1 // SPDX-License-Identifier: GPL-2.0
3 * Generic binary BCH encoding/decoding library
5 * Copyright © 2011 Parrot S.A.
7 * Author: Ivan Djelic <ivan.djelic@parrot.com>
11 * This library provides runtime configurable encoding/decoding of binary
12 * Bose-Chaudhuri-Hocquenghem (BCH) codes.
14 * Call init_bch to get a pointer to a newly allocated bch_control structure for
15 * the given m (Galois field order), t (error correction capability) and
16 * (optional) primitive polynomial parameters.
18 * Call encode_bch to compute and store ecc parity bytes to a given buffer.
19 * Call decode_bch to detect and locate errors in received data.
21 * On systems supporting hw BCH features, intermediate results may be provided
22 * to decode_bch in order to skip certain steps. See decode_bch() documentation
25 * Option CONFIG_BCH_CONST_PARAMS can be used to force fixed values of
26 * parameters m and t; thus allowing extra compiler optimizations and providing
27 * better (up to 2x) encoding performance. Using this option makes sense when
28 * (m,t) are fixed and known in advance, e.g. when using BCH error correction
29 * on a particular NAND flash device.
31 * Algorithmic details:
33 * Encoding is performed by processing 32 input bits in parallel, using 4
34 * remainder lookup tables.
36 * The final stage of decoding involves the following internal steps:
37 * a. Syndrome computation
38 * b. Error locator polynomial computation using Berlekamp-Massey algorithm
39 * c. Error locator root finding (by far the most expensive step)
41 * In this implementation, step c is not performed using the usual Chien search.
42 * Instead, an alternative approach described in [1] is used. It consists in
43 * factoring the error locator polynomial using the Berlekamp Trace algorithm
44 * (BTA) down to a certain degree (4), after which ad hoc low-degree polynomial
45 * solving techniques [2] are used. The resulting algorithm, called BTZ, yields
46 * much better performance than Chien search for usual (m,t) values (typically
47 * m >= 13, t < 32, see [1]).
49 * [1] B. Biswas, V. Herbert. Efficient root finding of polynomials over fields
50 * of characteristic 2, in: Western European Workshop on Research in Cryptology
51 * - WEWoRC 2009, Graz, Austria, LNCS, Springer, July 2009, to appear.
52 * [2] [Zin96] V.A. Zinoviev. On the solution of equations of degree 10 over
53 * finite fields GF(2^q). In Rapport de recherche INRIA no 2829, 1996.
58 #include <ubi_uboot.h>
60 #include <linux/bitops.h>
63 #if defined(__FreeBSD__)
64 #include <sys/endian.h>
73 #define cpu_to_be32 htobe32
74 #define DIV_ROUND_UP(n,d) (((n) + (d) - 1) / (d))
75 #define kmalloc(size, flags) malloc(size)
76 #define kzalloc(size, flags) calloc(1, size)
78 #define ARRAY_SIZE(arr) (sizeof(arr) / sizeof((arr)[0]))
81 #include <asm/byteorder.h>
82 #include <linux/bch.h>
84 #if defined(CONFIG_BCH_CONST_PARAMS)
85 #define GF_M(_p) (CONFIG_BCH_CONST_M)
