zmap.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Copyright (C) 2018-2019 HUAWEI, Inc.
  4 *             https://www.huawei.com/
  5 */
  6#include "internal.h"
  7#include <linux/unaligned.h>
  8#include <trace/events/erofs.h>
  9
 10struct z_erofs_maprecorder {
 11	struct inode *inode;
 12	struct erofs_map_blocks *map;
 13	unsigned long lcn;
 14	/* compression extent information gathered */
 15	u8  type, headtype;
 16	u16 clusterofs;
 17	u16 delta[2];
 18	erofs_blk_t pblk, compressedblks;
 19	erofs_off_t nextpackoff;
 20	bool partialref, in_mbox;
 21};
 22
 23static int z_erofs_load_full_lcluster(struct z_erofs_maprecorder *m,
 24				      unsigned long lcn)
 25{
 26	struct inode *const inode = m->inode;
 27	struct erofs_inode *const vi = EROFS_I(inode);
 28	const erofs_off_t pos = Z_EROFS_FULL_INDEX_START(erofs_iloc(inode) +
 29			vi->inode_isize + vi->xattr_isize) +
 30			lcn * sizeof(struct z_erofs_lcluster_index);
 31	struct z_erofs_lcluster_index *di;
 32	unsigned int advise;
 33
 34	di = erofs_read_metabuf(&m->map->buf, inode->i_sb, pos, m->in_mbox);
 35	if (IS_ERR(di))
 36		return PTR_ERR(di);
 37	m->lcn = lcn;
 38	m->nextpackoff = pos + sizeof(struct z_erofs_lcluster_index);
 39
 40	advise = le16_to_cpu(di->di_advise);
 41	m->type = advise & Z_EROFS_LI_LCLUSTER_TYPE_MASK;
 42	if (m->type == Z_EROFS_LCLUSTER_TYPE_NONHEAD) {
 43		m->clusterofs = 1 << vi->z_lclusterbits;
 44		m->delta[0] = le16_to_cpu(di->di_u.delta[0]);
 45		if (m->delta[0] & Z_EROFS_LI_D0_CBLKCNT) {
 46			if (!(vi->z_advise & (Z_EROFS_ADVISE_BIG_PCLUSTER_1 |
 47					Z_EROFS_ADVISE_BIG_PCLUSTER_2))) {
 48				DBG_BUGON(1);
 49				return -EFSCORRUPTED;
 50			}
 51			m->compressedblks = m->delta[0] & ~Z_EROFS_LI_D0_CBLKCNT;
 52			m->delta[0] = 1;
 53		}
 54		m->delta[1] = le16_to_cpu(di->di_u.delta[1]);
 55	} else {
 56		m->partialref = !!(advise & Z_EROFS_LI_PARTIAL_REF);
 57		m->clusterofs = le16_to_cpu(di->di_clusterofs);
 58		m->pblk = le32_to_cpu(di->di_u.blkaddr);
 59	}
 60	return 0;
 61}
 62
 63static unsigned int decode_compactedbits(unsigned int lobits,
 64					 u8 *in, unsigned int pos, u8 *type)
 65{
 66	const unsigned int v = get_unaligned_le32(in + pos / 8) >> (pos & 7);
 67	const unsigned int lo = v & ((1 << lobits) - 1);
 68
 69	*type = (v >> lobits) & 3;
 70	return lo;
 71}
 72
 73static int get_compacted_la_distance(unsigned int lobits,
 74				     unsigned int encodebits,
 75				     unsigned int vcnt, u8 *in, int i)
 76{
 77	unsigned int lo, d1 = 0;
 78	u8 type;
 79
 80	DBG_BUGON(i >= vcnt);
 81
 82	do {
 83		lo = decode_compactedbits(lobits, in, encodebits * i, &type);
 84
 85		if (type != Z_EROFS_LCLUSTER_TYPE_NONHEAD)
 86			return d1;
 87		++d1;
 88	} while (++i < vcnt);
 89
 90	/* vcnt - 1 (Z_EROFS_LCLUSTER_TYPE_NONHEAD) item */
 91	if (!(lo & Z_EROFS_LI_D0_CBLKCNT))
 92		d1 += lo - 1;
 93	return d1;
 94}
 95
 96static int z_erofs_load_compact_lcluster(struct z_erofs_maprecorder *m,
