xref: /linux/fs/xfs/libxfs/xfs_ag.c (revision 76f623d2d4283cc36a9c8a5b585df74638f1efa5)
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 		error = -EFSCORRUPTED;
221 		goto out;
222 	}
223 
224 	/* Overwrite incore superblock counters with just-read data */
225 	spin_lock(&mp->m_sb_lock);
226 	sbp->sb_ifree = ifree;
227 	sbp->sb_icount = ialloc;
228 	sbp->sb_fdblocks = fdblocks;
229 	spin_unlock(&mp->m_sb_lock);
230 
231 	xfs_reinit_percpu_counters(mp);
232 out:
233 	xfs_fs_mark_healthy(mp, XFS_SICK_FS_COUNTERS);
234 	return error;
235 }
236 
237 STATIC void
238 __xfs_free_perag(
239 	struct rcu_head	*head)
240 {
241 	struct xfs_perag *pag = container_of(head, struct xfs_perag, rcu_head);
242 
243 	ASSERT(!delayed_work_pending(&pag->pag_blockgc_work));
244 	kmem_free(pag);
245 }
246 
247 /*
248  * Free up the per-ag resources associated with the mount structure.
249  */
250 void
251 xfs_free_perag(
252 	struct xfs_mount	*mp)
253 {
254 	struct xfs_perag	*pag;
255 	xfs_agnumber_t		agno;
256 
257 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
258 		spin_lock(&mp->m_perag_lock);
259 		pag = radix_tree_delete(&mp->m_perag_tree, agno);
260 		spin_unlock(&mp->m_perag_lock);
261 		ASSERT(pag);
262 		XFS_IS_CORRUPT(pag->pag_mount, atomic_read(&pag->pag_ref) != 0);
263 		xfs_defer_drain_free(&pag->pag_intents_drain);
264 
265 		cancel_delayed_work_sync(&pag->pag_blockgc_work);
266 		xfs_buf_hash_destroy(pag);
267 
268 		/* drop the mount's active reference */
269 		xfs_perag_rele(pag);
270 		XFS_IS_CORRUPT(pag->pag_mount,
271 				atomic_read(&pag->pag_active_ref) != 0);
272 		call_rcu(&pag->rcu_head, __xfs_free_perag);
273 	}
274 }
275 
276 /* Find the size of the AG, in blocks. */
277 static xfs_agblock_t
278 __xfs_ag_block_count(
279 	struct xfs_mount	*mp,
280 	xfs_agnumber_t		agno,
281 	xfs_agnumber_t		agcount,
282 	xfs_rfsblock_t		dblocks)
283 {
284 	ASSERT(agno < agcount);
285 
286 	if (agno < agcount - 1)
287 		return mp->m_sb.sb_agblocks;
288 	return dblocks - (agno * mp->m_sb.sb_agblocks);
289 }
290 
291 xfs_agblock_t
292 xfs_ag_block_count(
293 	struct xfs_mount	*mp,
294 	xfs_agnumber_t		agno)
295 {
296 	return __xfs_ag_block_count(mp, agno, mp->m_sb.sb_agcount,
297 			mp->m_sb.sb_dblocks);
298 }
299 
300 /* Calculate the first and last possible inode number in an AG. */
301 static void
302 __xfs_agino_range(
303 	struct xfs_mount	*mp,
304 	xfs_agblock_t		eoag,
305 	xfs_agino_t		*first,
306 	xfs_agino_t		*last)
307 {
308 	xfs_agblock_t		bno;
309 
310 	/*
311 	 * Calculate the first inode, which will be in the first
312 	 * cluster-aligned block after the AGFL.
313 	 */
314 	bno = round_up(XFS_AGFL_BLOCK(mp) + 1, M_IGEO(mp)->cluster_align);
315 	*first = XFS_AGB_TO_AGINO(mp, bno);
316 
317 	/*
318 	 * Calculate the last inode, which will be at the end of the
319 	 * last (aligned) cluster that can be allocated in the AG.
320 	 */
321 	bno = round_down(eoag, M_IGEO(mp)->cluster_align);
322 	*last = XFS_AGB_TO_AGINO(mp, bno) - 1;
323 }
324 
325 void
326 xfs_agino_range(
327 	struct xfs_mount	*mp,
328 	xfs_agnumber_t		agno,
329 	xfs_agino_t		*first,
330 	xfs_agino_t		*last)
331 {
332 	return __xfs_agino_range(mp, xfs_ag_block_count(mp, agno), first, last);
333 }
334 
335 /*
336  * Free perag within the specified AG range, it is only used to free unused
337  * perags under the error handling path.
