xref: /linux/fs/xfs/libxfs/xfs_ag.c (revision 71dfa617ea9f18e4585fe78364217cd32b1fc382)
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 
48 	rcu_read_lock();
49 	pag = radix_tree_lookup(&mp->m_perag_tree, agno);
50 	if (pag) {
51 		trace_xfs_perag_get(pag, _RET_IP_);
52 		ASSERT(atomic_read(&pag->pag_ref) >= 0);
53 		atomic_inc(&pag->pag_ref);
54 	}
55 	rcu_read_unlock();
56 	return pag;
57 }
58 
59 /*
60  * search from @first to find the next perag with the given tag set.
61  */
62 struct xfs_perag *
63 xfs_perag_get_tag(
64 	struct xfs_mount	*mp,
65 	xfs_agnumber_t		first,
66 	unsigned int		tag)
67 {
68 	struct xfs_perag	*pag;
69 	int			found;
70 
71 	rcu_read_lock();
72 	found = radix_tree_gang_lookup_tag(&mp->m_perag_tree,
73 					(void **)&pag, first, 1, tag);
74 	if (found <= 0) {
75 		rcu_read_unlock();
76 		return NULL;
77 	}
78 	trace_xfs_perag_get_tag(pag, _RET_IP_);
79 	atomic_inc(&pag->pag_ref);
80 	rcu_read_unlock();
81 	return pag;
82 }
83 
84 /* Get a passive reference to the given perag. */
85 struct xfs_perag *
86 xfs_perag_hold(
87 	struct xfs_perag	*pag)
88 {
89 	ASSERT(atomic_read(&pag->pag_ref) > 0 ||
90 	       atomic_read(&pag->pag_active_ref) > 0);
91 
92 	trace_xfs_perag_hold(pag, _RET_IP_);
93 	atomic_inc(&pag->pag_ref);
94 	return pag;
95 }
96 
97 void
98 xfs_perag_put(
99 	struct xfs_perag	*pag)
100 {
101 	trace_xfs_perag_put(pag, _RET_IP_);
102 	ASSERT(atomic_read(&pag->pag_ref) > 0);
103 	atomic_dec(&pag->pag_ref);
104 }
105 
106 /*
107  * Active references for perag structures. This is for short term access to the
108  * per ag structures for walking trees or accessing state. If an AG is being
109  * shrunk or is offline, then this will fail to find that AG and return NULL
110  * instead.
111  */
112 struct xfs_perag *
113 xfs_perag_grab(
114 	struct xfs_mount	*mp,
115 	xfs_agnumber_t		agno)
116 {
117 	struct xfs_perag	*pag;
118 
119 	rcu_read_lock();
120 	pag = radix_tree_lookup(&mp->m_perag_tree, agno);
121 	if (pag) {
122 		trace_xfs_perag_grab(pag, _RET_IP_);
123 		if (!atomic_inc_not_zero(&pag->pag_active_ref))
124 			pag = NULL;
125 	}
126 	rcu_read_unlock();
127 	return pag;
128 }
129 
130 /*
131  * search from @first to find the next perag with the given tag set.
132  */
133 struct xfs_perag *
134 xfs_perag_grab_tag(
135 	struct xfs_mount	*mp,
136 	xfs_agnumber_t		first,
137 	int			tag)
138 {
139 	struct xfs_perag	*pag;
140 	int			found;
141 
142 	rcu_read_lock();
143 	found = radix_tree_gang_lookup_tag(&mp->m_perag_tree,
144 					(void **)&pag, first, 1, tag);
145 	if (found <= 0) {
146 		rcu_read_unlock();
147 		return NULL;
148 	}
149 	trace_xfs_perag_grab_tag(pag, _RET_IP_);
150 	if (!atomic_inc_not_zero(&pag->pag_active_ref))
151 		pag = NULL;
152 	rcu_read_unlock();
153 	return pag;
154 }
155 
156 void
157 xfs_perag_rele(
158 	struct xfs_perag	*pag)
159 {
160 	trace_xfs_perag_rele(pag, _RET_IP_);
161 	if (atomic_dec_and_test(&pag->pag_active_ref))
162 		wake_up(&pag->pag_active_wq);
163 }
164 
165 /*
166  * xfs_initialize_perag_data
167  *
168  * Read in each per-ag structure so we can count up the number of
169  * allocated inodes, free inodes and used filesystem blocks as this
170  * information is no longer persistent in the superblock. Once we have
171  * this information, write it into the in-core superblock structure.
172  */
173 int
174 xfs_initialize_perag_data(
175 	struct xfs_mount	*mp,
176 	xfs_agnumber_t		agcount)
177 {
178 	xfs_agnumber_t		index;
179 	struct xfs_perag	*pag;
180 	struct xfs_sb		*sbp = &mp->m_sb;
181 	uint64_t		ifree = 0;
182 	uint64_t		ialloc = 0;
183 	uint64_t		bfree = 0;
184 	uint64_t		bfreelst = 0;
185 	uint64_t		btree = 0;
186 	uint64_t		fdblocks;
187 	int			error = 0;
188 
189 	for (index = 0; index < agcount; index++) {
190 		/*
191 		 * Read the AGF and AGI buffers to populate the per-ag
192 		 * structures for us.