 97					 unsigned long lcn, bool lookahead)
 98{
 99	struct inode *const inode = m->inode;
100	struct erofs_inode *const vi = EROFS_I(inode);
101	const erofs_off_t ebase = Z_EROFS_MAP_HEADER_END(erofs_iloc(inode) +
102			vi->inode_isize + vi->xattr_isize);
103	const unsigned int lclusterbits = vi->z_lclusterbits;
104	const unsigned int totalidx = erofs_iblks(inode);
105	unsigned int compacted_4b_initial, compacted_2b, amortizedshift;
106	unsigned int vcnt, lo, lobits, encodebits, nblk, bytes;
107	bool big_pcluster = vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1;
108	erofs_off_t pos;
109	u8 *in, type;
110	int i;
111
112	if (lcn >= totalidx || lclusterbits > 14)
113		return -EINVAL;
114
115	m->lcn = lcn;
116	/* used to align to 32-byte (compacted_2b) alignment */
117	compacted_4b_initial = ((32 - ebase % 32) / 4) & 7;
118	compacted_2b = 0;
119	if ((vi->z_advise & Z_EROFS_ADVISE_COMPACTED_2B) &&
120	    compacted_4b_initial < totalidx)
121		compacted_2b = rounddown(totalidx - compacted_4b_initial, 16);
122
123	pos = ebase;
124	amortizedshift = 2;	/* compact_4b */
125	if (lcn >= compacted_4b_initial) {
126		pos += compacted_4b_initial * 4;
127		lcn -= compacted_4b_initial;
128		if (lcn < compacted_2b) {
129			amortizedshift = 1;
130		} else {
131			pos += compacted_2b * 2;
132			lcn -= compacted_2b;
133		}
134	}
135	pos += lcn * (1 << amortizedshift);
136
137	/* figure out the lcluster count in this pack */
138	if (1 << amortizedshift == 4 && lclusterbits <= 14)
139		vcnt = 2;
140	else if (1 << amortizedshift == 2 && lclusterbits <= 12)
141		vcnt = 16;
142	else
143		return -EOPNOTSUPP;
144
145	in = erofs_read_metabuf(&m->map->buf, inode->i_sb, pos, m->in_mbox);
146	if (IS_ERR(in))
147		return PTR_ERR(in);
148
149	/* it doesn't equal to round_up(..) */
150	m->nextpackoff = round_down(pos, vcnt << amortizedshift) +
151			 (vcnt << amortizedshift);
152	lobits = max(lclusterbits, ilog2(Z_EROFS_LI_D0_CBLKCNT) + 1U);
153	encodebits = ((vcnt << amortizedshift) - sizeof(__le32)) * 8 / vcnt;
154	bytes = pos & ((vcnt << amortizedshift) - 1);
155	in -= bytes;
156	i = bytes >> amortizedshift;
157
158	lo = decode_compactedbits(lobits, in, encodebits * i, &type);
159	m->type = type;
160	if (type == Z_EROFS_LCLUSTER_TYPE_NONHEAD) {
161		m->clusterofs = 1 << lclusterbits;
162
163		/* figure out lookahead_distance: delta[1] if needed */
164		if (lookahead)
165			m->delta[1] = get_compacted_la_distance(lobits,
166						encodebits, vcnt, in, i);
167		if (lo & Z_EROFS_LI_D0_CBLKCNT) {
168			if (!big_pcluster) {
169				DBG_BUGON(1);
170				return -EFSCORRUPTED;
171			}
172			m->compressedblks = lo & ~Z_EROFS_LI_D0_CBLKCNT;
173			m->delta[0] = 1;
174			return 0;
175		} else if (i + 1 != (int)vcnt) {
176			m->delta[0] = lo;
177			return 0;
178		}
179		/*
180		 * since the last lcluster in the pack is special,
181		 * of which lo saves delta[1] rather than delta[0].
182		 * Hence, get delta[0] by the previous lcluster indirectly.
