xref: /linux/fs/xfs/libxfs/xfs_ag.c (revision a8b3be2617d677796e576cc64d4ad9de45dfaf14)
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 *
xfs_perag_get(struct xfs_mount * mp,xfs_agnumber_t agno)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 = xa_load(&mp->m_perags, 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 /* Get a passive reference to the given perag. */
60 struct xfs_perag *
xfs_perag_hold(struct xfs_perag * pag)61 xfs_perag_hold(
62 	struct xfs_perag	*pag)
63 {
64 	ASSERT(atomic_read(&pag->pag_ref) > 0 ||
65 	       atomic_read(&pag->pag_active_ref) > 0);
66 
67 	trace_xfs_perag_hold(pag, _RET_IP_);
68 	atomic_inc(&pag->pag_ref);
69 	return pag;
70 }
71 
72 void
xfs_perag_put(struct xfs_perag * pag)73 xfs_perag_put(
74 	struct xfs_perag	*pag)
75 {
76 	trace_xfs_perag_put(pag, _RET_IP_);
77 	ASSERT(atomic_read(&pag->pag_ref) > 0);
78 	atomic_dec(&pag->pag_ref);
79 }
80 
81 /*
82  * Active references for perag structures. This is for short term access to the
83  * per ag structures for walking trees or accessing state. If an AG is being
84  * shrunk or is offline, then this will fail to find that AG and return NULL
85  * instead.
86  */
87 struct xfs_perag *
xfs_perag_grab(struct xfs_mount * mp,xfs_agnumber_t agno)88 xfs_perag_grab(
89 	struct xfs_mount	*mp,
90 	xfs_agnumber_t		agno)
91 {
92 	struct xfs_perag	*pag;
93 
94 	rcu_read_lock();
95 	pag = xa_load(&mp->m_perags, agno);
96 	if (pag) {
97 		trace_xfs_perag_grab(pag, _RET_IP_);
98 		if (!atomic_inc_not_zero(&pag->pag_active_ref))
99 			pag = NULL;
100 	}
101 	rcu_read_unlock();
102 	return pag;
103 }
104 
105 void
xfs_perag_rele(struct xfs_perag * pag)106 xfs_perag_rele(
107 	struct xfs_perag	*pag)
108 {
109 	trace_xfs_perag_rele(pag, _RET_IP_);
110 	if (atomic_dec_and_test(&pag->pag_active_ref))
111 		wake_up(&pag->pag_active_wq);
112 }
113 
114 /*
115  * xfs_initialize_perag_data
116  *
117  * Read in each per-ag structure so we can count up the number of
118  * allocated inodes, free inodes and used filesystem blocks as this
119  * information is no longer persistent in the superblock. Once we have
120  * this information, write it into the in-core superblock structure.
121  */
122 int
xfs_initialize_perag_data(struct xfs_mount * mp,xfs_agnumber_t agcount)123 xfs_initialize_perag_data(
124 	struct xfs_mount	*mp,
125 	xfs_agnumber_t		agcount)
126 {
127 	xfs_agnumber_t		index;
128 	struct xfs_perag	*pag;
129 	struct xfs_sb		*sbp = &mp->m_sb;
130 	uint64_t		ifree = 0;
131 	uint64_t		ialloc = 0;
132 	uint64_t		bfree = 0;
133 	uint64_t		bfreelst = 0;
134 	uint64_t		btree = 0;
135 	uint64_t		fdblocks;
136 	int			error = 0;
137 
138 	for (index = 0; index < agcount; index++) {
139 		/*
140 		 * Read the AGF and AGI buffers to populate the per-ag
141 		 * structures for us.
142 		 */
143 		pag = xfs_perag_get(mp, index);
144 		error = xfs_alloc_read_agf(pag, NULL, 0, NULL);
145 		if (!error)
146 			error = xfs_ialloc_read_agi(pag, NULL, 0, NULL);
147 		if (error) {
148 			xfs_perag_put(pag);
149 			return error;
150 		}
151 
152 		ifree += pag->pagi_freecount;
153 		ialloc += pag->pagi_count;
154 		bfree += pag->pagf_freeblks;
155 		bfreelst += pag->pagf_flcount;
156 		btree += pag->pagf_btreeblks;
157 		xfs_perag_put(pag);
158 	}
159 	fdblocks = bfree + bfreelst + btree;
160 
161 	/*
162 	 * If the new summary counts are obviously incorrect, fail the
163 	 * mount operation because that implies the AGFs are also corrupt.
164 	 * Clear FS_COUNTERS so that we don't unmount with a dirty log, which
165 	 * will prevent xfs_repair from fixing anything.
166 	 */
167 	if (fdblocks > sbp->sb_dblocks || ifree > ialloc) {
168 		xfs_alert(mp, "AGF corruption. Please run xfs_repair.");
169 		xfs_fs_mark_sick(mp, XFS_SICK_FS_COUNTERS);
170 		error = -EFSCORRUPTED;
171 		goto out;
172 	}
173 
174 	/* Overwrite incore superblock counters with just-read data */
175 	spin_lock(&mp->m_sb_lock);
176 	sbp->sb_ifree = ifree;
177 	sbp->sb_icount = ialloc;
178 	sbp->sb_fdblocks = fdblocks;
179 	spin_unlock(&mp->m_sb_lock);
180 
181 	xfs_reinit_percpu_counters(mp);
182 out:
183 	xfs_fs_mark_healthy(mp, XFS_SICK_FS_COUNTERS);
184 	return error;
185 }
186 
187 /*
188  * Free up the per-ag resources  within the specified AG range.