86 #define GF_T(_p) (CONFIG_BCH_CONST_T)
87 #define GF_N(_p) ((1 << (CONFIG_BCH_CONST_M))-1)
89 #define GF_M(_p) ((_p)->m)
90 #define GF_T(_p) ((_p)->t)
91 #define GF_N(_p) ((_p)->n)
94 #define BCH_ECC_WORDS(_p) DIV_ROUND_UP(GF_M(_p)*GF_T(_p), 32)
95 #define BCH_ECC_BYTES(_p) DIV_ROUND_UP(GF_M(_p)*GF_T(_p), 8)
98 #define dbg(_fmt, args...) do {} while (0)
102 * represent a polynomial over GF(2^m)
105 unsigned int deg; /* polynomial degree */
106 unsigned int c[0]; /* polynomial terms */
109 /* given its degree, compute a polynomial size in bytes */
110 #define GF_POLY_SZ(_d) (sizeof(struct gf_poly)+((_d)+1)*sizeof(unsigned int))
112 /* polynomial of degree 1 */
113 struct gf_poly_deg1 {
119 #if !defined(__DragonFly__) && !defined(__FreeBSD__)
120 static int fls(int x)
126 if (!(x & 0xffff0000u)) {
130 if (!(x & 0xff000000u)) {
134 if (!(x & 0xf0000000u)) {
138 if (!(x & 0xc0000000u)) {
142 if (!(x & 0x80000000u)) {
152 * same as encode_bch(), but process input data one byte at a time
154 static void encode_bch_unaligned(struct bch_control *bch,
155 const unsigned char *data, unsigned int len,
160 const int l = BCH_ECC_WORDS(bch)-1;
163 p = bch->mod8_tab + (l+1)*(((ecc[0] >> 24)^(*data++)) & 0xff);
165 for (i = 0; i < l; i++)
166 ecc[i] = ((ecc[i] << 8)|(ecc[i+1] >> 24))^(*p++);
168 ecc[l] = (ecc[l] << 8)^(*p);
173 * convert ecc bytes to aligned, zero-padded 32-bit ecc words
175 static void load_ecc8(struct bch_control *bch, uint32_t *dst,
178 uint8_t pad[4] = {0, 0, 0, 0};
179 unsigned int i, nwords = BCH_ECC_WORDS(bch)-1;
181 for (i = 0; i < nwords; i++, src += 4)
182 dst[i] = (src[0] << 24)|(src[1] << 16)|(src[2] << 8)|src[3];
184 memcpy(pad, src, BCH_ECC_BYTES(bch)-4*nwords);
185 dst[nwords] = (pad[0] << 24)|(pad[1] << 16)|(pad[2] << 8)|pad[3];
189 * convert 32-bit ecc words to ecc bytes
191 static void store_ecc8(struct bch_control *bch, uint8_t *dst,
195 unsigned int i, nwords = BCH_ECC_WORDS(bch)-1;
197 for (i = 0; i < nwords; i++) {
198 *dst++ = (src[i] >> 24);
199 *dst++ = (src[i] >> 16) & 0xff;
200 *dst++ = (src[i] >> 8) & 0xff;
201 *dst++ = (src[i] >> 0) & 0xff;
203 pad[0] = (src[nwords] >> 24);
204 pad[1] = (src[nwords] >> 16) & 0xff;
205 pad[2] = (src[nwords] >> 8) & 0xff;
206 pad[3] = (src[nwords] >> 0) & 0xff;
207 memcpy(dst, pad, BCH_ECC_BYTES(bch)-4*nwords);
211 * encode_bch - calculate BCH ecc parity of data
212 * @bch: BCH control structure
213 * @data: data to encode
214 * @len: data length in bytes
215 * @ecc: ecc parity data, must be initialized by caller
217 * The @ecc parity array is used both as input and output parameter, in order to
218 * allow incremental computations. It should be of the size indicated by member
219 * @ecc_bytes of @bch, and should be initialized to 0 before the first call.
221 * The exact number of computed ecc parity bits is given by member @ecc_bits of
222 * @bch; it may be less than m*t for large values of t.
