xref: /linux/fs/xfs/xfs_iwalk.c (revision bbeb1219eeeeab7ef302fdaedee71b08e413a04c)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright (C) 2019 Oracle.  All Rights Reserved.
4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_inode.h"
14 #include "xfs_btree.h"
15 #include "xfs_ialloc.h"
16 #include "xfs_ialloc_btree.h"
17 #include "xfs_iwalk.h"
18 #include "xfs_error.h"
19 #include "xfs_trace.h"
20 #include "xfs_icache.h"
21 #include "xfs_health.h"
22 #include "xfs_trans.h"
23 #include "xfs_pwork.h"
24 #include "xfs_ag.h"
25 #include "xfs_bit.h"
26 
27 /*
28  * Walking Inodes in the Filesystem
29  * ================================
30  *
31  * This iterator function walks a subset of filesystem inodes in increasing
32  * order from @startino until there are no more inodes.  For each allocated
33  * inode it finds, it calls a walk function with the relevant inode number and
34  * a pointer to caller-provided data.  The walk function can return the usual
35  * negative error code to stop the iteration; 0 to continue the iteration; or
36  * -ECANCELED to stop the iteration.  This return value is returned to the
37  * caller.
38  *
39  * Internally, we allow the walk function to do anything, which means that we
40  * cannot maintain the inobt cursor or our lock on the AGI buffer.  We
41  * therefore cache the inobt records in kernel memory and only call the walk
42  * function when our memory buffer is full.  @nr_recs is the number of records
43  * that we've cached, and @sz_recs is the size of our cache.
44  *
45  * It is the responsibility of the walk function to ensure it accesses
46  * allocated inodes, as the inobt records may be stale by the time they are
47  * acted upon.
48  */
49 
50 struct xfs_iwalk_ag {
51 	/* parallel work control data; will be null if single threaded */
52 	struct xfs_pwork		pwork;
53 
54 	struct xfs_mount		*mp;
55 	struct xfs_trans		*tp;
56 	struct xfs_perag		*pag;
57 
58 	/* Where do we start the traversal? */
59 	xfs_ino_t			startino;
60 
61 	/* What was the last inode number we saw when iterating the inobt? */
62 	xfs_ino_t			lastino;
63 
64 	/* Array of inobt records we cache. */
65 	struct xfs_inobt_rec_incore	*recs;
66 
67 	/* Number of entries allocated for the @recs array. */
68 	unsigned int			sz_recs;
69 
70 	/* Number of entries in the @recs array that are in use. */
71 	unsigned int			nr_recs;
72 
73 	/* Inode walk function and data pointer. */
74 	xfs_iwalk_fn			iwalk_fn;
75 	xfs_inobt_walk_fn		inobt_walk_fn;
76 	void				*data;
77 
78 	/*
79 	 * Make it look like the inodes up to startino are free so that
80 	 * bulkstat can start its inode iteration at the correct place without
81 	 * needing to special case everywhere.
82 	 */
83 	unsigned int			trim_start:1;
84 
85 	/* Skip empty inobt records? */
86 	unsigned int			skip_empty:1;
87 
88 	/* Drop the (hopefully empty) transaction when calling iwalk_fn. */
89 	unsigned int			drop_trans:1;
90 };
91 
92 /*
93  * Loop over all clusters in a chunk for a given incore inode allocation btree
94  * record.  Do a readahead if there are any allocated inodes in that cluster.
