xref: /linux/fs/erofs/decompressor.c (revision 662fa3d6099374c4615bf64d06895e3573b935b2)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2019 HUAWEI, Inc.
4  *             https://www.huawei.com/
5  */
6 #include "compress.h"
7 #include <linux/module.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_decompressor {
20 	/*
21 	 * if destpages have sparsed pages, fill them with bounce pages.
22 	 * it also check whether destpages indicate continuous physical memory.
23 	 */
24 	int (*prepare_destpages)(struct z_erofs_decompress_req *rq,
25 				 struct list_head *pagepool);
26 	int (*decompress)(struct z_erofs_decompress_req *rq, u8 *out);
27 	char *name;
28 };
29 
30 int z_erofs_load_lz4_config(struct super_block *sb,
31 			    struct erofs_super_block *dsb,
32 			    struct z_erofs_lz4_cfgs *lz4, int size)
33 {
34 	struct erofs_sb_info *sbi = EROFS_SB(sb);
35 	u16 distance;
36 
37 	if (lz4) {
38 		if (size < sizeof(struct z_erofs_lz4_cfgs)) {
39 			erofs_err(sb, "invalid lz4 cfgs, size=%u", size);
40 			return -EINVAL;
41 		}
42 		distance = le16_to_cpu(lz4->max_distance);
43 
44 		sbi->lz4.max_pclusterblks = le16_to_cpu(lz4->max_pclusterblks);
45 		if (!sbi->lz4.max_pclusterblks) {
46 			sbi->lz4.max_pclusterblks = 1;	/* reserved case */
47 		} else if (sbi->lz4.max_pclusterblks >
48 			   Z_EROFS_PCLUSTER_MAX_SIZE / EROFS_BLKSIZ) {
49 			erofs_err(sb, "too large lz4 pclusterblks %u",
50 				  sbi->lz4.max_pclusterblks);
51 			return -EINVAL;
52 		} else if (sbi->lz4.max_pclusterblks >= 2) {
53 			erofs_info(sb, "EXPERIMENTAL big pcluster feature in use. Use at your own risk!");
54 		}
55 	} else {
56 		distance = le16_to_cpu(dsb->u1.lz4_max_distance);
57 		sbi->lz4.max_pclusterblks = 1;
58 	}
59 
60 	sbi->lz4.max_distance_pages = distance ?
61 					DIV_ROUND_UP(distance, PAGE_SIZE) + 1 :
62 					LZ4_MAX_DISTANCE_PAGES;
63 	return erofs_pcpubuf_growsize(sbi->lz4.max_pclusterblks);
64 }
65 
66 static int z_erofs_lz4_prepare_destpages(struct z_erofs_decompress_req *rq,
67 					 struct list_head *pagepool)
68 {
69 	const unsigned int nr =
70 		PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT;
71 	struct page *availables[LZ4_MAX_DISTANCE_PAGES] = { NULL };
72 	unsigned long bounced[DIV_ROUND_UP(LZ4_MAX_DISTANCE_PAGES,
73 					   BITS_PER_LONG)] = { 0 };
74 	unsigned int lz4_max_distance_pages =
75 				EROFS_SB(rq->sb)->lz4.max_distance_pages;
76 	void *kaddr = NULL;
77 	unsigned int i, j, top;
78 
79 	top = 0;
80 	for (i = j = 0; i < nr; ++i, ++j) {
81 		struct page *const page = rq->out[i];
82 		struct page *victim;
83 
84 		if (j >= lz4_max_distance_pages)
85 			j = 0;
86 
87 		/* 'valid' bounced can only be tested after a complete round */
88 		if (test_bit(j, bounced)) {
89 			DBG_BUGON(i < lz4_max_distance_pages);
90 			DBG_BUGON(top >= lz4_max_distance_pages);
91 			availables[top++] = rq->out[i - lz4_max_distance_pages];
92 		}
93 
94 		if (page) {
95 			__clear_bit(j, bounced);
96 			if (kaddr) {
97 				if (kaddr + PAGE_SIZE == page_address(page))
98 					kaddr += PAGE_SIZE;
99 				else
100 					kaddr = NULL;
101 			} else if (!i) {
102 				kaddr = page_address(page);
