xref: /linux/lib/decompress_unxz.c (revision 55d0969c451159cff86949b38c39171cab962069)
1 // SPDX-License-Identifier: 0BSD
2 
3 /*
4  * Wrapper for decompressing XZ-compressed kernel, initramfs, and initrd
5  *
6  * Author: Lasse Collin <lasse.collin@tukaani.org>
7  */
8 
9 /*
10  * Important notes about in-place decompression
11  *
12  * At least on x86, the kernel is decompressed in place: the compressed data
13  * is placed to the end of the output buffer, and the decompressor overwrites
14  * most of the compressed data. There must be enough safety margin to
15  * guarantee that the write position is always behind the read position.
16  *
17  * The safety margin for XZ with LZMA2 or BCJ+LZMA2 is calculated below.
18  * Note that the margin with XZ is bigger than with Deflate (gzip)!
19  *
20  * The worst case for in-place decompression is that the beginning of
21  * the file is compressed extremely well, and the rest of the file is
22  * incompressible. Thus, we must look for worst-case expansion when the
23  * compressor is encoding incompressible data.
24  *
25  * The structure of the .xz file in case of a compressed kernel is as follows.
26  * Sizes (as bytes) of the fields are in parenthesis.
27  *
28  *    Stream Header (12)
29  *    Block Header:
30  *      Block Header (8-12)
31  *      Compressed Data (N)
32  *      Block Padding (0-3)
33  *      CRC32 (4)
34  *    Index (8-20)
35  *    Stream Footer (12)
36  *
37  * Normally there is exactly one Block, but let's assume that there are
38  * 2-4 Blocks just in case. Because Stream Header and also Block Header
39  * of the first Block don't make the decompressor produce any uncompressed
40  * data, we can ignore them from our calculations. Block Headers of possible
41  * additional Blocks have to be taken into account still. With these
42  * assumptions, it is safe to assume that the total header overhead is
43  * less than 128 bytes.
44  *
45  * Compressed Data contains LZMA2 or BCJ+LZMA2 encoded data. Since BCJ
46  * doesn't change the size of the data, it is enough to calculate the
47  * safety margin for LZMA2.
48  *
49  * LZMA2 stores the data in chunks. Each chunk has a header whose size is
50  * a maximum of 6 bytes, but to get round 2^n numbers, let's assume that
51  * the maximum chunk header size is 8 bytes. After the chunk header, there
52  * may be up to 64 KiB of actual payload in the chunk. Often the payload is
53  * quite a bit smaller though; to be safe, let's assume that an average
54  * chunk has only 32 KiB of payload.
55  *
56  * The maximum uncompressed size of the payload is 2 MiB. The minimum
57  * uncompressed size of the payload is in practice never less than the
58  * payload size itself. The LZMA2 format would allow uncompressed size
59  * to be less than the payload size, but no sane compressor creates such
60  * files. LZMA2 supports storing incompressible data in uncompressed form,
61  * so there's never a need to create payloads whose uncompressed size is
62  * smaller than the compressed size.
63  *
64  * The assumption, that the uncompressed size of the payload is never
65  * smaller than the payload itself, is valid only when talking about
66  * the payload as a whole. It is possible that the payload has parts where
67  * the decompressor consumes more input than it produces output. Calculating
68  * the worst case for this would be tricky. Instead of trying to do that,
69  * let's simply make sure that the decompressor never overwrites any bytes
70  * of the payload which it is currently reading.
71  *
72  * Now we have enough information to calculate the safety margin. We need
73  *   - 128 bytes for the .xz file format headers;
74  *   - 8 bytes per every 32 KiB of uncompressed size (one LZMA2 chunk header
75  *     per chunk, each chunk having average payload size of 32 KiB); and
76  *   - 64 KiB (biggest possible LZMA2 chunk payload size) to make sure that
77  *     the decompressor never overwrites anything from the LZMA2 chunk
78  *     payload it is currently reading.
79  *
80  * We get the following formula:
81  *
82  *    safety_margin = 128 + uncompressed_size * 8 / 32768 + 65536
83  *                  = 128 + (uncompressed_size >> 12) + 65536
84  *
85  * For comparison, according to arch/x86/boot/compressed/misc.c, the
86  * equivalent formula for Deflate is this:
87  *
88  *    safety_margin = 18 + (uncompressed_size >> 12) + 32768
89  *
90  * Thus, when updating Deflate-only in-place kernel decompressor to
91  * support XZ, the fixed overhead has to be increased from 18+32768 bytes
92  * to 128+65536 bytes.
93  */
94 
95 /*
96  * STATIC is defined to "static" if we are being built for kernel
97  * decompression (pre-boot code). <linux/decompress/mm.h> will define
98  * STATIC to empty if it wasn't already defined. Since we will need to
99  * know later if we are being used for kernel decompression, we define
100  * XZ_PREBOOT here.