183		 */
184		lo = decode_compactedbits(lobits, in,
185					  encodebits * (i - 1), &type);
186		if (type != Z_EROFS_LCLUSTER_TYPE_NONHEAD)
187			lo = 0;
188		else if (lo & Z_EROFS_LI_D0_CBLKCNT)
189			lo = 1;
190		m->delta[0] = lo + 1;
191		return 0;
192	}
193	m->clusterofs = lo;
194	m->delta[0] = 0;
195	/* figout out blkaddr (pblk) for HEAD lclusters */
196	if (!big_pcluster) {
197		nblk = 1;
198		while (i > 0) {
199			--i;
200			lo = decode_compactedbits(lobits, in,
201						  encodebits * i, &type);
202			if (type == Z_EROFS_LCLUSTER_TYPE_NONHEAD)
203				i -= lo;
204
205			if (i >= 0)
206				++nblk;
207		}
208	} else {
209		nblk = 0;
210		while (i > 0) {
211			--i;
212			lo = decode_compactedbits(lobits, in,
213						  encodebits * i, &type);
214			if (type == Z_EROFS_LCLUSTER_TYPE_NONHEAD) {
215				if (lo & Z_EROFS_LI_D0_CBLKCNT) {
216					--i;
217					nblk += lo & ~Z_EROFS_LI_D0_CBLKCNT;
218					continue;
219				}
220				/* bigpcluster shouldn't have plain d0 == 1 */
221				if (lo <= 1) {
222					DBG_BUGON(1);
223					return -EFSCORRUPTED;
224				}
225				i -= lo - 2;
226				continue;
227			}
228			++nblk;
229		}
230	}
231	in += (vcnt << amortizedshift) - sizeof(__le32);
232	m->pblk = le32_to_cpu(*(__le32 *)in) + nblk;
233	return 0;
234}
235
236static int z_erofs_load_lcluster_from_disk(struct z_erofs_maprecorder *m,
237					   unsigned int lcn, bool lookahead)
238{
239	struct erofs_inode *vi = EROFS_I(m->inode);
240	int err;
241
242	if (vi->datalayout == EROFS_INODE_COMPRESSED_COMPACT) {
243		err = z_erofs_load_compact_lcluster(m, lcn, lookahead);
244	} else {
245		DBG_BUGON(vi->datalayout != EROFS_INODE_COMPRESSED_FULL);
246		err = z_erofs_load_full_lcluster(m, lcn);
247	}
248	if (err)
249		return err;
250
251	if (m->type >= Z_EROFS_LCLUSTER_TYPE_MAX) {
252		erofs_err(m->inode->i_sb, "unknown type %u @ lcn %u of nid %llu",
253			  m->type, lcn, EROFS_I(m->inode)->nid);
254		DBG_BUGON(1);
255		return -EOPNOTSUPP;
256	} else if (m->type != Z_EROFS_LCLUSTER_TYPE_NONHEAD &&
257		   m->clusterofs >= (1 << vi->z_lclusterbits)) {
258		DBG_BUGON(1);
259		return -EFSCORRUPTED;
260	}
261	return 0;
262}
263
264static int z_erofs_extent_lookback(struct z_erofs_maprecorder *m,
265				   unsigned int lookback_distance)
266{
267	struct super_block *sb = m->inode->i_sb;
268	struct erofs_inode *const vi = EROFS_I(m->inode);
269	const unsigned int lclusterbits = vi->z_lclusterbits;
270
271	while (m->lcn >= lookback_distance) {
272		unsigned long lcn = m->lcn - lookback_distance;
273		int err;
274
275		if (!lookback_distance)
276			break;
277
278		err = z_erofs_load_lcluster_from_disk(m, lcn, false);
279		if (err)
280			return err;
281		if (m->type == Z_EROFS_LCLUSTER_TYPE_NONHEAD) {
282			lookback_distance = m->delta[0];
283			continue;
284		}
285		m->headtype = m->type;
286		m->map->m_la = (lcn << lclusterbits) | m->clusterofs;
287		return 0;
288	}
289	erofs_err(sb, "bogus lookback distance %u @ lcn %lu of nid %llu",
290		  lookback_distance, m->lcn, vi->nid);
291	DBG_BUGON(1);
292	return -EFSCORRUPTED;
293}
294
295static int z_erofs_get_extent_compressedlen(struct z_erofs_maprecorder *m,
296					    unsigned int initial_lcn)
297{
298	struct inode *inode = m->inode;
299	struct super_block *sb = inode->i_sb;
300	struct erofs_inode *vi = EROFS_I(inode);
301	bool bigpcl1 = vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1;
302	bool bigpcl2 = vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_2;
303	unsigned long lcn = m->lcn + 1;
304	int err;
305
306	DBG_BUGON(m->type == Z_EROFS_LCLUSTER_TYPE_NONHEAD);
307	DBG_BUGON(m->type != m->headtype);
308
309	if ((m->headtype == Z_EROFS_LCLUSTER_TYPE_HEAD1 && !bigpcl1) ||
310	    ((m->headtype == Z_EROFS_LCLUSTER_TYPE_PLAIN ||
311	      m->headtype == Z_EROFS_LCLUSTER_TYPE_HEAD2) && !bigpcl2) ||
312	    (lcn << vi->z_lclusterbits) >= inode->i_size)
313		m->compressedblks = 1;
314
315	if (m->compressedblks)
316		goto out;
317
318	err = z_erofs_load_lcluster_from_disk(m, lcn, false);
319	if (err)
320		return err;
321
322	/*
323	 * If the 1st NONHEAD lcluster has already been handled initially w/o
324	 * valid compressedblks, which means at least it mustn't be CBLKCNT, or
325	 * an internal implemenatation error is detected.
