,
246 				dblocks);
247 	pag_group(pag)->xg_min_gbno = XFS_AGFL_BLOCK(mp) + 1;
248 	__xfs_agino_range(mp, pag_group(pag)->xg_block_count, &pag->agino_min,
249 			&pag->agino_max);
250 
251 	error = xfs_group_insert(mp, pag_group(pag), index, XG_TYPE_AG);
252 	if (error)
253 		goto out_buf_cache_destroy;
254 
255 	return 0;
256 
257 out_buf_cache_destroy:
258 	xfs_buf_cache_destroy(&pag->pag_bcache);
259 out_free_perag:
260 	kfree(pag);
261 	return error;
262 }
263 
264 int
265 xfs_initialize_perag(
266 	struct xfs_mount	*mp,
267 	xfs_agnumber_t		orig_agcount,
268 	xfs_agnumber_t		new_agcount,
269 	xfs_rfsblock_t		dblocks,
270 	xfs_agnumber_t		*maxagi)
271 {
272 	xfs_agnumber_t		index;
273 	int			error;
274 
275 	if (orig_agcount >= new_agcount)
276 		return 0;
277 
278 	for (index = orig_agcount; index < new_agcount; index++) {
279 		error = xfs_perag_alloc(mp, index, new_agcount, dblocks);
280 		if (error)
281 			goto out_unwind_new_pags;
282 	}
283 
284 	*maxagi = xfs_set_inode_alloc(mp, new_agcount);
285 	mp->m_ag_prealloc_blocks = xfs_prealloc_blocks(mp);
286 	return 0;
287 
288 out_unwind_new_pags:
289 	xfs_free_perag_range(mp, orig_agcount, index);
290 	return error;
291 }
292 
293 static int
294 xfs_get_aghdr_buf(
295 	struct xfs_mount	*mp,
296 	xfs_daddr_t		blkno,
297 	size_t			numblks,
298 	struct xfs_buf		**bpp,
299 	const struct xfs_buf_ops *ops)
300 {
301 	struct xfs_buf		*bp;
302 	int			error;
303 
304 	error = xfs_buf_get_uncached(mp->m_ddev_targp, numblks, 0, &bp);
305 	if (error)
306 		return error;
307 
308 	bp->b_maps[0].bm_bn = blkno;
309 	bp->b_ops = ops;
310 
311 	*bpp = bp;
312 	return 0;
313 }
314 
315 /*
316  * Generic btree root block init function
317  */
318 static void
319 xfs_btroot_init(
320 	struct xfs_mount	*mp,
321 	struct xfs_buf		*bp,
322 	struct aghdr_init_data	*id)
323 {
324 	xfs_btree_init_buf(mp, bp, id->bc_ops, 0, 0, id->agno);
325 }
326 
327 /* Finish initializing a free space btree. */
328 static void
329 xfs_freesp_init_recs(
330 	struct xfs_mount	*mp,
331 	struct xfs_buf		*bp,
332 	struct aghdr_init_data	*id)
333 {
334 	struct xfs_alloc_rec	*arec;
335 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
336 
337 	arec = XFS_ALLOC_REC_ADDR(mp, XFS_BUF_TO_BLOCK(bp), 1);
338 	arec->ar_startblock = cpu_to_be32(mp->m_ag_prealloc_blocks);
339 
340 	if (xfs_ag_contains_log(mp, id->agno)) {
341 		struct xfs_alloc_rec	*nrec;
342 		xfs_agblock_t		start = XFS_FSB_TO_AGBNO(mp,
343 							mp->m_sb.sb_logstart);
344 
345 		ASSERT(start >= mp->m_ag_prealloc_blocks);
346 		if (start != mp->m_ag_prealloc_blocks) {
347 			/*
348 			 * Modify first record to pad stripe align of log and
349 			 * bump the record count.
350 			 */
351 			arec->ar_blockcount = cpu_to_be32(start -
352 						mp->m_ag_prealloc_blocks);
353 			be16_add_cpu(&block->bb_numrecs, 1);
354 			nrec = arec + 1;
355 
356 			/*
357 			 * Insert second record at start of internal log
358 			 * which then gets trimmed.
