1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) 2019 HUAWEI, Inc.
4 * https://www.huawei.com/
5 * Copyright (C) 2024 Alibaba Cloud
6 */
7 #include "compress.h"
8 #include <linux/lz4.h>
9
10 #ifndef LZ4_DISTANCE_MAX /* history window size */
11 #define LZ4_DISTANCE_MAX 65535 /* set to maximum value by default */
12 #endif
13
14 #define LZ4_MAX_DISTANCE_PAGES (DIV_ROUND_UP(LZ4_DISTANCE_MAX, PAGE_SIZE) + 1)
15 #ifndef LZ4_DECOMPRESS_INPLACE_MARGIN
16 #define LZ4_DECOMPRESS_INPLACE_MARGIN(srcsize) (((srcsize) >> 8) + 32)
17 #endif
18
19 struct z_erofs_lz4_decompress_ctx {
20 struct z_erofs_decompress_req *rq;
21 /* # of encoded, decoded pages */
22 unsigned int inpages, outpages;
23 /* decoded block total length (used for in-place decompression) */
24 unsigned int oend;
25 };
26
z_erofs_load_lz4_config(struct super_block * sb,struct erofs_super_block * dsb,void * data,int size)27 static int z_erofs_load_lz4_config(struct super_block *sb,
28 struct erofs_super_block *dsb, void *data, int size)
29 {
30 struct erofs_sb_info *sbi = EROFS_SB(sb);
31 struct z_erofs_lz4_cfgs *lz4 = data;
32 u16 distance;
33
34 if (lz4) {
35 if (size < sizeof(struct z_erofs_lz4_cfgs)) {
36 erofs_err(sb, "invalid lz4 cfgs, size=%u", size);
37 return -EINVAL;
38 }
39 distance = le16_to_cpu(lz4->max_distance);
40
41 sbi->lz4.max_pclusterblks = le16_to_cpu(lz4->max_pclusterblks);
42 if (!sbi->lz4.max_pclusterblks) {
43 sbi->lz4.max_pclusterblks = 1; /* reserved case */
44 } else if (sbi->lz4.max_pclusterblks >
45 erofs_blknr(sb, Z_EROFS_PCLUSTER_MAX_SIZE)) {
46 erofs_err(sb, "too large lz4 pclusterblks %u",
47 sbi->lz4.max_pclusterblks);
48 return -EINVAL;
49 }
50 } else {
51 distance = le16_to_cpu(dsb->u1.lz4_max_distance);
52 sbi->lz4.max_pclusterblks = 1;
53 }
54
55 sbi->lz4.max_distance_pages = distance ?
56 DIV_ROUND_UP(distance, PAGE_SIZE) + 1 :
57 LZ4_MAX_DISTANCE_PAGES;
58 return z_erofs_gbuf_growsize(sbi->lz4.max_pclusterblks);
59 }
60
61 /*
62 * Fill all gaps with bounce pages if it's a sparse page list. Also check if
63 * all physical pages are consecutive, which can be seen for moderate CR.