326	 *
327	 * The following code can also handle it properly anyway, but let's
328	 * BUG_ON in the debugging mode only for developers to notice that.
329	 */
330	DBG_BUGON(lcn == initial_lcn &&
331		  m->type == Z_EROFS_LCLUSTER_TYPE_NONHEAD);
332
333	if (m->type == Z_EROFS_LCLUSTER_TYPE_NONHEAD && m->delta[0] != 1) {
334		erofs_err(sb, "bogus CBLKCNT @ lcn %lu of nid %llu", lcn, vi->nid);
335		DBG_BUGON(1);
336		return -EFSCORRUPTED;
337	}
338
339	/*
340	 * if the 1st NONHEAD lcluster is actually PLAIN or HEAD type rather
341	 * than CBLKCNT, it's a 1 block-sized pcluster.
342	 */
343	if (m->type != Z_EROFS_LCLUSTER_TYPE_NONHEAD || !m->compressedblks)
344		m->compressedblks = 1;
345out:
346	m->map->m_plen = erofs_pos(sb, m->compressedblks);
347	return 0;
348}
349
350static int z_erofs_get_extent_decompressedlen(struct z_erofs_maprecorder *m)
351{
352	struct inode *inode = m->inode;
353	struct erofs_inode *vi = EROFS_I(inode);
354	struct erofs_map_blocks *map = m->map;
355	unsigned int lclusterbits = vi->z_lclusterbits;
356	u64 lcn = m->lcn, headlcn = map->m_la >> lclusterbits;
357	int err;
358
359	while (1) {
360		/* handle the last EOF pcluster (no next HEAD lcluster) */
361		if ((lcn << lclusterbits) >= inode->i_size) {
362			map->m_llen = inode->i_size - map->m_la;
363			return 0;
364		}
365
366		err = z_erofs_load_lcluster_from_disk(m, lcn, true);
367		if (err)
368			return err;
369
370		if (m->type == Z_EROFS_LCLUSTER_TYPE_NONHEAD) {
371			/* work around invalid d1 generated by pre-1.0 mkfs */
372			if (unlikely(!m->delta[1])) {
373				m->delta[1] = 1;
374				DBG_BUGON(1);
375			}
376		} else if (m->type < Z_EROFS_LCLUSTER_TYPE_MAX) {
377			if (lcn != headlcn)
378				break;	/* ends at the next HEAD lcluster */
379			m->delta[1] = 1;
380		}
381		lcn += m->delta[1];
382	}
383	map->m_llen = (lcn << lclusterbits) + m->clusterofs - map->m_la;
384	return 0;
385}
386
387static int z_erofs_map_blocks_fo(struct inode *inode,
388				 struct erofs_map_blocks *map, int flags)
389{
390	struct erofs_inode *vi = EROFS_I(inode);
391	struct super_block *sb = inode->i_sb;
392	bool fragment = vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER;
393	bool ztailpacking = vi->z_idata_size;
394	unsigned int lclusterbits = vi->z_lclusterbits;
395	struct z_erofs_maprecorder m = {
396		.inode = inode,
397		.map = map,
398		.in_mbox = erofs_inode_in_metabox(inode),
399	};
400	unsigned int endoff;
401	unsigned long initial_lcn;
402	unsigned long long ofs, end;
403	int err;
404
405	ofs = flags & EROFS_GET_BLOCKS_FINDTAIL ? inode->i_size - 1 : map->m_la;
406	if (fragment && !(flags & EROFS_GET_BLOCKS_FINDTAIL) &&
407	    !vi->z_tailextent_headlcn) {
408		map->m_la = 0;
409		map->m_llen = inode->i_size;
410		map->m_flags = EROFS_MAP_FRAGMENT;
411		return 0;
412	}
413	initial_lcn = ofs >> lclusterbits;
414	endoff = ofs & ((1 << lclusterbits) - 1);
415
416	err = z_erofs_load_lcluster_from_disk(&m, initial_lcn, false);
417	if (err)
418		goto unmap_out;
419
420	if ((flags & EROFS_GET_BLOCKS_FINDTAIL) && ztailpacking)
421		vi->z_fragmentoff = m.nextpackoff;
422	map->m_flags = EROFS_MAP_MAPPED | EROFS_MAP_ENCODED;
423	end = (m.lcn + 1ULL) << lclusterbits;
424
425	if (m.type != Z_EROFS_LCLUSTER_TYPE_NONHEAD && endoff >= m.clusterofs) {
426		m.headtype = m.type;
427		map->m_la = (m.lcn << lclusterbits) | m.clusterofs;
428		/*
429		 * For ztailpacking files, in order to inline data more
430		 * effectively, special EOF lclusters are now supported
431		 * which can have three parts at most.