359 			 */
360 			nrec->ar_startblock = cpu_to_be32(
361 					be32_to_cpu(arec->ar_startblock) +
362 					be32_to_cpu(arec->ar_blockcount));
363 			arec = nrec;
364 		}
365 		/*
366 		 * Change record start to after the internal log
367 		 */
368 		be32_add_cpu(&arec->ar_startblock, mp->m_sb.sb_logblocks);
369 	}
370 
371 	/*
372 	 * Calculate the block count of this record; if it is nonzero,
373 	 * increment the record count.
374 	 */
375 	arec->ar_blockcount = cpu_to_be32(id->agsize -
376 					  be32_to_cpu(arec->ar_startblock));
377 	if (arec->ar_blockcount)
378 		be16_add_cpu(&block->bb_numrecs, 1);
379 }
380 
381 /*
382  * bnobt/cntbt btree root block init functions
383  */
384 static void
385 xfs_bnoroot_init(
386 	struct xfs_mount	*mp,
387 	struct xfs_buf		*bp,
388 	struct aghdr_init_data	*id)
389 {
390 	xfs_btree_init_buf(mp, bp, id->bc_ops, 0, 0, id->agno);
391 	xfs_freesp_init_recs(mp, bp, id);
392 }
393 
394 /*
395  * Reverse map root block init
396  */
397 static void
398 xfs_rmaproot_init(
399 	struct xfs_mount	*mp,
400 	struct xfs_buf		*bp,
401 	struct aghdr_init_data	*id)
402 {
403 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
404 	struct xfs_rmap_rec	*rrec;
405 
406 	xfs_btree_init_buf(mp, bp, id->bc_ops, 0, 4, id->agno);
407 
408 	/*
409 	 * mark the AG header regions as static metadata The BNO
410 	 * btree block is the first block after the headers, so
411 	 * it's location defines the size of region the static
412 	 * metadata consumes.
413 	 *
414 	 * Note: unlike mkfs, we never have to account for log
415 	 * space when growing the data regions
416 	 */
417 	rrec = XFS_RMAP_REC_ADDR(block, 1);
418 	rrec->rm_startblock = 0;
419 	rrec->rm_blockcount = cpu_to_be32(XFS_BNO_BLOCK(mp));
420 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_FS);
421 	rrec->rm_offset = 0;
422 
423 	/* account freespace btree root blocks */
424 	rrec = XFS_RMAP_REC_ADDR(block, 2);
425 	rrec->rm_startblock = cpu_to_be32(XFS_BNO_BLOCK(mp));
426 	rrec->rm_blockcount = cpu_to_be32(2);
427 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_AG);
428 	rrec->rm_offset = 0;
429 
430 	/* account inode btree root blocks */
431 	rrec = XFS_RMAP_REC_ADDR(block, 3);
432 	rrec->rm_startblock = cpu_to_be32(XFS_IBT_BLOCK(mp));
433 	rrec->rm_blockcount = cpu_to_be32(XFS_RMAP_BLOCK(mp) -
434 					  XFS_IBT_BLOCK(mp));
435 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_INOBT);
436 	rrec->rm_offset = 0;
437 
438 	/* account for rmap btree root */
439 	rrec = XFS_RMAP_REC_ADDR(block, 4);
440 	rrec->rm_startblock = cpu_to_be32(XFS_RMAP_BLOCK(mp));
441 	rrec->rm_blockcount = cpu_to_be32(1);
442 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_AG);
443 	rrec->rm_offset = 0;
444 
445 	/* account for refc btree root */
446 	if (xfs_has_reflink(mp)) {
447 		rrec = XFS_RMAP_REC_ADDR(block, 5);
448 		rrec->rm_startblock = cpu_to_be32(xfs_refc_block(mp));
449 		rrec->rm_blockcount = cpu_to_be32(1);
450 		rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_REFC);
451 		rrec->rm_offset = 0;
452 		be16_add_cpu(&block->bb_numrecs, 1);
453 	}
454 
455 	/* account for the log space */
456 	if (xfs_ag_contains_log(mp, id->agno)) {
457 		rrec = XFS_RMAP_REC_ADDR(block,
458 				be16_to_cpu(block->bb_numrecs) + 1);
459 		rrec->rm_startblock = cpu_to_be32(
460 				XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart));
461 		rrec->rm_blockcount = cpu_to_be32(mp->m_sb.sb_logblocks);
462 		rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_LOG);
463 		rrec->rm_offset = 0;
464 		be16_add_cpu(&block->bb_numrecs, 1);
465 	}
466 }
467 
468 /*
469  * Initialise new secondary superblocks with the pre-grow geometry, but mark
470  * them as "in progress" so we know they haven't yet been activated. This will
471  * get cleared when the update with the new geometry information is done after
472  * changes to the primary are committed. This isn't strictly necessary, but we
473  * get it for free with the delayed buffer write lists and it means we can tell
474  * if a grow operation didn't complete properly after the fact.
