xref: /openbmc/linux/fs/xfs/libxfs/xfs_ag.c (revision b38269ec)
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 
34 
35 /*
36  * Passive reference counting access wrappers to the perag structures.  If the
37  * per-ag structure is to be freed, the freeing code is responsible for cleaning
38  * up objects with passive references before freeing the structure. This is
39  * things like cached buffers.
40  */
41 struct xfs_perag *
42 xfs_perag_get(
43 	struct xfs_mount	*mp,
44 	xfs_agnumber_t		agno)
45 {
46 	struct xfs_perag	*pag;
47 	int			ref = 0;
48 
49 	rcu_read_lock();
50 	pag = radix_tree_lookup(&mp->m_perag_tree, agno);
51 	if (pag) {
52 		ASSERT(atomic_read(&pag->pag_ref) >= 0);
53 		ref = atomic_inc_return(&pag->pag_ref);
54 	}
55 	rcu_read_unlock();
56 	trace_xfs_perag_get(mp, agno, ref, _RET_IP_);
57 	return pag;
58 }
59 
60 /*
61  * search from @first to find the next perag with the given tag set.
62  */
63 struct xfs_perag *
64 xfs_perag_get_tag(
65 	struct xfs_mount	*mp,
66 	xfs_agnumber_t		first,
67 	unsigned int		tag)
68 {
69 	struct xfs_perag	*pag;
70 	int			found;
71 	int			ref;
72 
73 	rcu_read_lock();
74 	found = radix_tree_gang_lookup_tag(&mp->m_perag_tree,
75 					(void **)&pag, first, 1, tag);
76 	if (found <= 0) {
77 		rcu_read_unlock();
78 		return NULL;
79 	}
80 	ref = atomic_inc_return(&pag->pag_ref);
81 	rcu_read_unlock();
82 	trace_xfs_perag_get_tag(mp, pag->pag_agno, ref, _RET_IP_);
83 	return pag;
84 }
85 
86 void
87 xfs_perag_put(
88 	struct xfs_perag	*pag)
89 {
90 	int	ref;
91 
92 	ASSERT(atomic_read(&pag->pag_ref) > 0);
93 	ref = atomic_dec_return(&pag->pag_ref);
94 	trace_xfs_perag_put(pag->pag_mount, pag->pag_agno, ref, _RET_IP_);
95 }
96 
97 /*
98  * xfs_initialize_perag_data
99  *
100  * Read in each per-ag structure so we can count up the number of
101  * allocated inodes, free inodes and used filesystem blocks as this
102  * information is no longer persistent in the superblock. Once we have
103  * this information, write it into the in-core superblock structure.
104  */
105 int
106 xfs_initialize_perag_data(
107 	struct xfs_mount	*mp,
108 	xfs_agnumber_t		agcount)
109 {
110 	xfs_agnumber_t		index;
111 	struct xfs_perag	*pag;
112 	struct xfs_sb		*sbp = &mp->m_sb;
113 	uint64_t		ifree = 0;
114 	uint64_t		ialloc = 0;
115 	uint64_t		bfree = 0;
116 	uint64_t		bfreelst = 0;
117 	uint64_t		btree = 0;
118 	uint64_t		fdblocks;
119 	int			error = 0;
120 
121 	for (index = 0; index < agcount; index++) {
122 		/*
123 		 * read the agf, then the agi. This gets us
124 		 * all the information we need and populates the
125 		 * per-ag structures for us.
126 		 */
127 		error = xfs_alloc_pagf_init(mp, NULL, index, 0);
128 		if (error)
129 			return error;
130 
131 		error = xfs_ialloc_pagi_init(mp, NULL, index);
132 		if (error)
133 			return error;
134 		pag = xfs_perag_get(mp, index);
135 		ifree += pag->pagi_freecount;
136 		ialloc += pag->pagi_count;
137 		bfree += pag->pagf_freeblks;
138 		bfreelst += pag->pagf_flcount;
139 		btree += pag->pagf_btreeblks;
140 		xfs_perag_put(pag);
141 	}
142 	fdblocks = bfree + bfreelst + btree;
143 
144 	/*
145 	 * If the new summary counts are obviously incorrect, fail the
146 	 * mount operation because that implies the AGFs are also corrupt.
147 	 * Clear FS_COUNTERS so that we don't unmount with a dirty log, which
148 	 * will prevent xfs_repair from fixing anything.