338  */
339 void
340 xfs_free_unused_perag_range(
341 	struct xfs_mount	*mp,
342 	xfs_agnumber_t		agstart,
343 	xfs_agnumber_t		agend)
344 {
345 	struct xfs_perag	*pag;
346 	xfs_agnumber_t		index;
347 
348 	for (index = agstart; index < agend; index++) {
349 		spin_lock(&mp->m_perag_lock);
350 		pag = radix_tree_delete(&mp->m_perag_tree, index);
351 		spin_unlock(&mp->m_perag_lock);
352 		if (!pag)
353 			break;
354 		xfs_buf_hash_destroy(pag);
355 		xfs_defer_drain_free(&pag->pag_intents_drain);
356 		kmem_free(pag);
357 	}
358 }
359 
360 int
361 xfs_initialize_perag(
362 	struct xfs_mount	*mp,
363 	xfs_agnumber_t		agcount,
364 	xfs_rfsblock_t		dblocks,
365 	xfs_agnumber_t		*maxagi)
366 {
367 	struct xfs_perag	*pag;
368 	xfs_agnumber_t		index;
369 	xfs_agnumber_t		first_initialised = NULLAGNUMBER;
370 	int			error;
371 
372 	/*
373 	 * Walk the current per-ag tree so we don't try to initialise AGs
374 	 * that already exist (growfs case). Allocate and insert all the
375 	 * AGs we don't find ready for initialisation.
376 	 */
377 	for (index = 0; index < agcount; index++) {
378 		pag = xfs_perag_get(mp, index);
379 		if (pag) {
380 			xfs_perag_put(pag);
381 			continue;
382 		}
383 
384 		pag = kmem_zalloc(sizeof(*pag), KM_MAYFAIL);
385 		if (!pag) {
386 			error = -ENOMEM;
387 			goto out_unwind_new_pags;
388 		}
389 		pag->pag_agno = index;
390 		pag->pag_mount = mp;
391 
392 		error = radix_tree_preload(GFP_NOFS);
393 		if (error)
394 			goto out_free_pag;
395 
396 		spin_lock(&mp->m_perag_lock);
397 		if (radix_tree_insert(&mp->m_perag_tree, index, pag)) {
398 			WARN_ON_ONCE(1);
399 			spin_unlock(&mp->m_perag_lock);
400 			radix_tree_preload_end();
401 			error = -EEXIST;
402 			goto out_free_pag;
403 		}
404 		spin_unlock(&mp->m_perag_lock);
405 		radix_tree_preload_end();
406 
407 #ifdef __KERNEL__
408 		/* Place kernel structure only init below this point. */
409 		spin_lock_init(&pag->pag_ici_lock);
410 		spin_lock_init(&pag->pagb_lock);
411 		spin_lock_init(&pag->pag_state_lock);
412 		INIT_DELAYED_WORK(&pag->pag_blockgc_work, xfs_blockgc_worker);
413 		INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
414 		xfs_defer_drain_init(&pag->pag_intents_drain);
415 		init_waitqueue_head(&pag->pagb_wait);
416 		init_waitqueue_head(&pag->pag_active_wq);
417 		pag->pagb_count = 0;
418 		pag->pagb_tree = RB_ROOT;
419 #endif /* __KERNEL__ */
420 
421 		error = xfs_buf_hash_init(pag);
422 		if (error)
423 			goto out_remove_pag;
424 
425 		/* Active ref owned by mount indicates AG is online. */
426 		atomic_set(&pag->pag_active_ref, 1);
427 
428 		/* first new pag is fully initialized */
429 		if (first_initialised == NULLAGNUMBER)
430 			first_initialised = index;
431 
432 		/*
433 		 * Pre-calculated geometry
434 		 */
435 		pag->block_count = __xfs_ag_block_count(mp, index, agcount,
436 				dblocks);
437 		pag->min_block = XFS_AGFL_BLOCK(mp);
438 		__xfs_agino_range(mp, pag->block_count, &pag->agino_min,
439 				&pag->agino_max);
440 	}
441 
442 	index = xfs_set_inode_alloc(mp, agcount);
443 
444 	if (maxagi)
445 		*maxagi = index;
446 
447 	mp->m_ag_prealloc_blocks = xfs_prealloc_blocks(mp);
448 	return 0;
449 
450 out_remove_pag:
451 	xfs_defer_drain_free(&pag->pag_intents_drain);
452 	spin_lock(&mp->m_perag_lock);
453 	radix_tree_delete(&mp->m_perag_tree, index);
454 	spin_unlock(&mp->m_perag_lock);
455 out_free_pag:
456 	kmem_free(pag);
457 out_unwind_new_pags:
458 	/* unwind any prior newly initialized pags */
459 	xfs_free_unused_perag_range(mp, first_initialised, agcount);