95  */
96 STATIC void
xfs_iwalk_ichunk_ra(struct xfs_mount * mp,struct xfs_perag * pag,struct xfs_inobt_rec_incore * irec)97 xfs_iwalk_ichunk_ra(
98 	struct xfs_mount		*mp,
99 	struct xfs_perag		*pag,
100 	struct xfs_inobt_rec_incore	*irec)
101 {
102 	struct xfs_ino_geometry		*igeo = M_IGEO(mp);
103 	xfs_agnumber_t			agno = pag->pag_agno;
104 	xfs_agblock_t			agbno;
105 	struct blk_plug			plug;
106 	int				i;	/* inode chunk index */
107 
108 	agbno = XFS_AGINO_TO_AGBNO(mp, irec->ir_startino);
109 
110 	blk_start_plug(&plug);
111 	for (i = 0; i < XFS_INODES_PER_CHUNK; i += igeo->inodes_per_cluster) {
112 		xfs_inofree_t	imask;
113 
114 		imask = xfs_inobt_maskn(i, igeo->inodes_per_cluster);
115 		if (imask & ~irec->ir_free) {
116 			xfs_buf_readahead(mp->m_ddev_targp,
117 					XFS_AGB_TO_DADDR(mp, agno, agbno),
118 					igeo->blocks_per_cluster * mp->m_bsize,
119 					&xfs_inode_buf_ops);
120 		}
121 		agbno += igeo->blocks_per_cluster;
122 	}
123 	blk_finish_plug(&plug);
124 }
125 
126 /*
127  * Set the bits in @irec's free mask that correspond to the inodes before
128  * @agino so that we skip them.  This is how we restart an inode walk that was
129  * interrupted in the middle of an inode record.
130  */
131 STATIC void
xfs_iwalk_adjust_start(xfs_agino_t agino,struct xfs_inobt_rec_incore * irec)132 xfs_iwalk_adjust_start(
133 	xfs_agino_t			agino,	/* starting inode of chunk */
134 	struct xfs_inobt_rec_incore	*irec)	/* btree record */
135 {
136 	int				idx;	/* index into inode chunk */
137 
138 	idx = agino - irec->ir_startino;
139 
140 	irec->ir_free |= xfs_inobt_maskn(0, idx);
141 	irec->ir_freecount = hweight64(irec->ir_free);
142 }
143 
144 /* Allocate memory for a walk. */
145 STATIC int
xfs_iwalk_alloc(struct xfs_iwalk_ag * iwag)146 xfs_iwalk_alloc(
147 	struct xfs_iwalk_ag	*iwag)
148 {
149 	size_t			size;
150 
151 	ASSERT(iwag->recs == NULL);
152 	iwag->nr_recs = 0;
153 
154 	/* Allocate a prefetch buffer for inobt records. */
155 	size = iwag->sz_recs * sizeof(struct xfs_inobt_rec_incore);
156 	iwag->recs = kmalloc(size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
157 	if (iwag->recs == NULL)
158 		return -ENOMEM;
159 
160 	return 0;
161 }
162 
163 /* Free memory we allocated for a walk. */
164 STATIC void
xfs_iwalk_free(struct xfs_iwalk_ag * iwag)165 xfs_iwalk_free(
166 	struct xfs_iwalk_ag	*iwag)
167 {
168 	kfree(iwag->recs);
169 	iwag->recs = NULL;
170 }
171 
172 /* For each inuse inode in each cached inobt record, call our function. */
173 STATIC int
xfs_iwalk_ag_recs(struct xfs_iwalk_ag * iwag)174 xfs_iwalk_ag_recs(
175 	struct xfs_iwalk_ag	*iwag)
176 {
177 	struct xfs_mount	*mp = iwag->mp;
178 	struct xfs_trans	*tp = iwag->tp;
179 	struct xfs_perag	*pag = iwag->pag;
180 	xfs_ino_t		ino;
181 	unsigned int		i, j;
182 	int			error;
183 
184 	for (i = 0; i < iwag->nr_recs; i++) {
185 		struct xfs_inobt_rec_incore	*irec = &iwag->recs[i];
186 
187 		trace_xfs_iwalk_ag_rec(mp, pag->pag_agno, irec);
188 
189 		if (xfs_pwork_want_abort(&iwag->pwork))
190 			return 0;
191 
192 		if (iwag->inobt_walk_fn) {
193 			error = iwag->inobt_walk_fn(mp, tp, pag->pag_agno, irec,
194 					iwag->data);
195 			if (error)
196 				return error;
197 		}
198 