189  */
190 void
xfs_free_perag_range(struct xfs_mount * mp,xfs_agnumber_t first_agno,xfs_agnumber_t end_agno)191 xfs_free_perag_range(
192 	struct xfs_mount	*mp,
193 	xfs_agnumber_t		first_agno,
194 	xfs_agnumber_t		end_agno)
195 
196 {
197 	xfs_agnumber_t		agno;
198 
199 	for (agno = first_agno; agno < end_agno; agno++) {
200 		struct xfs_perag	*pag = xa_erase(&mp->m_perags, agno);
201 
202 		ASSERT(pag);
203 		XFS_IS_CORRUPT(pag->pag_mount, atomic_read(&pag->pag_ref) != 0);
204 		xfs_defer_drain_free(&pag->pag_intents_drain);
205 
206 		cancel_delayed_work_sync(&pag->pag_blockgc_work);
207 		xfs_buf_cache_destroy(&pag->pag_bcache);
208 
209 		/* drop the mount's active reference */
210 		xfs_perag_rele(pag);
211 		XFS_IS_CORRUPT(pag->pag_mount,
212 				atomic_read(&pag->pag_active_ref) != 0);
213 		kfree_rcu_mightsleep(pag);
214 	}
215 }
216 
217 /* Find the size of the AG, in blocks. */
218 static xfs_agblock_t
__xfs_ag_block_count(struct xfs_mount * mp,xfs_agnumber_t agno,xfs_agnumber_t agcount,xfs_rfsblock_t dblocks)219 __xfs_ag_block_count(
220 	struct xfs_mount	*mp,
221 	xfs_agnumber_t		agno,
222 	xfs_agnumber_t		agcount,
223 	xfs_rfsblock_t		dblocks)
224 {
225 	ASSERT(agno < agcount);
226 
227 	if (agno < agcount - 1)
228 		return mp->m_sb.sb_agblocks;
229 	return dblocks - (agno * mp->m_sb.sb_agblocks);
230 }
231 
232 xfs_agblock_t
xfs_ag_block_count(struct xfs_mount * mp,xfs_agnumber_t agno)233 xfs_ag_block_count(
234 	struct xfs_mount	*mp,
235 	xfs_agnumber_t		agno)
236 {
237 	return __xfs_ag_block_count(mp, agno, mp->m_sb.sb_agcount,
238 			mp->m_sb.sb_dblocks);
239 }
240 
241 /* Calculate the first and last possible inode number in an AG. */
242 static void
__xfs_agino_range(struct xfs_mount * mp,xfs_agblock_t eoag,xfs_agino_t * first,xfs_agino_t * last)243 __xfs_agino_range(
244 	struct xfs_mount	*mp,
245 	xfs_agblock_t		eoag,
246 	xfs_agino_t		*first,
247 	xfs_agino_t		*last)
248 {
249 	xfs_agblock_t		bno;
250 
251 	/*
252 	 * Calculate the first inode, which will be in the first
253 	 * cluster-aligned block after the AGFL.
254 	 */
255 	bno = round_up(XFS_AGFL_BLOCK(mp) + 1, M_IGEO(mp)->cluster_align);
256 	*first = XFS_AGB_TO_AGINO(mp, bno);
257 
258 	/*
259 	 * Calculate the last inode, which will be at the end of the
260 	 * last (aligned) cluster that can be allocated in the AG.
261 	 */
262 	bno = round_down(eoag, M_IGEO(mp)->cluster_align);
263 	*last = XFS_AGB_TO_AGINO(mp, bno) - 1;
264 }
265 
266 void
xfs_agino_range(struct xfs_mount * mp,xfs_agnumber_t agno,xfs_agino_t * first,xfs_agino_t * last)267 xfs_agino_range(
268 	struct xfs_mount	*mp,
269 	xfs_agnumber_t		agno,
270 	xfs_agino_t		*first,
271 	xfs_agino_t		*last)
272 {
273 	return __xfs_agino_range(mp, xfs_ag_block_count(mp, agno), first, last);
274 }
275 
276 int
xfs_update_last_ag_size(struct xfs_mount * mp,xfs_agnumber_t prev_agcount)277 xfs_update_last_ag_size(
278 	struct xfs_mount	*mp,
279 	xfs_agnumber_t		prev_agcount)
280 {
281 	struct xfs_perag	*pag = xfs_perag_grab(mp, prev_agcount - 1);
282 
283 	if (!pag)
284 		return -EFSCORRUPTED;
285 	pag->block_count = __xfs_ag_block_count(mp, prev_agcount - 1,
286 			mp->m_sb.sb_agcount, mp->m_sb.sb_dblocks);
287 	__xfs_agino_range(mp, pag->block_count, &pag->agino_min,
288 			&pag->agino_max);
289 	xfs_perag_rele(pag);
290 	return 0;
291 }
292 
293 int
xfs_initialize_perag(struct xfs_mount * mp,xfs_agnumber_t old_agcount,xfs_agnumber_t new_agcount,xfs_rfsblock_t dblocks,xfs_agnumber_t * maxagi)294 xfs_initialize_perag(
295 	struct xfs_mount	*mp,
296 	xfs_agnumber_t		old_agcount,
297 	xfs_agnumber_t		new_agcount,
298 	xfs_rfsblock_t		dblocks,
299 	xfs_agnumber_t		*maxagi)
300 {
301 	struct xfs_perag	*pag;
302 	xfs_agnumber_t		index;
303 	int			error;
304 
305 	for (index = old_agcount; index < new_agcount; index++) {
306 		pag = kzalloc(sizeof(*pag), GFP_KERNEL);
307 		if (!pag) {