149 	 */
150 	if (fdblocks > sbp->sb_dblocks || ifree > ialloc) {
151 		xfs_alert(mp, "AGF corruption. Please run xfs_repair.");
152 		error = -EFSCORRUPTED;
153 		goto out;
154 	}
155 
156 	/* Overwrite incore superblock counters with just-read data */
157 	spin_lock(&mp->m_sb_lock);
158 	sbp->sb_ifree = ifree;
159 	sbp->sb_icount = ialloc;
160 	sbp->sb_fdblocks = fdblocks;
161 	spin_unlock(&mp->m_sb_lock);
162 
163 	xfs_reinit_percpu_counters(mp);
164 out:
165 	xfs_fs_mark_healthy(mp, XFS_SICK_FS_COUNTERS);
166 	return error;
167 }
168 
169 STATIC void
170 __xfs_free_perag(
171 	struct rcu_head	*head)
172 {
173 	struct xfs_perag *pag = container_of(head, struct xfs_perag, rcu_head);
174 
175 	ASSERT(!delayed_work_pending(&pag->pag_blockgc_work));
176 	ASSERT(atomic_read(&pag->pag_ref) == 0);
177 	kmem_free(pag);
178 }
179 
180 /*
181  * Free up the per-ag resources associated with the mount structure.
182  */
183 void
184 xfs_free_perag(
185 	struct xfs_mount	*mp)
186 {
187 	struct xfs_perag	*pag;
188 	xfs_agnumber_t		agno;
189 
190 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
191 		spin_lock(&mp->m_perag_lock);
192 		pag = radix_tree_delete(&mp->m_perag_tree, agno);
193 		spin_unlock(&mp->m_perag_lock);
194 		ASSERT(pag);
195 		ASSERT(atomic_read(&pag->pag_ref) == 0);
196 
197 		cancel_delayed_work_sync(&pag->pag_blockgc_work);
198 		xfs_iunlink_destroy(pag);
199 		xfs_buf_hash_destroy(pag);
200 
201 		call_rcu(&pag->rcu_head, __xfs_free_perag);
202 	}
203 }
204 
205 int
206 xfs_initialize_perag(
207 	struct xfs_mount	*mp,
208 	xfs_agnumber_t		agcount,
209 	xfs_agnumber_t		*maxagi)
210 {
211 	struct xfs_perag	*pag;
212 	xfs_agnumber_t		index;
213 	xfs_agnumber_t		first_initialised = NULLAGNUMBER;
214 	int			error;
215 
216 	/*
217 	 * Walk the current per-ag tree so we don't try to initialise AGs
218 	 * that already exist (growfs case). Allocate and insert all the
219 	 * AGs we don't find ready for initialisation.
220 	 */
221 	for (index = 0; index < agcount; index++) {
222 		pag = xfs_perag_get(mp, index);
223 		if (pag) {
224 			xfs_perag_put(pag);
225 			continue;
226 		}
227 
228 		pag = kmem_zalloc(sizeof(*pag), KM_MAYFAIL);
229 		if (!pag) {
230 			error = -ENOMEM;
231 			goto out_unwind_new_pags;
232 		}
233 		pag->pag_agno = index;
234 		pag->pag_mount = mp;
235 
236 		error = radix_tree_preload(GFP_NOFS);
237 		if (error)
238 			goto out_free_pag;
239 
240 		spin_lock(&mp->m_perag_lock);
241 		if (radix_tree_insert(&mp->m_perag_tree, index, pag)) {
242 			WARN_ON_ONCE(1);
243 			spin_unlock(&mp->m_perag_lock);
244 			radix_tree_preload_end();
245 			error = -EEXIST;
246 			goto out_free_pag;
247 		}
248 		spin_unlock(&mp->m_perag_lock);
249 		radix_tree_preload_end();
250 
251 		/* Place kernel structure only init below this point. */
252 		spin_lock_init(&pag->pag_ici_lock);
253 		spin_lock_init(&pag->pagb_lock);
254 		spin_lock_init(&pag->pag_state_lock);
255 		INIT_DELAYED_WORK(&pag->pag_blockgc_work, xfs_blockgc_worker);
256 		INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
257 		init_waitqueue_head(&pag->pagb_wait);
258 		pag->pagb_count = 0;
259 		pag->pagb_tree = RB_ROOT;
260 
261 		error = xfs_buf_hash_init(pag);
262 		if (error)
263 			goto out_remove_pag;
264 
265 		error = xfs_iunlink_init(pag);
266 		if (error)
267 			goto out_hash_destroy;
268 
269 		/* first new pag is fully initialized */
270 		if (first_initialised == NULLAGNUMBER)
271 			first_initialised = index;
272 	}
273 
274 	index = xfs_set_inode_alloc(mp, agcount);
275 
276 	if (maxagi)
277 		*maxagi = index;
278 
279 	mp->m_ag_prealloc_blocks = xfs_prealloc_blocks(mp);
280 	return 0;
281 
282 out_hash_destroy:
283 	xfs_buf_hash_destroy(pag);
284 out_remove_pag:
285 	radix_tree_delete(&mp->m_perag_tree, index);
286 out_free_pag:
287 	kmem_free(pag);
288 out_unwind_new_pags:
289 	/* unwind any prior newly initialized pags */
290 	for (index = first_initialised; index < agcount; index++) {