199 		if (!iwag->iwalk_fn)
200 			continue;
201 
202 		for (j = 0; j < XFS_INODES_PER_CHUNK; j++) {
203 			if (xfs_pwork_want_abort(&iwag->pwork))
204 				return 0;
205 
206 			/* Skip if this inode is free */
207 			if (XFS_INOBT_MASK(j) & irec->ir_free)
208 				continue;
209 
210 			/* Otherwise call our function. */
211 			ino = XFS_AGINO_TO_INO(mp, pag->pag_agno,
212 						irec->ir_startino + j);
213 			error = iwag->iwalk_fn(mp, tp, ino, iwag->data);
214 			if (error)
215 				return error;
216 		}
217 	}
218 
219 	return 0;
220 }
221 
222 /* Delete cursor and let go of AGI. */
223 static inline void
xfs_iwalk_del_inobt(struct xfs_trans * tp,struct xfs_btree_cur ** curpp,struct xfs_buf ** agi_bpp,int error)224 xfs_iwalk_del_inobt(
225 	struct xfs_trans	*tp,
226 	struct xfs_btree_cur	**curpp,
227 	struct xfs_buf		**agi_bpp,
228 	int			error)
229 {
230 	if (*curpp) {
231 		xfs_btree_del_cursor(*curpp, error);
232 		*curpp = NULL;
233 	}
234 	if (*agi_bpp) {
235 		xfs_trans_brelse(tp, *agi_bpp);
236 		*agi_bpp = NULL;
237 	}
238 }
239 
240 /*
241  * Set ourselves up for walking inobt records starting from a given point in
242  * the filesystem.
243  *
244  * If caller passed in a nonzero start inode number, load the record from the
245  * inobt and make the record look like all the inodes before agino are free so
246  * that we skip them, and then move the cursor to the next inobt record.  This
247  * is how we support starting an iwalk in the middle of an inode chunk.
248  *
249  * If the caller passed in a start number of zero, move the cursor to the first
250  * inobt record.
251  *
252  * The caller is responsible for cleaning up the cursor and buffer pointer
253  * regardless of the error status.
254  */
255 STATIC int
xfs_iwalk_ag_start(struct xfs_iwalk_ag * iwag,xfs_agino_t agino,struct xfs_btree_cur ** curpp,struct xfs_buf ** agi_bpp,int * has_more)256 xfs_iwalk_ag_start(
257 	struct xfs_iwalk_ag	*iwag,
258 	xfs_agino_t		agino,
259 	struct xfs_btree_cur	**curpp,
260 	struct xfs_buf		**agi_bpp,
261 	int			*has_more)
262 {
263 	struct xfs_mount	*mp = iwag->mp;
264 	struct xfs_trans	*tp = iwag->tp;
265 	struct xfs_perag	*pag = iwag->pag;
266 	struct xfs_inobt_rec_incore *irec;
267 	int			error;
268 
269 	/* Set up a fresh cursor and empty the inobt cache. */
270 	iwag->nr_recs = 0;
271 	error = xfs_ialloc_read_agi(pag, tp, 0, agi_bpp);
272 	if (error)
273 		return error;
274 	*curpp = xfs_inobt_init_cursor(pag, tp, *agi_bpp);
275 
276 	/* Starting at the beginning of the AG?  That's easy! */
277 	if (agino == 0)
278 		return xfs_inobt_lookup(*curpp, 0, XFS_LOOKUP_GE, has_more);
279 
280 	/*
281 	 * Otherwise, we have to grab the inobt record where we left off, stuff
282 	 * the record into our cache, and then see if there are more records.
283 	 * We require a lookup cache of at least two elements so that the
284 	 * caller doesn't have to deal with tearing down the cursor to walk the
285 	 * records.
286 	 */
287 	error = xfs_inobt_lookup(*curpp, agino, XFS_LOOKUP_LE, has_more);
288 	if (error)
289 		return error;
290 
291 	/*
292 	 * If the LE lookup at @agino yields no records, jump ahead to the
293 	 * inobt cursor increment to see if there are more records to process.