308 			error = -ENOMEM;
309 			goto out_unwind_new_pags;
310 		}
311 		pag->pag_agno = index;
312 		pag->pag_mount = mp;
313 
314 		error = xa_insert(&mp->m_perags, index, pag, GFP_KERNEL);
315 		if (error) {
316 			WARN_ON_ONCE(error == -EBUSY);
317 			goto out_free_pag;
318 		}
319 
320 #ifdef __KERNEL__
321 		/* Place kernel structure only init below this point. */
322 		spin_lock_init(&pag->pag_ici_lock);
323 		spin_lock_init(&pag->pagb_lock);
324 		spin_lock_init(&pag->pag_state_lock);
325 		INIT_DELAYED_WORK(&pag->pag_blockgc_work, xfs_blockgc_worker);
326 		INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
327 		xfs_defer_drain_init(&pag->pag_intents_drain);
328 		init_waitqueue_head(&pag->pagb_wait);
329 		init_waitqueue_head(&pag->pag_active_wq);
330 		pag->pagb_count = 0;
331 		pag->pagb_tree = RB_ROOT;
332 		xfs_hooks_init(&pag->pag_rmap_update_hooks);
333 #endif /* __KERNEL__ */
334 
335 		error = xfs_buf_cache_init(&pag->pag_bcache);
336 		if (error)
337 			goto out_remove_pag;
338 
339 		/* Active ref owned by mount indicates AG is online. */
340 		atomic_set(&pag->pag_active_ref, 1);
341 
342 		/*
343 		 * Pre-calculated geometry
344 		 */
345 		pag->block_count = __xfs_ag_block_count(mp, index, new_agcount,
346 				dblocks);
347 		pag->min_block = XFS_AGFL_BLOCK(mp);
348 		__xfs_agino_range(mp, pag->block_count, &pag->agino_min,
349 				&pag->agino_max);
350 	}
351 
352 	index = xfs_set_inode_alloc(mp, new_agcount);
353 
354 	if (maxagi)
355 		*maxagi = index;
356 
357 	mp->m_ag_prealloc_blocks = xfs_prealloc_blocks(mp);
358 	return 0;
359 
360 out_remove_pag:
361 	xfs_defer_drain_free(&pag->pag_intents_drain);
362 	pag = xa_erase(&mp->m_perags, index);
363 out_free_pag:
364 	kfree(pag);
365 out_unwind_new_pags:
366 	xfs_free_perag_range(mp, old_agcount, index);
367 	return error;
368 }
369 
370 static int
xfs_get_aghdr_buf(struct xfs_mount * mp,xfs_daddr_t blkno,size_t numblks,struct xfs_buf ** bpp,const struct xfs_buf_ops * ops)371 xfs_get_aghdr_buf(
372 	struct xfs_mount	*mp,
373 	xfs_daddr_t		blkno,
374 	size_t			numblks,
375 	struct xfs_buf		**bpp,
376 	const struct xfs_buf_ops *ops)
377 {
378 	struct xfs_buf		*bp;
379 	int			error;
380 
381 	error = xfs_buf_get_uncached(mp->m_ddev_targp, numblks, 0, &bp);
382 	if (error)
383 		return error;
384 
385 	bp->b_maps[0].bm_bn = blkno;
386 	bp->b_ops = ops;
387 
388 	*bpp = bp;
389 	return 0;
390 }
391 
392 /*
393  * Generic btree root block init function
394  */
395 static void
xfs_btroot_init(struct xfs_mount * mp,struct xfs_buf * bp,struct aghdr_init_data * id)396 xfs_btroot_init(
397 	struct xfs_mount	*mp,
398 	struct xfs_buf		*bp,
399 	struct aghdr_init_data	*id)
400 {
401 	xfs_btree_init_buf(mp, bp, id->bc_ops, 0, 0, id->agno);
402 }
403 
404 /* Finish initializing a free space btree. */
405 static void
xfs_freesp_init_recs(struct xfs_mount * mp,struct xfs_buf * bp,struct aghdr_init_data * id)406 xfs_freesp_init_recs(
407 	struct xfs_mount	*mp,
408 	struct xfs_buf		*bp,
409 	struct aghdr_init_data	*id)
410 {
411 	struct xfs_alloc_rec	*arec;
412 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
413 
414 	arec = XFS_ALLOC_REC_ADDR(mp, XFS_BUF_TO_BLOCK(bp), 1);
415 	arec->ar_startblock = cpu_to_be32(mp->m_ag_prealloc_blocks);
416 
417 	if (xfs_ag_contains_log(mp, id->agno)) {
418 		struct xfs_alloc_rec	*nrec;
419 		xfs_agblock_t		start = XFS_FSB_TO_AGBNO(mp,
420 							mp->m_sb.sb_logstart);
421 
422 		ASSERT(start >= mp->m_ag_prealloc_blocks);
423 		if (start != mp->m_ag_prealloc_blocks) {
424 			/*
425 			 * Modify first record to pad stripe align of log and
426 			 * bump the record count.
427 			 */
428 			arec->ar_blockcount = cpu_to_be32(start -
429 						mp->m_ag_prealloc_blocks);
430 			be16_add_cpu(&block->bb_numrecs, 1);
431 			nrec = arec + 1;
432 
433 			/*
434 			 * Insert second record at start of internal log
435 			 * which then gets trimmed.