224 void encode_bch(struct bch_control *bch, const uint8_t *data,
225 unsigned int len, uint8_t *ecc)
227 const unsigned int l = BCH_ECC_WORDS(bch)-1;
228 unsigned int i, mlen;
231 const uint32_t * const tab0 = bch->mod8_tab;
232 const uint32_t * const tab1 = tab0 + 256*(l+1);
233 const uint32_t * const tab2 = tab1 + 256*(l+1);
234 const uint32_t * const tab3 = tab2 + 256*(l+1);
235 const uint32_t *pdata, *p0, *p1, *p2, *p3;
238 /* load ecc parity bytes into internal 32-bit buffer */
239 load_ecc8(bch, bch->ecc_buf, ecc);
241 memset(bch->ecc_buf, 0, sizeof(r));
244 /* process first unaligned data bytes */
245 m = ((unsigned long)data) & 3;
247 mlen = (len < (4-m)) ? len : 4-m;
248 encode_bch_unaligned(bch, data, mlen, bch->ecc_buf);
253 /* process 32-bit aligned data words */
254 pdata = (uint32_t *)data;
258 memcpy(r, bch->ecc_buf, sizeof(r));
261 * split each 32-bit word into 4 polynomials of weight 8 as follows:
263 * 31 ...24 23 ...16 15 ... 8 7 ... 0
264 * xxxxxxxx yyyyyyyy zzzzzzzz tttttttt
265 * tttttttt mod g = r0 (precomputed)
266 * zzzzzzzz 00000000 mod g = r1 (precomputed)
267 * yyyyyyyy 00000000 00000000 mod g = r2 (precomputed)
268 * xxxxxxxx 00000000 00000000 00000000 mod g = r3 (precomputed)
269 * xxxxxxxx yyyyyyyy zzzzzzzz tttttttt mod g = r0^r1^r2^r3
272 /* input data is read in big-endian format */
273 w = r[0]^cpu_to_be32(*pdata++);
274 p0 = tab0 + (l+1)*((w >> 0) & 0xff);
275 p1 = tab1 + (l+1)*((w >> 8) & 0xff);
276 p2 = tab2 + (l+1)*((w >> 16) & 0xff);
277 p3 = tab3 + (l+1)*((w >> 24) & 0xff);
279 for (i = 0; i < l; i++)
280 r[i] = r[i+1]^p0[i]^p1[i]^p2[i]^p3[i];
282 r[l] = p0[l]^p1[l]^p2[l]^p3[l];
284 memcpy(bch->ecc_buf, r, sizeof(r));
286 /* process last unaligned bytes */
288 encode_bch_unaligned(bch, data, len, bch->ecc_buf);
290 /* store ecc parity bytes into original parity buffer */
292 store_ecc8(bch, ecc, bch->ecc_buf);
295 static inline int modulo(struct bch_control *bch, unsigned int v)
297 const unsigned int n = GF_N(bch);
300 v = (v & n) + (v >> GF_M(bch));
306 * shorter and faster modulo function, only works when v < 2N.
308 static inline int mod_s(struct bch_control *bch, unsigned int v)
310 const unsigned int n = GF_N(bch);
311 return (v < n) ? v : v-n;
314 static inline int deg(unsigned int poly)
316 /* polynomial degree is the most-significant bit index */
320 static inline int parity(unsigned int x)
323 * public domain code snippet, lifted from
324 * http://www-graphics.stanford.edu/~seander/bithacks.html
328 x = (x & 0x11111111U) * 0x11111111U;
329 return (x >> 28) & 1;
332 /* Galois field basic operations: multiply, divide, inverse, etc. */
334 static inline unsigned int gf_mul(struct bch_control *bch, unsigned int a,
337 return (a && b) ? bch->a_pow_tab[mod_s(bch, bch->a_log_tab[a]+
338 bch->a_log_tab[b])] : 0;
341 static inline unsigned int gf_sqr(struct bch_control *bch, unsigned int a)
343 return a ? bch->a_pow_tab[mod_s(bch, 2*bch->a_log_tab[a])] : 0;
346 static inline unsigned int gf_div(struct bch_control *bch, unsigned int a,
349 return a ? bch->a_pow_tab[mod_s(bch, bch->a_log_tab[a]+
350 GF_N(bch)-bch->a_log_tab[b])] : 0;