432		 */
433		if (ztailpacking && end > inode->i_size)
434			end = inode->i_size;
435	} else {
436		if (m.type != Z_EROFS_LCLUSTER_TYPE_NONHEAD) {
437			end = (m.lcn << lclusterbits) | m.clusterofs;
438			map->m_flags |= EROFS_MAP_FULL_MAPPED;
439			m.delta[0] = 1;
440		}
441		/* get the corresponding first chunk */
442		err = z_erofs_extent_lookback(&m, m.delta[0]);
443		if (err)
444			goto unmap_out;
445	}
446	if (m.partialref)
447		map->m_flags |= EROFS_MAP_PARTIAL_REF;
448	map->m_llen = end - map->m_la;
449
450	if (flags & EROFS_GET_BLOCKS_FINDTAIL) {
451		vi->z_tailextent_headlcn = m.lcn;
452		/* for non-compact indexes, fragmentoff is 64 bits */
453		if (fragment && vi->datalayout == EROFS_INODE_COMPRESSED_FULL)
454			vi->z_fragmentoff |= (u64)m.pblk << 32;
455	}
456	if (ztailpacking && m.lcn == vi->z_tailextent_headlcn) {
457		map->m_flags |= EROFS_MAP_META;
458		map->m_pa = vi->z_fragmentoff;
459		map->m_plen = vi->z_idata_size;
460		if (erofs_blkoff(sb, map->m_pa) + map->m_plen > sb->s_blocksize) {
461			erofs_err(sb, "ztailpacking inline data across blocks @ nid %llu",
462				  vi->nid);
463			err = -EFSCORRUPTED;
464			goto unmap_out;
465		}
466	} else if (fragment && m.lcn == vi->z_tailextent_headlcn) {
467		map->m_flags = EROFS_MAP_FRAGMENT;
468	} else {
469		map->m_pa = erofs_pos(sb, m.pblk);
470		err = z_erofs_get_extent_compressedlen(&m, initial_lcn);
471		if (err)
472			goto unmap_out;
473	}
474
475	if (m.headtype == Z_EROFS_LCLUSTER_TYPE_PLAIN) {
476		if (map->m_llen > map->m_plen) {
477			DBG_BUGON(1);
478			err = -EFSCORRUPTED;
479			goto unmap_out;
480		}
481		if (vi->z_advise & Z_EROFS_ADVISE_INTERLACED_PCLUSTER)
482			map->m_algorithmformat = Z_EROFS_COMPRESSION_INTERLACED;
483		else
484			map->m_algorithmformat = Z_EROFS_COMPRESSION_SHIFTED;
485	} else if (m.headtype == Z_EROFS_LCLUSTER_TYPE_HEAD2) {
486		map->m_algorithmformat = vi->z_algorithmtype[1];
487	} else {
488		map->m_algorithmformat = vi->z_algorithmtype[0];
489	}
490
491	if ((flags & EROFS_GET_BLOCKS_FIEMAP) ||
492	    ((flags & EROFS_GET_BLOCKS_READMORE) &&
493	     (map->m_algorithmformat == Z_EROFS_COMPRESSION_LZMA ||
494	      map->m_algorithmformat == Z_EROFS_COMPRESSION_DEFLATE ||
495	      map->m_algorithmformat == Z_EROFS_COMPRESSION_ZSTD) &&
496	      map->m_llen >= i_blocksize(inode))) {
497		err = z_erofs_get_extent_decompressedlen(&m);
498		if (!err)
499			map->m_flags |= EROFS_MAP_FULL_MAPPED;
500	}
501
502unmap_out:
503	erofs_unmap_metabuf(&m.map->buf);
504	return err;
505}
506
507static int z_erofs_map_blocks_ext(struct inode *inode,
508				  struct erofs_map_blocks *map, int flags)
509{
510	struct erofs_inode *vi = EROFS_I(inode);
511	struct super_block *sb = inode->i_sb;
512	bool interlaced = vi->z_advise & Z_EROFS_ADVISE_INTERLACED_PCLUSTER;
513	unsigned int recsz = z_erofs_extent_recsize(vi->z_advise);
514	erofs_off_t pos = round_up(Z_EROFS_MAP_HEADER_END(erofs_iloc(inode) +
515				   vi->inode_isize + vi->xattr_isize), recsz);
516	unsigned int bmask = sb->s_blocksize - 1;
517	bool in_mbox = erofs_inode_in_metabox(inode);
518	erofs_off_t lend = inode->i_size;
519	erofs_off_t l, r, mid, pa, la, lstart;
520	struct z_erofs_extent *ext;
521	unsigned int fmt;
522	bool last;
523
524	map->m_flags = 0;
525	if (recsz <= offsetof(struct z_erofs_extent, pstart_hi)) {