64 */
z_erofs_lz4_prepare_dstpages(struct z_erofs_lz4_decompress_ctx * ctx,struct page ** pagepool)65 static int z_erofs_lz4_prepare_dstpages(struct z_erofs_lz4_decompress_ctx *ctx,
66 struct page **pagepool)
67 {
68 struct z_erofs_decompress_req *rq = ctx->rq;
69 struct page *availables[LZ4_MAX_DISTANCE_PAGES] = { NULL };
70 unsigned long bounced[DIV_ROUND_UP(LZ4_MAX_DISTANCE_PAGES,
71 BITS_PER_LONG)] = { 0 };
72 unsigned int lz4_max_distance_pages =
73 EROFS_SB(rq->sb)->lz4.max_distance_pages;
74 void *kaddr = NULL;
75 unsigned int i, j, top;
76
77 top = 0;
78 for (i = j = 0; i < ctx->outpages; ++i, ++j) {
79 struct page *const page = rq->out[i];
80 struct page *victim;
81
82 if (j >= lz4_max_distance_pages)
83 j = 0;
84
85 /* 'valid' bounced can only be tested after a complete round */
86 if (!rq->fillgaps && test_bit(j, bounced)) {
87 DBG_BUGON(i < lz4_max_distance_pages);
88 DBG_BUGON(top >= lz4_max_distance_pages);
89 availables[top++] = rq->out[i - lz4_max_distance_pages];
90 }
91
92 if (page) {
93 __clear_bit(j, bounced);
94 if (!PageHighMem(page)) {
95 if (!i) {
96 kaddr = page_address(page);
97 continue;
98 }
99 if (kaddr &&
100 kaddr + PAGE_SIZE == page_address(page)) {
101 kaddr += PAGE_SIZE;
102 continue;
103 }
104 }
105 kaddr = NULL;
106 continue;
107 }
108 kaddr = NULL;
109 __set_bit(j, bounced);
110
111 if (top) {
112 victim = availables[--top];
113 } else {
114 victim = __erofs_allocpage(pagepool, rq->gfp, true);
115 if (!victim)
116 return -ENOMEM;
117 set_page_private(victim, Z_EROFS_SHORTLIVED_PAGE);
118 }
119 rq->out[i] = victim;
120 }
121 return kaddr ? 1 : 0;
122 }
123
z_erofs_lz4_handle_overlap(struct z_erofs_lz4_decompress_ctx * ctx,void * inpage,void * out,unsigned int * inputmargin,int * maptype,bool may_inplace)124 static void *z_erofs_lz4_handle_overlap(struct z_erofs_lz4_decompress_ctx *ctx,
125 void *inpage, void *out, unsigned int *inputmargin,
126 int *maptype, bool may_inplace)
127 {
128 struct z_erofs_decompress_req *rq = ctx->rq;
129 unsigned int omargin, total, i;
130 struct page **in;
131 void *src, *tmp;
132
133 if (rq->inplace_io) {
134 omargin = PAGE_ALIGN(ctx->oend) - ctx->oend;
135 if (rq->partial_decoding || !may_inplace ||
136 omargin < LZ4_DECOMPRESS_INPLACE_MARGIN(rq->inputsize))
137 goto docopy;
138
139 for (i = 0; i < ctx->inpages; ++i)
140 if (rq->out[ctx->outpages - ctx->inpages + i] !=
141 rq->in[i])
142 goto docopy;
143 kunmap_local(inpage);
144 *maptype = 3;
145 return out + ((ctx->outpages - ctx->inpages) << PAGE_SHIFT);
146 }
147
148 if (ctx->inpages <= 1) {
149 *maptype = 0;
150 return inpage;
151 }
152 kunmap_local(inpage);
153 src = erofs_vm_map_ram(rq->in, ctx->inpages);
154 if (!src)
155 return ERR_PTR(-ENOMEM);
156 *maptype = 1;
157 return src;
158
159 docopy:
160 /* Or copy compressed data which can be overlapped to per-CPU buffer */
161 in = rq->in;
162 src = z_erofs_get_gbuf(ctx->inpages);
163 if (!src) {
164 DBG_BUGON(1);
165 kunmap_local(inpage);
166 return ERR_PTR(-EFAULT);
167 }
168
169 tmp = src;
170 total = rq->inputsize;
171 while (total) {
172 unsigned int page_copycnt =
173 min_t(unsigned int, total, PAGE_SIZE - *inputmargin);
174
175 if (!inpage)
176 inpage = kmap_local_page(*in);
177 memcpy(tmp, inpage + *inputmargin, page_copycnt);
178 kunmap_local(inpage);
179 inpage = NULL;
180 tmp += page_copycnt;
181 total -= page_copycnt;
182 ++in;
183 *inputmargin = 0;
184 }
185 *maptype = 2;
186 return src;
187 }
188
189 /*
190 * Get the exact inputsize with zero_padding feature.
191 * - For LZ4, it should work if zero_padding feature is on (5.3+);
192 * - For MicroLZMA, it'd be enabled all the time.