475  */
476 static void
477 xfs_sbblock_init(
478 	struct xfs_mount	*mp,
479 	struct xfs_buf		*bp,
480 	struct aghdr_init_data	*id)
481 {
482 	struct xfs_dsb		*dsb = bp->b_addr;
483 
484 	xfs_sb_to_disk(dsb, &mp->m_sb);
485 	dsb->sb_inprogress = 1;
486 }
487 
488 static void
489 xfs_agfblock_init(
490 	struct xfs_mount	*mp,
491 	struct xfs_buf		*bp,
492 	struct aghdr_init_data	*id)
493 {
494 	struct xfs_agf		*agf = bp->b_addr;
495 	xfs_extlen_t		tmpsize;
496 
497 	agf->agf_magicnum = cpu_to_be32(XFS_AGF_MAGIC);
498 	agf->agf_versionnum = cpu_to_be32(XFS_AGF_VERSION);
499 	agf->agf_seqno = cpu_to_be32(id->agno);
500 	agf->agf_length = cpu_to_be32(id->agsize);
501 	agf->agf_bno_root = cpu_to_be32(XFS_BNO_BLOCK(mp));
502 	agf->agf_cnt_root = cpu_to_be32(XFS_CNT_BLOCK(mp));
503 	agf->agf_bno_level = cpu_to_be32(1);
504 	agf->agf_cnt_level = cpu_to_be32(1);
505 	if (xfs_has_rmapbt(mp)) {
506 		agf->agf_rmap_root = cpu_to_be32(XFS_RMAP_BLOCK(mp));
507 		agf->agf_rmap_level = cpu_to_be32(1);
508 		agf->agf_rmap_blocks = cpu_to_be32(1);
509 	}
510 
511 	agf->agf_flfirst = cpu_to_be32(1);
512 	agf->agf_fllast = 0;
513 	agf->agf_flcount = 0;
514 	tmpsize = id->agsize - mp->m_ag_prealloc_blocks;
515 	agf->agf_freeblks = cpu_to_be32(tmpsize);
516 	agf->agf_longest = cpu_to_be32(tmpsize);
517 	if (xfs_has_crc(mp))
518 		uuid_copy(&agf->agf_uuid, &mp->m_sb.sb_meta_uuid);
519 	if (xfs_has_reflink(mp)) {
520 		agf->agf_refcount_root = cpu_to_be32(
521 				xfs_refc_block(mp));
522 		agf->agf_refcount_level = cpu_to_be32(1);
523 		agf->agf_refcount_blocks = cpu_to_be32(1);
524 	}
525 
526 	if (xfs_ag_contains_log(mp, id->agno)) {
527 		int64_t	logblocks = mp->m_sb.sb_logblocks;
528 
529 		be32_add_cpu(&agf->agf_freeblks, -logblocks);
530 		agf->agf_longest = cpu_to_be32(id->agsize -
531 			XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart) - logblocks);
532 	}
533 }
534 
535 static void
536 xfs_agflblock_init(
537 	struct xfs_mount	*mp,
538 	struct xfs_buf		*bp,
539 	struct aghdr_init_data	*id)
540 {
541 	struct xfs_agfl		*agfl = XFS_BUF_TO_AGFL(bp);
542 	__be32			*agfl_bno;
543 	int			bucket;
544 
545 	if (xfs_has_crc(mp)) {
546 		agfl->agfl_magicnum = cpu_to_be32(XFS_AGFL_MAGIC);
547 		agfl->agfl_seqno = cpu_to_be32(id->agno);
548 		uuid_copy(&agfl->agfl_uuid, &mp->m_sb.sb_meta_uuid);
549 	}
550 
551 	agfl_bno = xfs_buf_to_agfl_bno(bp);
552 	for (bucket = 0; bucket < xfs_agfl_size(mp); bucket++)
553 		agfl_bno[bucket] = cpu_to_be32(NULLAGBLOCK);
554 }
555 
556 static void