436 			 */
437 			nrec->ar_startblock = cpu_to_be32(
438 					be32_to_cpu(arec->ar_startblock) +
439 					be32_to_cpu(arec->ar_blockcount));
440 			arec = nrec;
441 		}
442 		/*
443 		 * Change record start to after the internal log
444 		 */
445 		be32_add_cpu(&arec->ar_startblock, mp->m_sb.sb_logblocks);
446 	}
447 
448 	/*
449 	 * Calculate the block count of this record; if it is nonzero,
450 	 * increment the record count.
451 	 */
452 	arec->ar_blockcount = cpu_to_be32(id->agsize -
453 					  be32_to_cpu(arec->ar_startblock));
454 	if (arec->ar_blockcount)
455 		be16_add_cpu(&block->bb_numrecs, 1);
456 }
457 
458 /*
459  * bnobt/cntbt btree root block init functions
460  */
461 static void
xfs_bnoroot_init(struct xfs_mount * mp,struct xfs_buf * bp,struct aghdr_init_data * id)462 xfs_bnoroot_init(
463 	struct xfs_mount	*mp,
464 	struct xfs_buf		*bp,
465 	struct aghdr_init_data	*id)
466 {
467 	xfs_btree_init_buf(mp, bp, id->bc_ops, 0, 0, id->agno);
468 	xfs_freesp_init_recs(mp, bp, id);
469 }
470 
471 /*
472  * Reverse map root block init
473  */
474 static void
xfs_rmaproot_init(struct xfs_mount * mp,struct xfs_buf * bp,struct aghdr_init_data * id)475 xfs_rmaproot_init(
476 	struct xfs_mount	*mp,
477 	struct xfs_buf		*bp,
478 	struct aghdr_init_data	*id)
479 {
480 	struct xfs_btree_block	*block = XFS_BUF_TO_BLOCK(bp);
481 	struct xfs_rmap_rec	*rrec;
482 
483 	xfs_btree_init_buf(mp, bp, id->bc_ops, 0, 4, id->agno);
484 
485 	/*
486 	 * mark the AG header regions as static metadata The BNO
487 	 * btree block is the first block after the headers, so
488 	 * it's location defines the size of region the static
489 	 * metadata consumes.
490 	 *
491 	 * Note: unlike mkfs, we never have to account for log
492 	 * space when growing the data regions
493 	 */
494 	rrec = XFS_RMAP_REC_ADDR(block, 1);
495 	rrec->rm_startblock = 0;
496 	rrec->rm_blockcount = cpu_to_be32(XFS_BNO_BLOCK(mp));
497 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_FS);
498 	rrec->rm_offset = 0;
499 
500 	/* account freespace btree root blocks */
501 	rrec = XFS_RMAP_REC_ADDR(block, 2);
502 	rrec->rm_startblock = cpu_to_be32(XFS_BNO_BLOCK(mp));
503 	rrec->rm_blockcount = cpu_to_be32(2);
504 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_AG);
505 	rrec->rm_offset = 0;
506 
507 	/* account inode btree root blocks */
508 	rrec = XFS_RMAP_REC_ADDR(block, 3);
509 	rrec->rm_startblock = cpu_to_be32(XFS_IBT_BLOCK(mp));
510 	rrec->rm_blockcount = cpu_to_be32(XFS_RMAP_BLOCK(mp) -
511 					  XFS_IBT_BLOCK(mp));
512 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_INOBT);
513 	rrec->rm_offset = 0;
514 
515 	/* account for rmap btree root */
516 	rrec = XFS_RMAP_REC_ADDR(block, 4);
517 	rrec->rm_startblock = cpu_to_be32(XFS_RMAP_BLOCK(mp));
518 	rrec->rm_blockcount = cpu_to_be32(1);
519 	rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_AG);
520 	rrec->rm_offset = 0;
521 
522 	/* account for refc btree root */
523 	if (xfs_has_reflink(mp)) {
524 		rrec = XFS_RMAP_REC_ADDR(block, 5);
525 		rrec->rm_startblock = cpu_to_be32(xfs_refc_block(mp));
526 		rrec->rm_blockcount = cpu_to_be32(1);
527 		rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_REFC);
528 		rrec->rm_offset = 0;
529 		be16_add_cpu(&block->bb_numrecs, 1);
530 	}
531 
532 	/* account for the log space */
533 	if (xfs_ag_contains_log(mp, id->agno)) {
534 		rrec = XFS_RMAP_REC_ADDR(block,
535 				be16_to_cpu(block->bb_numrecs) + 1);
536 		rrec->rm_startblock = cpu_to_be32(
537 				XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart));
538 		rrec->rm_blockcount = cpu_to_be32(mp->m_sb.sb_logblocks);
539 		rrec->rm_owner = cpu_to_be64(XFS_RMAP_OWN_LOG);
540 		rrec->rm_offset = 0;
541 		be16_add_cpu(&block->bb_numrecs, 1);
542 	}
543 }
544 
545 /*
546  * Initialise new secondary superblocks with the pre-grow geometry, but mark
547  * them as "in progress" so we know they haven't yet been activated. This will
548  * get cleared when the update with the new geometry information is done after
549  * changes to the primary are committed. This isn't strictly necessary, but we
550  * get it for free with the delayed buffer write lists and it means we can tell
551  * if a grow operation didn't complete properly after the fact.