460 	return error;
461 }
462 
463 static int
464 xfs_get_aghdr_buf(
465 	struct xfs_mount	*mp,
466 	xfs_daddr_t		blkno,
467 	size_t			numblks,
468 	struct xfs_buf		**bpp,
469 	const struct xfs_buf_ops *ops)
470 {
471 	struct xfs_buf		*bp;
472 	int			error;
473 
474 	error = xfs_buf_get_uncached(mp->m_ddev_targp, numblks, 0, &bp);
475 	if (error)
476 		return error;
477 
478 	bp->b_maps[0].bm_bn = blkno;
479 	bp->b_ops = ops;
480 
481 	*bpp = bp;
482 	return 0;
483 }
484 
485 /*
486  * Generic btree root block init function
487  */
488 static void
489 xfs_btroot_init(
490 	struct xfs_mount	*mp,
491 	struct xfs_buf		*bp,
492 	struct aghdr_init_data	*id)
493 {
494 	xfs_btree_init_block(mp, bp, id->type, 0, 0, id->agno);
495 }
496 
497 /* Finish initializing a free space btree. */
498 static void
499 xfs_freesp_init_recs(
500 	struct xfs_mount	*mp,
501 	struct xfs_buf		*bp,
502 	struct aghdr_init_data	*id)
503 {
504 	struct xfs_alloc_rec	*arec;
505 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
506 
507 	arec = XFS_ALLOC_REC_ADDR(mp, XFS_BUF_TO_BLOCK(bp), 1);
508 	arec->ar_startblock = cpu_to_be32(mp->m_ag_prealloc_blocks);
509 
510 	if (xfs_ag_contains_log(mp, id->agno)) {
511 		struct xfs_alloc_rec	*nrec;
512 		xfs_agblock_t		start = XFS_FSB_TO_AGBNO(mp,
513 							mp->m_sb.sb_logstart);
514 
515 		ASSERT(start >= mp->m_ag_prealloc_blocks);
516 		if (start != mp->m_ag_prealloc_blocks) {
517 			/*
518 			 * Modify first record to pad stripe align of log and
519 			 * bump the record count.
520 			 */
521 			arec->ar_blockcount = cpu_to_be32(start -
522 						mp->m_ag_prealloc_blocks);
523 			be16_add_cpu(&block->bb_numrecs, 1);
524 			nrec = arec + 1;
525 
526 			/*
527 			 * Insert second record at start of internal log
528 			 * which then gets trimmed.
529 			 */
530 			nrec->ar_startblock = cpu_to_be32(
531 					be32_to_cpu(arec->ar_startblock) +
532 					be32_to_cpu(arec->ar_blockcount));
533 			arec = nrec;
534 		}
535 		/*
536 		 * Change record start to after the internal log
537 		 */
538 		be32_add_cpu(&arec->ar_startblock, mp->m_sb.sb_logblocks);
539 	}
540 
541 	/*
542 	 * Calculate the block count of this record; if it is nonzero,
543 	 * increment the record count.
544 	 */
545 	arec->ar_blockcount = cpu_to_be32(id->agsize -
546 					  be32_to_cpu(arec->ar_startblock));
547 	if (arec->ar_blockcount)
548 		be16_add_cpu(&block->bb_numrecs, 1);
549 }
550 
551 /*
552  * Alloc btree root block init functions
553  */
554 static void
555 xfs_bnoroot_init(
556 	struct xfs_mount	*mp,
557 	struct xfs_buf		*bp,
558 	struct aghdr_init_data	*id)
559 {
560 	xfs_btree_init_block(mp, bp, XFS_BTNUM_BNO, 0, 0, id->agno);
561 	xfs_freesp_init_recs(mp, bp, id);
562 }
563 
564 static void
565 xfs_cntroot_init(
566 	struct xfs_mount	*mp,
567 	struct xfs_buf		*bp,
568 	struct aghdr_init_data	*id)
569 {
570 	xfs_btree_init_block(mp, bp, XFS_BTNUM_CNT, 0, 0, id->agno);
571 	xfs_freesp_init_recs(mp, bp, id);
572 }
573 
574 /*
575  * Reverse map root block init
576  */
577 static void
578 xfs_rmaproot_init(
579 	struct xfs_mount	*mp,
580 	struct xfs_buf		*bp,
581 	struct aghdr_init_data	*id)
582 {
583 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
584 	struct xfs_rmap_rec	*rrec;
585 
586 	xfs_btree_init_block(mp, bp, XFS_BTNUM_RMAP, 0, 4, id->agno);
587 
588 	/*
589 	 * mark the AG header regions as static metadata The BNO
590 	 * btree block is the first block after the headers, so
591 	 * it's location defines the size of region the static
592 	 * metadata consumes.