294 	 */
295 	if (!*has_more)
296 		goto out_advance;
297 
298 	/* Get the record, should always work */
299 	irec = &iwag->recs[iwag->nr_recs];
300 	error = xfs_inobt_get_rec(*curpp, irec, has_more);
301 	if (error)
302 		return error;
303 	if (XFS_IS_CORRUPT(mp, *has_more != 1)) {
304 		xfs_btree_mark_sick(*curpp);
305 		return -EFSCORRUPTED;
306 	}
307 
308 	iwag->lastino = XFS_AGINO_TO_INO(mp, pag->pag_agno,
309 				irec->ir_startino + XFS_INODES_PER_CHUNK - 1);
310 
311 	/*
312 	 * If the LE lookup yielded an inobt record before the cursor position,
313 	 * skip it and see if there's another one after it.
314 	 */
315 	if (irec->ir_startino + XFS_INODES_PER_CHUNK <= agino)
316 		goto out_advance;
317 
318 	/*
319 	 * If agino fell in the middle of the inode record, make it look like
320 	 * the inodes up to agino are free so that we don't return them again.
321 	 */
322 	if (iwag->trim_start)
323 		xfs_iwalk_adjust_start(agino, irec);
324 
325 	/*
326 	 * The prefetch calculation is supposed to give us a large enough inobt
327 	 * record cache that grab_ichunk can stage a partial first record and
328 	 * the loop body can cache a record without having to check for cache
329 	 * space until after it reads an inobt record.
330 	 */
331 	iwag->nr_recs++;
332 	ASSERT(iwag->nr_recs < iwag->sz_recs);
333 
334 out_advance:
335 	return xfs_btree_increment(*curpp, 0, has_more);
336 }
337 
338 /*
339  * The inobt record cache is full, so preserve the inobt cursor state and
340  * run callbacks on the cached inobt records.  When we're done, restore the
341  * cursor state to wherever the cursor would have been had the cache not been
342  * full (and therefore we could've just incremented the cursor) if *@has_more
343  * is true.  On exit, *@has_more will indicate whether or not the caller should
344  * try for more inode records.
345  */
346 STATIC int
xfs_iwalk_run_callbacks(struct xfs_iwalk_ag * iwag,struct xfs_btree_cur ** curpp,struct xfs_buf ** agi_bpp,int * has_more)347 xfs_iwalk_run_callbacks(
348 	struct xfs_iwalk_ag		*iwag,
349 	struct xfs_btree_cur		**curpp,
350 	struct xfs_buf			**agi_bpp,
351 	int				*has_more)
352 {
353 	struct xfs_mount		*mp = iwag->mp;
354 	xfs_agino_t			next_agino;
355 	int				error;
356 
357 	next_agino = XFS_INO_TO_AGINO(mp, iwag->lastino) + 1;
358 
359 	ASSERT(iwag->nr_recs > 0);
360 
361 	/* Delete cursor but remember the last record we cached... */
362 	xfs_iwalk_del_inobt(iwag->tp, curpp, agi_bpp, 0);
363 	ASSERT(next_agino >= iwag->recs[iwag->nr_recs - 1].ir_startino +
364 			XFS_INODES_PER_CHUNK);
365 
366 	if (iwag->drop_trans) {
367 		xfs_trans_cancel(iwag->tp);
368 		iwag->tp = NULL;
369 	}
370 
371 	error = xfs_iwalk_ag_recs(iwag);
372 	if (error)
373 		return error;
374 
375 	/* ...empty the cache... */
376 	iwag->nr_recs = 0;
377 
378 	if (!has_more)
379 		return 0;
380 
381 	if (iwag->drop_trans) {
382 		error = xfs_trans_alloc_empty(mp, &iwag->tp);
383 		if (error)
384 			return error;
385 	}
386 
387 	/* ...and recreate the cursor just past where we left off. */
388 	error = xfs_ialloc_read_agi(iwag->pag, iwag->tp, 0, agi_bpp);
389 	if (error)
390 		return error;
