Merge tag 'fscache-next-20210829' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-2.6-microblaze.git] / fs / erofs / 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 <asm/unaligned.h>
8 #include <trace/events/erofs.h>
9
10 int z_erofs_fill_inode(struct inode *inode)
11 {
12         struct erofs_inode *const vi = EROFS_I(inode);
13         struct erofs_sb_info *sbi = EROFS_SB(inode->i_sb);
14
15         if (!erofs_sb_has_big_pcluster(sbi) &&
16             vi->datalayout == EROFS_INODE_FLAT_COMPRESSION_LEGACY) {
17                 vi->z_advise = 0;
18                 vi->z_algorithmtype[0] = 0;
19                 vi->z_algorithmtype[1] = 0;
20                 vi->z_logical_clusterbits = LOG_BLOCK_SIZE;
21                 set_bit(EROFS_I_Z_INITED_BIT, &vi->flags);
22         }
23         inode->i_mapping->a_ops = &z_erofs_aops;
24         return 0;
25 }
26
27 static int z_erofs_fill_inode_lazy(struct inode *inode)
28 {
29         struct erofs_inode *const vi = EROFS_I(inode);
30         struct super_block *const sb = inode->i_sb;
31         int err;
32         erofs_off_t pos;
33         struct page *page;
34         void *kaddr;
35         struct z_erofs_map_header *h;
36
37         if (test_bit(EROFS_I_Z_INITED_BIT, &vi->flags)) {
38                 /*
39                  * paired with smp_mb() at the end of the function to ensure
40                  * fields will only be observed after the bit is set.
41                  */
42                 smp_mb();
43                 return 0;
44         }
45
46         if (wait_on_bit_lock(&vi->flags, EROFS_I_BL_Z_BIT, TASK_KILLABLE))
47                 return -ERESTARTSYS;
48
49         err = 0;
50         if (test_bit(EROFS_I_Z_INITED_BIT, &vi->flags))
51                 goto out_unlock;
52
53         DBG_BUGON(!erofs_sb_has_big_pcluster(EROFS_SB(sb)) &&
54                   vi->datalayout == EROFS_INODE_FLAT_COMPRESSION_LEGACY);
55
56         pos = ALIGN(iloc(EROFS_SB(sb), vi->nid) + vi->inode_isize +
57                     vi->xattr_isize, 8);
58         page = erofs_get_meta_page(sb, erofs_blknr(pos));
59         if (IS_ERR(page)) {
60                 err = PTR_ERR(page);
61                 goto out_unlock;
62         }
63
64         kaddr = kmap_atomic(page);
65
66         h = kaddr + erofs_blkoff(pos);
67         vi->z_advise = le16_to_cpu(h->h_advise);
68         vi->z_algorithmtype[0] = h->h_algorithmtype & 15;
69         vi->z_algorithmtype[1] = h->h_algorithmtype >> 4;
70
71         if (vi->z_algorithmtype[0] >= Z_EROFS_COMPRESSION_MAX) {
72                 erofs_err(sb, "unknown compression format %u for nid %llu, please upgrade kernel",
73                           vi->z_algorithmtype[0], vi->nid);
74                 err = -EOPNOTSUPP;
75                 goto unmap_done;
76         }
77
78         vi->z_logical_clusterbits = LOG_BLOCK_SIZE + (h->h_clusterbits & 7);
79         if (!erofs_sb_has_big_pcluster(EROFS_SB(sb)) &&
80             vi->z_advise & (Z_EROFS_ADVISE_BIG_PCLUSTER_1 |
81                             Z_EROFS_ADVISE_BIG_PCLUSTER_2)) {
82                 erofs_err(sb, "per-inode big pcluster without sb feature for nid %llu",
83                           vi->nid);
84                 err = -EFSCORRUPTED;
85                 goto unmap_done;
86         }
87         if (vi->datalayout == EROFS_INODE_FLAT_COMPRESSION &&
88             !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1) ^
89             !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_2)) {