291 		pag = radix_tree_delete(&mp->m_perag_tree, index);
292 		if (!pag)
293 			break;
294 		xfs_buf_hash_destroy(pag);
295 		xfs_iunlink_destroy(pag);
296 		kmem_free(pag);
297 	}
298 	return error;
299 }
300 
301 static int
302 xfs_get_aghdr_buf(
303 	struct xfs_mount	*mp,
304 	xfs_daddr_t		blkno,
305 	size_t			numblks,
306 	struct xfs_buf		**bpp,
307 	const struct xfs_buf_ops *ops)
308 {
309 	struct xfs_buf		*bp;
310 	int			error;
311 
312 	error = xfs_buf_get_uncached(mp->m_ddev_targp, numblks, 0, &bp);
313 	if (error)
314 		return error;
315 
316 	bp->b_maps[0].bm_bn = blkno;
317 	bp->b_ops = ops;
318 
319 	*bpp = bp;
320 	return 0;
321 }
322 
323 static inline bool is_log_ag(struct xfs_mount *mp, struct aghdr_init_data *id)
324 {
325 	return mp->m_sb.sb_logstart > 0 &&
326 	       id->agno == XFS_FSB_TO_AGNO(mp, mp->m_sb.sb_logstart);
327 }
328 
329 /*
330  * Generic btree root block init function
331  */
332 static void
333 xfs_btroot_init(
334 	struct xfs_mount	*mp,
335 	struct xfs_buf		*bp,
336 	struct aghdr_init_data	*id)
337 {
338 	xfs_btree_init_block(mp, bp, id->type, 0, 0, id->agno);
339 }
340 
341 /* Finish initializing a free space btree. */
342 static void
343 xfs_freesp_init_recs(
344 	struct xfs_mount	*mp,
345 	struct xfs_buf		*bp,
346 	struct aghdr_init_data	*id)
347 {
348 	struct xfs_alloc_rec	*arec;
349 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
350 
351 	arec = XFS_ALLOC_REC_ADDR(mp, XFS_BUF_TO_BLOCK(bp), 1);
352 	arec->ar_startblock = cpu_to_be32(mp->m_ag_prealloc_blocks);
353 
354 	if (is_log_ag(mp, id)) {
355 		struct xfs_alloc_rec	*nrec;
356 		xfs_agblock_t		start = XFS_FSB_TO_AGBNO(mp,
357 							mp->m_sb.sb_logstart);
358 
359 		ASSERT(start >= mp->m_ag_prealloc_blocks);
360 		if (start != mp->m_ag_prealloc_blocks) {
361 			/*
362 			 * Modify first record to pad stripe align of log
363 			 */
364 			arec->ar_blockcount = cpu_to_be32(start -
365 						mp->m_ag_prealloc_blocks);
366 			nrec = arec + 1;
367 
368 			/*
369 			 * Insert second record at start of internal log
370 			 * which then gets trimmed.
371 			 */
372 			nrec->ar_startblock = cpu_to_be32(
373 					be32_to_cpu(arec->ar_startblock) +
374 					be32_to_cpu(arec->ar_blockcount));
375 			arec = nrec;
376 			be16_add_cpu(&block->bb_numrecs, 1);
377 		}
378 		/*
379 		 * Change record start to after the internal log
380 		 */
381 		be32_add_cpu(&arec->ar_startblock, mp->m_sb.sb_logblocks);
382 	}
383 
384 	/*
385 	 * Calculate the record block count and check for the case where
386 	 * the log might have consumed all available space in the AG. If
387 	 * so, reset the record count to 0 to avoid exposure of an invalid
388 	 * record start block.
389 	 */
390 	arec->ar_blockcount = cpu_to_be32(id->agsize -
391 					  be32_to_cpu(arec->ar_startblock));
392 	if (!arec->ar_blockcount)
393 		block->bb_numrecs = 0;
394 }
395 
396 /*
397  * Alloc btree root block init functions
398  */
399 static void
400 xfs_bnoroot_init(
401 	struct xfs_mount	*mp,
402 	struct xfs_buf		*bp,
403 	struct aghdr_init_data	*id)
404 {
405 	xfs_btree_init_block(mp, bp, XFS_BTNUM_BNO, 0, 1, id->agno);
406 	xfs_freesp_init_recs(mp, bp, id);
407 }
408 
409 static void
410 xfs_cntroot_init(
411 	struct xfs_mount	*mp,
412 	struct xfs_buf		*bp,
413 	struct aghdr_init_data	*id)
414 {
415 	xfs_btree_init_block(mp, bp, XFS_BTNUM_CNT, 0, 1, id->agno);
416 	xfs_freesp_init_recs(mp, bp, id);
417 }
418 
419 /*
420  * Reverse map root block init
421  */
422 static void
423 xfs_rmaproot_init(
424 	struct xfs_mount	*mp,
425 	struct xfs_buf		*bp,
426 	struct aghdr_init_data	*id)
427 {
428 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
429 	struct xfs_rmap_rec	*rrec;
430 
431 	xfs_btree_init_block(mp, bp, XFS_BTNUM_RMAP, 0, 4, id->agno);
432 
433 	/*
434 	 * mark the AG header regions as static metadata The BNO
435 	 * btree block is the first block after the headers, so
436 	 * it's location defines the size of region the static
437 	 * metadata consumes.