391 	*curpp = xfs_inobt_init_cursor(iwag->pag, iwag->tp, *agi_bpp);
392 	return xfs_inobt_lookup(*curpp, next_agino, XFS_LOOKUP_GE, has_more);
393 }
394 
395 /* Walk all inodes in a single AG, from @iwag->startino to the end of the AG. */
396 STATIC int
xfs_iwalk_ag(struct xfs_iwalk_ag * iwag)397 xfs_iwalk_ag(
398 	struct xfs_iwalk_ag		*iwag)
399 {
400 	struct xfs_mount		*mp = iwag->mp;
401 	struct xfs_perag		*pag = iwag->pag;
402 	struct xfs_buf			*agi_bp = NULL;
403 	struct xfs_btree_cur		*cur = NULL;
404 	xfs_agino_t			agino;
405 	int				has_more;
406 	int				error = 0;
407 
408 	/* Set up our cursor at the right place in the inode btree. */
409 	ASSERT(pag->pag_agno == XFS_INO_TO_AGNO(mp, iwag->startino));
410 	agino = XFS_INO_TO_AGINO(mp, iwag->startino);
411 	error = xfs_iwalk_ag_start(iwag, agino, &cur, &agi_bp, &has_more);
412 
413 	while (!error && has_more) {
414 		struct xfs_inobt_rec_incore	*irec;
415 		xfs_ino_t			rec_fsino;
416 
417 		cond_resched();
418 		if (xfs_pwork_want_abort(&iwag->pwork))
419 			goto out;
420 
421 		/* Fetch the inobt record. */
422 		irec = &iwag->recs[iwag->nr_recs];
423 		error = xfs_inobt_get_rec(cur, irec, &has_more);
424 		if (error || !has_more)
425 			break;
426 
427 		/* Make sure that we always move forward. */
428 		rec_fsino = XFS_AGINO_TO_INO(mp, pag->pag_agno, irec->ir_startino);
429 		if (iwag->lastino != NULLFSINO &&
430 		    XFS_IS_CORRUPT(mp, iwag->lastino >= rec_fsino)) {
431 			xfs_btree_mark_sick(cur);
432 			error = -EFSCORRUPTED;
433 			goto out;
434 		}
435 		iwag->lastino = rec_fsino + XFS_INODES_PER_CHUNK - 1;
436 
437 		/* No allocated inodes in this chunk; skip it. */
438 		if (iwag->skip_empty && irec->ir_freecount == irec->ir_count) {
439 			error = xfs_btree_increment(cur, 0, &has_more);
440 			if (error)
441 				break;
442 			continue;
443 		}
444 
445 		/*
446 		 * Start readahead for this inode chunk in anticipation of
447 		 * walking the inodes.
448 		 */
449 		if (iwag->iwalk_fn)
450 			xfs_iwalk_ichunk_ra(mp, pag, irec);
451 
452 		/*
453 		 * If there's space in the buffer for more records, increment
454 		 * the btree cursor and grab more.
455 		 */
456 		if (++iwag->nr_recs < iwag->sz_recs) {
457 			error = xfs_btree_increment(cur, 0, &has_more);
458 			if (error || !has_more)
459 				break;
460 			continue;
461 		}
462 
463 		/*
464 		 * Otherwise, we need to save cursor state and run the callback
465 		 * function on the cached records.  The run_callbacks function
466 		 * is supposed to return a cursor pointing to the record where
467 		 * we would be if we had been able to increment like above.
468 		 */
469 		ASSERT(has_more);
470 		error = xfs_iwalk_run_callbacks(iwag, &cur, &agi_bp, &has_more);
471 	}
472 
473 	if (iwag->nr_recs == 0 || error)
474 		goto out;
475 
476 	/* Walk the unprocessed records in the cache. */
477 	error = xfs_iwalk_run_callbacks(iwag, &cur, &agi_bp, &has_more);
478 
479 out:
480 	xfs_iwalk_del_inobt(iwag->tp, &cur, &agi_bp, error);
481 	return error;
482 }
483 
484 /*
485  * We experimentally determined that the reduction in ioctl call overhead
486  * diminishes when userspace asks for more than 2048 inodes, so we'll cap
487  * prefetch at this point.