103 			}
104 			continue;
105 		}
106 		kaddr = NULL;
107 		__set_bit(j, bounced);
108 
109 		if (top) {
110 			victim = availables[--top];
111 			get_page(victim);
112 		} else {
113 			victim = erofs_allocpage(pagepool,
114 						 GFP_KERNEL | __GFP_NOFAIL);
115 			set_page_private(victim, Z_EROFS_SHORTLIVED_PAGE);
116 		}
117 		rq->out[i] = victim;
118 	}
119 	return kaddr ? 1 : 0;
120 }
121 
122 static void *z_erofs_handle_inplace_io(struct z_erofs_decompress_req *rq,
123 			void *inpage, unsigned int *inputmargin, int *maptype,
124 			bool support_0padding)
125 {
126 	unsigned int nrpages_in, nrpages_out;
127 	unsigned int ofull, oend, inputsize, total, i, j;
128 	struct page **in;
129 	void *src, *tmp;
130 
131 	inputsize = rq->inputsize;
132 	nrpages_in = PAGE_ALIGN(inputsize) >> PAGE_SHIFT;
133 	oend = rq->pageofs_out + rq->outputsize;
134 	ofull = PAGE_ALIGN(oend);
135 	nrpages_out = ofull >> PAGE_SHIFT;
136 
137 	if (rq->inplace_io) {
138 		if (rq->partial_decoding || !support_0padding ||
139 		    ofull - oend < LZ4_DECOMPRESS_INPLACE_MARGIN(inputsize))
140 			goto docopy;
141 
142 		for (i = 0; i < nrpages_in; ++i) {
143 			DBG_BUGON(rq->in[i] == NULL);
144 			for (j = 0; j < nrpages_out - nrpages_in + i; ++j)
145 				if (rq->out[j] == rq->in[i])
146 					goto docopy;
147 		}
148 	}
149 
150 	if (nrpages_in <= 1) {
151 		*maptype = 0;
152 		return inpage;
153 	}
154 	kunmap_atomic(inpage);
155 	might_sleep();
156 	src = erofs_vm_map_ram(rq->in, nrpages_in);
157 	if (!src)
158 		return ERR_PTR(-ENOMEM);
159 	*maptype = 1;
160 	return src;
161 
162 docopy:
163 	/* Or copy compressed data which can be overlapped to per-CPU buffer */
164 	in = rq->in;
165 	src = erofs_get_pcpubuf(nrpages_in);
166 	if (!src) {
167 		DBG_BUGON(1);
168 		kunmap_atomic(inpage);
169 		return ERR_PTR(-EFAULT);
170 	}
171 
172 	tmp = src;
173 	total = rq->inputsize;
174 	while (total) {
175 		unsigned int page_copycnt =
176 			min_t(unsigned int, total, PAGE_SIZE - *inputmargin);
177 
178 		if (!inpage)
179 			inpage = kmap_atomic(*in);
180 		memcpy(tmp, inpage + *inputmargin, page_copycnt);
181 		kunmap_atomic(inpage);
182 		inpage = NULL;
183 		tmp += page_copycnt;
184 		total -= page_copycnt;
185 		++in;
186 		*inputmargin = 0;
187 	}
188 	*maptype = 2;
189 	return src;
190 }
191 
192 static int z_erofs_lz4_decompress(struct z_erofs_decompress_req *rq, u8 *out)
193 {
194 	unsigned int inputmargin;
195 	u8 *headpage, *src;
196 	bool support_0padding;
197 	int ret, maptype;
198 
199 	DBG_BUGON(*rq->in == NULL);
200 	headpage = kmap_atomic(*rq->in);
201 	inputmargin = 0;
202 	support_0padding = false;
203 
204 	/* decompression inplace is only safe when 0padding is enabled */
205 	if (erofs_sb_has_lz4_0padding(EROFS_SB(rq->sb))) {
206 		support_0padding = true;
207 
208 		while (!headpage[inputmargin & ~PAGE_MASK])
209 			if (!(++inputmargin & ~PAGE_MASK))
210 				break;
211 
212 		if (inputmargin >= rq->inputsize) {
213 			kunmap_atomic(headpage);
214 			return -EIO;
215 		}
216 	}
217 
218 	rq->inputsize -= inputmargin;
219 	src = z_erofs_handle_inplace_io(rq, headpage, &inputmargin, &maptype,