552  */
553 static void
xfs_sbblock_init(struct xfs_mount * mp,struct xfs_buf * bp,struct aghdr_init_data * id)554 xfs_sbblock_init(
555 	struct xfs_mount	*mp,
556 	struct xfs_buf		*bp,
557 	struct aghdr_init_data	*id)
558 {
559 	struct xfs_dsb		*dsb = bp->b_addr;
560 
561 	xfs_sb_to_disk(dsb, &mp->m_sb);
562 	dsb->sb_inprogress = 1;
563 }
564 
565 static void
xfs_agfblock_init(struct xfs_mount * mp,struct xfs_buf * bp,struct aghdr_init_data * id)566 xfs_agfblock_init(
567 	struct xfs_mount	*mp,
568 	struct xfs_buf		*bp,
569 	struct aghdr_init_data	*id)
570 {
571 	struct xfs_agf		*agf = bp->b_addr;
572 	xfs_extlen_t		tmpsize;
573 
574 	agf->agf_magicnum = cpu_to_be32(XFS_AGF_MAGIC);
575 	agf->agf_versionnum = cpu_to_be32(XFS_AGF_VERSION);
576 	agf->agf_seqno = cpu_to_be32(id->agno);
577 	agf->agf_length = cpu_to_be32(id->agsize);
578 	agf->agf_bno_root = cpu_to_be32(XFS_BNO_BLOCK(mp));
579 	agf->agf_cnt_root = cpu_to_be32(XFS_CNT_BLOCK(mp));
580 	agf->agf_bno_level = cpu_to_be32(1);
581 	agf->agf_cnt_level = cpu_to_be32(1);
582 	if (xfs_has_rmapbt(mp)) {
583 		agf->agf_rmap_root = cpu_to_be32(XFS_RMAP_BLOCK(mp));
584 		agf->agf_rmap_level = cpu_to_be32(1);
585 		agf->agf_rmap_blocks = cpu_to_be32(1);
586 	}
587 
588 	agf->agf_flfirst = cpu_to_be32(1);
589 	agf->agf_fllast = 0;
590 	agf->agf_flcount = 0;
591 	tmpsize = id->agsize - mp->m_ag_prealloc_blocks;
592 	agf->agf_freeblks = cpu_to_be32(tmpsize);
593 	agf->agf_longest = cpu_to_be32(tmpsize);
594 	if (xfs_has_crc(mp))
595 		uuid_copy(&agf->agf_uuid, &mp->m_sb.sb_meta_uuid);
596 	if (xfs_has_reflink(mp)) {
597 		agf->agf_refcount_root = cpu_to_be32(
598 				xfs_refc_block(mp));
599 		agf->agf_refcount_level = cpu_to_be32(1);
600 		agf->agf_refcount_blocks = cpu_to_be32(1);
601 	}
602 
603 	if (xfs_ag_contains_log(mp, id->agno)) {
604 		int64_t	logblocks = mp->m_sb.sb_logblocks;
605 
606 		be32_add_cpu(&agf->agf_freeblks, -logblocks);
607 		agf->agf_longest = cpu_to_be32(id->agsize -
608 			XFS_FSB_TO_AGBNO(mp, mp->m_sb.sb_logstart) - logblocks);
609 	}
610 }
611 
612 static void
xfs_agflblock_init(struct xfs_mount * mp,struct xfs_buf * bp,struct aghdr_init_data * id)613 xfs_agflblock_init(
614 	struct xfs_mount	*mp,
615 	struct xfs_buf		*bp,
616 	struct aghdr_init_data	*id)
617 {
618 	struct xfs_agfl		*agfl = XFS_BUF_TO_AGFL(bp);
619 	__be32			*agfl_bno;
620 	int			bucket;
621 
622 	if (xfs_has_crc(mp)) {
623 		agfl->agfl_magicnum = cpu_to_be32(XFS_AGFL_MAGIC);
624 		agfl->agfl_seqno = cpu_to_be32(id->agno);
625 		uuid_copy(&agfl->agfl_uuid, &mp->m_sb.sb_meta_uuid);
626 	}
627 
628 	agfl_bno = xfs_buf_to_agfl_bno(bp);
629 	for (bucket = 0; bucket < xfs_agfl_size(mp); bucket++)
630 		agfl_bno[bucket] = cpu_to_be32(NULLAGBLOCK);
631 }
632 
633 static void
xfs_agiblock_init(struct xfs_mount * mp,struct xfs_buf * bp,struct aghdr_init_data * id)634 xfs_agiblock_init(
635 	struct xfs_mount	*mp,
636 	struct xfs_buf		*bp,
637 	struct aghdr_init_data	*id)
638 {
639 	struct xfs_agi		*agi = bp->b_addr;
640 	int			bucket;
641 
642 	agi->agi_magicnum = cpu_to_be32(XFS_AGI_MAGIC);
643 	agi->agi_versionnum = cpu_to_be32(XFS_AGI_VERSION);
644 	agi->agi_seqno = cpu_to_be32(id->agno);
645 	agi->agi_length = cpu_to_be32(id->agsize);
646 	agi->agi_count = 0;
647 	agi->agi_root = cpu_to_be32(XFS_IBT_BLOCK(mp));
648 	agi->agi_level = cpu_to_be32(1);
649 	agi->agi_freecount = 0;
650 	agi->agi_newino = cpu_to_be32(NULLAGINO);
651 	agi->agi_dirino = cpu_to_be32(NULLAGINO);
652 	if (xfs_has_crc(mp))
653 		uuid_copy(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid);
654 	if (xfs_has_finobt(mp)) {
655 		agi->agi_free_root = cpu_to_be32(XFS_FIBT_BLOCK(mp));
656 		agi->agi_free_level = cpu_to_be32(1);
657 	}
658 	for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++)
659 		agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
660 	if (xfs_has_inobtcounts(mp)) {
661 		agi->agi_iblocks = cpu_to_be32(1);
662 		if (xfs_has_finobt(mp))
663 			agi->agi_fblocks = cpu_to_be32(1);
664 	}
665 }
666 
667 typedef void (*aghdr_init_work_f)(struct xfs_mount *mp, struct xfs_buf *bp,
668 				  struct aghdr_init_data *id);
669 static int