90                 erofs_err(sb, "big pcluster head1/2 of compact indexes should be consistent for nid %llu",
91                           vi->nid);
92                 err = -EFSCORRUPTED;
93                 goto unmap_done;
94         }
95         /* paired with smp_mb() at the beginning of the function */
96         smp_mb();
97         set_bit(EROFS_I_Z_INITED_BIT, &vi->flags);
98 unmap_done:
99         kunmap_atomic(kaddr);
100         unlock_page(page);
101         put_page(page);
102 out_unlock:
103         clear_and_wake_up_bit(EROFS_I_BL_Z_BIT, &vi->flags);
104         return err;
105 }
106
107 struct z_erofs_maprecorder {
108         struct inode *inode;
109         struct erofs_map_blocks *map;
110         void *kaddr;
111
112         unsigned long lcn;
113         /* compression extent information gathered */
114         u8  type;
115         u16 clusterofs;
116         u16 delta[2];
117         erofs_blk_t pblk, compressedlcs;
118 };
119
120 static int z_erofs_reload_indexes(struct z_erofs_maprecorder *m,
121                                   erofs_blk_t eblk)
122 {
123         struct super_block *const sb = m->inode->i_sb;
124         struct erofs_map_blocks *const map = m->map;
125         struct page *mpage = map->mpage;
126
127         if (mpage) {
128                 if (mpage->index == eblk) {
129                         if (!m->kaddr)
130                                 m->kaddr = kmap_atomic(mpage);
131                         return 0;
132                 }
133
134                 if (m->kaddr) {
135                         kunmap_atomic(m->kaddr);
136                         m->kaddr = NULL;
137                 }
138                 put_page(mpage);
139         }
140
141         mpage = erofs_get_meta_page(sb, eblk);
142         if (IS_ERR(mpage)) {
143                 map->mpage = NULL;
144                 return PTR_ERR(mpage);
145         }
146         m->kaddr = kmap_atomic(mpage);
147         unlock_page(mpage);
148         map->mpage = mpage;
149         return 0;
150 }
151
152 static int legacy_load_cluster_from_disk(struct z_erofs_maprecorder *m,
153                                          unsigned long lcn)
154 {
155         struct inode *const inode = m->inode;
156         struct erofs_inode *const vi = EROFS_I(inode);
157         const erofs_off_t ibase = iloc(EROFS_I_SB(inode), vi->nid);
158         const erofs_off_t pos =
159                 Z_EROFS_VLE_LEGACY_INDEX_ALIGN(ibase + vi->inode_isize +
160                                                vi->xattr_isize) +
161                 lcn * sizeof(struct z_erofs_vle_decompressed_index);
162         struct z_erofs_vle_decompressed_index *di;
163         unsigned int advise, type;
164         int err;
165
166         err = z_erofs_reload_indexes(m, erofs_blknr(pos));
167         if (err)
168                 return err;
169
170         m->lcn = lcn;
171         di = m->kaddr + erofs_blkoff(pos);
172
173         advise = le16_to_cpu(di->di_advise);
174         type = (advise >> Z_EROFS_VLE_DI_CLUSTER_TYPE_BIT) &
175                 ((1 << Z_EROFS_VLE_DI_CLUSTER_TYPE_BITS) - 1);
176         switch (type) {
177         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
178                 m->clusterofs = 1 << vi->z_logical_clusterbits;
179                 m->delta[0] = le16_to_cpu(di->di_u.delta[0]);
180                 if (m->delta[0] & Z_EROFS_VLE_DI_D0_CBLKCNT) {
181                         if (!(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1)) {