101  */
102 #ifdef STATIC
103 #	define XZ_PREBOOT
104 #else
105 #	include <linux/decompress/unxz.h>
106 #endif
107 #ifdef __KERNEL__
108 #	include <linux/decompress/mm.h>
109 #endif
110 
111 #ifndef XZ_PREBOOT
112 #	include <linux/slab.h>
113 #	include <linux/xz.h>
114 #else
115 /*
116  * Use the internal CRC32 code instead of kernel's CRC32 module, which
117  * is not available in early phase of booting.
118  */
119 #define XZ_INTERNAL_CRC32 1
120 
121 /*
122  * For boot time use, we enable only the BCJ filter of the current
123  * architecture or none if no BCJ filter is available for the architecture.
124  */
125 #ifdef CONFIG_X86
126 #	define XZ_DEC_X86
127 #endif
128 #if defined(CONFIG_PPC) && defined(CONFIG_CPU_BIG_ENDIAN)
129 #	define XZ_DEC_POWERPC
130 #endif
131 #ifdef CONFIG_ARM
132 #	ifdef CONFIG_THUMB2_KERNEL
133 #		define XZ_DEC_ARMTHUMB
134 #	else
135 #		define XZ_DEC_ARM
136 #	endif
137 #endif
138 #ifdef CONFIG_ARM64
139 #	define XZ_DEC_ARM64
140 #endif
141 #ifdef CONFIG_RISCV
142 #	define XZ_DEC_RISCV
143 #endif
144 #ifdef CONFIG_SPARC
145 #	define XZ_DEC_SPARC
146 #endif
147 
148 /*
149  * This will get the basic headers so that memeq() and others
150  * can be defined.
151  */
152 #include "xz/xz_private.h"
153 
154 /*
155  * Replace the normal allocation functions with the versions from
156  * <linux/decompress/mm.h>. vfree() needs to support vfree(NULL)
157  * when XZ_DYNALLOC is used, but the pre-boot free() doesn't support it.
158  * Workaround it here because the other decompressors don't need it.
159  */
160 #undef kmalloc
161 #undef kfree
162 #undef vmalloc
163 #undef vfree
164 #define kmalloc(size, flags) malloc(size)
165 #define kfree(ptr) free(ptr)
166 #define vmalloc(size) malloc(size)
167 #define vfree(ptr) do { if (ptr != NULL) free(ptr); } while (0)
168 
169 /*
170  * FIXME: Not all basic memory functions are provided in architecture-specific
171  * files (yet). We define our own versions here for now, but this should be
172  * only a temporary solution.
173  *
174  * memeq and memzero are not used much and any remotely sane implementation
175  * is fast enough. memcpy/memmove speed matters in multi-call mode, but
176  * the kernel image is decompressed in single-call mode, in which only
177  * memmove speed can matter and only if there is a lot of incompressible data
178  * (LZMA2 stores incompressible chunks in uncompressed form). Thus, the
179  * functions below should just be kept small; it's probably not worth
180  * optimizing for speed.
181  */
182 
183 #ifndef memeq
184 static bool memeq(const void *a, const void *b, size_t size)
185 {
186 	const uint8_t *x = a;
187 	const uint8_t *y = b;
188 	size_t i;
189 
190 	for (i = 0; i < size; ++i)
191 		if (x[i] != y[i])
192 			return false;
193 
194 	return true;
195 }
196 #endif
197 
198 #ifndef memzero
199 static void memzero(void *buf, size_t size)
200 {
201 	uint8_t *b = buf;
202 	uint8_t *e = b + size;
203 
204 	while (b != e)
205 		*b++ = '\0';
206 }
207 #endif
208 
209 #ifndef memmove
210 /* Not static to avoid a conflict with the prototype in the Linux headers. */
211 void *memmove(void *dest, const void *src, size_t size)
212 {
213 	uint8_t *d = dest;
214 	const uint8_t *s = src;
215 	size_t i;
216 
217 	if (d < s) {
218 		for (i = 0; i < size; ++i)
219 			d[i] = s[i];
220 	} else if (d > s) {
221 		i = size;
222 		while (i-- > 0)
223 			d[i] = s[i];
224 	}
225 
226 	return dest;
227 }
228 #endif
229 
230 /*
231  * Since we need memmove anyway, we could use it as memcpy too.
232  * Commented out for now to avoid breaking things.
233  */
234 /*
235 #ifndef memcpy
236 #	define memcpy memmove
237 #endif
238 */
239 
240 #include "xz/xz_crc32.c"
241 #include "xz/xz_dec_stream.c"
242 #include "xz/xz_dec_lzma2.c"
243 #include "xz/xz_dec_bcj.c"
244 
245 #endif /* XZ_PREBOOT */
246 
247 /* Size of the input and output buffers in multi-call mode */
248 #define XZ_IOBUF_SIZE 4096
249 
250 /*
251  * This function implements the API defined in <linux/decompress/generic.h>.
252  *
253  * This wrapper will automatically choose single-call or multi-call mode
254  * of the native XZ decoder API. The single-call mode can be used only when
255  * both input and output buffers are available as a single chunk, i.e. when
256  * fill() and flush() won't be used.