353 static inline unsigned int gf_inv(struct bch_control *bch, unsigned int a)
355 return bch->a_pow_tab[GF_N(bch)-bch->a_log_tab[a]];
358 static inline unsigned int a_pow(struct bch_control *bch, int i)
360 return bch->a_pow_tab[modulo(bch, i)];
363 static inline int a_log(struct bch_control *bch, unsigned int x)
365 return bch->a_log_tab[x];
368 static inline int a_ilog(struct bch_control *bch, unsigned int x)
370 return mod_s(bch, GF_N(bch)-bch->a_log_tab[x]);
374 * compute 2t syndromes of ecc polynomial, i.e. ecc(a^j) for j=1..2t
376 static void compute_syndromes(struct bch_control *bch, uint32_t *ecc,
382 const int t = GF_T(bch);
386 /* make sure extra bits in last ecc word are cleared */
387 m = ((unsigned int)s) & 31;
389 ecc[s/32] &= ~((1u << (32-m))-1);
390 memset(syn, 0, 2*t*sizeof(*syn));
392 /* compute v(a^j) for j=1 .. 2t-1 */
398 for (j = 0; j < 2*t; j += 2)
399 syn[j] ^= a_pow(bch, (j+1)*(i+s));
405 /* v(a^(2j)) = v(a^j)^2 */
406 for (j = 0; j < t; j++)
407 syn[2*j+1] = gf_sqr(bch, syn[j]);
410 static void gf_poly_copy(struct gf_poly *dst, struct gf_poly *src)
412 memcpy(dst, src, GF_POLY_SZ(src->deg));
415 static int compute_error_locator_polynomial(struct bch_control *bch,
416 const unsigned int *syn)
418 const unsigned int t = GF_T(bch);
419 const unsigned int n = GF_N(bch);
420 unsigned int i, j, tmp, l, pd = 1, d = syn[0];
421 struct gf_poly *elp = bch->elp;
422 struct gf_poly *pelp = bch->poly_2t[0];
423 struct gf_poly *elp_copy = bch->poly_2t[1];
426 memset(pelp, 0, GF_POLY_SZ(2*t));
427 memset(elp, 0, GF_POLY_SZ(2*t));
434 /* use simplified binary Berlekamp-Massey algorithm */
435 for (i = 0; (i < t) && (elp->deg <= t); i++) {
438 gf_poly_copy(elp_copy, elp);
439 /* e[i+1](X) = e[i](X)+di*dp^-1*X^2(i-p)*e[p](X) */
440 tmp = a_log(bch, d)+n-a_log(bch, pd);
441 for (j = 0; j <= pelp->deg; j++) {
443 l = a_log(bch, pelp->c[j]);
444 elp->c[j+k] ^= a_pow(bch, tmp+l);
447 /* compute l[i+1] = max(l[i]->c[l[p]+2*(i-p]) */
449 if (tmp > elp->deg) {
451 gf_poly_copy(pelp, elp_copy);
456 /* di+1 = S(2i+3)+elp[i+1].1*S(2i+2)+...+elp[i+1].lS(2i+3-l) */
459 for (j = 1; j <= elp->deg; j++)
460 d ^= gf_mul(bch, elp->c[j], syn[2*i+2-j]);
463 dbg("elp=%s\n", gf_poly_str(elp));
464 return (elp->deg > t) ? -1 : (int)elp->deg;
468 * solve a m x m linear system in GF(2) with an expected number of solutions,
469 * and return the number of found solutions
471 static int solve_linear_system(struct bch_control *bch, unsigned int *rows,
472 unsigned int *sol, int nsol)
474 const int m = GF_M(bch);
475 unsigned int tmp, mask;
476 int rem, c, r, p, k, param[m];
481 /* Gaussian elimination */
482 for (c = 0; c < m; c++) {
485 /* find suitable row for elimination */
486 for (r = p; r < m; r++) {
487 if (rows[r] & mask) {
498 /* perform elimination on remaining rows */
500 for (r = rem; r < m; r++) {
505 /* elimination not needed, store defective row index */
510 /* rewrite system, inserting fake parameter rows */
513 for (r = m-1; r >= 0; r--) {
514 if ((r > m-1-k) && rows[r])
515 /* system has no solution */
518 rows[r] = (p && (r == param[p-1])) ?