182                                 DBG_BUGON(1);
183                                 return -EFSCORRUPTED;
184                         }
185                         m->compressedlcs = m->delta[0] &
186                                 ~Z_EROFS_VLE_DI_D0_CBLKCNT;
187                         m->delta[0] = 1;
188                 }
189                 m->delta[1] = le16_to_cpu(di->di_u.delta[1]);
190                 break;
191         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
192         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
193                 m->clusterofs = le16_to_cpu(di->di_clusterofs);
194                 m->pblk = le32_to_cpu(di->di_u.blkaddr);
195                 break;
196         default:
197                 DBG_BUGON(1);
198                 return -EOPNOTSUPP;
199         }
200         m->type = type;
201         return 0;
202 }
203
204 static unsigned int decode_compactedbits(unsigned int lobits,
205                                          unsigned int lomask,
206                                          u8 *in, unsigned int pos, u8 *type)
207 {
208         const unsigned int v = get_unaligned_le32(in + pos / 8) >> (pos & 7);
209         const unsigned int lo = v & lomask;
210
211         *type = (v >> lobits) & 3;
212         return lo;
213 }
214
215 static int unpack_compacted_index(struct z_erofs_maprecorder *m,
216                                   unsigned int amortizedshift,
217                                   unsigned int eofs)
218 {
219         struct erofs_inode *const vi = EROFS_I(m->inode);
220         const unsigned int lclusterbits = vi->z_logical_clusterbits;
221         const unsigned int lomask = (1 << lclusterbits) - 1;
222         unsigned int vcnt, base, lo, encodebits, nblk;
223         int i;
224         u8 *in, type;
225         bool big_pcluster;
226
227         if (1 << amortizedshift == 4)
228                 vcnt = 2;
229         else if (1 << amortizedshift == 2 && lclusterbits == 12)
230                 vcnt = 16;
231         else
232                 return -EOPNOTSUPP;
233
234         big_pcluster = vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1;
235         encodebits = ((vcnt << amortizedshift) - sizeof(__le32)) * 8 / vcnt;
236         base = round_down(eofs, vcnt << amortizedshift);
237         in = m->kaddr + base;
238
239         i = (eofs - base) >> amortizedshift;
240
241         lo = decode_compactedbits(lclusterbits, lomask,
242                                   in, encodebits * i, &type);
243         m->type = type;
244         if (type == Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD) {
245                 m->clusterofs = 1 << lclusterbits;
246                 if (lo & Z_EROFS_VLE_DI_D0_CBLKCNT) {
247                         if (!big_pcluster) {
248                                 DBG_BUGON(1);
249                                 return -EFSCORRUPTED;
250                         }
251                         m->compressedlcs = lo & ~Z_EROFS_VLE_DI_D0_CBLKCNT;
252                         m->delta[0] = 1;
253                         return 0;
254                 } else if (i + 1 != (int)vcnt) {
255                         m->delta[0] = lo;
256                         return 0;
257                 }
258                 /*
259                  * since the last lcluster in the pack is special,
260                  * of which lo saves delta[1] rather than delta[0].
261                  * Hence, get delta[0] by the previous lcluster indirectly.