220 					support_0padding);
221 	if (IS_ERR(src))
222 		return PTR_ERR(src);
223 
224 	/* legacy format could compress extra data in a pcluster. */
225 	if (rq->partial_decoding || !support_0padding)
226 		ret = LZ4_decompress_safe_partial(src + inputmargin, out,
227 				rq->inputsize, rq->outputsize, rq->outputsize);
228 	else
229 		ret = LZ4_decompress_safe(src + inputmargin, out,
230 					  rq->inputsize, rq->outputsize);
231 
232 	if (ret != rq->outputsize) {
233 		erofs_err(rq->sb, "failed to decompress %d in[%u, %u] out[%u]",
234 			  ret, rq->inputsize, inputmargin, rq->outputsize);
235 
236 		WARN_ON(1);
237 		print_hex_dump(KERN_DEBUG, "[ in]: ", DUMP_PREFIX_OFFSET,
238 			       16, 1, src + inputmargin, rq->inputsize, true);
239 		print_hex_dump(KERN_DEBUG, "[out]: ", DUMP_PREFIX_OFFSET,
240 			       16, 1, out, rq->outputsize, true);
241 
242 		if (ret >= 0)
243 			memset(out + ret, 0, rq->outputsize - ret);
244 		ret = -EIO;
245 	}
246 
247 	if (maptype == 0) {
248 		kunmap_atomic(src);
249 	} else if (maptype == 1) {
250 		vm_unmap_ram(src, PAGE_ALIGN(rq->inputsize) >> PAGE_SHIFT);
251 	} else if (maptype == 2) {
252 		erofs_put_pcpubuf(src);
253 	} else {
254 		DBG_BUGON(1);
255 		return -EFAULT;
256 	}
257 	return ret;
258 }
259 
260 static struct z_erofs_decompressor decompressors[] = {
261 	[Z_EROFS_COMPRESSION_SHIFTED] = {
262 		.name = "shifted"
263 	},
264 	[Z_EROFS_COMPRESSION_LZ4] = {
265 		.prepare_destpages = z_erofs_lz4_prepare_destpages,
266 		.decompress = z_erofs_lz4_decompress,
267 		.name = "lz4"
268 	},
269 };
270 
271 static void copy_from_pcpubuf(struct page **out, const char *dst,
272 			      unsigned short pageofs_out,
273 			      unsigned int outputsize)
274 {
275 	const char *end = dst + outputsize;
276 	const unsigned int righthalf = PAGE_SIZE - pageofs_out;
277 	const char *cur = dst - pageofs_out;
278 
279 	while (cur < end) {
280 		struct page *const page = *out++;
281 
282 		if (page) {
283 			char *buf = kmap_atomic(page);
284 
285 			if (cur >= dst) {
286 				memcpy(buf, cur, min_t(uint, PAGE_SIZE,
287 						       end - cur));
288 			} else {
289 				memcpy(buf + pageofs_out, cur + pageofs_out,
290 				       min_t(uint, righthalf, end - cur));
291 			}
292 			kunmap_atomic(buf);
293 		}
294 		cur += PAGE_SIZE;
295 	}
296 }
297 
298 static int z_erofs_decompress_generic(struct z_erofs_decompress_req *rq,
299 				      struct list_head *pagepool)
300 {
301 	const unsigned int nrpages_out =
302 		PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT;
303 	const struct z_erofs_decompressor *alg = decompressors + rq->alg;
304 	unsigned int dst_maptype;
305 	void *dst;
306 	int ret;
307 
308 	/* two optimized fast paths only for non bigpcluster cases yet */
309 	if (rq->inputsize <= PAGE_SIZE) {
310 		if (nrpages_out == 1 && !rq->inplace_io) {
311 			DBG_BUGON(!*rq->out);
312 			dst = kmap_atomic(*rq->out);
313 			dst_maptype = 0;
314 			goto dstmap_out;
315 		}
316 
317 		/*
318 		 * For the case of small output size (especially much less
319 		 * than PAGE_SIZE), memcpy the decompressed data rather than
320 		 * compressed data is preferred.