557 xfs_agiblock_init(
558 	struct xfs_mount	*mp,
559 	struct xfs_buf		*bp,
560 	struct aghdr_init_data	*id)
561 {
562 	struct xfs_agi		*agi = bp->b_addr;
563 	int			bucket;
564 
565 	agi->agi_magicnum = cpu_to_be32(XFS_AGI_MAGIC);
566 	agi->agi_versionnum = cpu_to_be32(XFS_AGI_VERSION);
567 	agi->agi_seqno = cpu_to_be32(id->agno);
568 	agi->agi_length = cpu_to_be32(id->agsize);
569 	agi->agi_count = 0;
570 	agi->agi_root = cpu_to_be32(XFS_IBT_BLOCK(mp));
571 	agi->agi_level = cpu_to_be32(1);
572 	agi->agi_freecount = 0;
573 	agi->agi_newino = cpu_to_be32(NULLAGINO);
574 	agi->agi_dirino = cpu_to_be32(NULLAGINO);
575 	if (xfs_has_crc(mp))
576 		uuid_copy(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid);
577 	if (xfs_has_finobt(mp)) {
578 		agi->agi_free_root = cpu_to_be32(XFS_FIBT_BLOCK(mp));
579 		agi->agi_free_level = cpu_to_be32(1);
580 	}
581 	for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++)
582 		agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
583 	if (xfs_has_inobtcounts(mp)) {
584 		agi->agi_iblocks = cpu_to_be32(1);
585 		if (xfs_has_finobt(mp))
586 			agi->agi_fblocks = cpu_to_be32(1);
587 	}
588 }
589 
590 typedef void (*aghdr_init_work_f)(struct xfs_mount *mp, struct xfs_buf *bp,
591 				  struct aghdr_init_data *id);
592 static int
593 xfs_ag_init_hdr(
594 	struct xfs_mount	*mp,
595 	struct aghdr_init_data	*id,
596 	aghdr_init_work_f	work,
597 	const struct xfs_buf_ops *ops)
598 {
599 	struct xfs_buf		*bp;
600 	int			error;
601 
602 	error = xfs_get_aghdr_buf(mp, id->daddr, id->numblks, &bp, ops);
603 	if (error)
604 		return error;
605 
606 	(*work)(mp, bp, id);
607 
608 	xfs_buf_delwri_queue(bp, &id->buffer_list);
609 	xfs_buf_relse(bp);
610 	return 0;
611 }
612 
613 struct xfs_aghdr_grow_data {
614 	xfs_daddr_t		daddr;
615 	size_t			numblks;
616 	const struct xfs_buf_ops *ops;
617 	aghdr_init_work_f	work;
618 	const struct xfs_btree_ops *bc_ops;
619 	bool			need_init;
620 };
621 
622 /*
623  * Prepare new AG headers to be written to disk. We use uncached buffers here,
624  * as it is assumed these new AG headers are currently beyond the currently
625  * valid filesystem address space. Using cached buffers would trip over EOFS
626  * corruption detection alogrithms in the buffer cache lookup routines.
627  *
628  * This is a non-transactional function, but the prepared buffers are added to a
629  * delayed write buffer list supplied by the caller so they can submit them to
630  * disk and wait on them as required.