262                  */
263                 lo = decode_compactedbits(lclusterbits, lomask,
264                                           in, encodebits * (i - 1), &type);
265                 if (type != Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD)
266                         lo = 0;
267                 else if (lo & Z_EROFS_VLE_DI_D0_CBLKCNT)
268                         lo = 1;
269                 m->delta[0] = lo + 1;
270                 return 0;
271         }
272         m->clusterofs = lo;
273         m->delta[0] = 0;
274         /* figout out blkaddr (pblk) for HEAD lclusters */
275         if (!big_pcluster) {
276                 nblk = 1;
277                 while (i > 0) {
278                         --i;
279                         lo = decode_compactedbits(lclusterbits, lomask,
280                                                   in, encodebits * i, &type);
281                         if (type == Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD)
282                                 i -= lo;
283
284                         if (i >= 0)
285                                 ++nblk;
286                 }
287         } else {
288                 nblk = 0;
289                 while (i > 0) {
290                         --i;
291                         lo = decode_compactedbits(lclusterbits, lomask,
292                                                   in, encodebits * i, &type);
293                         if (type == Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD) {
294                                 if (lo & Z_EROFS_VLE_DI_D0_CBLKCNT) {
295                                         --i;
296                                         nblk += lo & ~Z_EROFS_VLE_DI_D0_CBLKCNT;
297                                         continue;
298                                 }
299                                 /* bigpcluster shouldn't have plain d0 == 1 */
300                                 if (lo <= 1) {
301                                         DBG_BUGON(1);
302                                         return -EFSCORRUPTED;
303                                 }
304                                 i -= lo - 2;
305                                 continue;
306                         }
307                         ++nblk;
308                 }
309         }
310         in += (vcnt << amortizedshift) - sizeof(__le32);
311         m->pblk = le32_to_cpu(*(__le32 *)in) + nblk;
312         return 0;
313 }
314
315 static int compacted_load_cluster_from_disk(struct z_erofs_maprecorder *m,
316                                             unsigned long lcn)
317 {
318         struct inode *const inode = m->inode;
319         struct erofs_inode *const vi = EROFS_I(inode);
320         const unsigned int lclusterbits = vi->z_logical_clusterbits;
321         const erofs_off_t ebase = ALIGN(iloc(EROFS_I_SB(inode), vi->nid) +
322                                         vi->inode_isize + vi->xattr_isize, 8) +
323                 sizeof(struct z_erofs_map_header);
324         const unsigned int totalidx = DIV_ROUND_UP(inode->i_size, EROFS_BLKSIZ);
325         unsigned int compacted_4b_initial, compacted_2b;
326         unsigned int amortizedshift;
327         erofs_off_t pos;
328         int err;
329
330         if (lclusterbits != 12)
331                 return -EOPNOTSUPP;
332
333         if (lcn >= totalidx)
334                 return -EINVAL;
335
336         m->lcn = lcn;
337         /* used to align to 32-byte (compacted_2b) alignment */
338         compacted_4b_initial = (32 - ebase % 32) / 4;
339         if (compacted_4b_initial == 32 / 4)
340                 compacted_4b_initial = 0;
341
342         if (vi->z_advise & Z_EROFS_ADVISE_COMPACTED_2B)
343                 compacted_2b = rounddown(totalidx - compacted_4b_initial, 16);
344         else
345                 compacted_2b = 0;
346
347         pos = ebase;
348         if (lcn < compacted_4b_initial) {
349                 amortizedshift = 2;
350                 goto out;
351         }
352         pos += compacted_4b_initial * 4;
353         lcn -= compacted_4b_initial;
354
355         if (lcn < compacted_2b) {
356                 amortizedshift = 1;
357                 goto out;
358         }
359         pos += compacted_2b * 2;
360         lcn -= compacted_2b;
361         amortizedshift = 2;
362 out:
363         pos += lcn * (1 << amortizedshift);
364         err = z_erofs_reload_indexes(m, erofs_blknr(pos));
365         if (err)
366                 return err;