438 	 *
439 	 * Note: unlike mkfs, we never have to account for log
440 	 * space when growing the data regions
441 	 */
442 	rrec = XFS_RMAP_REC_ADDR(block, 1);
443 	rrec->rm_startblock = 0;
444 	rrec->rm_blockcount = cpu_to_be32(XFS_BNO_BLOCK(mp));
445 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_FS);
446 	rrec->rm_offset = 0;
447 
448 	/* account freespace btree root blocks */
449 	rrec = XFS_RMAP_REC_ADDR(block, 2);
450 	rrec->rm_startblock = cpu_to_be32(XFS_BNO_BLOCK(mp));
451 	rrec->rm_blockcount = cpu_to_be32(2);
452 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_AG);
453 	rrec->rm_offset = 0;
454 
455 	/* account inode btree root blocks */
456 	rrec = XFS_RMAP_REC_ADDR(block, 3);
457 	rrec->rm_startblock = cpu_to_be32(XFS_IBT_BLOCK(mp));
458 	rrec->rm_blockcount = cpu_to_be32(XFS_RMAP_BLOCK(mp) -
459 					  XFS_IBT_BLOCK(mp));
460 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_INOBT);
461 	rrec->rm_offset = 0;
462 
463 	/* account for rmap btree root */
464 	rrec = XFS_RMAP_REC_ADDR(block, 4);
465 	rrec->rm_startblock = cpu_to_be32(XFS_RMAP_BLOCK(mp));
466 	rrec->rm_blockcount = cpu_to_be32(1);
467 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_AG);
468 	rrec->rm_offset = 0;
469 
470 	/* account for refc btree root */
471 	if (xfs_has_reflink(mp)) {
472 		rrec = XFS_RMAP_REC_ADDR(block, 5);
473 		rrec->rm_startblock = cpu_to_be32(xfs_refc_block(mp));
474 		rrec->rm_blockcount = cpu_to_be32(1);
475 		rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_REFC);
476 		rrec->rm_offset = 0;
477 		be16_add_cpu(&block->bb_numrecs, 1);
478 	}
479 
480 	/* account for the log space */
481 	if (is_log_ag(mp, id)) {
482 		rrec = XFS_RMAP_REC_ADDR(block,
483 				be16_to_cpu(block->bb_numrecs) + 1);
484 		rrec->rm_startblock = cpu_to_be32(
485 				XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart));
486 		rrec->rm_blockcount = cpu_to_be32(mp->m_sb.sb_logblocks);
487 		rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_LOG);
488 		rrec->rm_offset = 0;
489 		be16_add_cpu(&block->bb_numrecs, 1);
490 	}
491 }
492 
493 /*
494  * Initialise new secondary superblocks with the pre-grow geometry, but mark
495  * them as "in progress" so we know they haven't yet been activated. This will
496  * get cleared when the update with the new geometry information is done after
497  * changes to the primary are committed. This isn't strictly necessary, but we
498  * get it for free with the delayed buffer write lists and it means we can tell
499  * if a grow operation didn't complete properly after the fact.