526		if (recsz <= offsetof(struct z_erofs_extent, pstart_lo)) {
527			ext = erofs_read_metabuf(&map->buf, sb, pos, in_mbox);
528			if (IS_ERR(ext))
529				return PTR_ERR(ext);
530			pa = le64_to_cpu(*(__le64 *)ext);
531			pos += sizeof(__le64);
532			lstart = 0;
533		} else {
534			lstart = round_down(map->m_la, 1 << vi->z_lclusterbits);
535			pos += (lstart >> vi->z_lclusterbits) * recsz;
536			pa = EROFS_NULL_ADDR;
537		}
538
539		for (; lstart <= map->m_la; lstart += 1 << vi->z_lclusterbits) {
540			ext = erofs_read_metabuf(&map->buf, sb, pos, in_mbox);
541			if (IS_ERR(ext))
542				return PTR_ERR(ext);
543			map->m_plen = le32_to_cpu(ext->plen);
544			if (pa != EROFS_NULL_ADDR) {
545				map->m_pa = pa;
546				pa += map->m_plen & Z_EROFS_EXTENT_PLEN_MASK;
547			} else {
548				map->m_pa = le32_to_cpu(ext->pstart_lo);
549			}
550			pos += recsz;
551		}
552		last = (lstart >= round_up(lend, 1 << vi->z_lclusterbits));
553		lend = min(lstart, lend);
554		lstart -= 1 << vi->z_lclusterbits;
555	} else {
556		lstart = lend;
557		for (l = 0, r = vi->z_extents; l < r; ) {
558			mid = l + (r - l) / 2;
559			ext = erofs_read_metabuf(&map->buf, sb,
560						 pos + mid * recsz, in_mbox);
561			if (IS_ERR(ext))
562				return PTR_ERR(ext);
563
564			la = le32_to_cpu(ext->lstart_lo);
565			pa = le32_to_cpu(ext->pstart_lo) |
566				(u64)le32_to_cpu(ext->pstart_hi) << 32;
567			if (recsz > offsetof(struct z_erofs_extent, lstart_hi))
568				la |= (u64)le32_to_cpu(ext->lstart_hi) << 32;
569
570			if (la > map->m_la) {
571				r = mid;
572				if (la > lend) {
573					DBG_BUGON(1);
574					return -EFSCORRUPTED;
575				}
576				lend = la;
577			} else {
578				l = mid + 1;
579				if (map->m_la == la)
580					r = min(l + 1, r);
581				lstart = la;
582				map->m_plen = le32_to_cpu(ext->plen);
583				map->m_pa = pa;
584			}
585		}
586		last = (l >= vi->z_extents);
587	}
588
589	if (lstart < lend) {
590		map->m_la = lstart;
591		if (last && (vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER)) {
592			map->m_flags = EROFS_MAP_FRAGMENT;
593			vi->z_fragmentoff = map->m_plen;
594			if (recsz > offsetof(struct z_erofs_extent, pstart_lo))
595				vi->z_fragmentoff |= map->m_pa << 32;
596		} else if (map->m_plen & Z_EROFS_EXTENT_PLEN_MASK) {
597			map->m_flags |= EROFS_MAP_MAPPED |
598				EROFS_MAP_FULL_MAPPED | EROFS_MAP_ENCODED;
599			fmt = map->m_plen >> Z_EROFS_EXTENT_PLEN_FMT_BIT;
600			if (map->m_plen & Z_EROFS_EXTENT_PLEN_PARTIAL)
601				map->m_flags |= EROFS_MAP_PARTIAL_REF;
602			map->m_plen &= Z_EROFS_EXTENT_PLEN_MASK;
603			if (fmt)
604				map->m_algorithmformat = fmt - 1;
605			else if (interlaced && !((map->m_pa | map->m_plen) & bmask))
606				map->m_algorithmformat =
607					Z_EROFS_COMPRESSION_INTERLACED;
608			else
609				map->m_algorithmformat =
610					Z_EROFS_COMPRESSION_SHIFTED;
611		}
612	}
613	map->m_llen = lend - map->m_la;
614	return 0;
615}
616
617static int z_erofs_fill_inode(struct inode *inode, struct erofs_map_blocks *map)
618{
619	struct erofs_inode *const vi = EROFS_I(inode);
620	struct super_block *const sb = inode->i_sb;
621	struct z_erofs_map_header *h;
622	erofs_off_t pos;
623	int err = 0;
624
625	if (test_bit(EROFS_I_Z_INITED_BIT, &vi->flags)) {
626		/*
627		 * paired with smp_mb() at the end of the function to ensure
628		 * fields will only be observed after the bit is set.