593 	 *
594 	 * Note: unlike mkfs, we never have to account for log
595 	 * space when growing the data regions
596 	 */
597 	rrec = XFS_RMAP_REC_ADDR(block, 1);
598 	rrec->rm_startblock = 0;
599 	rrec->rm_blockcount = cpu_to_be32(XFS_BNO_BLOCK(mp));
600 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_FS);
601 	rrec->rm_offset = 0;
602 
603 	/* account freespace btree root blocks */
604 	rrec = XFS_RMAP_REC_ADDR(block, 2);
605 	rrec->rm_startblock = cpu_to_be32(XFS_BNO_BLOCK(mp));
606 	rrec->rm_blockcount = cpu_to_be32(2);
607 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_AG);
608 	rrec->rm_offset = 0;
609 
610 	/* account inode btree root blocks */
611 	rrec = XFS_RMAP_REC_ADDR(block, 3);
612 	rrec->rm_startblock = cpu_to_be32(XFS_IBT_BLOCK(mp));
613 	rrec->rm_blockcount = cpu_to_be32(XFS_RMAP_BLOCK(mp) -
614 					  XFS_IBT_BLOCK(mp));
615 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_INOBT);
616 	rrec->rm_offset = 0;
617 
618 	/* account for rmap btree root */
619 	rrec = XFS_RMAP_REC_ADDR(block, 4);
620 	rrec->rm_startblock = cpu_to_be32(XFS_RMAP_BLOCK(mp));
621 	rrec->rm_blockcount = cpu_to_be32(1);
622 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_AG);
623 	rrec->rm_offset = 0;
624 
625 	/* account for refc btree root */
626 	if (xfs_has_reflink(mp)) {
627 		rrec = XFS_RMAP_REC_ADDR(block, 5);
628 		rrec->rm_startblock = cpu_to_be32(xfs_refc_block(mp));
629 		rrec->rm_blockcount = cpu_to_be32(1);
630 		rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_REFC);
631 		rrec->rm_offset = 0;
632 		be16_add_cpu(&block->bb_numrecs, 1);
633 	}
634 
635 	/* account for the log space */
636 	if (xfs_ag_contains_log(mp, id->agno)) {
637 		rrec = XFS_RMAP_REC_ADDR(block,
638 				be16_to_cpu(block->bb_numrecs) + 1);
639 		rrec->rm_startblock = cpu_to_be32(
640 				XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart));
641 		rrec->rm_blockcount = cpu_to_be32(mp->m_sb.sb_logblocks);
642 		rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_LOG);
643 		rrec->rm_offset = 0;
644 		be16_add_cpu(&block->bb_numrecs, 1);
645 	}
646 }
647 
648 /*
649  * Initialise new secondary superblocks with the pre-grow geometry, but mark
650  * them as "in progress" so we know they haven't yet been activated. This will
651  * get cleared when the update with the new geometry information is done after
652  * changes to the primary are committed. This isn't strictly necessary, but we
653  * get it for free with the delayed buffer write lists and it means we can tell
654  * if a grow operation didn't complete properly after the fact.
655  */
656 static void
657 xfs_sbblock_init(
658 	struct xfs_mount	*mp,
659 	struct xfs_buf		*bp,
660 	struct aghdr_init_data	*id)
661 {
662 	struct xfs_dsb		*dsb = bp->b_addr;
663 
664 	xfs_sb_to_disk(dsb, &mp->m_sb);
665 	dsb->sb_inprogress = 1;
666 }
667 
668 static void
669 xfs_agfblock_init(
670 	struct xfs_mount	*mp,
671 	struct xfs_buf		*bp,
672 	struct aghdr_init_data	*id)
673 {
674 	struct xfs_agf		*agf = bp->b_addr;
675 	xfs_extlen_t		tmpsize;
676 
677 	agf->agf_magicnum = cpu_to_be32(XFS_AGF_MAGIC);
678 	agf->agf_versionnum = cpu_to_be32(XFS_AGF_VERSION);
679 	agf->agf_seqno = cpu_to_be32(id->agno);
680 	agf->agf_length = cpu_to_be32(id->agsize);
681 	agf->agf_roots[XFS_BTNUM_BNOi] = cpu_to_be32(XFS_BNO_BLOCK(mp));
682 	agf->agf_roots[XFS_BTNUM_CNTi] = cpu_to_be32(XFS_CNT_BLOCK(mp));
683 	agf->agf_levels[XFS_BTNUM_BNOi] = cpu_to_be32(1);
684 	agf->agf_levels[XFS_BTNUM_CNTi] = cpu_to_be32(1);
685 	if (xfs_has_rmapbt(mp)) {
686 		agf->agf_roots[XFS_BTNUM_RMAPi] =
687 					cpu_to_be32(XFS_RMAP_BLOCK(mp));
688 		agf->agf_levels[XFS_BTNUM_RMAPi] = cpu_to_be32(1);
689 		agf->agf_rmap_blocks = cpu_to_be32(1);
690 	}