500  */
501 static void
502 xfs_sbblock_init(
503 	struct xfs_mount	*mp,
504 	struct xfs_buf		*bp,
505 	struct aghdr_init_data	*id)
506 {
507 	struct xfs_dsb		*dsb = bp->b_addr;
508 
509 	xfs_sb_to_disk(dsb, &mp->m_sb);
510 	dsb->sb_inprogress = 1;
511 }
512 
513 static void
514 xfs_agfblock_init(
515 	struct xfs_mount	*mp,
516 	struct xfs_buf		*bp,
517 	struct aghdr_init_data	*id)
518 {
519 	struct xfs_agf		*agf = bp->b_addr;
520 	xfs_extlen_t		tmpsize;
521 
522 	agf->agf_magicnum = cpu_to_be32(XFS_AGF_MAGIC);
523 	agf->agf_versionnum = cpu_to_be32(XFS_AGF_VERSION);
524 	agf->agf_seqno = cpu_to_be32(id->agno);
525 	agf->agf_length = cpu_to_be32(id->agsize);
526 	agf->agf_roots[XFS_BTNUM_BNOi] = cpu_to_be32(XFS_BNO_BLOCK(mp));
527 	agf->agf_roots[XFS_BTNUM_CNTi] = cpu_to_be32(XFS_CNT_BLOCK(mp));
528 	agf->agf_levels[XFS_BTNUM_BNOi] = cpu_to_be32(1);
529 	agf->agf_levels[XFS_BTNUM_CNTi] = cpu_to_be32(1);
530 	if (xfs_has_rmapbt(mp)) {
531 		agf->agf_roots[XFS_BTNUM_RMAPi] =
532 					cpu_to_be32(XFS_RMAP_BLOCK(mp));
533 		agf->agf_levels[XFS_BTNUM_RMAPi] = cpu_to_be32(1);
534 		agf->agf_rmap_blocks = cpu_to_be32(1);
535 	}
536 
537 	agf->agf_flfirst = cpu_to_be32(1);
538 	agf->agf_fllast = 0;
539 	agf->agf_flcount = 0;
540 	tmpsize = id->agsize - mp->m_ag_prealloc_blocks;
541 	agf->agf_freeblks = cpu_to_be32(tmpsize);
542 	agf->agf_longest = cpu_to_be32(tmpsize);
543 	if (xfs_has_crc(mp))
544 		uuid_copy(&agf->agf_uuid, &mp->m_sb.sb_meta_uuid);
545 	if (xfs_has_reflink(mp)) {
546 		agf->agf_refcount_root = cpu_to_be32(
547 				xfs_refc_block(mp));
548 		agf->agf_refcount_level = cpu_to_be32(1);
549 		agf->agf_refcount_blocks = cpu_to_be32(1);
550 	}
551 
552 	if (is_log_ag(mp, id)) {
553 		int64_t	logblocks = mp->m_sb.sb_logblocks;
554 
555 		be32_add_cpu(&agf->agf_freeblks, -logblocks);
556 		agf->agf_longest = cpu_to_be32(id->agsize -
557 			XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart) - logblocks);
558 	}
559 }
560 
561 static void
562 xfs_agflblock_init(
563 	struct xfs_mount	*mp,
564 	struct xfs_buf		*bp,
565 	struct aghdr_init_data	*id)
566 {
567 	struct xfs_agfl		*agfl = XFS_BUF_TO_AGFL(bp);
568 	__be32			*agfl_bno;
569 	int			bucket;
570 
571 	if (xfs_has_crc(mp)) {
572 		agfl->agfl_magicnum = cpu_to_be32(XFS_AGFL_MAGIC);
573 		agfl->agfl_seqno = cpu_to_be32(id->agno);
574 		uuid_copy(&agfl->agfl_uuid, &mp->m_sb.sb_meta_uuid);
575 	}
576 
577 	agfl_bno = xfs_buf_to_agfl_bno(bp);
578 	for (bucket = 0; bucket < xfs_agfl_size(mp); bucket++)
579 		agfl_bno[bucket] = cpu_to_be32(NULLAGBLOCK);
580 }
581 
582 static void
583 xfs_agiblock_init(
584 	struct xfs_mount	*mp,
585 	struct xfs_buf		*bp,
586 	struct aghdr_init_data	*id)
587 {
588 	struct xfs_agi		*agi = bp->b_addr;
589 	int			bucket;
590 
591 	agi->agi_magicnum = cpu_to_be32(XFS_AGI_MAGIC);
592 	agi->agi_versionnum = cpu_to_be32(XFS_AGI_VERSION);
593 	agi->agi_seqno = cpu_to_be32(id->agno);
594 	agi->agi_length = cpu_to_be32(id->agsize);
595 	agi->agi_count = 0;
596 	agi->agi_root = cpu_to_be32(XFS_IBT_BLOCK(mp));
597 	agi->agi_level = cpu_to_be32(1);
598 	agi->agi_freecount = 0;
599 	agi->agi_newino = cpu_to_be32(NULLAGINO);
600 	agi->agi_dirino = cpu_to_be32(NULLAGINO);
601 	if (xfs_has_crc(mp))
602 		uuid_copy(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid);
603 	if (xfs_has_finobt(mp)) {
604 		agi->agi_free_root = cpu_to_be32(XFS_FIBT_BLOCK(mp));
605 		agi->agi_free_level = cpu_to_be32(1);
606 	}
607 	for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++)
608 		agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
609 	if (xfs_has_inobtcounts(mp)) {
610 		agi->agi_iblocks = cpu_to_be32(1);
611 		if (xfs_has_finobt(mp))
612 			agi->agi_fblocks = cpu_to_be32(1);
613 	}
614 }
615 
616 typedef void (*aghdr_init_work_f)(struct xfs_mount *mp, struct xfs_buf *bp,
617 				  struct aghdr_init_data *id);
618 static int
619 xfs_ag_init_hdr(
620 	struct xfs_mount	*mp,
621 	struct aghdr_init_data	*id,
622 	aghdr_init_work_f	work,
623 	const struct xfs_buf_ops *ops)
624 {
625 	struct xfs_buf		*bp;
626 	int			error;
627 
628 	error = xfs_get_aghdr_buf(mp, id->daddr, id->numblks, &bp, ops);
629 	if (error)
630 		return error;
631 
632 	(*work)(mp, bp, id);
633 
634 	xfs_buf_delwri_queue(bp, &id->buffer_list);
635 	xfs_buf_relse(bp);
636 	return 0;
637 }
638 
639 struct xfs_aghdr_grow_data {
640 	xfs_daddr_t		daddr;
641 	size_t			numblks;
642 	const struct xfs_buf_ops *ops;
643 	aghdr_init_work_f	work;
644 	xfs_btnum_t		type;
645 	bool			need_init;
646 };
647 
648 /*
649  * Prepare new AG headers to be written to disk. We use uncached buffers here,
650  * as it is assumed these new AG headers are currently beyond the currently
651  * valid filesystem address space. Using cached buffers would trip over EOFS
652  * corruption detection alogrithms in the buffer cache lookup routines.