193 		 */
194 		pag = xfs_perag_get(mp, index);
195 		error = xfs_alloc_read_agf(pag, NULL, 0, NULL);
196 		if (!error)
197 			error = xfs_ialloc_read_agi(pag, NULL, NULL);
198 		if (error) {
199 			xfs_perag_put(pag);
200 			return error;
201 		}
202 
203 		ifree += pag->pagi_freecount;
204 		ialloc += pag->pagi_count;
205 		bfree += pag->pagf_freeblks;
206 		bfreelst += pag->pagf_flcount;
207 		btree += pag->pagf_btreeblks;
208 		xfs_perag_put(pag);
209 	}
210 	fdblocks = bfree + bfreelst + btree;
211 
212 	/*
213 	 * If the new summary counts are obviously incorrect, fail the
214 	 * mount operation because that implies the AGFs are also corrupt.
215 	 * Clear FS_COUNTERS so that we don't unmount with a dirty log, which
216 	 * will prevent xfs_repair from fixing anything.
217 	 */
218 	if (fdblocks > sbp->sb_dblocks || ifree > ialloc) {
219 		xfs_alert(mp, "AGF corruption. Please run xfs_repair.");
220 		xfs_fs_mark_sick(mp, XFS_SICK_FS_COUNTERS);
221 		error = -EFSCORRUPTED;
222 		goto out;
223 	}
224 
225 	/* Overwrite incore superblock counters with just-read data */
226 	spin_lock(&mp->m_sb_lock);
227 	sbp->sb_ifree = ifree;
228 	sbp->sb_icount = ialloc;
229 	sbp->sb_fdblocks = fdblocks;
230 	spin_unlock(&mp->m_sb_lock);
231 
232 	xfs_reinit_percpu_counters(mp);
233 out:
234 	xfs_fs_mark_healthy(mp, XFS_SICK_FS_COUNTERS);
235 	return error;
236 }
237 
238 STATIC void
239 __xfs_free_perag(
240 	struct rcu_head	*head)
241 {
242 	struct xfs_perag *pag = container_of(head, struct xfs_perag, rcu_head);
243 
244 	ASSERT(!delayed_work_pending(&pag->pag_blockgc_work));
245 	kfree(pag);
246 }
247 
248 /*
249  * Free up the per-ag resources associated with the mount structure.
250  */
251 void
252 xfs_free_perag(
253 	struct xfs_mount	*mp)
254 {
255 	struct xfs_perag	*pag;
256 	xfs_agnumber_t		agno;
257 
258 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
259 		spin_lock(&mp->m_perag_lock);
260 		pag = radix_tree_delete(&mp->m_perag_tree, agno);
261 		spin_unlock(&mp->m_perag_lock);
262 		ASSERT(pag);
263 		XFS_IS_CORRUPT(pag->pag_mount, atomic_read(&pag->pag_ref) != 0);
264 		xfs_defer_drain_free(&pag->pag_intents_drain);
265 
266 		cancel_delayed_work_sync(&pag->pag_blockgc_work);
267 		xfs_buf_cache_destroy(&pag->pag_bcache);
268 
269 		/* drop the mount's active reference */
270 		xfs_perag_rele(pag);
271 		XFS_IS_CORRUPT(pag->pag_mount,
272 				atomic_read(&pag->pag_active_ref) != 0);
273 		call_rcu(&pag->rcu_head, __xfs_free_perag);
274 	}
275 }
276 
277 /* Find the size of the AG, in blocks. */
278 static xfs_agblock_t
279 __xfs_ag_block_count(
280 	struct xfs_mount	*mp,
281 	xfs_agnumber_t		agno,
282 	xfs_agnumber_t		agcount,
283 	xfs_rfsblock_t		dblocks)
284 {
285 	ASSERT(agno < agcount);
286 
287 	if (agno < agcount - 1)
288 		return mp->m_sb.sb_agblocks;
289 	return dblocks - (agno * mp->m_sb.sb_agblocks);
290 }
291 
292 xfs_agblock_t
293 xfs_ag_block_count(
294 	struct xfs_mount	*mp,
295 	xfs_agnumber_t		agno)
296 {
297 	return __xfs_ag_block_count(mp, agno, mp->m_sb.sb_agcount,
298 			mp->m_sb.sb_dblocks);
299 }
300 
301 /* Calculate the first and last possible inode number in an AG. */
302 static void
303 __xfs_agino_range(
304 	struct xfs_mount	*mp,
305 	xfs_agblock_t		eoag,
306 	xfs_agino_t		*first,
307 	xfs_agino_t		*last)
308 {
309 	xfs_agblock_t		bno;
310 
311 	/*
312 	 * Calculate the first inode, which will be in the first
313 	 * cluster-aligned block after the AGFL.
314 	 */
315 	bno = round_up(XFS_AGFL_BLOCK(mp) + 1, M_IGEO(mp)->cluster_align);
316 	*first = XFS_AGB_TO_AGINO(mp, bno);
317 
318 	/*
319 	 * Calculate the last inode, which will be at the end of the
320 	 * last (aligned) cluster that can be allocated in the AG.
321 	 */
322 	bno = round_down(eoag, M_IGEO(mp)->cluster_align);
323 	*last = XFS_AGB_TO_AGINO(mp, bno) - 1;
324 }
325 
326 void
327 xfs_agino_range(
328 	struct xfs_mount	*mp,
329 	xfs_agnumber_t		agno,
330 	xfs_agino_t		*first,
331 	xfs_agino_t		*last)
332 {
333 	return __xfs_agino_range(mp, xfs_ag_block_count(mp, agno), first, last);
334 }
335 
336 /*
337  * Free perag within the specified AG range, it is only used to free unused
338  * perags under the error handling path.