631  */
632 int
633 xfs_ag_init_headers(
634 	struct xfs_mount	*mp,
635 	struct aghdr_init_data	*id)
636 
637 {
638 	struct xfs_aghdr_grow_data aghdr_data[] = {
639 	{ /* SB */
640 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_SB_DADDR),
641 		.numblks = XFS_FSS_TO_BB(mp, 1),
642 		.ops = &xfs_sb_buf_ops,
643 		.work = &xfs_sbblock_init,
644 		.need_init = true
645 	},
646 	{ /* AGF */
647 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGF_DADDR(mp)),
648 		.numblks = XFS_FSS_TO_BB(mp, 1),
649 		.ops = &xfs_agf_buf_ops,
650 		.work = &xfs_agfblock_init,
651 		.need_init = true
652 	},
653 	{ /* AGFL */
654 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGFL_DADDR(mp)),
655 		.numblks = XFS_FSS_TO_BB(mp, 1),
656 		.ops = &xfs_agfl_buf_ops,
657 		.work = &xfs_agflblock_init,
658 		.need_init = true
659 	},
660 	{ /* AGI */
661 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGI_DADDR(mp)),
662 		.numblks = XFS_FSS_TO_BB(mp, 1),
663 		.ops = &xfs_agi_buf_ops,
664 		.work = &xfs_agiblock_init,
665 		.need_init = true
666 	},
667 	{ /* BNO root block */
668 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_BNO_BLOCK(mp)),
669 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
670 		.ops = &xfs_bnobt_buf_ops,
671 		.work = &xfs_bnoroot_init,
672 		.bc_ops = &xfs_bnobt_ops,
673 		.need_init = true
674 	},
675 	{ /* CNT root block */
676 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_CNT_BLOCK(mp)),
677 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
678 		.ops = &xfs_cntbt_buf_ops,
679 		.work = &xfs_bnoroot_init,
680 		.bc_ops = &xfs_cntbt_ops,
681 		.need_init = true
682 	},
683 	{ /* INO root block */
684 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_IBT_BLOCK(mp)),
685 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
686 		.ops = &xfs_inobt_buf_ops,
687 		.work = &xfs_btroot_init,
688 		.bc_ops = &xfs_inobt_ops,
689 		.need_init = true
690 	},
691 	{ /* FINO root block */
692 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_FIBT_BLOCK(mp)),
693 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
694 		.ops = &xfs_finobt_buf_ops,
695 		.work = &xfs_btroot_init,
696 		.bc_ops = &xfs_finobt_ops,
697 		.need_init =  xfs_has_finobt(mp)
698 	},
699 	{ /* RMAP root block */
700 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_RMAP_BLOCK(mp)),
701 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
702 		.ops = &xfs_rmapbt_buf_ops,
703 		.work = &xfs_rmaproot_init,
704 		.bc_ops = &xfs_rmapbt_ops,
705 		.need_init = xfs_has_rmapbt(mp)
706 	},
707 	{ /* REFC root block */
708 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, xfs_refc_block(mp)),
709 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
710 		.ops = &xfs_refcountbt_buf_ops,
711 		.work = &xfs_btroot_init,
712 		.bc_ops = &xfs_refcountbt_ops,
713 		.need_init = xfs_has_reflink(mp)
714 	},
715 	{ /* NULL terminating block */
716 		.daddr = XFS_BUF_DADDR_NULL,
717 	}
718 	};
719 	struct  xfs_aghdr_grow_data *dp;
720 	int			error = 0;
721 
722 	/* Account for AG free space in new AG */
723 	id->nfree += id->agsize - mp->m_ag_prealloc_blocks;
724 	for (dp = &aghdr_data[0]; dp->daddr != XFS_BUF_DADDR_NULL; dp++) {
725 		if (!dp->need_init)
726 			continue;
727 
728 		id->daddr = dp->daddr;
729 		id->numblks = dp->numblks;
730 		id->bc_ops = dp->bc_ops;
731 		error = xfs_ag_init_hdr(mp, id, dp->work, dp->ops);
732 		if (error)
733 			break;
734 	}
735 	return error;
736 }
737 
738 int
739 xfs_ag_shrink_space(
740 	struct xfs_perag	*pag,
741 	struct xfs_trans	**tpp,
742 	xfs_extlen_t		delta)
743 {
744 	struct xfs_mount	*mp = pag_mount(pag);
745 	struct xfs_alloc_arg	args = {
746 		.tp	= *tpp,
747 		.mp	= mp,
748 		.pag	= pag,
749 		.minlen = delta,
750 		.maxlen = delta,
751 		.oinfo	= XFS_RMAP_OINFO_SKIP_UPDATE,
752 		.resv	= XFS_AG_RESV_NONE,
753 		.prod	= 1
754 	};
755 	struct xfs_buf		*agibp, *agfbp;
756 	struct xfs_agi		*agi;
757 	struct xfs_agf		*agf;
758 	xfs_agblock_t		aglen;
759 	int			error, err2;
760 
761 	ASSERT(pag_agno(pag) == mp->m_sb.sb_agcount - 1);
762 	error = xfs_ialloc_read_agi(pag, *tpp, 0, &agibp);
763 	if (error)
764 		return error;
765 
766 	agi = agibp->b_addr;
767 
768 	error = xfs_alloc_read_agf(pag, *tpp, 0, &agfbp);
769 	if (error)
770 		return error;
771 
772 	agf = agfbp->b_addr;
773 	aglen = be32_to_cpu(agi->agi_length);
774 	/* some extra paranoid checks before we shrink the ag */
775 	if (XFS_IS_CORRUPT(mp, agf->agf_length != agi->agi_length)) {
776 		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGF);
777 		return -EFSCORRUPTED;
778 	}
779 	if (delta >= aglen)
780 		return -EINVAL;
781 
782 	/*
783 	 * Make sure that the last inode cluster cannot overlap with the new
784 	 * end of the AG, even if it's sparse.