653  *
654  * This is a non-transactional function, but the prepared buffers are added to a
655  * delayed write buffer list supplied by the caller so they can submit them to
656  * disk and wait on them as required.
657  */
658 int
659 xfs_ag_init_headers(
660 	struct xfs_mount	*mp,
661 	struct aghdr_init_data	*id)
662 
663 {
664 	struct xfs_aghdr_grow_data aghdr_data[] = {
665 	{ /* SB */
666 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_SB_DADDR),
667 		.numblks = XFS_FSS_TO_BB(mp, 1),
668 		.ops = &xfs_sb_buf_ops,
669 		.work = &xfs_sbblock_init,
670 		.need_init = true
671 	},
672 	{ /* AGF */
673 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGF_DADDR(mp)),
674 		.numblks = XFS_FSS_TO_BB(mp, 1),
675 		.ops = &xfs_agf_buf_ops,
676 		.work = &xfs_agfblock_init,
677 		.need_init = true
678 	},
679 	{ /* AGFL */
680 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGFL_DADDR(mp)),
681 		.numblks = XFS_FSS_TO_BB(mp, 1),
682 		.ops = &xfs_agfl_buf_ops,
683 		.work = &xfs_agflblock_init,
684 		.need_init = true
685 	},
686 	{ /* AGI */
687 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGI_DADDR(mp)),
688 		.numblks = XFS_FSS_TO_BB(mp, 1),
689 		.ops = &xfs_agi_buf_ops,
690 		.work = &xfs_agiblock_init,
691 		.need_init = true
692 	},
693 	{ /* BNO root block */
694 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_BNO_BLOCK(mp)),
695 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
696 		.ops = &xfs_bnobt_buf_ops,
697 		.work = &xfs_bnoroot_init,
698 		.need_init = true
699 	},
700 	{ /* CNT root block */
701 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_CNT_BLOCK(mp)),
702 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
703 		.ops = &xfs_cntbt_buf_ops,
704 		.work = &xfs_cntroot_init,
705 		.need_init = true
706 	},
707 	{ /* INO root block */
708 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_IBT_BLOCK(mp)),
709 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
710 		.ops = &xfs_inobt_buf_ops,
711 		.work = &xfs_btroot_init,
712 		.type = XFS_BTNUM_INO,
713 		.need_init = true
714 	},
715 	{ /* FINO root block */
716 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_FIBT_BLOCK(mp)),
717 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
718 		.ops = &xfs_finobt_buf_ops,
719 		.work = &xfs_btroot_init,
720 		.type = XFS_BTNUM_FINO,
721 		.need_init =  xfs_has_finobt(mp)
722 	},
723 	{ /* RMAP root block */
724 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_RMAP_BLOCK(mp)),
725 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
726 		.ops = &xfs_rmapbt_buf_ops,
727 		.work = &xfs_rmaproot_init,
728 		.need_init = xfs_has_rmapbt(mp)
729 	},
730 	{ /* REFC root block */
731 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, xfs_refc_block(mp)),
732 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
733 		.ops = &xfs_refcountbt_buf_ops,
734 		.work = &xfs_btroot_init,
735 		.type = XFS_BTNUM_REFC,
736 		.need_init = xfs_has_reflink(mp)
737 	},
738 	{ /* NULL terminating block */
739 		.daddr = XFS_BUF_DADDR_NULL,
740 	}
741 	};
742 	struct  xfs_aghdr_grow_data *dp;
743 	int			error = 0;
744 
745 	/* Account for AG free space in new AG */
746 	id->nfree += id->agsize - mp->m_ag_prealloc_blocks;