519 p--, 1u << (m-r) : rows[r-p];
523 if (nsol != (1 << k))
524 /* unexpected number of solutions */
527 for (p = 0; p < nsol; p++) {
528 /* set parameters for p-th solution */
529 for (c = 0; c < k; c++)
530 rows[param[c]] = (rows[param[c]] & ~1)|((p >> c) & 1);
532 /* compute unique solution */
534 for (r = m-1; r >= 0; r--) {
535 mask = rows[r] & (tmp|1);
536 tmp |= parity(mask) << (m-r);
544 * this function builds and solves a linear system for finding roots of a degree
545 * 4 affine monic polynomial X^4+aX^2+bX+c over GF(2^m).
547 static int find_affine4_roots(struct bch_control *bch, unsigned int a,
548 unsigned int b, unsigned int c,
552 const int m = GF_M(bch);
553 unsigned int mask = 0xff, t, rows[16] = {0,};
559 /* buid linear system to solve X^4+aX^2+bX+c = 0 */
560 for (i = 0; i < m; i++) {
561 rows[i+1] = bch->a_pow_tab[4*i]^
562 (a ? bch->a_pow_tab[mod_s(bch, k)] : 0)^
563 (b ? bch->a_pow_tab[mod_s(bch, j)] : 0);
568 * transpose 16x16 matrix before passing it to linear solver
569 * warning: this code assumes m < 16
571 for (j = 8; j != 0; j >>= 1, mask ^= (mask << j)) {
572 for (k = 0; k < 16; k = (k+j+1) & ~j) {
573 t = ((rows[k] >> j)^rows[k+j]) & mask;
578 return solve_linear_system(bch, rows, roots, 4);
582 * compute root r of a degree 1 polynomial over GF(2^m) (returned as log(1/r))
584 static int find_poly_deg1_roots(struct bch_control *bch, struct gf_poly *poly,
590 /* poly[X] = bX+c with c!=0, root=c/b */
591 roots[n++] = mod_s(bch, GF_N(bch)-bch->a_log_tab[poly->c[0]]+
592 bch->a_log_tab[poly->c[1]]);
597 * compute roots of a degree 2 polynomial over GF(2^m)
599 static int find_poly_deg2_roots(struct bch_control *bch, struct gf_poly *poly,
602 int n = 0, i, l0, l1, l2;
603 unsigned int u, v, r;
605 if (poly->c[0] && poly->c[1]) {
607 l0 = bch->a_log_tab[poly->c[0]];
608 l1 = bch->a_log_tab[poly->c[1]];
609 l2 = bch->a_log_tab[poly->c[2]];
611 /* using z=a/bX, transform aX^2+bX+c into z^2+z+u (u=ac/b^2) */
612 u = a_pow(bch, l0+l2+2*(GF_N(bch)-l1));
614 * let u = sum(li.a^i) i=0..m-1; then compute r = sum(li.xi):
615 * r^2+r = sum(li.(xi^2+xi)) = sum(li.(a^i+Tr(a^i).a^k)) =
616 * u + sum(li.Tr(a^i).a^k) = u+a^k.Tr(sum(li.a^i)) = u+a^k.Tr(u)
617 * i.e. r and r+1 are roots iff Tr(u)=0
627 if ((gf_sqr(bch, r)^r) == u) {
628 /* reverse z=a/bX transformation and compute log(1/r) */
629 roots[n++] = modulo(bch, 2*GF_N(bch)-l1-
630 bch->a_log_tab[r]+l2);
631 roots[n++] = modulo(bch, 2*GF_N(bch)-l1-
632 bch->a_log_tab[r^1]+l2);
639 * compute roots of a degree 3 polynomial over GF(2^m)
641 static int find_poly_deg3_roots(struct bch_control *bch, struct gf_poly *poly,
645 unsigned int a, b, c, a2, b2, c2, e3, tmp[4];
648 /* transform polynomial into monic X^3 + a2X^2 + b2X + c2 */