193 */
z_erofs_fixup_insize(struct z_erofs_decompress_req * rq,const char * padbuf,unsigned int padbufsize)194 int z_erofs_fixup_insize(struct z_erofs_decompress_req *rq, const char *padbuf,
195 unsigned int padbufsize)
196 {
197 const char *padend;
198
199 padend = memchr_inv(padbuf, 0, padbufsize);
200 if (!padend)
201 return -EFSCORRUPTED;
202 rq->inputsize -= padend - padbuf;
203 rq->pageofs_in += padend - padbuf;
204 return 0;
205 }
206
z_erofs_lz4_decompress_mem(struct z_erofs_lz4_decompress_ctx * ctx,u8 * dst)207 static int z_erofs_lz4_decompress_mem(struct z_erofs_lz4_decompress_ctx *ctx,
208 u8 *dst)
209 {
210 struct z_erofs_decompress_req *rq = ctx->rq;
211 bool support_0padding = false, may_inplace = false;
212 unsigned int inputmargin;
213 u8 *out, *headpage, *src;
214 int ret, maptype;
215
216 DBG_BUGON(*rq->in == NULL);
217 headpage = kmap_local_page(*rq->in);
218
219 /* LZ4 decompression inplace is only safe if zero_padding is enabled */
220 if (erofs_sb_has_zero_padding(EROFS_SB(rq->sb))) {
221 support_0padding = true;
222 ret = z_erofs_fixup_insize(rq, headpage + rq->pageofs_in,
223 min_t(unsigned int, rq->inputsize,
224 rq->sb->s_blocksize - rq->pageofs_in));
225 if (ret) {
226 kunmap_local(headpage);
227 return ret;
228 }
229 may_inplace = !((rq->pageofs_in + rq->inputsize) &
230 (rq->sb->s_blocksize - 1));
231 }
232
233 inputmargin = rq->pageofs_in;
234 src = z_erofs_lz4_handle_overlap(ctx, headpage, dst, &inputmargin,
235 &maptype, may_inplace);
236 if (IS_ERR(src))
237 return PTR_ERR(src);
238
239 out = dst + rq->pageofs_out;
240 /* legacy format could compress extra data in a pcluster. */
241 if (rq->partial_decoding || !support_0padding)
242 ret = LZ4_decompress_safe_partial(src + inputmargin, out,
243 rq->inputsize, rq->outputsize, rq->outputsize);
244 else
245 ret = LZ4_decompress_safe(src + inputmargin, out,
246 rq->inputsize, rq->outputsize);
247
248 if (ret != rq->outputsize) {
249 erofs_err(rq->sb, "failed to decompress %d in[%u, %u] out[%u]",
250 ret, rq->inputsize, inputmargin, rq->outputsize);
251 if (ret >= 0)
252 memset(out + ret, 0, rq->outputsize - ret);
253 ret = -EFSCORRUPTED;
254 } else {
255 ret = 0;
256 }
257
258 if (maptype == 0) {
259 kunmap_local(headpage);
260 } else if (maptype == 1) {
261 vm_unmap_ram(src, ctx->inpages);
262 } else if (maptype == 2) {
263 z_erofs_put_gbuf(src);
264 } else if (maptype != 3) {
265 DBG_BUGON(1);
266 return -EFAULT;
267 }
268 return ret;
269 }
270
z_erofs_lz4_decompress(struct z_erofs_decompress_req * rq,struct page ** pagepool)271 static int z_erofs_lz4_decompress(struct z_erofs_decompress_req *rq,
272 struct page **pagepool)
273 {
274 struct z_erofs_lz4_decompress_ctx ctx;
275 unsigned int dst_maptype;
276 void *dst;
277 int ret;
278
279 ctx.rq = rq;
280 ctx.oend = rq->pageofs_out + rq->outputsize;