xfs_ag_init_hdr(struct xfs_mount * mp,struct aghdr_init_data * id,aghdr_init_work_f work,const struct xfs_buf_ops * ops)670 xfs_ag_init_hdr(
671 	struct xfs_mount	*mp,
672 	struct aghdr_init_data	*id,
673 	aghdr_init_work_f	work,
674 	const struct xfs_buf_ops *ops)
675 {
676 	struct xfs_buf		*bp;
677 	int			error;
678 
679 	error = xfs_get_aghdr_buf(mp, id->daddr, id->numblks, &bp, ops);
680 	if (error)
681 		return error;
682 
683 	(*work)(mp, bp, id);
684 
685 	xfs_buf_delwri_queue(bp, &id->buffer_list);
686 	xfs_buf_relse(bp);
687 	return 0;
688 }
689 
690 struct xfs_aghdr_grow_data {
691 	xfs_daddr_t		daddr;
692 	size_t			numblks;
693 	const struct xfs_buf_ops *ops;
694 	aghdr_init_work_f	work;
695 	const struct xfs_btree_ops *bc_ops;
696 	bool			need_init;
697 };
698 
699 /*
700  * Prepare new AG headers to be written to disk. We use uncached buffers here,
701  * as it is assumed these new AG headers are currently beyond the currently
702  * valid filesystem address space. Using cached buffers would trip over EOFS
703  * corruption detection alogrithms in the buffer cache lookup routines.
704  *
705  * This is a non-transactional function, but the prepared buffers are added to a
706  * delayed write buffer list supplied by the caller so they can submit them to
707  * disk and wait on them as required.
708  */
709 int
xfs_ag_init_headers(struct xfs_mount * mp,struct aghdr_init_data * id)710 xfs_ag_init_headers(
711 	struct xfs_mount	*mp,
712 	struct aghdr_init_data	*id)
713 
714 {
715 	struct xfs_aghdr_grow_data aghdr_data[] = {
716 	{ /* SB */
717 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_SB_DADDR),
718 		.numblks = XFS_FSS_TO_BB(mp, 1),
719 		.ops = &xfs_sb_buf_ops,
720 		.work = &xfs_sbblock_init,
721 		.need_init = true
722 	},
723 	{ /* AGF */
724 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGF_DADDR(mp)),
725 		.numblks = XFS_FSS_TO_BB(mp, 1),
726 		.ops = &xfs_agf_buf_ops,
727 		.work = &xfs_agfblock_init,
728 		.need_init = true
729 	},
730 	{ /* AGFL */
731 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGFL_DADDR(mp)),
732 		.numblks = XFS_FSS_TO_BB(mp, 1),
733 		.ops = &xfs_agfl_buf_ops,
734 		.work = &xfs_agflblock_init,
735 		.need_init = true
736 	},
737 	{ /* AGI */
738 		.daddr = XFS_AG_DADDR(mp, id->agno, XFS_AGI_DADDR(mp)),
739 		.numblks = XFS_FSS_TO_BB(mp, 1),
740 		.ops = &xfs_agi_buf_ops,
741 		.work = &xfs_agiblock_init,
742 		.need_init = true
743 	},
744 	{ /* BNO root block */
745 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_BNO_BLOCK(mp)),
746 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
747 		.ops = &xfs_bnobt_buf_ops,
748 		.work = &xfs_bnoroot_init,
749 		.bc_ops = &xfs_bnobt_ops,
750 		.need_init = true
751 	},
752 	{ /* CNT root block */
753 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_CNT_BLOCK(mp)),
754 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
755 		.ops = &xfs_cntbt_buf_ops,
756 		.work = &xfs_bnoroot_init,
757 		.bc_ops = &xfs_cntbt_ops,
758 		.need_init = true
759 	},
760 	{ /* INO root block */
761 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_IBT_BLOCK(mp)),
762 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
763 		.ops = &xfs_inobt_buf_ops,
764 		.work = &xfs_btroot_init,
765 		.bc_ops = &xfs_inobt_ops,
766 		.need_init = true
767 	},
768 	{ /* FINO root block */
769 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_FIBT_BLOCK(mp)),
770 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
771 		.ops = &xfs_finobt_buf_ops,
772 		.work = &xfs_btroot_init,
773 		.bc_ops = &xfs_finobt_ops,
774 		.need_init =  xfs_has_finobt(mp)
775 	},
776 	{ /* RMAP root block */
777 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, XFS_RMAP_BLOCK(mp)),
778 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
779 		.ops = &xfs_rmapbt_buf_ops,
780 		.work = &xfs_rmaproot_init,
781 		.bc_ops = &xfs_rmapbt_ops,
782 		.need_init = xfs_has_rmapbt(mp)
783 	},
784 	{ /* REFC root block */
785 		.daddr = XFS_AGB_TO_DADDR(mp, id->agno, xfs_refc_block(mp)),
786 		.numblks = BTOBB(mp->m_sb.sb_blocksize),
787 		.ops = &xfs_refcountbt_buf_ops,
788 		.work = &xfs_btroot_init,
789 		.bc_ops = &xfs_refcountbt_ops,