629		 */
630		smp_mb();
631		return 0;
632	}
633
634	if (wait_on_bit_lock(&vi->flags, EROFS_I_BL_Z_BIT, TASK_KILLABLE))
635		return -ERESTARTSYS;
636
637	if (test_bit(EROFS_I_Z_INITED_BIT, &vi->flags))
638		goto out_unlock;
639
640	pos = ALIGN(erofs_iloc(inode) + vi->inode_isize + vi->xattr_isize, 8);
641	h = erofs_read_metabuf(&map->buf, sb, pos, erofs_inode_in_metabox(inode));
642	if (IS_ERR(h)) {
643		err = PTR_ERR(h);
644		goto out_unlock;
645	}
646
647	/*
648	 * if the highest bit of the 8-byte map header is set, the whole file
649	 * is stored in the packed inode. The rest bits keeps z_fragmentoff.
650	 */
651	if (h->h_clusterbits >> Z_EROFS_FRAGMENT_INODE_BIT) {
652		vi->z_advise = Z_EROFS_ADVISE_FRAGMENT_PCLUSTER;
653		vi->z_fragmentoff = le64_to_cpu(*(__le64 *)h) ^ (1ULL << 63);
654		vi->z_tailextent_headlcn = 0;
655		goto done;
656	}
657	vi->z_advise = le16_to_cpu(h->h_advise);
658	vi->z_lclusterbits = sb->s_blocksize_bits + (h->h_clusterbits & 15);
659	if (vi->datalayout == EROFS_INODE_COMPRESSED_FULL &&
660	    (vi->z_advise & Z_EROFS_ADVISE_EXTENTS)) {
661		vi->z_extents = le32_to_cpu(h->h_extents_lo) |
662			((u64)le16_to_cpu(h->h_extents_hi) << 32);
663		goto done;
664	}
665
666	vi->z_algorithmtype[0] = h->h_algorithmtype & 15;
667	vi->z_algorithmtype[1] = h->h_algorithmtype >> 4;
668	if (vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER)
669		vi->z_fragmentoff = le32_to_cpu(h->h_fragmentoff);
670	else if (vi->z_advise & Z_EROFS_ADVISE_INLINE_PCLUSTER)
671		vi->z_idata_size = le16_to_cpu(h->h_idata_size);
672
673	if (!erofs_sb_has_big_pcluster(EROFS_SB(sb)) &&
674	    vi->z_advise & (Z_EROFS_ADVISE_BIG_PCLUSTER_1 |
675			    Z_EROFS_ADVISE_BIG_PCLUSTER_2)) {
676		erofs_err(sb, "per-inode big pcluster without sb feature for nid %llu",
677			  vi->nid);
678		err = -EFSCORRUPTED;
679		goto out_unlock;
680	}
681	if (vi->datalayout == EROFS_INODE_COMPRESSED_COMPACT &&
682	    !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1) ^
683	    !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_2)) {
684		erofs_err(sb, "big pcluster head1/2 of compact indexes should be consistent for nid %llu",
685			  vi->nid);
686		err = -EFSCORRUPTED;
687		goto out_unlock;
688	}
689
690	if (vi->z_idata_size ||
691	    (vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER)) {
692		struct erofs_map_blocks tm = {
693			.buf = __EROFS_BUF_INITIALIZER
694		};
695
696		err = z_erofs_map_blocks_fo(inode, &tm,
697					    EROFS_GET_BLOCKS_FINDTAIL);
698		erofs_put_metabuf(&tm.buf);
699		if (err < 0)
700			goto out_unlock;
701	}
702done:
703	/* paired with smp_mb() at the beginning of the function */
704	smp_mb();
705	set_bit(EROFS_I_Z_INITED_BIT, &vi->flags);
706out_unlock:
707	clear_and_wake_up_bit(EROFS_I_BL_Z_BIT, &vi->flags);
708	return err;
709}
710
711static int z_erofs_map_sanity_check(struct inode *inode,
712				    struct erofs_map_blocks *map)
713{
714	struct erofs_sb_info *sbi = EROFS_I_SB(inode);
715	u64 pend;
716
717	if (!(map->m_flags & EROFS_MAP_ENCODED))
718		return 0;
719	if (unlikely(map->m_algorithmformat >= Z_EROFS_COMPRESSION_RUNTIME_MAX)) {