488  */
489 #define IWALK_MAX_INODE_PREFETCH	(2048U)
490 
491 /*
492  * Given the number of inodes to prefetch, set the number of inobt records that
493  * we cache in memory, which controls the number of inodes we try to read
494  * ahead.  Set the maximum if @inodes == 0.
495  */
496 static inline unsigned int
xfs_iwalk_prefetch(unsigned int inodes)497 xfs_iwalk_prefetch(
498 	unsigned int		inodes)
499 {
500 	unsigned int		inobt_records;
501 
502 	/*
503 	 * If the caller didn't tell us the number of inodes they wanted,
504 	 * assume the maximum prefetch possible for best performance.
505 	 * Otherwise, cap prefetch at that maximum so that we don't start an
506 	 * absurd amount of prefetch.
507 	 */
508 	if (inodes == 0)
509 		inodes = IWALK_MAX_INODE_PREFETCH;
510 	inodes = min(inodes, IWALK_MAX_INODE_PREFETCH);
511 
512 	/* Round the inode count up to a full chunk. */
513 	inodes = round_up(inodes, XFS_INODES_PER_CHUNK);
514 
515 	/*
516 	 * In order to convert the number of inodes to prefetch into an
517 	 * estimate of the number of inobt records to cache, we require a
518 	 * conversion factor that reflects our expectations of the average
519 	 * loading factor of an inode chunk.  Based on data gathered, most
520 	 * (but not all) filesystems manage to keep the inode chunks totally
521 	 * full, so we'll underestimate slightly so that our readahead will
522 	 * still deliver the performance we want on aging filesystems:
523 	 *
524 	 * inobt = inodes / (INODES_PER_CHUNK * (4 / 5));
525 	 *
526 	 * The funny math is to avoid integer division.
527 	 */
528 	inobt_records = (inodes * 5) / (4 * XFS_INODES_PER_CHUNK);
529 
530 	/*
531 	 * Allocate enough space to prefetch at least two inobt records so that
532 	 * we can cache both the record where the iwalk started and the next
533 	 * record.  This simplifies the AG inode walk loop setup code.
534 	 */
535 	return max(inobt_records, 2U);
536 }
537 
538 /*
539  * Walk all inodes in the filesystem starting from @startino.  The @iwalk_fn
540  * will be called for each allocated inode, being passed the inode's number and
541  * @data.  @max_prefetch controls how many inobt records' worth of inodes we
542  * try to readahead.
543  */
544 int
xfs_iwalk(struct xfs_mount * mp,struct xfs_trans * tp,xfs_ino_t startino,unsigned int flags,xfs_iwalk_fn iwalk_fn,unsigned int inode_records,void * data)545 xfs_iwalk(
546 	struct xfs_mount	*mp,
547 	struct xfs_trans	*tp,
548 	xfs_ino_t		startino,
549 	unsigned int		flags,
550 	xfs_iwalk_fn		iwalk_fn,
551 	unsigned int		inode_records,
552 	void			*data)
553 {
554 	struct xfs_iwalk_ag	iwag = {
555 		.mp		= mp,
556 		.tp		= tp,
557 		.iwalk_fn	= iwalk_fn,
558 		.data		= data,
559 		.startino	= startino,
560 		.sz_recs	= xfs_iwalk_prefetch(inode_records),
561 		.trim_start	= 1,
562 		.skip_empty	= 1,
563 		.pwork		= XFS_PWORK_SINGLE_THREADED,
564 		.lastino	= NULLFSINO,
565 	};
566 	struct xfs_perag	*pag;
567 	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, startino);
568 	int			error;
569 
570 	ASSERT(agno < mp->m_sb.sb_agcount);
571 	ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
572 
573 	error = xfs_iwalk_alloc(&iwag);
574 	if (error)
575 		return error;
576 
577 	for_each_perag_from(mp, agno, pag) {
578 		iwag.pag = pag;
579 		error = xfs_iwalk_ag(&iwag);
580 		if (error)
581 			break;
582 		iwag.startino = XFS_AGINO_TO_INO(mp, agno + 1, 0);
583 		if (flags & XFS_INOBT_WALK_SAME_AG)
584 			break;
585 		iwag.pag = NULL;
586 	}
587 
588 	if (iwag.pag)
589 		xfs_perag_rele(pag);
590 	xfs_iwalk_free(&iwag);
591 	return error;
592 }
593 
594 /* Run per-thread iwalk work. */
595 static int
xfs_iwalk_ag_work(struct xfs_mount * mp,struct xfs_pwork * pwork)596 xfs_iwalk_ag_work(
597 	struct xfs_mount	*mp,
598 	struct xfs_pwork	*pwork)
599 {
600 	struct xfs_iwalk_ag	*iwag;
601 	int			error = 0;
602 
603 	iwag = container_of(pwork, struct xfs_iwalk_ag, pwork);
604 	if (xfs_pwork_want_abort(pwork))
605 		goto out;
606 
607 	error = xfs_iwalk_alloc(iwag);
608 	if (error)
609 		goto out;
610 	/*
611 	 * Grab an empty transaction so that we can use its recursive buffer
612 	 * locking abilities to detect cycles in the inobt without deadlocking.