339  */
340 void
341 xfs_free_unused_perag_range(
342 	struct xfs_mount	*mp,
343 	xfs_agnumber_t		agstart,
344 	xfs_agnumber_t		agend)
345 {
346 	struct xfs_perag	*pag;
347 	xfs_agnumber_t		index;
348 
349 	for (index = agstart; index < agend; index++) {
350 		spin_lock(&mp->m_perag_lock);
351 		pag = radix_tree_delete(&mp->m_perag_tree, index);
352 		spin_unlock(&mp->m_perag_lock);
353 		if (!pag)
354 			break;
355 		xfs_buf_cache_destroy(&pag->pag_bcache);
356 		xfs_defer_drain_free(&pag->pag_intents_drain);
357 		kfree(pag);
358 	}
359 }
360 
361 int
362 xfs_initialize_perag(
363 	struct xfs_mount	*mp,
364 	xfs_agnumber_t		agcount,
365 	xfs_rfsblock_t		dblocks,
366 	xfs_agnumber_t		*maxagi)
367 {
368 	struct xfs_perag	*pag;
369 	xfs_agnumber_t		index;
370 	xfs_agnumber_t		first_initialised = NULLAGNUMBER;
371 	int			error;
372 
373 	/*
374 	 * Walk the current per-ag tree so we don't try to initialise AGs
375 	 * that already exist (growfs case). Allocate and insert all the
376 	 * AGs we don't find ready for initialisation.
377 	 */
378 	for (index = 0; index < agcount; index++) {
379 		pag = xfs_perag_get(mp, index);
380 		if (pag) {
381 			xfs_perag_put(pag);
382 			continue;
383 		}
384 
385 		pag = kzalloc(sizeof(*pag), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
386 		if (!pag) {
387 			error = -ENOMEM;
388 			goto out_unwind_new_pags;
389 		}
390 		pag->pag_agno = index;
391 		pag->pag_mount = mp;
392 
393 		error = radix_tree_preload(GFP_KERNEL | __GFP_RETRY_MAYFAIL);
394 		if (error)
395 			goto out_free_pag;
396 
397 		spin_lock(&mp->m_perag_lock);
398 		if (radix_tree_insert(&mp->m_perag_tree, index, pag)) {
399 			WARN_ON_ONCE(1);
400 			spin_unlock(&mp->m_perag_lock);
401 			radix_tree_preload_end();
402 			error = -EEXIST;
403 			goto out_free_pag;
404 		}
405 		spin_unlock(&mp->m_perag_lock);
406 		radix_tree_preload_end();
407 
408 #ifdef __KERNEL__
409 		/* Place kernel structure only init below this point. */
410 		spin_lock_init(&pag->pag_ici_lock);
411 		spin_lock_init(&pag->pagb_lock);
412 		spin_lock_init(&pag->pag_state_lock);
413 		INIT_DELAYED_WORK(&pag->pag_blockgc_work, xfs_blockgc_worker);
414 		INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
415 		xfs_defer_drain_init(&pag->pag_intents_drain);
416 		init_waitqueue_head(&pag->pagb_wait);
417 		init_waitqueue_head(&pag->pag_active_wq);
418 		pag->pagb_count = 0;
419 		pag->pagb_tree = RB_ROOT;
420 		xfs_hooks_init(&pag->pag_rmap_update_hooks);
421 #endif /* __KERNEL__ */
422 
423 		error = xfs_buf_cache_init(&pag->pag_bcache);
424 		if (error)
425 			goto out_remove_pag;
426 
427 		/* Active ref owned by mount indicates AG is online. */
428 		atomic_set(&pag->pag_active_ref, 1);
429 
430 		/* first new pag is fully initialized */
431 		if (first_initialised == NULLAGNUMBER)
432 			first_initialised = index;
433 
434 		/*
435 		 * Pre-calculated geometry
436 		 */
437 		pag->block_count = __xfs_ag_block_count(mp, index, agcount,
438 				dblocks);
439 		pag->min_block = XFS_AGFL_BLOCK(mp);
440 		__xfs_agino_range(mp, pag->block_count, &pag->agino_min,
441 				&pag->agino_max);
442 	}
443 
444 	index = xfs_set_inode_alloc(mp, agcount);
445 
446 	if (maxagi)
447 		*maxagi = index;
448 
449 	mp->m_ag_prealloc_blocks = xfs_prealloc_blocks(mp);
450 	return 0;
451 
452 out_remove_pag:
453 	xfs_defer_drain_free(&pag->pag_intents_drain);
454 	spin_lock(&mp->m_perag_lock);
455 	radix_tree_delete(&mp->m_perag_tree, index);
456 	spin_unlock(&mp->m_perag_lock);
457 out_free_pag:
458 	kfree(pag);
459 out_unwind_new_pags:
460 	/* unwind any prior newly initialized pags */
461 	xfs_free_unused_perag_range(mp, first_initialised, agcount);
462 	return error;
463 }
464 
465 static int
466 xfs_get_aghdr_buf(
467 	struct xfs_mount	*mp,
468 	xfs_daddr_t		blkno,
469 	size_t			numblks,
470 	struct xfs_buf		**bpp,
471 	const struct xfs_buf_ops *ops)
472 {
473 	struct xfs_buf		*bp;
474 	int			error;
475 
476 	error = xfs_buf_get_uncached(mp->m_ddev_targp, numblks, 0, &bp);
477 	if (error)
478 		return error;
479 
480 	bp->b_maps[0].bm_bn = blkno;
481 	bp->b_ops = ops;
482 
483 	*bpp = bp;
484 	return 0;
485 }
486 
487 /*
488  * Generic btree root block init function
489  */
490 static void
491 xfs_btroot_init(
492 	struct xfs_mount	*mp,
493 	struct xfs_buf		*bp,
494 	struct aghdr_init_data	*id)
495 {
496 	xfs_btree_init_buf(mp, bp, id->bc_ops, 0, 0, id->agno);
497 }
498 
499 /* Finish initializing a free space btree. */
500 static void
501 xfs_freesp_init_recs(
502 	struct xfs_mount	*mp,
503 	struct xfs_buf		*bp,
504 	struct aghdr_init_data	*id)
505 {
506 	struct xfs_alloc_rec	*arec;
507 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
508 
509 	arec = XFS_ALLOC_REC_ADDR(mp, XFS_BUF_TO_BLOCK(bp), 1);
510 	arec->ar_startblock = cpu_to_be32(mp->m_ag_prealloc_blocks);
511 
512 	if (xfs_ag_contains_log(mp, id->agno)) {
513 		struct xfs_alloc_rec	*nrec;
514 		xfs_agblock_t		start = XFS_FSB_TO_AGBNO(mp,
515 							mp->m_sb.sb_logstart);
516 
517 		ASSERT(start >= mp->m_ag_prealloc_blocks);
518 		if (start != mp->m_ag_prealloc_blocks) {
519 			/*
520 			 * Modify first record to pad stripe align of log and
521 			 * bump the record count.