367         return unpack_compacted_index(m, amortizedshift, erofs_blkoff(pos));
368 }
369
370 static int z_erofs_load_cluster_from_disk(struct z_erofs_maprecorder *m,
371                                           unsigned int lcn)
372 {
373         const unsigned int datamode = EROFS_I(m->inode)->datalayout;
374
375         if (datamode == EROFS_INODE_FLAT_COMPRESSION_LEGACY)
376                 return legacy_load_cluster_from_disk(m, lcn);
377
378         if (datamode == EROFS_INODE_FLAT_COMPRESSION)
379                 return compacted_load_cluster_from_disk(m, lcn);
380
381         return -EINVAL;
382 }
383
384 static int z_erofs_extent_lookback(struct z_erofs_maprecorder *m,
385                                    unsigned int lookback_distance)
386 {
387         struct erofs_inode *const vi = EROFS_I(m->inode);
388         struct erofs_map_blocks *const map = m->map;
389         const unsigned int lclusterbits = vi->z_logical_clusterbits;
390         unsigned long lcn = m->lcn;
391         int err;
392
393         if (lcn < lookback_distance) {
394                 erofs_err(m->inode->i_sb,
395                           "bogus lookback distance @ nid %llu", vi->nid);
396                 DBG_BUGON(1);
397                 return -EFSCORRUPTED;
398         }
399
400         /* load extent head logical cluster if needed */
401         lcn -= lookback_distance;
402         err = z_erofs_load_cluster_from_disk(m, lcn);
403         if (err)
404                 return err;
405
406         switch (m->type) {
407         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
408                 if (!m->delta[0]) {
409                         erofs_err(m->inode->i_sb,
410                                   "invalid lookback distance 0 @ nid %llu",
411                                   vi->nid);
412                         DBG_BUGON(1);
413                         return -EFSCORRUPTED;
414                 }
415                 return z_erofs_extent_lookback(m, m->delta[0]);
416         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
417                 map->m_flags &= ~EROFS_MAP_ZIPPED;
418                 fallthrough;
419         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
420                 map->m_la = (lcn << lclusterbits) | m->clusterofs;
421                 break;
422         default:
423                 erofs_err(m->inode->i_sb,
424                           "unknown type %u @ lcn %lu of nid %llu",
425                           m->type, lcn, vi->nid);
426                 DBG_BUGON(1);
427                 return -EOPNOTSUPP;
428         }
429         return 0;
430 }
431
432 static int z_erofs_get_extent_compressedlen(struct z_erofs_maprecorder *m,
433                                             unsigned int initial_lcn)
434 {
435         struct erofs_inode *const vi = EROFS_I(m->inode);
436         struct erofs_map_blocks *const map = m->map;
437         const unsigned int lclusterbits = vi->z_logical_clusterbits;
438         unsigned long lcn;
439         int err;
440
441         DBG_BUGON(m->type != Z_EROFS_VLE_CLUSTER_TYPE_PLAIN &&
442                   m->type != Z_EROFS_VLE_CLUSTER_TYPE_HEAD);
443         if (!(map->m_flags & EROFS_MAP_ZIPPED) ||
444             !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1)) {
445                 map->m_plen = 1 << lclusterbits;
446                 return 0;
447         }
448
449         lcn = m->lcn + 1;
450         if (m->compressedlcs)
451                 goto out;
452
453         err = z_erofs_load_cluster_from_disk(m, lcn);
454         if (err)
455                 return err;
456
457         /*
458          * If the 1st NONHEAD lcluster has already been handled initially w/o
459          * valid compressedlcs, which means at least it mustn't be CBLKCNT, or
460          * an internal implemenatation error is detected.
461          *
462          * The following code can also handle it properly anyway, but let's
463          * BUG_ON in the debugging mode only for developers to notice that.
464          */
465         DBG_BUGON(lcn == initial_lcn &&
466                   m->type == Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD);
467
468         switch (m->type) {
469         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
470         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
471                 /*
472                  * if the 1st NONHEAD lcluster is actually PLAIN or HEAD type
473                  * rather than CBLKCNT, it's a 1 lcluster-sized pcluster.