281 ctx.outpages = PAGE_ALIGN(ctx.oend) >> PAGE_SHIFT;
282 ctx.inpages = PAGE_ALIGN(rq->inputsize) >> PAGE_SHIFT;
283
284 /* one optimized fast path only for non bigpcluster cases yet */
285 if (ctx.inpages == 1 && ctx.outpages == 1 && !rq->inplace_io) {
286 DBG_BUGON(!*rq->out);
287 dst = kmap_local_page(*rq->out);
288 dst_maptype = 0;
289 goto dstmap_out;
290 }
291
292 /* general decoding path which can be used for all cases */
293 ret = z_erofs_lz4_prepare_dstpages(&ctx, pagepool);
294 if (ret < 0) {
295 return ret;
296 } else if (ret > 0) {
297 dst = page_address(*rq->out);
298 dst_maptype = 1;
299 } else {
300 dst = erofs_vm_map_ram(rq->out, ctx.outpages);
301 if (!dst)
302 return -ENOMEM;
303 dst_maptype = 2;
304 }
305
306 dstmap_out:
307 ret = z_erofs_lz4_decompress_mem(&ctx, dst);
308 if (!dst_maptype)
309 kunmap_local(dst);
310 else if (dst_maptype == 2)
311 vm_unmap_ram(dst, ctx.outpages);
312 return ret;
313 }
314
z_erofs_transform_plain(struct z_erofs_decompress_req * rq,struct page ** pagepool)315 static int z_erofs_transform_plain(struct z_erofs_decompress_req *rq,
316 struct page **pagepool)
317 {
318 const unsigned int nrpages_in =
319 PAGE_ALIGN(rq->pageofs_in + rq->inputsize) >> PAGE_SHIFT;
320 const unsigned int nrpages_out =
321 PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT;
322 const unsigned int bs = rq->sb->s_blocksize;
323 unsigned int cur = 0, ni = 0, no, pi, po, insz, cnt;
324 u8 *kin;
325
326 if (rq->outputsize > rq->inputsize)
327 return -EOPNOTSUPP;
328 if (rq->alg == Z_EROFS_COMPRESSION_INTERLACED) {
329 cur = bs - (rq->pageofs_out & (bs - 1));
330 pi = (rq->pageofs_in + rq->inputsize - cur) & ~PAGE_MASK;
331 cur = min(cur, rq->outputsize);
332 if (cur && rq->out[0]) {
333 kin = kmap_local_page(rq->in[nrpages_in - 1]);
334 if (rq->out[0] == rq->in[nrpages_in - 1]) {
335 memmove(kin + rq->pageofs_out, kin + pi, cur);
336 flush_dcache_page(rq->out[0]);
337 } else {
338 memcpy_to_page(rq->out[0], rq->pageofs_out,
339 kin + pi, cur);
340 }
341 kunmap_local(kin);
342 }
343 rq->outputsize -= cur;
344 }
345
346 for (; rq->outputsize; rq->pageofs_in = 0, cur += PAGE_SIZE, ni++) {
347 insz = min(PAGE_SIZE - rq->pageofs_in, rq->outputsize);
348 rq->outputsize -= insz;
349 if (!rq->in[ni])
350 continue;
351 kin = kmap_local_page(rq->in[ni]);
352 pi = 0;
353 do {
354 no = (rq->pageofs_out + cur + pi) >> PAGE_SHIFT;
355 po = (rq->pageofs_out + cur + pi) & ~PAGE_MASK;
356 DBG_BUGON(no >= nrpages_out);
357 cnt = min(insz - pi, PAGE_SIZE - po);
358 if (rq->out[no] == rq->in[ni]) {
359 memmove(kin + po,
360 kin + rq->pageofs_in + pi, cnt);
361 flush_dcache_page(rq->out[no]);
362 } else if (rq->out[no]) {
363 memcpy_to_page(rq->out[no], po,
364 kin + rq->pageofs_in + pi, cnt);
365 }
366 pi += cnt;
367 } while (pi < insz);
368 kunmap_local(kin);
369 }
370 DBG_BUGON(ni > nrpages_in);
371 return 0;