720		erofs_err(inode->i_sb, "unknown algorithm %d @ pos %llu for nid %llu, please upgrade kernel",
721			  map->m_algorithmformat, map->m_la, EROFS_I(inode)->nid);
722		return -EOPNOTSUPP;
723	}
724	if (unlikely(map->m_algorithmformat < Z_EROFS_COMPRESSION_MAX &&
725		     !(sbi->available_compr_algs & (1 << map->m_algorithmformat)))) {
726		erofs_err(inode->i_sb, "inconsistent algorithmtype %u for nid %llu",
727			  map->m_algorithmformat, EROFS_I(inode)->nid);
728		return -EFSCORRUPTED;
729	}
730	if (unlikely(map->m_plen > Z_EROFS_PCLUSTER_MAX_SIZE ||
731		     map->m_llen > Z_EROFS_PCLUSTER_MAX_DSIZE))
732		return -EOPNOTSUPP;
733	/* Filesystems beyond 48-bit physical block addresses are invalid */
734	if (unlikely(check_add_overflow(map->m_pa, map->m_plen, &pend) ||
735		     (pend >> sbi->blkszbits) >= BIT_ULL(48)))
736		return -EFSCORRUPTED;
737	return 0;
738}
739
740int z_erofs_map_blocks_iter(struct inode *inode, struct erofs_map_blocks *map,
741			    int flags)
742{
743	struct erofs_inode *const vi = EROFS_I(inode);
744	int err = 0;
745
746	trace_erofs_map_blocks_enter(inode, map, flags);
747	if (map->m_la >= inode->i_size) {	/* post-EOF unmapped extent */
748		map->m_llen = map->m_la + 1 - inode->i_size;
749		map->m_la = inode->i_size;
750		map->m_flags = 0;
751	} else {
752		err = z_erofs_fill_inode(inode, map);
753		if (!err) {
754			if (vi->datalayout == EROFS_INODE_COMPRESSED_FULL &&
755			    (vi->z_advise & Z_EROFS_ADVISE_EXTENTS))
756				err = z_erofs_map_blocks_ext(inode, map, flags);
757			else
758				err = z_erofs_map_blocks_fo(inode, map, flags);
759		}
760		if (!err)
761			err = z_erofs_map_sanity_check(inode, map);
762		if (err)
763			map->m_llen = 0;
764	}
765	trace_erofs_map_blocks_exit(inode, map, flags, err);
766	return err;
767}
768
769static int z_erofs_iomap_begin_report(struct inode *inode, loff_t offset,
770				loff_t length, unsigned int flags,
771				struct iomap *iomap, struct iomap *srcmap)
772{
773	int ret;
774	struct erofs_map_blocks map = { .m_la = offset };
775
776	ret = z_erofs_map_blocks_iter(inode, &map, EROFS_GET_BLOCKS_FIEMAP);
777	erofs_put_metabuf(&map.buf);
778	if (ret < 0)
779		return ret;
780
781	iomap->bdev = inode->i_sb->s_bdev;
782	iomap->offset = map.m_la;
783	iomap->length = map.m_llen;
784	if (map.m_flags & EROFS_MAP_MAPPED) {
785		iomap->type = IOMAP_MAPPED;
786		iomap->addr = map.m_flags & __EROFS_MAP_FRAGMENT ?
787			      IOMAP_NULL_ADDR : map.m_pa;
788	} else {
789		iomap->type = IOMAP_HOLE;
790		iomap->addr = IOMAP_NULL_ADDR;
791		/*
792		 * No strict rule on how to describe extents for post EOF, yet
793		 * we need to do like below. Otherwise, iomap itself will get
794		 * into an endless loop on post EOF.
795		 *
796		 * Calculate the effective offset by subtracting extent start
797		 * (map.m_la) from the requested offset, and add it to length.
798		 * (NB: offset >= map.m_la always)
799		 */
800		if (iomap->offset >= inode->i_size)
801			iomap->length = length + offset - map.m_la;
802	}
803	iomap->flags = 0;
804	return 0;
805}
806
807const struct iomap_ops z_erofs_iomap_report_ops = {
808	.iomap_begin = z_erofs_iomap_begin_report,
809};