522 			 */
523 			arec->ar_blockcount = cpu_to_be32(start -
524 						mp->m_ag_prealloc_blocks);
525 			be16_add_cpu(&block->bb_numrecs, 1);
526 			nrec = arec + 1;
527 
528 			/*
529 			 * Insert second record at start of internal log
530 			 * which then gets trimmed.
531 			 */
532 			nrec->ar_startblock = cpu_to_be32(
533 					be32_to_cpu(arec->ar_startblock) +
534 					be32_to_cpu(arec->ar_blockcount));
535 			arec = nrec;
536 		}
537 		/*
538 		 * Change record start to after the internal log
539 		 */
540 		be32_add_cpu(&arec->ar_startblock, mp->m_sb.sb_logblocks);
541 	}
542 
543 	/*
544 	 * Calculate the block count of this record; if it is nonzero,
545 	 * increment the record count.
546 	 */
547 	arec->ar_blockcount = cpu_to_be32(id->agsize -
548 					  be32_to_cpu(arec->ar_startblock));
549 	if (arec->ar_blockcount)
550 		be16_add_cpu(&block->bb_numrecs, 1);
551 }
552 
553 /*
554  * bnobt/cntbt btree root block init functions
555  */
556 static void
557 xfs_bnoroot_init(
558 	struct xfs_mount	*mp,
559 	struct xfs_buf		*bp,
560 	struct aghdr_init_data	*id)
561 {
562 	xfs_btree_init_buf(mp, bp, id->bc_ops, 0, 0, id->agno);
563 	xfs_freesp_init_recs(mp, bp, id);
564 }
565 
566 /*
567  * Reverse map root block init
568  */
569 static void
570 xfs_rmaproot_init(
571 	struct xfs_mount	*mp,
572 	struct xfs_buf		*bp,
573 	struct aghdr_init_data	*id)
574 {
575 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
576 	struct xfs_rmap_rec	*rrec;
577 
578 	xfs_btree_init_buf(mp, bp, id->bc_ops, 0, 4, id->agno);
579 
580 	/*
581 	 * mark the AG header regions as static metadata The BNO
582 	 * btree block is the first block after the headers, so
583 	 * it's location defines the size of region the static
584 	 * metadata consumes.
585 	 *
586 	 * Note: unlike mkfs, we never have to account for log
587 	 * space when growing the data regions
588 	 */
589 	rrec = XFS_RMAP_REC_ADDR(block, 1);
590 	rrec->rm_startblock = 0;
591 	rrec->rm_blockcount = cpu_to_be32(XFS_BNO_BLOCK(mp));
592 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_FS);
593 	rrec->rm_offset = 0;
594 
595 	/* account freespace btree root blocks */
596 	rrec = XFS_RMAP_REC_ADDR(block, 2);
597 	rrec->rm_startblock = cpu_to_be32(XFS_BNO_BLOCK(mp));
598 	rrec->rm_blockcount = cpu_to_be32(2);
599 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_AG);
600 	rrec->rm_offset = 0;
601 
602 	/* account inode btree root blocks */
603 	rrec = XFS_RMAP_REC_ADDR(block, 3);
604 	rrec->rm_startblock = cpu_to_be32(XFS_IBT_BLOCK(mp));
605 	rrec->rm_blockcount = cpu_to_be32(XFS_RMAP_BLOCK(mp) -
606 					  XFS_IBT_BLOCK(mp));
607 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_INOBT);
608 	rrec->rm_offset = 0;
609 
610 	/* account for rmap btree root */
611 	rrec = XFS_RMAP_REC_ADDR(block, 4);
612 	rrec->rm_startblock = cpu_to_be32(XFS_RMAP_BLOCK(mp));
613 	rrec->rm_blockcount = cpu_to_be32(1);
614 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_AG);
615 	rrec->rm_offset = 0;
616 
617 	/* account for refc btree root */
618 	if (xfs_has_reflink(mp)) {
619 		rrec = XFS_RMAP_REC_ADDR(block, 5);
620 		rrec->rm_startblock = cpu_to_be32(xfs_refc_block(mp));
621 		rrec->rm_blockcount = cpu_to_be32(1);
622 		rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_REFC);
623 		rrec->rm_offset = 0;
624 		be16_add_cpu(&block->bb_numrecs, 1);
625 	}
626 
627 	/* account for the log space */
628 	if (xfs_ag_contains_log(mp, id->agno)) {
629 		rrec = XFS_RMAP_REC_ADDR(block,
630 				be16_to_cpu(block->bb_numrecs) + 1);
631 		rrec->rm_startblock = cpu_to_be32(
632 				XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart));
633 		rrec->rm_blockcount = cpu_to_be32(mp->m_sb.sb_logblocks);
634 		rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_LOG);
635 		rrec->rm_offset = 0;
636 		be16_add_cpu(&block->bb_numrecs, 1);
637 	}
638 }
639 
640 /*
641  * Initialise new secondary superblocks with the pre-grow geometry, but mark
642  * them as "in progress" so we know they haven't yet been activated. This will
643  * get cleared when the update with the new geometry information is done after
644  * changes to the primary are committed. This isn't strictly necessary, but we
645  * get it for free with the delayed buffer write lists and it means we can tell
646  * if a grow operation didn't complete properly after the fact.