650 c2 = gf_div(bch, poly->c[0], e3);
651 b2 = gf_div(bch, poly->c[1], e3);
652 a2 = gf_div(bch, poly->c[2], e3);
654 /* (X+a2)(X^3+a2X^2+b2X+c2) = X^4+aX^2+bX+c (affine) */
655 c = gf_mul(bch, a2, c2); /* c = a2c2 */
656 b = gf_mul(bch, a2, b2)^c2; /* b = a2b2 + c2 */
657 a = gf_sqr(bch, a2)^b2; /* a = a2^2 + b2 */
659 /* find the 4 roots of this affine polynomial */
660 if (find_affine4_roots(bch, a, b, c, tmp) == 4) {
661 /* remove a2 from final list of roots */
662 for (i = 0; i < 4; i++) {
664 roots[n++] = a_ilog(bch, tmp[i]);
672 * compute roots of a degree 4 polynomial over GF(2^m)
674 static int find_poly_deg4_roots(struct bch_control *bch, struct gf_poly *poly,
678 unsigned int a, b, c, d, e = 0, f, a2, b2, c2, e4;
683 /* transform polynomial into monic X^4 + aX^3 + bX^2 + cX + d */
685 d = gf_div(bch, poly->c[0], e4);
686 c = gf_div(bch, poly->c[1], e4);
687 b = gf_div(bch, poly->c[2], e4);
688 a = gf_div(bch, poly->c[3], e4);
690 /* use Y=1/X transformation to get an affine polynomial */
692 /* first, eliminate cX by using z=X+e with ae^2+c=0 */
694 /* compute e such that e^2 = c/a */
695 f = gf_div(bch, c, a);
697 l += (l & 1) ? GF_N(bch) : 0;
700 * use transformation z=X+e:
701 * z^4+e^4 + a(z^3+ez^2+e^2z+e^3) + b(z^2+e^2) +cz+ce+d
702 * z^4 + az^3 + (ae+b)z^2 + (ae^2+c)z+e^4+be^2+ae^3+ce+d
703 * z^4 + az^3 + (ae+b)z^2 + e^4+be^2+d
704 * z^4 + az^3 + b'z^2 + d'
706 d = a_pow(bch, 2*l)^gf_mul(bch, b, f)^d;
707 b = gf_mul(bch, a, e)^b;
709 /* now, use Y=1/X to get Y^4 + b/dY^2 + a/dY + 1/d */
711 /* assume all roots have multiplicity 1 */
715 b2 = gf_div(bch, a, d);
716 a2 = gf_div(bch, b, d);
718 /* polynomial is already affine */
723 /* find the 4 roots of this affine polynomial */
724 if (find_affine4_roots(bch, a2, b2, c2, roots) == 4) {
725 for (i = 0; i < 4; i++) {
726 /* post-process roots (reverse transformations) */
727 f = a ? gf_inv(bch, roots[i]) : roots[i];
728 roots[i] = a_ilog(bch, f^e);
736 * build monic, log-based representation of a polynomial
738 static void gf_poly_logrep(struct bch_control *bch,
739 const struct gf_poly *a, int *rep)
741 int i, d = a->deg, l = GF_N(bch)-a_log(bch, a->c[a->deg]);
743 /* represent 0 values with -1; warning, rep[d] is not set to 1 */
744 for (i = 0; i < d; i++)
745 rep[i] = a->c[i] ? mod_s(bch, a_log(bch, a->c[i])+l) : -1;
749 * compute polynomial Euclidean division remainder in GF(2^m)[X]
751 static void gf_poly_mod(struct bch_control *bch, struct gf_poly *a,
752 const struct gf_poly *b, int *rep)
755 unsigned int i, j, *c = a->c;
756 const unsigned int d = b->deg;
761 /* reuse or compute log representation of denominator */
764 gf_poly_logrep(bch, b, rep);
767 for (j = a->deg; j >= d; j--) {
769 la = a_log(bch, c[j]);