257  */
258 STATIC int INIT unxz(unsigned char *in, long in_size,
259 		     long (*fill)(void *dest, unsigned long size),
260 		     long (*flush)(void *src, unsigned long size),
261 		     unsigned char *out, long *in_used,
262 		     void (*error)(char *x))
263 {
264 	struct xz_buf b;
265 	struct xz_dec *s;
266 	enum xz_ret ret;
267 	bool must_free_in = false;
268 
269 #if XZ_INTERNAL_CRC32
270 	xz_crc32_init();
271 #endif
272 
273 	if (in_used != NULL)
274 		*in_used = 0;
275 
276 	if (fill == NULL && flush == NULL)
277 		s = xz_dec_init(XZ_SINGLE, 0);
278 	else
279 		s = xz_dec_init(XZ_DYNALLOC, (uint32_t)-1);
280 
281 	if (s == NULL)
282 		goto error_alloc_state;
283 
284 	if (flush == NULL) {
285 		b.out = out;
286 		b.out_size = (size_t)-1;
287 	} else {
288 		b.out_size = XZ_IOBUF_SIZE;
289 		b.out = malloc(XZ_IOBUF_SIZE);
290 		if (b.out == NULL)
291 			goto error_alloc_out;
292 	}
293 
294 	if (in == NULL) {
295 		must_free_in = true;
296 		in = malloc(XZ_IOBUF_SIZE);
297 		if (in == NULL)
298 			goto error_alloc_in;
299 	}
300 
301 	b.in = in;
302 	b.in_pos = 0;
303 	b.in_size = in_size;
304 	b.out_pos = 0;
305 
306 	if (fill == NULL && flush == NULL) {
307 		ret = xz_dec_run(s, &b);
308 	} else {
309 		do {
310 			if (b.in_pos == b.in_size && fill != NULL) {
311 				if (in_used != NULL)
312 					*in_used += b.in_pos;
313 
314 				b.in_pos = 0;
315 
316 				in_size = fill(in, XZ_IOBUF_SIZE);
317 				if (in_size < 0) {
318 					/*
319 					 * This isn't an optimal error code
320 					 * but it probably isn't worth making
321 					 * a new one either.
322 					 */
323 					ret = XZ_BUF_ERROR;
324 					break;
325 				}
326 
327 				b.in_size = in_size;
328 			}
329 
330 			ret = xz_dec_run(s, &b);
331 
332 			if (flush != NULL && (b.out_pos == b.out_size
333 					|| (ret != XZ_OK && b.out_pos > 0))) {
334 				/*
335 				 * Setting ret here may hide an error
336 				 * returned by xz_dec_run(), but probably
337 				 * it's not too bad.
338 				 */
339 				if (flush(b.out, b.out_pos) != (long)b.out_pos)
340 					ret = XZ_BUF_ERROR;
341 
342 				b.out_pos = 0;
343 			}
344 		} while (ret == XZ_OK);
345 
346 		if (must_free_in)
347 			free(in);
348 
349 		if (flush != NULL)
350 			free(b.out);
351 	}
352 
353 	if (in_used != NULL)
354 		*in_used += b.in_pos;
355 
356 	xz_dec_end(s);
357 
358 	switch (ret) {
359 	case XZ_STREAM_END:
360 		return 0;
361 
362 	case XZ_MEM_ERROR:
363 		/* This can occur only in multi-call mode. */
364 		error("XZ decompressor ran out of memory");
365 		break;
366 
367 	case XZ_FORMAT_ERROR:
368 		error("Input is not in the XZ format (wrong magic bytes)");
369 		break;
370 
371 	case XZ_OPTIONS_ERROR:
372 		error("Input was encoded with settings that are not "
373 				"supported by this XZ decoder");
374 		break;
375 
376 	case XZ_DATA_ERROR:
377 	case XZ_BUF_ERROR:
378 		error("XZ-compressed data is corrupt");
379 		break;
380 
381 	default:
382 		error("Bug in the XZ decompressor");
383 		break;
384 	}
385 
386 	return -1;
387 
388 error_alloc_in:
389 	if (flush != NULL)
390 		free(b.out);
391 
392 error_alloc_out:
393 	xz_dec_end(s);
394 
395 error_alloc_state:
396 	error("XZ decompressor ran out of memory");
397 	return -1;
398 }
399 
400 /*
401  * This function is used by architecture-specific files to decompress
402  * the kernel image.
403  */
404 #ifdef XZ_PREBOOT
405 STATIC int INIT __decompress(unsigned char *in, long in_size,
406 			     long (*fill)(void *dest, unsigned long size),
407 			     long (*flush)(void *src, unsigned long size),
408 			     unsigned char *out, long out_size,
409 			     long *in_used,
410 			     void (*error)(char *x))
411 {
412 	return unxz(in, in_size, fill, flush, out, in_used, error);
413 }
414 #endif
415