647  */
648 static void
649 xfs_sbblock_init(
650 	struct xfs_mount	*mp,
651 	struct xfs_buf		*bp,
652 	struct aghdr_init_data	*id)
653 {
654 	struct xfs_dsb		*dsb = bp->b_addr;
655 
656 	xfs_sb_to_disk(dsb, &mp->m_sb);
657 	dsb->sb_inprogress = 1;
658 }
659 
660 static void
661 xfs_agfblock_init(
662 	struct xfs_mount	*mp,
663 	struct xfs_buf		*bp,
664 	struct aghdr_init_data	*id)
665 {
666 	struct xfs_agf		*agf = bp->b_addr;
667 	xfs_extlen_t		tmpsize;
668 
669 	agf->agf_magicnum = cpu_to_be32(XFS_AGF_MAGIC);
670 	agf->agf_versionnum = cpu_to_be32(XFS_AGF_VERSION);
671 	agf->agf_seqno = cpu_to_be32(id->agno);
672 	agf->agf_length = cpu_to_be32(id->agsize);
673 	agf->agf_bno_root = cpu_to_be32(XFS_BNO_BLOCK(mp));
674 	agf->agf_cnt_root = cpu_to_be32(XFS_CNT_BLOCK(mp));
675 	agf->agf_bno_level = cpu_to_be32(1);
676 	agf->agf_cnt_level = cpu_to_be32(1);
677 	if (xfs_has_rmapbt(mp)) {
678 		agf->agf_rmap_root = cpu_to_be32(XFS_RMAP_BLOCK(mp));
679 		agf->agf_rmap_level = cpu_to_be32(1);
680 		agf->agf_rmap_blocks = cpu_to_be32(1);
681 	}
682 
683 	agf->agf_flfirst = cpu_to_be32(1);
684 	agf->agf_fllast = 0;
685 	agf->agf_flcount = 0;
686 	tmpsize = id->agsize - mp->m_ag_prealloc_blocks;
687 	agf->agf_freeblks = cpu_to_be32(tmpsize);
688 	agf->agf_longest = cpu_to_be32(tmpsize);
689 	if (xfs_has_crc(mp))
690 		uuid_copy(&agf->agf_uuid, &mp->m_sb.sb_meta_uuid);
691 	if (xfs_has_reflink(mp)) {
692 		agf->agf_refcount_root = cpu_to_be32(
693 				xfs_refc_block(mp));
694 		agf->agf_refcount_level = cpu_to_be32(1);
695 		agf->agf_refcount_blocks = cpu_to_be32(1);
696 	}
697 
698 	if (xfs_ag_contains_log(mp, id->agno)) {
699 		int64_t	logblocks = mp->m_sb.sb_logblocks;
700 
701 		be32_add_cpu(&agf->agf_freeblks, -logblocks);
702 		agf->agf_longest = cpu_to_be32(id->agsize -
703 			XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart) - logblocks);
704 	}
705 }
706 
707 static void
708 xfs_agflblock_init(
709 	struct xfs_mount	*mp,
710 	struct xfs_buf		*bp,
711 	struct aghdr_init_data	*id)
712 {
713 	struct xfs_agfl		*agfl = XFS_BUF_TO_AGFL(bp);
714 	__be32			*agfl_bno;
715 	int			bucket;
716 
717 	if (xfs_has_crc(mp)) {
718 		agfl->agfl_magicnum = cpu_to_be32(XFS_AGFL_MAGIC);
719 		agfl->agfl_seqno = cpu_to_be32(id->agno);
720 		uuid_copy(&agfl->agfl_uuid, &mp->m_sb.sb_meta_uuid);
721 	}
722 
723 	agfl_bno = xfs_buf_to_agfl_bno(bp);
724 	for (bucket = 0; bucket < xfs_agfl_size(mp); bucket++)
725 		agfl_bno[bucket] = cpu_to_be32(NULLAGBLOCK);
726 }
727 
728 static void
729 xfs_agiblock_init(
730 	struct xfs_mount	*mp,
731 	struct xfs_buf		*bp,
732 	struct aghdr_init_data	*id)
733 {
734 	struct xfs_agi		*agi = bp->b_addr;
735 	int			bucket;
736 
737 	agi->agi_magicnum = cpu_to_be32(XFS_AGI_MAGIC);
738 	agi->agi_versionnum = cpu_to_be32(XFS_AGI_VERSION);
739 	agi->agi_seqno = cpu_to_be32(id->agno);
740 	agi->agi_length = cpu_to_be32(id->agsize);
741 	agi->agi_count = 0;
742 	agi->agi_root = cpu_to_be32(XFS_IBT_BLOCK(mp));
743 	agi->agi_level = cpu_to_be32(1);
744 	agi->agi_freecount = 0;
745 	agi->agi_newino = cpu_to_be32(NULLAGINO);
746 	agi->agi_dirino = cpu_to_be32(NULLAGINO);
747 	if (xfs_has_crc(mp))
748 		uuid_copy(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid);
749 	if (xfs_has_finobt(mp)) {
750 		agi->agi_free_root = cpu_to_be32(XFS_FIBT_BLOCK(mp));
751 		agi->agi_free_level = cpu_to_be32(1);
752 	}
753 	for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++)
754 		agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
755 	if (xfs_has_inobtcounts(mp)) {
756 		agi->agi_iblocks = cpu_to_be32(1);
757 		if (xfs_has_finobt(mp))
758 			agi->agi_fblocks = cpu_to_be32(1);
759 	}
760 }
761 
762 typedef void (*aghdr_init_work_f)(struct xfs_mount *mp, struct xfs_buf *bp,
763 				  struct aghdr_init_data *id);
764 static int
765 xfs_ag_init_hdr(
766 	struct xfs_mount	*mp,
767 	struct aghdr_init_data	*id,
768 	aghdr_init_work_f	work,
769 	const struct xfs_buf_ops *ops)
770 {
771 	struct xfs_buf		*bp;