691 
692 	agf->agf_flfirst = cpu_to_be32(1);
693 	agf->agf_fllast = 0;
694 	agf->agf_flcount = 0;
695 	tmpsize = id->agsize - mp->m_ag_prealloc_blocks;
696 	agf->agf_freeblks = cpu_to_be32(tmpsize);
697 	agf->agf_longest = cpu_to_be32(tmpsize);
698 	if (xfs_has_crc(mp))
699 		uuid_copy(&agf->agf_uuid, &mp->m_sb.sb_meta_uuid);
700 	if (xfs_has_reflink(mp)) {
701 		agf->agf_refcount_root = cpu_to_be32(
702 				xfs_refc_block(mp));
703 		agf->agf_refcount_level = cpu_to_be32(1);
704 		agf->agf_refcount_blocks = cpu_to_be32(1);
705 	}
706 
707 	if (xfs_ag_contains_log(mp, id->agno)) {
708 		int64_t	logblocks = mp->m_sb.sb_logblocks;
709 
710 		be32_add_cpu(&agf->agf_freeblks, -logblocks);
711 		agf->agf_longest = cpu_to_be32(id->agsize -
712 			XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart) - logblocks);
713 	}
714 }
715 
716 static void
717 xfs_agflblock_init(
718 	struct xfs_mount	*mp,
719 	struct xfs_buf		*bp,
720 	struct aghdr_init_data	*id)
721 {
722 	struct xfs_agfl		*agfl = XFS_BUF_TO_AGFL(bp);
723 	__be32			*agfl_bno;
724 	int			bucket;
725 
726 	if (xfs_has_crc(mp)) {
727 		agfl->agfl_magicnum = cpu_to_be32(XFS_AGFL_MAGIC);
728 		agfl->agfl_seqno = cpu_to_be32(id->agno);
729 		uuid_copy(&agfl->agfl_uuid, &mp->m_sb.sb_meta_uuid);
730 	}
731 
732 	agfl_bno = xfs_buf_to_agfl_bno(bp);
733 	for (bucket = 0; bucket < xfs_agfl_size(mp); bucket++)
734 		agfl_bno[bucket] = cpu_to_be32(NULLAGBLOCK);
735 }
736 
737 static void
738 xfs_agiblock_init(
739 	struct xfs_mount	*mp,
740 	struct xfs_buf		*bp,
741 	struct aghdr_init_data	*id)
742 {
743 	struct xfs_agi		*agi = bp->b_addr;
744 	int			bucket;
745 
746 	agi->agi_magicnum = cpu_to_be32(XFS_AGI_MAGIC);
747 	agi->agi_versionnum = cpu_to_be32(XFS_AGI_VERSION);
748 	agi->agi_seqno = cpu_to_be32(id->agno);
749 	agi->agi_length = cpu_to_be32(id->agsize);
750 	agi->agi_count = 0;
751 	agi->agi_root = cpu_to_be32(XFS_IBT_BLOCK(mp));
752 	agi->agi_level = cpu_to_be32(1);
753 	agi->agi_freecount = 0;
754 	agi->agi_newino = cpu_to_be32(NULLAGINO);
755 	agi->agi_dirino = cpu_to_be32(NULLAGINO);
756 	if (xfs_has_crc(mp))
757 		uuid_copy(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid);
758 	if (xfs_has_finobt(mp)) {
759 		agi->agi_free_root = cpu_to_be32(XFS_FIBT_BLOCK(mp));
760 		agi->agi_free_level = cpu_to_be32(1);
761 	}
762 	for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++)
763 		agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
764 	if (xfs_has_inobtcounts(mp)) {
765 		agi->agi_iblocks = cpu_to_be32(1);
766 		if (xfs_has_finobt(mp))
767 			agi->agi_fblocks = cpu_to_be32(1);
768 	}
769 }
770 
771 typedef void (*aghdr_init_work_f)(struct xfs_mount *mp, struct xfs_buf *bp,
772 				  struct aghdr_init_data *id);
773 static int
774 xfs_ag_init_hdr(
775 	struct xfs_mount	*mp,
776 	struct aghdr_init_data	*id,
777 	aghdr_init_work_f	work,
778 	const struct xfs_buf_ops *ops)
779 {
780 	struct xfs_buf		*bp;
781 	int			error;
782 
783 	error = xfs_get_aghdr_buf(mp, id->daddr, id->numblks, &bp, ops);
784 	if (error)
785 		return error;
786 
787 	(*work)(mp, bp, id);
788 
789 	xfs_buf_delwri_queue(bp, &id->buffer_list);
790 	xfs_buf_relse(bp);
791 	return 0;
792 }
793 
794 struct xfs_aghdr_grow_data {
795 	xfs_daddr_t		daddr;
796 	size_t			numblks;
797 	const struct xfs_buf_ops *ops;
798 	aghdr_init_work_f	work;
799 	xfs_btnum_t		type;
800 	bool			need_init;
801 };
802 
803 /*
804  * Prepare new AG headers to be written to disk. We use uncached buffers here,
805  * as it is assumed these new AG headers are currently beyond the currently
806  * valid filesystem address space. Using cached buffers would trip over EOFS
807  * corruption detection alogrithms in the buffer cache lookup routines.