790 		.need_init = xfs_has_reflink(mp)
791 	},
792 	{ /* NULL terminating block */
793 		.daddr = XFS_BUF_DADDR_NULL,
794 	}
795 	};
796 	struct  xfs_aghdr_grow_data *dp;
797 	int			error = 0;
798 
799 	/* Account for AG free space in new AG */
800 	id->nfree += id->agsize - mp->m_ag_prealloc_blocks;
801 	for (dp = &aghdr_data[0]; dp->daddr != XFS_BUF_DADDR_NULL; dp++) {
802 		if (!dp->need_init)
803 			continue;
804 
805 		id->daddr = dp->daddr;
806 		id->numblks = dp->numblks;
807 		id->bc_ops = dp->bc_ops;
808 		error = xfs_ag_init_hdr(mp, id, dp->work, dp->ops);
809 		if (error)
810 			break;
811 	}
812 	return error;
813 }
814 
815 int
xfs_ag_shrink_space(struct xfs_perag * pag,struct xfs_trans ** tpp,xfs_extlen_t delta)816 xfs_ag_shrink_space(
817 	struct xfs_perag	*pag,
818 	struct xfs_trans	**tpp,
819 	xfs_extlen_t		delta)
820 {
821 	struct xfs_mount	*mp = pag->pag_mount;
822 	struct xfs_alloc_arg	args = {
823 		.tp	= *tpp,
824 		.mp	= mp,
825 		.pag	= pag,
826 		.minlen = delta,
827 		.maxlen = delta,
828 		.oinfo	= XFS_RMAP_OINFO_SKIP_UPDATE,
829 		.resv	= XFS_AG_RESV_NONE,
830 		.prod	= 1
831 	};
832 	struct xfs_buf		*agibp, *agfbp;
833 	struct xfs_agi		*agi;
834 	struct xfs_agf		*agf;
835 	xfs_agblock_t		aglen;
836 	int			error, err2;
837 
838 	ASSERT(pag->pag_agno == mp->m_sb.sb_agcount - 1);
839 	error = xfs_ialloc_read_agi(pag, *tpp, 0, &agibp);
840 	if (error)
841 		return error;
842 
843 	agi = agibp->b_addr;
844 
845 	error = xfs_alloc_read_agf(pag, *tpp, 0, &agfbp);
846 	if (error)
847 		return error;
848 
849 	agf = agfbp->b_addr;
850 	aglen = be32_to_cpu(agi->agi_length);
851 	/* some extra paranoid checks before we shrink the ag */
852 	if (XFS_IS_CORRUPT(mp, agf->agf_length != agi->agi_length)) {
853 		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGF);
854 		return -EFSCORRUPTED;
855 	}
856 	if (delta >= aglen)
857 		return -EINVAL;
858 
859 	/*
860 	 * Make sure that the last inode cluster cannot overlap with the new
861 	 * end of the AG, even if it's sparse.
862 	 */
863 	error = xfs_ialloc_check_shrink(pag, *tpp, agibp, aglen - delta);
864 	if (error)
865 		return error;
866 
867 	/*
868 	 * Disable perag reservations so it doesn't cause the allocation request
869 	 * to fail. We'll reestablish reservation before we return.
870 	 */
871 	xfs_ag_resv_free(pag);
872 
873 	/* internal log shouldn't also show up in the free space btrees */
874 	error = xfs_alloc_vextent_exact_bno(&args,
875 			XFS_AGB_TO_FSB(mp, pag->pag_agno, aglen - delta));
876 	if (!error && args.agbno == NULLAGBLOCK)
877 		error = -ENOSPC;
878 
879 	if (error) {
880 		/*
881 		 * If extent allocation fails, need to roll the transaction to
882 		 * ensure that the AGFL fixup has been committed anyway.
883 		 *
884 		 * We need to hold the AGF across the roll to ensure nothing can
885 		 * access the AG for allocation until the shrink is fully
886 		 * cleaned up. And due to the resetting of the AG block
887 		 * reservation space needing to lock the AGI, we also have to
888 		 * hold that so we don't get AGI/AGF lock order inversions in
889 		 * the error handling path.
890 		 */
891 		xfs_trans_bhold(*tpp, agfbp);
892 		xfs_trans_bhold(*tpp, agibp);
893 		err2 = xfs_trans_roll(tpp);
894 		if (err2)
895 			return err2;
896 		xfs_trans_bjoin(*tpp, agfbp);
897 		xfs_trans_bjoin(*tpp, agibp);
898 		goto resv_init_out;
899 	}
900 
901 	/*
902 	 * if successfully deleted from freespace btrees, need to confirm
903 	 * per-AG reservation works as expected.
904 	 */
905 	be32_add_cpu(&agi->agi_length, -delta);
906 	be32_add_cpu(&agf->agf_length, -delta);
907 
908 	err2 = xfs_ag_resv_init(pag, *tpp);
909 	if (err2) {
910 		be32_add_cpu(&agi->agi_length, delta);
911 		be32_add_cpu(&agf->agf_length, delta);
912 		if (err2 != -ENOSPC)
913 			goto resv_err;
914 
915 		err2 = xfs_free_extent_later(*tpp, args.fsbno, delta, NULL,
916 				XFS_AG_RESV_NONE, XFS_FREE_EXTENT_SKIP_DISCARD);
917 		if (err2)
918 			goto resv_err;
919 
920 		/*
921 		 * Roll the transaction before trying to re-init the per-ag
922 		 * reservation. The new transaction is clean so it will cancel
923 		 * without any side effects.