772 	int			error;
773 
774 	error = xfs_get_aghdr_buf(mp, id->daddr, id->numblks, &bp, ops);
775 	if (error)
776 		return error;
777 
778 	(*work)(mp, bp, id);
779 
780 	xfs_buf_delwri_queue(bp, &id->buffer_list);
781 	xfs_buf_relse(bp);
782 	return 0;
783 }
784 
785 struct xfs_aghdr_grow_data {
786 	xfs_daddr_t		daddr;
787 	size_t			numblks;
788 	const struct xfs_buf_ops *ops;
789 	aghdr_init_work_f	work;
790 	const struct xfs_btree_ops *bc_ops;
791 	bool			need_init;
792 };
793 
794 /*
795  * Prepare new AG headers to be written to disk. We use uncached buffers here,
796  * as it is assumed these new AG headers are currently beyond the currently
797  * valid filesystem address space. Using cached buffers would trip over EOFS
798  * corruption detection alogrithms in the buffer cache lookup routines.
799  *
800  * This is a non-transactional function, but the prepared buffers are added to a
801  * delayed write buffer list supplied by the caller so they can submit them to
802  * disk and wait on them as required.
803  */
804 int
805 xfs_ag_init_headers(
806 	struct xfs_mount	*mp,
807 	struct aghdr_init_data	*id)
808 
809 {
810 	struct xfs_aghdr_grow_data aghdr_data[] = {
811 	{ /* SB */
812 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_SB_DADDR),
813 		.numblks = XFS_FSS_TO_BB(mp, 1),
814 		.ops = &xfs_sb_buf_ops,
815 		.work = &xfs_sbblock_init,
816 		.need_init = true
817 	},
818 	{ /* AGF */
819 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGF_DADDR(mp)),
820 		.numblks = XFS_FSS_TO_BB(mp, 1),
821 		.ops = &xfs_agf_buf_ops,
822 		.work = &xfs_agfblock_init,
823 		.need_init = true
824 	},
825 	{ /* AGFL */
826 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGFL_DADDR(mp)),
827 		.numblks = XFS_FSS_TO_BB(mp, 1),
828 		.ops = &xfs_agfl_buf_ops,
829 		.work = &xfs_agflblock_init,
830 		.need_init = true
831 	},
832 	{ /* AGI */
833 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGI_DADDR(mp)),
834 		.numblks = XFS_FSS_TO_BB(mp, 1),
835 		.ops = &xfs_agi_buf_ops,
836 		.work = &xfs_agiblock_init,
837 		.need_init = true
838 	},
839 	{ /* BNO root block */
840 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_BNO_BLOCK(mp)),
841 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
842 		.ops = &xfs_bnobt_buf_ops,
843 		.work = &xfs_bnoroot_init,
844 		.bc_ops = &xfs_bnobt_ops,
845 		.need_init = true
846 	},
847 	{ /* CNT root block */
848 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_CNT_BLOCK(mp)),
849 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
850 		.ops = &xfs_cntbt_buf_ops,
851 		.work = &xfs_bnoroot_init,
852 		.bc_ops = &xfs_cntbt_ops,
853 		.need_init = true
854 	},
855 	{ /* INO root block */
856 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_IBT_BLOCK(mp)),
857 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
858 		.ops = &xfs_inobt_buf_ops,
859 		.work = &xfs_btroot_init,
860 		.bc_ops = &xfs_inobt_ops,
861 		.need_init = true
862 	},
863 	{ /* FINO root block */
864 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_FIBT_BLOCK(mp)),
865 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
866 		.ops = &xfs_finobt_buf_ops,
867 		.work = &xfs_btroot_init,
868 		.bc_ops = &xfs_finobt_ops,
869 		.need_init =  xfs_has_finobt(mp)
870 	},
871 	{ /* RMAP root block */
872 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_RMAP_BLOCK(mp)),
873 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
874 		.ops = &xfs_rmapbt_buf_ops,
875 		.work = &xfs_rmaproot_init,
876 		.bc_ops = &xfs_rmapbt_ops,
877 		.need_init = xfs_has_rmapbt(mp)
878 	},
879 	{ /* REFC root block */
880 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, xfs_refc_block(mp)),
881 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
882 		.ops = &xfs_refcountbt_buf_ops,
883 		.work = &xfs_btroot_init,
884 		.bc_ops = &xfs_refcountbt_ops,
885 		.need_init = xfs_has_reflink(mp)
886 	},
887 	{ /* NULL terminating block */
888 		.daddr = XFS_BUF_DADDR_NULL,
889 	}
890 	};