808  *
809  * This is a non-transactional function, but the prepared buffers are added to a
810  * delayed write buffer list supplied by the caller so they can submit them to
811  * disk and wait on them as required.
812  */
813 int
814 xfs_ag_init_headers(
815 	struct xfs_mount	*mp,
816 	struct aghdr_init_data	*id)
817 
818 {
819 	struct xfs_aghdr_grow_data aghdr_data[] = {
820 	{ /* SB */
821 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_SB_DADDR),
822 		.numblks = XFS_FSS_TO_BB(mp, 1),
823 		.ops = &xfs_sb_buf_ops,
824 		.work = &xfs_sbblock_init,
825 		.need_init = true
826 	},
827 	{ /* AGF */
828 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGF_DADDR(mp)),
829 		.numblks = XFS_FSS_TO_BB(mp, 1),
830 		.ops = &xfs_agf_buf_ops,
831 		.work = &xfs_agfblock_init,
832 		.need_init = true
833 	},
834 	{ /* AGFL */
835 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGFL_DADDR(mp)),
836 		.numblks = XFS_FSS_TO_BB(mp, 1),
837 		.ops = &xfs_agfl_buf_ops,
838 		.work = &xfs_agflblock_init,
839 		.need_init = true
840 	},
841 	{ /* AGI */
842 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGI_DADDR(mp)),
843 		.numblks = XFS_FSS_TO_BB(mp, 1),
844 		.ops = &xfs_agi_buf_ops,
845 		.work = &xfs_agiblock_init,
846 		.need_init = true
847 	},
848 	{ /* BNO root block */
849 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_BNO_BLOCK(mp)),
850 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
851 		.ops = &xfs_bnobt_buf_ops,
852 		.work = &xfs_bnoroot_init,
853 		.need_init = true
854 	},
855 	{ /* CNT root block */
856 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_CNT_BLOCK(mp)),
857 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
858 		.ops = &xfs_cntbt_buf_ops,
859 		.work = &xfs_cntroot_init,
860 		.need_init = true
861 	},
862 	{ /* INO root block */
863 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_IBT_BLOCK(mp)),
864 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
865 		.ops = &xfs_inobt_buf_ops,
866 		.work = &xfs_btroot_init,
867 		.type = XFS_BTNUM_INO,
868 		.need_init = true
869 	},
870 	{ /* FINO root block */
871 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_FIBT_BLOCK(mp)),
872 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
873 		.ops = &xfs_finobt_buf_ops,
874 		.work = &xfs_btroot_init,
875 		.type = XFS_BTNUM_FINO,
876 		.need_init =  xfs_has_finobt(mp)
877 	},
878 	{ /* RMAP root block */
879 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_RMAP_BLOCK(mp)),
880 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
881 		.ops = &xfs_rmapbt_buf_ops,
882 		.work = &xfs_rmaproot_init,
883 		.need_init = xfs_has_rmapbt(mp)
884 	},
885 	{ /* REFC root block */
886 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, xfs_refc_block(mp)),
887 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
888 		.ops = &xfs_refcountbt_buf_ops,
889 		.work = &xfs_btroot_init,
890 		.type = XFS_BTNUM_REFC,
891 		.need_init = xfs_has_reflink(mp)
892 	},
893 	{ /* NULL terminating block */
894 		.daddr = XFS_BUF_DADDR_NULL,
895 	}
896 	};
897 	struct  xfs_aghdr_grow_data *dp;
898 	int			error = 0;
899 
900 	/* Account for AG free space in new AG */
901 	id->nfree += id->agsize - mp->m_ag_prealloc_blocks;