747 	for (dp = &aghdr_data[0]; dp->daddr != XFS_BUF_DADDR_NULL; dp++) {
748 		if (!dp->need_init)
749 			continue;
750 
751 		id->daddr = dp->daddr;
752 		id->numblks = dp->numblks;
753 		id->type = dp->type;
754 		error = xfs_ag_init_hdr(mp, id, dp->work, dp->ops);
755 		if (error)
756 			break;
757 	}
758 	return error;
759 }
760 
761 int
762 xfs_ag_shrink_space(
763 	struct xfs_mount	*mp,
764 	struct xfs_trans	**tpp,
765 	xfs_agnumber_t		agno,
766 	xfs_extlen_t		delta)
767 {
768 	struct xfs_alloc_arg	args = {
769 		.tp	= *tpp,
770 		.mp	= mp,
771 		.type	= XFS_ALLOCTYPE_THIS_BNO,
772 		.minlen = delta,
773 		.maxlen = delta,
774 		.oinfo	= XFS_RMAP_OINFO_SKIP_UPDATE,
775 		.resv	= XFS_AG_RESV_NONE,
776 		.prod	= 1
777 	};
778 	struct xfs_buf		*agibp, *agfbp;
779 	struct xfs_agi		*agi;
780 	struct xfs_agf		*agf;
781 	xfs_agblock_t		aglen;
782 	int			error, err2;
783 
784 	ASSERT(agno == mp->m_sb.sb_agcount - 1);
785 	error = xfs_ialloc_read_agi(mp, *tpp, agno, &agibp);
786 	if (error)
787 		return error;
788 
789 	agi = agibp->b_addr;
790 
791 	error = xfs_alloc_read_agf(mp, *tpp, agno, 0, &agfbp);
792 	if (error)
793 		return error;
794 
795 	agf = agfbp->b_addr;
796 	aglen = be32_to_cpu(agi->agi_length);
797 	/* some extra paranoid checks before we shrink the ag */
798 	if (XFS_IS_CORRUPT(mp, agf->agf_length != agi->agi_length))
799 		return -EFSCORRUPTED;
800 	if (delta >= aglen)
801 		return -EINVAL;
802 
803 	args.fsbno = XFS_AGB_TO_FSB(mp, agno, aglen - delta);
804 
805 	/*
806 	 * Make sure that the last inode cluster cannot overlap with the new
807 	 * end of the AG, even if it's sparse.
808 	 */
809 	error = xfs_ialloc_check_shrink(*tpp, agno, agibp, aglen - delta);
810 	if (error)
811 		return error;
812 
813 	/*
814 	 * Disable perag reservations so it doesn't cause the allocation request
815 	 * to fail. We'll reestablish reservation before we return.
816 	 */
817 	error = xfs_ag_resv_free(agibp->b_pag);
818 	if (error)
819 		return error;
820 
821 	/* internal log shouldn't also show up in the free space btrees */
822 	error = xfs_alloc_vextent(&args);
823 	if (!error && args.agbno == NULLAGBLOCK)
824 		error = -ENOSPC;
825 
826 	if (error) {
827 		/*
828 		 * if extent allocation fails, need to roll the transaction to
829 		 * ensure that the AGFL fixup has been committed anyway.
830 		 */
831 		xfs_trans_bhold(*tpp, agfbp);
832 		err2 = xfs_trans_roll(tpp);
833 		if (err2)
834 			return err2;
835 		xfs_trans_bjoin(*tpp, agfbp);
836 		goto resv_init_out;
837 	}
838 
839 	/*
840 	 * if successfully deleted from freespace btrees, need to confirm
841 	 * per-AG reservation works as expected.
842 	 */
843 	be32_add_cpu(&agi->agi_length, -delta);
844 	be32_add_cpu(&agf->agf_length, -delta);
845 
846 	err2 = xfs_ag_resv_init(agibp->b_pag, *tpp);
847 	if (err2) {
848 		be32_add_cpu(&agi->agi_length, delta);
849 		be32_add_cpu(&agf->agf_length, delta);
850 		if (err2 != -ENOSPC)
851 			goto resv_err;
852 
853 		__xfs_bmap_add_free(*tpp, args.fsbno, delta, NULL, true);
854 
855 		/*
856 		 * Roll the transaction before trying to re-init the per-ag
857 		 * reservation. The new transaction is clean so it will cancel
858 		 * without any side effects.