891 	struct  xfs_aghdr_grow_data *dp;
892 	int			error = 0;
893 
894 	/* Account for AG free space in new AG */
895 	id->nfree += id->agsize - mp->m_ag_prealloc_blocks;
896 	for (dp = &aghdr_data[0]; dp->daddr != XFS_BUF_DADDR_NULL; dp++) {
897 		if (!dp->need_init)
898 			continue;
899 
900 		id->daddr = dp->daddr;
901 		id->numblks = dp->numblks;
902 		id->bc_ops = dp->bc_ops;
903 		error = xfs_ag_init_hdr(mp, id, dp->work, dp->ops);
904 		if (error)
905 			break;
906 	}
907 	return error;
908 }
909 
910 int
911 xfs_ag_shrink_space(
912 	struct xfs_perag	*pag,
913 	struct xfs_trans	**tpp,
914 	xfs_extlen_t		delta)
915 {
916 	struct xfs_mount	*mp = pag->pag_mount;
917 	struct xfs_alloc_arg	args = {
918 		.tp	= *tpp,
919 		.mp	= mp,
920 		.pag	= pag,
921 		.minlen = delta,
922 		.maxlen = delta,
923 		.oinfo	= XFS_RMAP_OINFO_SKIP_UPDATE,
924 		.resv	= XFS_AG_RESV_NONE,
925 		.prod	= 1
926 	};
927 	struct xfs_buf		*agibp, *agfbp;
928 	struct xfs_agi		*agi;
929 	struct xfs_agf		*agf;
930 	xfs_agblock_t		aglen;
931 	int			error, err2;
932 
933 	ASSERT(pag->pag_agno == mp->m_sb.sb_agcount - 1);
934 	error = xfs_ialloc_read_agi(pag, *tpp, &agibp);
935 	if (error)
936 		return error;
937 
938 	agi = agibp->b_addr;
939 
940 	error = xfs_alloc_read_agf(pag, *tpp, 0, &agfbp);
941 	if (error)
942 		return error;
943 
944 	agf = agfbp->b_addr;
945 	aglen = be32_to_cpu(agi->agi_length);
946 	/* some extra paranoid checks before we shrink the ag */
947 	if (XFS_IS_CORRUPT(mp, agf->agf_length != agi->agi_length)) {
948 		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGF);
949 		return -EFSCORRUPTED;
950 	}
951 	if (delta >= aglen)
952 		return -EINVAL;
953 
954 	/*
955 	 * Make sure that the last inode cluster cannot overlap with the new
956 	 * end of the AG, even if it's sparse.
957 	 */
958 	error = xfs_ialloc_check_shrink(pag, *tpp, agibp, aglen - delta);
959 	if (error)
960 		return error;
961 
962 	/*
963 	 * Disable perag reservations so it doesn't cause the allocation request
964 	 * to fail. We'll reestablish reservation before we return.
965 	 */
966 	error = xfs_ag_resv_free(pag);
967 	if (error)
968 		return error;
969 
970 	/* internal log shouldn't also show up in the free space btrees */
971 	error = xfs_alloc_vextent_exact_bno(&args,
972 			XFS_AGB_TO_FSB(mp, pag->pag_agno, aglen - delta));
973 	if (!error && args.agbno == NULLAGBLOCK)
974 		error = -ENOSPC;
975 
976 	if (error) {
977 		/*
978 		 * If extent allocation fails, need to roll the transaction to
979 		 * ensure that the AGFL fixup has been committed anyway.
980 		 *
981 		 * We need to hold the AGF across the roll to ensure nothing can
982 		 * access the AG for allocation until the shrink is fully
983 		 * cleaned up. And due to the resetting of the AG block
984 		 * reservation space needing to lock the AGI, we also have to
985 		 * hold that so we don't get AGI/AGF lock order inversions in
986 		 * the error handling path.
987 		 */
988 		xfs_trans_bhold(*tpp, agfbp);
989 		xfs_trans_bhold(*tpp, agibp);
990 		err2 = xfs_trans_roll(tpp);
991 		if (err2)
992 			return err2;
993 		xfs_trans_bjoin(*tpp, agfbp);
994 		xfs_trans_bjoin(*tpp, agibp);
995 		goto resv_init_out;
996 	}
997 
998 	/*
999 	 * if successfully deleted from freespace btrees, need to confirm
1000 	 * per-AG reservation works as expected.
1001 	 */
1002 	be32_add_cpu(&agi->agi_length, -delta);
1003 	be32_add_cpu(&agf->agf_length, -delta);
1004 
1005 	err2 = xfs_ag_resv_init(pag, *tpp);
1006 	if (err2) {
1007 		be32_add_cpu(&agi->agi_length, delta);
1008 		be32_add_cpu(&agf->agf_length, delta);
1009 		if (err2 != -ENOSPC)
1010 			goto resv_err;
1011 
1012 		err2 = xfs_free_extent_later(*tpp, args.fsbno, delta, NULL,
1013 				XFS_AG_RESV_NONE, true);
1014 		if (err2)
1015 			goto resv_err;
1016 
1017 		/*
1018 		 * Roll the transaction before trying to re-init the per-ag
1019 		 * reservation. The new transaction is clean so it will cancel
1020 		 * without any side effects.