372 }
373
z_erofs_stream_switch_bufs(struct z_erofs_stream_dctx * dctx,void ** dst,void ** src,struct page ** pgpl)374 int z_erofs_stream_switch_bufs(struct z_erofs_stream_dctx *dctx, void **dst,
375 void **src, struct page **pgpl)
376 {
377 struct z_erofs_decompress_req *rq = dctx->rq;
378 struct super_block *sb = rq->sb;
379 struct page **pgo, *tmppage;
380 unsigned int j;
381
382 if (!dctx->avail_out) {
383 if (++dctx->no >= dctx->outpages || !rq->outputsize) {
384 erofs_err(sb, "insufficient space for decompressed data");
385 return -EFSCORRUPTED;
386 }
387
388 if (dctx->kout)
389 kunmap_local(dctx->kout);
390 dctx->avail_out = min(rq->outputsize, PAGE_SIZE - rq->pageofs_out);
391 rq->outputsize -= dctx->avail_out;
392 pgo = &rq->out[dctx->no];
393 if (!*pgo && rq->fillgaps) { /* deduped */
394 *pgo = erofs_allocpage(pgpl, rq->gfp);
395 if (!*pgo) {
396 dctx->kout = NULL;
397 return -ENOMEM;
398 }
399 set_page_private(*pgo, Z_EROFS_SHORTLIVED_PAGE);
400 }
401 if (*pgo) {
402 dctx->kout = kmap_local_page(*pgo);
403 *dst = dctx->kout + rq->pageofs_out;
404 } else {
405 *dst = dctx->kout = NULL;
406 }
407 rq->pageofs_out = 0;
408 }
409
410 if (dctx->inbuf_pos == dctx->inbuf_sz && rq->inputsize) {
411 if (++dctx->ni >= dctx->inpages) {
412 erofs_err(sb, "invalid compressed data");
413 return -EFSCORRUPTED;
414 }
415 if (dctx->kout) /* unlike kmap(), take care of the orders */
416 kunmap_local(dctx->kout);
417 kunmap_local(dctx->kin);
418
419 dctx->inbuf_sz = min_t(u32, rq->inputsize, PAGE_SIZE);
420 rq->inputsize -= dctx->inbuf_sz;
421 dctx->kin = kmap_local_page(rq->in[dctx->ni]);
422 *src = dctx->kin;
423 dctx->bounced = false;
424 if (dctx->kout) {
425 j = (u8 *)*dst - dctx->kout;
426 dctx->kout = kmap_local_page(rq->out[dctx->no]);
427 *dst = dctx->kout + j;
428 }
429 dctx->inbuf_pos = 0;
430 }
431
432 /*
433 * Handle overlapping: Use the given bounce buffer if the input data is
434 * under processing; Or utilize short-lived pages from the on-stack page
435 * pool, where pages are shared among the same request. Note that only
436 * a few inplace I/O pages need to be doubled.
437 */
438 if (!dctx->bounced && rq->out[dctx->no] == rq->in[dctx->ni]) {
439 memcpy(dctx->bounce, *src, dctx->inbuf_sz);
440 *src = dctx->bounce;
441 dctx->bounced = true;
442 }
443
444 for (j = dctx->ni + 1; j < dctx->inpages; ++j) {
445 if (rq->out[dctx->no] != rq->in[j])
446 continue;
447 tmppage = erofs_allocpage(pgpl, rq->gfp);
448 if (!tmppage)
449 return -ENOMEM;
450 set_page_private(tmppage, Z_EROFS_SHORTLIVED_PAGE);
451 copy_highpage(tmppage, rq->in[j]);
452 rq->in[j] = tmppage;
453 }
454 return 0;
455 }
456
457 const struct z_erofs_decompressor *z_erofs_decomp[] = {
458 [Z_EROFS_COMPRESSION_SHIFTED] = &(const struct z_erofs_decompressor) {
459 .decompress = z_erofs_transform_plain,
460 .name = "shifted"
461 },