859 		 */
860 		error = xfs_defer_finish(tpp);
861 		if (error)
862 			return error;
863 
864 		error = -ENOSPC;
865 		goto resv_init_out;
866 	}
867 	xfs_ialloc_log_agi(*tpp, agibp, XFS_AGI_LENGTH);
868 	xfs_alloc_log_agf(*tpp, agfbp, XFS_AGF_LENGTH);
869 	return 0;
870 resv_init_out:
871 	err2 = xfs_ag_resv_init(agibp->b_pag, *tpp);
872 	if (!err2)
873 		return error;
874 resv_err:
875 	xfs_warn(mp, "Error %d reserving per-AG metadata reserve pool.", err2);
876 	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
877 	return err2;
878 }
879 
880 /*
881  * Extent the AG indicated by the @id by the length passed in
882  */
883 int
884 xfs_ag_extend_space(
885 	struct xfs_mount	*mp,
886 	struct xfs_trans	*tp,
887 	struct aghdr_init_data	*id,
888 	xfs_extlen_t		len)
889 {
890 	struct xfs_buf		*bp;
891 	struct xfs_agi		*agi;
892 	struct xfs_agf		*agf;
893 	int			error;
894 
895 	/*
896 	 * Change the agi length.
897 	 */
898 	error = xfs_ialloc_read_agi(mp, tp, id->agno, &bp);
899 	if (error)
900 		return error;
901 
902 	agi = bp->b_addr;
903 	be32_add_cpu(&agi->agi_length, len);
904 	ASSERT(id->agno == mp->m_sb.sb_agcount - 1 ||
905 	       be32_to_cpu(agi->agi_length) == mp->m_sb.sb_agblocks);
906 	xfs_ialloc_log_agi(tp, bp, XFS_AGI_LENGTH);
907 
908 	/*
909 	 * Change agf length.
910 	 */
911 	error = xfs_alloc_read_agf(mp, tp, id->agno, 0, &bp);
912 	if (error)
913 		return error;
914 
915 	agf = bp->b_addr;
916 	be32_add_cpu(&agf->agf_length, len);
917 	ASSERT(agf->agf_length == agi->agi_length);
918 	xfs_alloc_log_agf(tp, bp, XFS_AGF_LENGTH);
919 
920 	/*
921 	 * Free the new space.
922 	 *
923 	 * XFS_RMAP_OINFO_SKIP_UPDATE is used here to tell the rmap btree that
924 	 * this doesn't actually exist in the rmap btree.
925 	 */
926 	error = xfs_rmap_free(tp, bp, bp->b_pag,
927 				be32_to_cpu(agf->agf_length) - len,
928 				len, &XFS_RMAP_OINFO_SKIP_UPDATE);
929 	if (error)
930 		return error;
931 
932 	return  xfs_free_extent(tp, XFS_AGB_TO_FSB(mp, id->agno,
933 					be32_to_cpu(agf->agf_length) - len),
934 				len, &XFS_RMAP_OINFO_SKIP_UPDATE,
935 				XFS_AG_RESV_NONE);
936 }
937 
938 /* Retrieve AG geometry. */
939 int
940 xfs_ag_get_geometry(
941 	struct xfs_mount	*mp,
942 	xfs_agnumber_t		agno,
943 	struct xfs_ag_geometry	*ageo)
944 {
945 	struct xfs_buf		*agi_bp;
946 	struct xfs_buf		*agf_bp;
947 	struct xfs_agi		*agi;
948 	struct xfs_agf		*agf;
949 	struct xfs_perag	*pag;
950 	unsigned int		freeblks;
951 	int			error;
952 
953 	if (agno >= mp->m_sb.sb_agcount)
954 		return -EINVAL;
955 
956 	/* Lock the AG headers. */
957 	error = xfs_ialloc_read_agi(mp, NULL, agno, &agi_bp);
958 	if (error)
959 		return error;
960 	error = xfs_alloc_read_agf(mp, NULL, agno, 0, &agf_bp);
961 	if (error)
962 		goto out_agi;
963 
964 	pag = agi_bp->b_pag;
965 
966 	/* Fill out form. */
967 	memset(ageo, 0, sizeof(*ageo));
968 	ageo->ag_number = agno;
969 
970 	agi = agi_bp->b_addr;
971 	ageo->ag_icount = be32_to_cpu(agi->agi_count);
972 	ageo->ag_ifree = be32_to_cpu(agi->agi_freecount);
973 
974 	agf = agf_bp->b_addr;
975 	ageo->ag_length = be32_to_cpu(agf->agf_length);
976 	freeblks = pag->pagf_freeblks +
977 		   pag->pagf_flcount +
978 		   pag->pagf_btreeblks -
979 		   xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE);
980 	ageo->ag_freeblks = freeblks;
981 	xfs_ag_geom_health(pag, ageo);
982 
983 	/* Release resources. */
984 	xfs_buf_relse(agf_bp);
985 out_agi:
986 	xfs_buf_relse(agi_bp);
987 	return error;
988 }
989