1021 		 */
1022 		error = xfs_defer_finish(tpp);
1023 		if (error)
1024 			return error;
1025 
1026 		error = -ENOSPC;
1027 		goto resv_init_out;
1028 	}
1029 
1030 	/* Update perag geometry */
1031 	pag->block_count -= delta;
1032 	__xfs_agino_range(pag->pag_mount, pag->block_count, &pag->agino_min,
1033 				&pag->agino_max);
1034 
1035 	xfs_ialloc_log_agi(*tpp, agibp, XFS_AGI_LENGTH);
1036 	xfs_alloc_log_agf(*tpp, agfbp, XFS_AGF_LENGTH);
1037 	return 0;
1038 
1039 resv_init_out:
1040 	err2 = xfs_ag_resv_init(pag, *tpp);
1041 	if (!err2)
1042 		return error;
1043 resv_err:
1044 	xfs_warn(mp, "Error %d reserving per-AG metadata reserve pool.", err2);
1045 	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1046 	return err2;
1047 }
1048 
1049 /*
1050  * Extent the AG indicated by the @id by the length passed in
1051  */
1052 int
1053 xfs_ag_extend_space(
1054 	struct xfs_perag	*pag,
1055 	struct xfs_trans	*tp,
1056 	xfs_extlen_t		len)
1057 {
1058 	struct xfs_buf		*bp;
1059 	struct xfs_agi		*agi;
1060 	struct xfs_agf		*agf;
1061 	int			error;
1062 
1063 	ASSERT(pag->pag_agno == pag->pag_mount->m_sb.sb_agcount - 1);
1064 
1065 	error = xfs_ialloc_read_agi(pag, tp, &bp);
1066 	if (error)
1067 		return error;
1068 
1069 	agi = bp->b_addr;
1070 	be32_add_cpu(&agi->agi_length, len);
1071 	xfs_ialloc_log_agi(tp, bp, XFS_AGI_LENGTH);
1072 
1073 	/*
1074 	 * Change agf length.
1075 	 */
1076 	error = xfs_alloc_read_agf(pag, tp, 0, &bp);
1077 	if (error)
1078 		return error;
1079 
1080 	agf = bp->b_addr;
1081 	be32_add_cpu(&agf->agf_length, len);
1082 	ASSERT(agf->agf_length == agi->agi_length);
1083 	xfs_alloc_log_agf(tp, bp, XFS_AGF_LENGTH);
1084 
1085 	/*
1086 	 * Free the new space.
1087 	 *
1088 	 * XFS_RMAP_OINFO_SKIP_UPDATE is used here to tell the rmap btree that
1089 	 * this doesn't actually exist in the rmap btree.
1090 	 */
1091 	error = xfs_rmap_free(tp, bp, pag, be32_to_cpu(agf->agf_length) - len,
1092 				len, &XFS_RMAP_OINFO_SKIP_UPDATE);
1093 	if (error)
1094 		return error;
1095 
1096 	error = xfs_free_extent(tp, pag, be32_to_cpu(agf->agf_length) - len,
1097 			len, &XFS_RMAP_OINFO_SKIP_UPDATE, XFS_AG_RESV_NONE);
1098 	if (error)
1099 		return error;
1100 
1101 	/* Update perag geometry */
1102 	pag->block_count = be32_to_cpu(agf->agf_length);
1103 	__xfs_agino_range(pag->pag_mount, pag->block_count, &pag->agino_min,
1104 				&pag->agino_max);
1105 	return 0;
1106 }
1107 
1108 /* Retrieve AG geometry. */
1109 int
1110 xfs_ag_get_geometry(
1111 	struct xfs_perag	*pag,
1112 	struct xfs_ag_geometry	*ageo)
1113 {
1114 	struct xfs_buf		*agi_bp;
1115 	struct xfs_buf		*agf_bp;
1116 	struct xfs_agi		*agi;
1117 	struct xfs_agf		*agf;
1118 	unsigned int		freeblks;
1119 	int			error;
1120 
1121 	/* Lock the AG headers. */
1122 	error = xfs_ialloc_read_agi(pag, NULL, &agi_bp);
1123 	if (error)
1124 		return error;
1125 	error = xfs_alloc_read_agf(pag, NULL, 0, &agf_bp);
1126 	if (error)
1127 		goto out_agi;
1128 
1129 	/* Fill out form. */
1130 	memset(ageo, 0, sizeof(*ageo));
1131 	ageo->ag_number = pag->pag_agno;
1132 
1133 	agi = agi_bp->b_addr;
1134 	ageo->ag_icount = be32_to_cpu(agi->agi_count);
1135 	ageo->ag_ifree = be32_to_cpu(agi->agi_freecount);
1136 
1137 	agf = agf_bp->b_addr;
1138 	ageo->ag_length = be32_to_cpu(agf->agf_length);
1139 	freeblks = pag->pagf_freeblks +
1140 		   pag->pagf_flcount +
1141 		   pag->pagf_btreeblks -
1142 		   xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE);
1143 	ageo->ag_freeblks = freeblks;
1144 	xfs_ag_geom_health(pag, ageo);
1145 
1146 	/* Release resources. */
1147 	xfs_buf_relse(agf_bp);
1148 out_agi:
1149 	xfs_buf_relse(agi_bp);
1150 	return error;
1151 }
1152