771 for (i = 0; i < d; i++, p++) {
774 c[p] ^= bch->a_pow_tab[mod_s(bch,
780 while (!c[a->deg] && a->deg)
785 * compute polynomial Euclidean division quotient in GF(2^m)[X]
787 static void gf_poly_div(struct bch_control *bch, struct gf_poly *a,
788 const struct gf_poly *b, struct gf_poly *q)
790 if (a->deg >= b->deg) {
791 q->deg = a->deg-b->deg;
792 /* compute a mod b (modifies a) */
793 gf_poly_mod(bch, a, b, NULL);
794 /* quotient is stored in upper part of polynomial a */
795 memcpy(q->c, &a->c[b->deg], (1+q->deg)*sizeof(unsigned int));
803 * compute polynomial GCD (Greatest Common Divisor) in GF(2^m)[X]
805 static struct gf_poly *gf_poly_gcd(struct bch_control *bch, struct gf_poly *a,
810 dbg("gcd(%s,%s)=", gf_poly_str(a), gf_poly_str(b));
812 if (a->deg < b->deg) {
819 gf_poly_mod(bch, a, b, NULL);
825 dbg("%s\n", gf_poly_str(a));
831 * Given a polynomial f and an integer k, compute Tr(a^kX) mod f
832 * This is used in Berlekamp Trace algorithm for splitting polynomials
834 static void compute_trace_bk_mod(struct bch_control *bch, int k,
835 const struct gf_poly *f, struct gf_poly *z,
838 const int m = GF_M(bch);
841 /* z contains z^2j mod f */
844 z->c[1] = bch->a_pow_tab[k];
847 memset(out, 0, GF_POLY_SZ(f->deg));
849 /* compute f log representation only once */
850 gf_poly_logrep(bch, f, bch->cache);
852 for (i = 0; i < m; i++) {
853 /* add a^(k*2^i)(z^(2^i) mod f) and compute (z^(2^i) mod f)^2 */
854 for (j = z->deg; j >= 0; j--) {
855 out->c[j] ^= z->c[j];
856 z->c[2*j] = gf_sqr(bch, z->c[j]);
859 if (z->deg > out->deg)
864 /* z^(2(i+1)) mod f = (z^(2^i) mod f)^2 mod f */
865 gf_poly_mod(bch, z, f, bch->cache);
868 while (!out->c[out->deg] && out->deg)
871 dbg("Tr(a^%d.X) mod f = %s\n", k, gf_poly_str(out));
875 * factor a polynomial using Berlekamp Trace algorithm (BTA)
877 static void factor_polynomial(struct bch_control *bch, int k, struct gf_poly *f,
878 struct gf_poly **g, struct gf_poly **h)
880 struct gf_poly *f2 = bch->poly_2t[0];
881 struct gf_poly *q = bch->poly_2t[1];
882 struct gf_poly *tk = bch->poly_2t[2];
883 struct gf_poly *z = bch->poly_2t[3];
886 dbg("factoring %s...\n", gf_poly_str(f));
891 /* tk = Tr(a^k.X) mod f */
892 compute_trace_bk_mod(bch, k, f, z, tk);
895 /* compute g = gcd(f, tk) (destructive operation) */
897 gcd = gf_poly_gcd(bch, f2, tk);
898 if (gcd->deg < f->deg) {
899 /* compute h=f/gcd(f,tk); this will modify f and q */
900 gf_poly_div(bch, f, gcd, q);
901 /* store g and h in-place (clobbering f) */
902 *h = &((struct gf_poly_deg1 *)f)[gcd->deg].poly;
903 gf_poly_copy(*g, gcd);
910 * find roots of a polynomial, using BTZ algorithm; see the beginning of this
913 static int find_poly_roots(struct bch_control *bch, unsigned int k,