462 [Z_EROFS_COMPRESSION_INTERLACED] = &(const struct z_erofs_decompressor) {
463 .decompress = z_erofs_transform_plain,
464 .name = "interlaced"
465 },
466 [Z_EROFS_COMPRESSION_LZ4] = &(const struct z_erofs_decompressor) {
467 .config = z_erofs_load_lz4_config,
468 .decompress = z_erofs_lz4_decompress,
469 .init = z_erofs_gbuf_init,
470 .exit = z_erofs_gbuf_exit,
471 .name = "lz4"
472 },
473 #ifdef CONFIG_EROFS_FS_ZIP_LZMA
474 [Z_EROFS_COMPRESSION_LZMA] = &z_erofs_lzma_decomp,
475 #endif
476 #ifdef CONFIG_EROFS_FS_ZIP_DEFLATE
477 [Z_EROFS_COMPRESSION_DEFLATE] = &z_erofs_deflate_decomp,
478 #endif
479 #ifdef CONFIG_EROFS_FS_ZIP_ZSTD
480 [Z_EROFS_COMPRESSION_ZSTD] = &z_erofs_zstd_decomp,
481 #endif
482 };
483
z_erofs_parse_cfgs(struct super_block * sb,struct erofs_super_block * dsb)484 int z_erofs_parse_cfgs(struct super_block *sb, struct erofs_super_block *dsb)
485 {
486 struct erofs_sb_info *sbi = EROFS_SB(sb);
487 struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
488 unsigned int algs, alg;
489 erofs_off_t offset;
490 int size, ret = 0;
491
492 if (!erofs_sb_has_compr_cfgs(sbi)) {
493 sbi->available_compr_algs = 1 << Z_EROFS_COMPRESSION_LZ4;
494 return z_erofs_load_lz4_config(sb, dsb, NULL, 0);
495 }
496
497 sbi->available_compr_algs = le16_to_cpu(dsb->u1.available_compr_algs);
498 if (sbi->available_compr_algs & ~Z_EROFS_ALL_COMPR_ALGS) {
499 erofs_err(sb, "unidentified algorithms %x, please upgrade kernel",
500 sbi->available_compr_algs & ~Z_EROFS_ALL_COMPR_ALGS);
501 return -EOPNOTSUPP;
502 }
503
504 erofs_init_metabuf(&buf, sb);
505 offset = EROFS_SUPER_OFFSET + sbi->sb_size;
506 alg = 0;
507 for (algs = sbi->available_compr_algs; algs; algs >>= 1, ++alg) {
508 const struct z_erofs_decompressor *dec = z_erofs_decomp[alg];
509 void *data;
510
511 if (!(algs & 1))
512 continue;
513
514 data = erofs_read_metadata(sb, &buf, &offset, &size);
515 if (IS_ERR(data)) {
516 ret = PTR_ERR(data);
517 break;
518 }
519
520 if (alg < Z_EROFS_COMPRESSION_MAX && dec && dec->config) {
521 ret = dec->config(sb, dsb, data, size);
522 } else {
523 erofs_err(sb, "algorithm %d isn't enabled on this kernel",
524 alg);
525 ret = -EOPNOTSUPP;
526 }
527 kfree(data);
528 if (ret)
529 break;
530 }
531 erofs_put_metabuf(&buf);
532 return ret;
533 }
534
z_erofs_init_decompressor(void)535 int __init z_erofs_init_decompressor(void)
536 {
537 int i, err;
538
539 for (i = 0; i < Z_EROFS_COMPRESSION_MAX; ++i) {
540 err = z_erofs_decomp[i] ? z_erofs_decomp[i]->init() : 0;
541 if (err) {
542 while (i--)
543 if (z_erofs_decomp[i])
544 z_erofs_decomp[i]->exit();
545 return err;
546 }
547 }
548 return 0;
549 }
550
z_erofs_exit_decompressor(void)551 void z_erofs_exit_decompressor(void)
552 {
553 int i;
554
555 for (i = 0; i < Z_EROFS_COMPRESSION_MAX; ++i)
556 if (z_erofs_decomp[i])
557 z_erofs_decomp[i]->exit();
558 }
559