613 	 */
614 	error = xfs_trans_alloc_empty(mp, &iwag->tp);
615 	if (error)
616 		goto out;
617 	iwag->drop_trans = 1;
618 
619 	error = xfs_iwalk_ag(iwag);
620 	if (iwag->tp)
621 		xfs_trans_cancel(iwag->tp);
622 	xfs_iwalk_free(iwag);
623 out:
624 	xfs_perag_put(iwag->pag);
625 	kfree(iwag);
626 	return error;
627 }
628 
629 /*
630  * Walk all the inodes in the filesystem using multiple threads to process each
631  * AG.
632  */
633 int
xfs_iwalk_threaded(struct xfs_mount * mp,xfs_ino_t startino,unsigned int flags,xfs_iwalk_fn iwalk_fn,unsigned int inode_records,bool polled,void * data)634 xfs_iwalk_threaded(
635 	struct xfs_mount	*mp,
636 	xfs_ino_t		startino,
637 	unsigned int		flags,
638 	xfs_iwalk_fn		iwalk_fn,
639 	unsigned int		inode_records,
640 	bool			polled,
641 	void			*data)
642 {
643 	struct xfs_pwork_ctl	pctl;
644 	struct xfs_perag	*pag;
645 	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, startino);
646 	int			error;
647 
648 	ASSERT(agno < mp->m_sb.sb_agcount);
649 	ASSERT(!(flags & ~XFS_IWALK_FLAGS_ALL));
650 
651 	error = xfs_pwork_init(mp, &pctl, xfs_iwalk_ag_work, "xfs_iwalk");
652 	if (error)
653 		return error;
654 
655 	for_each_perag_from(mp, agno, pag) {
656 		struct xfs_iwalk_ag	*iwag;
657 
658 		if (xfs_pwork_ctl_want_abort(&pctl))
659 			break;
660 
661 		iwag = kzalloc(sizeof(struct xfs_iwalk_ag),
662 				GFP_KERNEL | __GFP_NOFAIL);
663 		iwag->mp = mp;
664 
665 		/*
666 		 * perag is being handed off to async work, so take a passive
667 		 * reference for the async work to release.
668 		 */
669 		iwag->pag = xfs_perag_hold(pag);
670 		iwag->iwalk_fn = iwalk_fn;
671 		iwag->data = data;
672 		iwag->startino = startino;
673 		iwag->sz_recs = xfs_iwalk_prefetch(inode_records);
674 		iwag->lastino = NULLFSINO;
675 		xfs_pwork_queue(&pctl, &iwag->pwork);
676 		startino = XFS_AGINO_TO_INO(mp, pag->pag_agno + 1, 0);
677 		if (flags & XFS_INOBT_WALK_SAME_AG)
678 			break;
679 	}
680 	if (pag)
681 		xfs_perag_rele(pag);
682 	if (polled)
683 		xfs_pwork_poll(&pctl);
684 	return xfs_pwork_destroy(&pctl);
685 }
686 
687 /*
688  * Allow callers to cache up to a page's worth of inobt records.  This reflects
689  * the existing inumbers prefetching behavior.  Since the inobt walk does not
690  * itself do anything with the inobt records, we can set a fairly high limit
691  * here.