924 		 */
925 		error = xfs_defer_finish(tpp);
926 		if (error)
927 			return error;
928 
929 		error = -ENOSPC;
930 		goto resv_init_out;
931 	}
932 
933 	/* Update perag geometry */
934 	pag->block_count -= delta;
935 	__xfs_agino_range(pag->pag_mount, pag->block_count, &pag->agino_min,
936 				&pag->agino_max);
937 
938 	xfs_ialloc_log_agi(*tpp, agibp, XFS_AGI_LENGTH);
939 	xfs_alloc_log_agf(*tpp, agfbp, XFS_AGF_LENGTH);
940 	return 0;
941 
942 resv_init_out:
943 	err2 = xfs_ag_resv_init(pag, *tpp);
944 	if (!err2)
945 		return error;
946 resv_err:
947 	xfs_warn(mp, "Error %d reserving per-AG metadata reserve pool.", err2);
948 	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
949 	return err2;
950 }
951 
952 /*
953  * Extent the AG indicated by the @id by the length passed in
954  */
955 int
xfs_ag_extend_space(struct xfs_perag * pag,struct xfs_trans * tp,xfs_extlen_t len)956 xfs_ag_extend_space(
957 	struct xfs_perag	*pag,
958 	struct xfs_trans	*tp,
959 	xfs_extlen_t		len)
960 {
961 	struct xfs_buf		*bp;
962 	struct xfs_agi		*agi;
963 	struct xfs_agf		*agf;
964 	int			error;
965 
966 	ASSERT(pag->pag_agno == pag->pag_mount->m_sb.sb_agcount - 1);
967 
968 	error = xfs_ialloc_read_agi(pag, tp, 0, &bp);
969 	if (error)
970 		return error;
971 
972 	agi = bp->b_addr;
973 	be32_add_cpu(&agi->agi_length, len);
974 	xfs_ialloc_log_agi(tp, bp, XFS_AGI_LENGTH);
975 
976 	/*
977 	 * Change agf length.
978 	 */
979 	error = xfs_alloc_read_agf(pag, tp, 0, &bp);
980 	if (error)
981 		return error;
982 
983 	agf = bp->b_addr;
984 	be32_add_cpu(&agf->agf_length, len);
985 	ASSERT(agf->agf_length == agi->agi_length);
986 	xfs_alloc_log_agf(tp, bp, XFS_AGF_LENGTH);
987 
988 	/*
989 	 * Free the new space.
990 	 *
991 	 * XFS_RMAP_OINFO_SKIP_UPDATE is used here to tell the rmap btree that
992 	 * this doesn't actually exist in the rmap btree.
993 	 */
994 	error = xfs_rmap_free(tp, bp, pag, be32_to_cpu(agf->agf_length) - len,
995 				len, &XFS_RMAP_OINFO_SKIP_UPDATE);
996 	if (error)
997 		return error;
998 
999 	error = xfs_free_extent(tp, pag, be32_to_cpu(agf->agf_length) - len,
1000 			len, &XFS_RMAP_OINFO_SKIP_UPDATE, XFS_AG_RESV_NONE);
1001 	if (error)
1002 		return error;
1003 
1004 	/* Update perag geometry */
1005 	pag->block_count = be32_to_cpu(agf->agf_length);
1006 	__xfs_agino_range(pag->pag_mount, pag->block_count, &pag->agino_min,
1007 				&pag->agino_max);
1008 	return 0;
1009 }
1010 
1011 /* Retrieve AG geometry. */
1012 int
xfs_ag_get_geometry(struct xfs_perag * pag,struct xfs_ag_geometry * ageo)1013 xfs_ag_get_geometry(
1014 	struct xfs_perag	*pag,
1015 	struct xfs_ag_geometry	*ageo)
1016 {
1017 	struct xfs_buf		*agi_bp;
1018 	struct xfs_buf		*agf_bp;
1019 	struct xfs_agi		*agi;
1020 	struct xfs_agf		*agf;
1021 	unsigned int		freeblks;
1022 	int			error;
1023 
1024 	/* Lock the AG headers. */
1025 	error = xfs_ialloc_read_agi(pag, NULL, 0, &agi_bp);
1026 	if (error)
1027 		return error;
1028 	error = xfs_alloc_read_agf(pag, NULL, 0, &agf_bp);
1029 	if (error)
1030 		goto out_agi;
1031 
1032 	/* Fill out form. */
1033 	memset(ageo, 0, sizeof(*ageo));
1034 	ageo->ag_number = pag->pag_agno;
1035 
1036 	agi = agi_bp->b_addr;
1037 	ageo->ag_icount = be32_to_cpu(agi->agi_count);
1038 	ageo->ag_ifree = be32_to_cpu(agi->agi_freecount);
1039 
1040 	agf = agf_bp->b_addr;
1041 	ageo->ag_length = be32_to_cpu(agf->agf_length);
1042 	freeblks = pag->pagf_freeblks +
1043 		   pag->pagf_flcount +
1044 		   pag->pagf_btreeblks -
1045 		   xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE);
1046 	ageo->ag_freeblks = freeblks;
1047 	xfs_ag_geom_health(pag, ageo);
1048 
1049 	/* Release resources. */
1050 	xfs_buf_relse(agf_bp);
1051 out_agi:
1052 	xfs_buf_relse(agi_bp);
1053 	return error;
1054 }
1055