785 	 */
786 	error = xfs_ialloc_check_shrink(pag, *tpp, agibp, aglen - delta);
787 	if (error)
788 		return error;
789 
790 	/*
791 	 * Disable perag reservations so it doesn't cause the allocation request
792 	 * to fail. We'll reestablish reservation before we return.
793 	 */
794 	xfs_ag_resv_free(pag);
795 
796 	/* internal log shouldn't also show up in the free space btrees */
797 	error = xfs_alloc_vextent_exact_bno(&args,
798 			xfs_agbno_to_fsb(pag, aglen - delta));
799 	if (!error && args.agbno == NULLAGBLOCK)
800 		error = -ENOSPC;
801 
802 	if (error) {
803 		/*
804 		 * If extent allocation fails, need to roll the transaction to
805 		 * ensure that the AGFL fixup has been committed anyway.
806 		 *
807 		 * We need to hold the AGF across the roll to ensure nothing can
808 		 * access the AG for allocation until the shrink is fully
809 		 * cleaned up. And due to the resetting of the AG block
810 		 * reservation space needing to lock the AGI, we also have to
811 		 * hold that so we don't get AGI/AGF lock order inversions in
812 		 * the error handling path.
813 		 */
814 		xfs_trans_bhold(*tpp, agfbp);
815 		xfs_trans_bhold(*tpp, agibp);
816 		err2 = xfs_trans_roll(tpp);
817 		if (err2)
818 			return err2;
819 		xfs_trans_bjoin(*tpp, agfbp);
820 		xfs_trans_bjoin(*tpp, agibp);
821 		goto resv_init_out;
822 	}
823 
824 	/*
825 	 * if successfully deleted from freespace btrees, need to confirm
826 	 * per-AG reservation works as expected.
827 	 */
828 	be32_add_cpu(&agi->agi_length, -delta);
829 	be32_add_cpu(&agf->agf_length, -delta);
830 
831 	err2 = xfs_ag_resv_init(pag, *tpp);
832 	if (err2) {
833 		be32_add_cpu(&agi->agi_length, delta);
834 		be32_add_cpu(&agf->agf_length, delta);
835 		if (err2 != -ENOSPC)
836 			goto resv_err;
837 
838 		err2 = xfs_free_extent_later(*tpp, args.fsbno, delta, NULL,
839 				XFS_AG_RESV_NONE, XFS_FREE_EXTENT_SKIP_DISCARD);
840 		if (err2)
841 			goto resv_err;
842 
843 		/*
844 		 * Roll the transaction before trying to re-init the per-ag
845 		 * reservation. The new transaction is clean so it will cancel
846 		 * without any side effects.