321 		 */
322 		if (rq->outputsize <= PAGE_SIZE * 7 / 8) {
323 			dst = erofs_get_pcpubuf(1);
324 			if (IS_ERR(dst))
325 				return PTR_ERR(dst);
326 
327 			rq->inplace_io = false;
328 			ret = alg->decompress(rq, dst);
329 			if (!ret)
330 				copy_from_pcpubuf(rq->out, dst, rq->pageofs_out,
331 						  rq->outputsize);
332 
333 			erofs_put_pcpubuf(dst);
334 			return ret;
335 		}
336 	}
337 
338 	/* general decoding path which can be used for all cases */
339 	ret = alg->prepare_destpages(rq, pagepool);
340 	if (ret < 0)
341 		return ret;
342 	if (ret) {
343 		dst = page_address(*rq->out);
344 		dst_maptype = 1;
345 		goto dstmap_out;
346 	}
347 
348 	dst = erofs_vm_map_ram(rq->out, nrpages_out);
349 	if (!dst)
350 		return -ENOMEM;
351 	dst_maptype = 2;
352 
353 dstmap_out:
354 	ret = alg->decompress(rq, dst + rq->pageofs_out);
355 
356 	if (!dst_maptype)
357 		kunmap_atomic(dst);
358 	else if (dst_maptype == 2)
359 		vm_unmap_ram(dst, nrpages_out);
360 	return ret;
361 }
362 
363 static int z_erofs_shifted_transform(const struct z_erofs_decompress_req *rq,
364 				     struct list_head *pagepool)
365 {
366 	const unsigned int nrpages_out =
367 		PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT;
368 	const unsigned int righthalf = PAGE_SIZE - rq->pageofs_out;
369 	unsigned char *src, *dst;
370 
371 	if (nrpages_out > 2) {
372 		DBG_BUGON(1);
373 		return -EIO;
374 	}
375 
376 	if (rq->out[0] == *rq->in) {
377 		DBG_BUGON(nrpages_out != 1);
378 		return 0;
379 	}
380 
381 	src = kmap_atomic(*rq->in);
382 	if (rq->out[0]) {
383 		dst = kmap_atomic(rq->out[0]);
384 		memcpy(dst + rq->pageofs_out, src, righthalf);
385 		kunmap_atomic(dst);
386 	}
387 
388 	if (nrpages_out == 2) {
389 		DBG_BUGON(!rq->out[1]);
390 		if (rq->out[1] == *rq->in) {
391 			memmove(src, src + righthalf, rq->pageofs_out);
392 		} else {
393 			dst = kmap_atomic(rq->out[1]);
394 			memcpy(dst, src + righthalf, rq->pageofs_out);
395 			kunmap_atomic(dst);
396 		}
397 	}
398 	kunmap_atomic(src);
399 	return 0;
400 }
401 
402 int z_erofs_decompress(struct z_erofs_decompress_req *rq,
403 		       struct list_head *pagepool)
404 {
405 	if (rq->alg == Z_EROFS_COMPRESSION_SHIFTED)
406 		return z_erofs_shifted_transform(rq, pagepool);
407 	return z_erofs_decompress_generic(rq, pagepool);
408 }
409