902 	for (dp = &aghdr_data[0]; dp->daddr != XFS_BUF_DADDR_NULL; dp++) {
903 		if (!dp->need_init)
904 			continue;
905 
906 		id->daddr = dp->daddr;
907 		id->numblks = dp->numblks;
908 		id->type = dp->type;
909 		error = xfs_ag_init_hdr(mp, id, dp->work, dp->ops);
910 		if (error)
911 			break;
912 	}
913 	return error;
914 }
915 
916 int
917 xfs_ag_shrink_space(
918 	struct xfs_perag	*pag,
919 	struct xfs_trans	**tpp,
920 	xfs_extlen_t		delta)
921 {
922 	struct xfs_mount	*mp = pag->pag_mount;
923 	struct xfs_alloc_arg	args = {
924 		.tp	= *tpp,
925 		.mp	= mp,
926 		.pag	= pag,
927 		.minlen = delta,
928 		.maxlen = delta,
929 		.oinfo	= XFS_RMAP_OINFO_SKIP_UPDATE,
930 		.resv	= XFS_AG_RESV_NONE,
931 		.prod	= 1
932 	};
933 	struct xfs_buf		*agibp, *agfbp;
934 	struct xfs_agi		*agi;
935 	struct xfs_agf		*agf;
936 	xfs_agblock_t		aglen;
937 	int			error, err2;
938 
939 	ASSERT(pag->pag_agno == mp->m_sb.sb_agcount - 1);
940 	error = xfs_ialloc_read_agi(pag, *tpp, &agibp);
941 	if (error)
942 		return error;
943 
944 	agi = agibp->b_addr;
945 
946 	error = xfs_alloc_read_agf(pag, *tpp, 0, &agfbp);
947 	if (error)
948 		return error;
949 
950 	agf = agfbp->b_addr;
951 	aglen = be32_to_cpu(agi->agi_length);
952 	/* some extra paranoid checks before we shrink the ag */
953 	if (XFS_IS_CORRUPT(mp, agf->agf_length != agi->agi_length))
954 		return -EFSCORRUPTED;
955 	if (delta >= aglen)
956 		return -EINVAL;
957 
958 	/*
959 	 * Make sure that the last inode cluster cannot overlap with the new
960 	 * end of the AG, even if it's sparse.
961 	 */
962 	error = xfs_ialloc_check_shrink(pag, *tpp, agibp, aglen - delta);
963 	if (error)
964 		return error;
965 
966 	/*
967 	 * Disable perag reservations so it doesn't cause the allocation request
968 	 * to fail. We'll reestablish reservation before we return.
969 	 */
970 	error = xfs_ag_resv_free(pag);
971 	if (error)
972 		return error;
973 
974 	/* internal log shouldn't also show up in the free space btrees */
975 	error = xfs_alloc_vextent_exact_bno(&args,
976 			XFS_AGB_TO_FSB(mp, pag->pag_agno, aglen - delta));
977 	if (!error && args.agbno == NULLAGBLOCK)
978 		error = -ENOSPC;
979 
980 	if (error) {
981 		/*
982 		 * if extent allocation fails, need to roll the transaction to
983 		 * ensure that the AGFL fixup has been committed anyway.
984 		 */
985 		xfs_trans_bhold(*tpp, agfbp);
986 		err2 = xfs_trans_roll(tpp);
987 		if (err2)
988 			return err2;
989 		xfs_trans_bjoin(*tpp, agfbp);
990 		goto resv_init_out;
991 	}
992 
993 	/*
994 	 * if successfully deleted from freespace btrees, need to confirm
995 	 * per-AG reservation works as expected.
996 	 */
997 	be32_add_cpu(&agi->agi_length, -delta);
998 	be32_add_cpu(&agf->agf_length, -delta);
999 
1000 	err2 = xfs_ag_resv_init(pag, *tpp);
1001 	if (err2) {
1002 		be32_add_cpu(&agi->agi_length, delta);
1003 		be32_add_cpu(&agf->agf_length, delta);
1004 		if (err2 != -ENOSPC)
1005 			goto resv_err;
1006 
1007 		err2 = xfs_free_extent_later(*tpp, args.fsbno, delta, NULL,
1008 				XFS_AG_RESV_NONE, true);
1009 		if (err2)
1010 			goto resv_err;
1011 
1012 		/*
1013 		 * Roll the transaction before trying to re-init the per-ag
1014 		 * reservation. The new transaction is clean so it will cancel
1015 		 * without any side effects.