847 		 */
848 		error = xfs_defer_finish(tpp);
849 		if (error)
850 			return error;
851 
852 		error = -ENOSPC;
853 		goto resv_init_out;
854 	}
855 
856 	/* Update perag geometry */
857 	pag_group(pag)->xg_block_count -= delta;
858 	__xfs_agino_range(mp, pag_group(pag)->xg_block_count, &pag->agino_min,
859 			&pag->agino_max);
860 
861 	xfs_ialloc_log_agi(*tpp, agibp, XFS_AGI_LENGTH);
862 	xfs_alloc_log_agf(*tpp, agfbp, XFS_AGF_LENGTH);
863 	return 0;
864 
865 resv_init_out:
866 	err2 = xfs_ag_resv_init(pag, *tpp);
867 	if (!err2)
868 		return error;
869 resv_err:
870 	xfs_warn(mp, "Error %d reserving per-AG metadata reserve pool.", err2);
871 	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
872 	return err2;
873 }
874 
875 /*
876  * Extent the AG indicated by the @id by the length passed in
877  */
878 int
879 xfs_ag_extend_space(
880 	struct xfs_perag	*pag,
881 	struct xfs_trans	*tp,
882 	xfs_extlen_t		len)
883 {
884 	struct xfs_mount	*mp = pag_mount(pag);
885 	struct xfs_buf		*bp;
886 	struct xfs_agi		*agi;
887 	struct xfs_agf		*agf;
888 	int			error;
889 
890 	ASSERT(pag_agno(pag) == mp->m_sb.sb_agcount - 1);
891 
892 	error = xfs_ialloc_read_agi(pag, tp, 0, &bp);
893 	if (error)
894 		return error;
895 
896 	agi = bp->b_addr;
897 	be32_add_cpu(&agi->agi_length, len);
898 	xfs_ialloc_log_agi(tp, bp, XFS_AGI_LENGTH);
899 
900 	/*
901 	 * Change agf length.
902 	 */
903 	error = xfs_alloc_read_agf(pag, tp, 0, &bp);
904 	if (error)
905 		return error;
906 
907 	agf = bp->b_addr;
908 	be32_add_cpu(&agf->agf_length, len);
909 	ASSERT(agf->agf_length == agi->agi_length);
910 	xfs_alloc_log_agf(tp, bp, XFS_AGF_LENGTH);
911 
912 	/*
913 	 * Free the new space.
914 	 *
915 	 * XFS_RMAP_OINFO_SKIP_UPDATE is used here to tell the rmap btree that
916 	 * this doesn't actually exist in the rmap btree.
917 	 */
918 	error = xfs_rmap_free(tp, bp, pag, be32_to_cpu(agf->agf_length) - len,
919 				len, &XFS_RMAP_OINFO_SKIP_UPDATE);
920 	if (error)
921 		return error;
922 
923 	error = xfs_free_extent(tp, pag, be32_to_cpu(agf->agf_length) - len,
924 			len, &XFS_RMAP_OINFO_SKIP_UPDATE, XFS_AG_RESV_NONE);
925 	if (error)
926 		return error;
927 
928 	/* Update perag geometry */
929 	pag_group(pag)->xg_block_count = be32_to_cpu(agf->agf_length);
930 	__xfs_agino_range(mp, pag_group(pag)->xg_block_count, &pag->agino_min,
931 			&pag->agino_max);
932 	return 0;
933 }
934 
935 /* Retrieve AG geometry. */
936 int
937 xfs_ag_get_geometry(
938 	struct xfs_perag	*pag,
939 	struct xfs_ag_geometry	*ageo)
940 {
941 	struct xfs_buf		*agi_bp;
942 	struct xfs_buf		*agf_bp;
943 	struct xfs_agi		*agi;
944 	struct xfs_agf		*agf;
945 	unsigned int		freeblks;
946 	int			error;
947 
948 	/* Lock the AG headers. */
949 	error = xfs_ialloc_read_agi(pag, NULL, 0, &agi_bp);
950 	if (error)
951 		return error;
952 	error = xfs_alloc_read_agf(pag, NULL, 0, &agf_bp);
953 	if (error)
954 		goto out_agi;
955 
956 	/* Fill out form. */
957 	memset(ageo, 0, sizeof(*ageo));
958 	ageo->ag_number = pag_agno(pag);
959 
960 	agi = agi_bp->b_addr;
961 	ageo->ag_icount = be32_to_cpu(agi->agi_count);
962 	ageo->ag_ifree = be32_to_cpu(agi->agi_freecount);
963 
964 	agf = agf_bp->b_addr;
965 	ageo->ag_length = be32_to_cpu(agf->agf_length);
966 	freeblks = pag->pagf_freeblks +
967 		   pag->pagf_flcount +
968 		   pag->pagf_btreeblks -
969 		   xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE);
970 	ageo->ag_freeblks = freeblks;
971 	xfs_ag_geom_health(pag, ageo);
972 
973 	/* Release resources. */
974 	xfs_buf_relse(agf_bp);
975 out_agi:
976 	xfs_buf_relse(agi_bp);
977 	return error;
978 }
979