474                  */
475                 m->compressedlcs = 1;
476                 break;
477         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
478                 if (m->delta[0] != 1)
479                         goto err_bonus_cblkcnt;
480                 if (m->compressedlcs)
481                         break;
482                 fallthrough;
483         default:
484                 erofs_err(m->inode->i_sb,
485                           "cannot found CBLKCNT @ lcn %lu of nid %llu",
486                           lcn, vi->nid);
487                 DBG_BUGON(1);
488                 return -EFSCORRUPTED;
489         }
490 out:
491         map->m_plen = m->compressedlcs << lclusterbits;
492         return 0;
493 err_bonus_cblkcnt:
494         erofs_err(m->inode->i_sb,
495                   "bogus CBLKCNT @ lcn %lu of nid %llu",
496                   lcn, vi->nid);
497         DBG_BUGON(1);
498         return -EFSCORRUPTED;
499 }
500
501 int z_erofs_map_blocks_iter(struct inode *inode,
502                             struct erofs_map_blocks *map,
503                             int flags)
504 {
505         struct erofs_inode *const vi = EROFS_I(inode);
506         struct z_erofs_maprecorder m = {
507                 .inode = inode,
508                 .map = map,
509         };
510         int err = 0;
511         unsigned int lclusterbits, endoff;
512         unsigned long initial_lcn;
513         unsigned long long ofs, end;
514
515         trace_z_erofs_map_blocks_iter_enter(inode, map, flags);
516
517         /* when trying to read beyond EOF, leave it unmapped */
518         if (map->m_la >= inode->i_size) {
519                 map->m_llen = map->m_la + 1 - inode->i_size;
520                 map->m_la = inode->i_size;
521                 map->m_flags = 0;
522                 goto out;
523         }
524
525         err = z_erofs_fill_inode_lazy(inode);
526         if (err)
527                 goto out;
528
529         lclusterbits = vi->z_logical_clusterbits;
530         ofs = map->m_la;
531         initial_lcn = ofs >> lclusterbits;
532         endoff = ofs & ((1 << lclusterbits) - 1);
533
534         err = z_erofs_load_cluster_from_disk(&m, initial_lcn);
535         if (err)
536                 goto unmap_out;
537
538         map->m_flags = EROFS_MAP_ZIPPED;        /* by default, compressed */
539         end = (m.lcn + 1ULL) << lclusterbits;
540
541         switch (m.type) {
542         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
543                 if (endoff >= m.clusterofs)
544                         map->m_flags &= ~EROFS_MAP_ZIPPED;
545                 fallthrough;
546         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
547                 if (endoff >= m.clusterofs) {
548                         map->m_la = (m.lcn << lclusterbits) | m.clusterofs;
549                         break;
550                 }
551                 /* m.lcn should be >= 1 if endoff < m.clusterofs */
552                 if (!m.lcn) {
553                         erofs_err(inode->i_sb,
554                                   "invalid logical cluster 0 at nid %llu",
555                                   vi->nid);
556                         err = -EFSCORRUPTED;
557                         goto unmap_out;
558                 }
559                 end = (m.lcn << lclusterbits) | m.clusterofs;
560                 map->m_flags |= EROFS_MAP_FULL_MAPPED;
561                 m.delta[0] = 1;
562                 fallthrough;
563         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
564                 /* get the corresponding first chunk */
565                 err = z_erofs_extent_lookback(&m, m.delta[0]);
566                 if (err)
567                         goto unmap_out;
568                 break;
569         default:
570                 erofs_err(inode->i_sb,
571                           "unknown type %u @ offset %llu of nid %llu",
572                           m.type, ofs, vi->nid);
573                 err = -EOPNOTSUPP;
574                 goto unmap_out;
575         }
576
577         map->m_llen = end - map->m_la;
578         map->m_pa = blknr_to_addr(m.pblk);
579         map->m_flags |= EROFS_MAP_MAPPED;
580
581         err = z_erofs_get_extent_compressedlen(&m, initial_lcn);
582         if (err)
583                 goto out;
584 unmap_out:
585         if (m.kaddr)
586                 kunmap_atomic(m.kaddr);
587
588 out:
589         erofs_dbg("%s, m_la %llu m_pa %llu m_llen %llu m_plen %llu m_flags 0%o",
590                   __func__, map->m_la, map->m_pa,
591                   map->m_llen, map->m_plen, map->m_flags);
592
593         trace_z_erofs_map_blocks_iter_exit(inode, map, flags, err);
594
595         /* aggressively BUG_ON iff CONFIG_EROFS_FS_DEBUG is on */
596         DBG_BUGON(err < 0 && err != -ENOMEM);
597         return err;
598 }