1016 		 */
1017 		error = xfs_defer_finish(tpp);
1018 		if (error)
1019 			return error;
1020 
1021 		error = -ENOSPC;
1022 		goto resv_init_out;
1023 	}
1024 
1025 	/* Update perag geometry */
1026 	pag->block_count -= delta;
1027 	__xfs_agino_range(pag->pag_mount, pag->block_count, &pag->agino_min,
1028 				&pag->agino_max);
1029 
1030 	xfs_ialloc_log_agi(*tpp, agibp, XFS_AGI_LENGTH);
1031 	xfs_alloc_log_agf(*tpp, agfbp, XFS_AGF_LENGTH);
1032 	return 0;
1033 
1034 resv_init_out:
1035 	err2 = xfs_ag_resv_init(pag, *tpp);
1036 	if (!err2)
1037 		return error;
1038 resv_err:
1039 	xfs_warn(mp, "Error %d reserving per-AG metadata reserve pool.", err2);
1040 	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1041 	return err2;
1042 }
1043 
1044 /*
1045  * Extent the AG indicated by the @id by the length passed in
1046  */
1047 int
1048 xfs_ag_extend_space(
1049 	struct xfs_perag	*pag,
1050 	struct xfs_trans	*tp,
1051 	xfs_extlen_t		len)
1052 {
1053 	struct xfs_buf		*bp;
1054 	struct xfs_agi		*agi;
1055 	struct xfs_agf		*agf;
1056 	int			error;
1057 
1058 	ASSERT(pag->pag_agno == pag->pag_mount->m_sb.sb_agcount - 1);
1059 
1060 	error = xfs_ialloc_read_agi(pag, tp, &bp);
1061 	if (error)
1062 		return error;
1063 
1064 	agi = bp->b_addr;
1065 	be32_add_cpu(&agi->agi_length, len);
1066 	xfs_ialloc_log_agi(tp, bp, XFS_AGI_LENGTH);
1067 
1068 	/*
1069 	 * Change agf length.
1070 	 */
1071 	error = xfs_alloc_read_agf(pag, tp, 0, &bp);
1072 	if (error)
1073 		return error;
1074 
1075 	agf = bp->b_addr;
1076 	be32_add_cpu(&agf->agf_length, len);
1077 	ASSERT(agf->agf_length == agi->agi_length);
1078 	xfs_alloc_log_agf(tp, bp, XFS_AGF_LENGTH);
1079 
1080 	/*
1081 	 * Free the new space.
1082 	 *
1083 	 * XFS_RMAP_OINFO_SKIP_UPDATE is used here to tell the rmap btree that
1084 	 * this doesn't actually exist in the rmap btree.
1085 	 */
1086 	error = xfs_rmap_free(tp, bp, pag, be32_to_cpu(agf->agf_length) - len,
1087 				len, &XFS_RMAP_OINFO_SKIP_UPDATE);
1088 	if (error)
1089 		return error;
1090 
1091 	error = xfs_free_extent(tp, pag, be32_to_cpu(agf->agf_length) - len,
1092 			len, &XFS_RMAP_OINFO_SKIP_UPDATE, XFS_AG_RESV_NONE);
1093 	if (error)
1094 		return error;
1095 
1096 	/* Update perag geometry */
1097 	pag->block_count = be32_to_cpu(agf->agf_length);
1098 	__xfs_agino_range(pag->pag_mount, pag->block_count, &pag->agino_min,
1099 				&pag->agino_max);
1100 	return 0;
1101 }
1102 
1103 /* Retrieve AG geometry. */
1104 int
1105 xfs_ag_get_geometry(
1106 	struct xfs_perag	*pag,
1107 	struct xfs_ag_geometry	*ageo)
1108 {
1109 	struct xfs_buf		*agi_bp;
1110 	struct xfs_buf		*agf_bp;
1111 	struct xfs_agi		*agi;
1112 	struct xfs_agf		*agf;
1113 	unsigned int		freeblks;
1114 	int			error;
1115 
1116 	/* Lock the AG headers. */
1117 	error = xfs_ialloc_read_agi(pag, NULL, &agi_bp);
1118 	if (error)
1119 		return error;
1120 	error = xfs_alloc_read_agf(pag, NULL, 0, &agf_bp);
1121 	if (error)
1122 		goto out_agi;
1123 
1124 	/* Fill out form. */
1125 	memset(ageo, 0, sizeof(*ageo));
1126 	ageo->ag_number = pag->pag_agno;
1127 
1128 	agi = agi_bp->b_addr;
1129 	ageo->ag_icount = be32_to_cpu(agi->agi_count);
1130 	ageo->ag_ifree = be32_to_cpu(agi->agi_freecount);
1131 
1132 	agf = agf_bp->b_addr;
1133 	ageo->ag_length = be32_to_cpu(agf->agf_length);
1134 	freeblks = pag->pagf_freeblks +
1135 		   pag->pagf_flcount +
1136 		   pag->pagf_btreeblks -
1137 		   xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE);
1138 	ageo->ag_freeblks = freeblks;
1139 	xfs_ag_geom_health(pag, ageo);
1140 
1141 	/* Release resources. */
1142 	xfs_buf_relse(agf_bp);
1143 out_agi:
1144 	xfs_buf_relse(agi_bp);
1145 	return error;
1146 }
1147