692  */
693 #define MAX_INOBT_WALK_PREFETCH	\
694 	(PAGE_SIZE / sizeof(struct xfs_inobt_rec_incore))
695 
696 /*
697  * Given the number of records that the user wanted, set the number of inobt
698  * records that we buffer in memory.  Set the maximum if @inobt_records == 0.
699  */
700 static inline unsigned int
xfs_inobt_walk_prefetch(unsigned int inobt_records)701 xfs_inobt_walk_prefetch(
702 	unsigned int		inobt_records)
703 {
704 	/*
705 	 * If the caller didn't tell us the number of inobt records they
706 	 * wanted, assume the maximum prefetch possible for best performance.
707 	 */
708 	if (inobt_records == 0)
709 		inobt_records = MAX_INOBT_WALK_PREFETCH;
710 
711 	/*
712 	 * Allocate enough space to prefetch at least two inobt records so that
713 	 * we can cache both the record where the iwalk started and the next
714 	 * record.  This simplifies the AG inode walk loop setup code.
715 	 */
716 	inobt_records = max(inobt_records, 2U);
717 
718 	/*
719 	 * Cap prefetch at that maximum so that we don't use an absurd amount
720 	 * of memory.
721 	 */
722 	return min_t(unsigned int, inobt_records, MAX_INOBT_WALK_PREFETCH);
723 }
724 
725 /*
726  * Walk all inode btree records in the filesystem starting from @startino.  The
727  * @inobt_walk_fn will be called for each btree record, being passed the incore
728  * record and @data.  @max_prefetch controls how many inobt records we try to
729  * cache ahead of time.
730  */
731 int
xfs_inobt_walk(struct xfs_mount * mp,struct xfs_trans * tp,xfs_ino_t startino,unsigned int flags,xfs_inobt_walk_fn inobt_walk_fn,unsigned int inobt_records,void * data)732 xfs_inobt_walk(
733 	struct xfs_mount	*mp,
734 	struct xfs_trans	*tp,
735 	xfs_ino_t		startino,
736 	unsigned int		flags,
737 	xfs_inobt_walk_fn	inobt_walk_fn,
738 	unsigned int		inobt_records,
739 	void			*data)
740 {
741 	struct xfs_iwalk_ag	iwag = {
742 		.mp		= mp,
743 		.tp		= tp,
744 		.inobt_walk_fn	= inobt_walk_fn,
745 		.data		= data,
746 		.startino	= startino,
747 		.sz_recs	= xfs_inobt_walk_prefetch(inobt_records),
748 		.pwork		= XFS_PWORK_SINGLE_THREADED,
749 		.lastino	= NULLFSINO,
750 	};
751 	struct xfs_perag	*pag;
752 	xfs_agnumber_t		agno = XFS_INO_TO_AGNO(mp, startino);
753 	int			error;
754 
755 	ASSERT(agno < mp->m_sb.sb_agcount);
756 	ASSERT(!(flags & ~XFS_INOBT_WALK_FLAGS_ALL));
757 
758 	error = xfs_iwalk_alloc(&iwag);
759 	if (error)
760 		return error;
761 
762 	for_each_perag_from(mp, agno, pag) {
763 		iwag.pag = pag;
764 		error = xfs_iwalk_ag(&iwag);
765 		if (error)
766 			break;
767 		iwag.startino = XFS_AGINO_TO_INO(mp, pag->pag_agno + 1, 0);
768 		if (flags & XFS_INOBT_WALK_SAME_AG)
769 			break;
770 		iwag.pag = NULL;
771 	}
772 
773 	if (iwag.pag)
774 		xfs_perag_rele(pag);
775 	xfs_iwalk_free(&iwag);
776 	return error;
777 }
778