914 struct gf_poly *poly, unsigned int *roots)
917 struct gf_poly *f1, *f2;
920 /* handle low degree polynomials with ad hoc techniques */
922 cnt = find_poly_deg1_roots(bch, poly, roots);
925 cnt = find_poly_deg2_roots(bch, poly, roots);
928 cnt = find_poly_deg3_roots(bch, poly, roots);
931 cnt = find_poly_deg4_roots(bch, poly, roots);
934 /* factor polynomial using Berlekamp Trace Algorithm (BTA) */
936 if (poly->deg && (k <= GF_M(bch))) {
937 factor_polynomial(bch, k, poly, &f1, &f2);
939 cnt += find_poly_roots(bch, k+1, f1, roots);
941 cnt += find_poly_roots(bch, k+1, f2, roots+cnt);
948 #if defined(USE_CHIEN_SEARCH)
950 * exhaustive root search (Chien) implementation - not used, included only for
951 * reference/comparison tests
953 static int chien_search(struct bch_control *bch, unsigned int len,
954 struct gf_poly *p, unsigned int *roots)
957 unsigned int i, j, syn, syn0, count = 0;
958 const unsigned int k = 8*len+bch->ecc_bits;
960 /* use a log-based representation of polynomial */
961 gf_poly_logrep(bch, p, bch->cache);
962 bch->cache[p->deg] = 0;
963 syn0 = gf_div(bch, p->c[0], p->c[p->deg]);
965 for (i = GF_N(bch)-k+1; i <= GF_N(bch); i++) {
966 /* compute elp(a^i) */
967 for (j = 1, syn = syn0; j <= p->deg; j++) {
970 syn ^= a_pow(bch, m+j*i);
973 roots[count++] = GF_N(bch)-i;
978 return (count == p->deg) ? count : 0;
980 #define find_poly_roots(_p, _k, _elp, _loc) chien_search(_p, len, _elp, _loc)
981 #endif /* USE_CHIEN_SEARCH */
984 * decode_bch - decode received codeword and find bit error locations
985 * @bch: BCH control structure
986 * @data: received data, ignored if @calc_ecc is provided
987 * @len: data length in bytes, must always be provided
988 * @recv_ecc: received ecc, if NULL then assume it was XORed in @calc_ecc
989 * @calc_ecc: calculated ecc, if NULL then calc_ecc is computed from @data
990 * @syn: hw computed syndrome data (if NULL, syndrome is calculated)
991 * @errloc: output array of error locations
994 * The number of errors found, or -EBADMSG if decoding failed, or -EINVAL if
995 * invalid parameters were provided
997 * Depending on the available hw BCH support and the need to compute @calc_ecc
998 * separately (using encode_bch()), this function should be called with one of
999 * the following parameter configurations -
1001 * by providing @data and @recv_ecc only:
1002 * decode_bch(@bch, @data, @len, @recv_ecc, NULL, NULL, @errloc)
1004 * by providing @recv_ecc and @calc_ecc:
1005 * decode_bch(@bch, NULL, @len, @recv_ecc, @calc_ecc, NULL, @errloc)
1007 * by providing ecc = recv_ecc XOR calc_ecc:
1008 * decode_bch(@bch, NULL, @len, NULL, ecc, NULL, @errloc)
1010 * by providing syndrome results @syn:
1011 * decode_bch(@bch, NULL, @len, NULL, NULL, @syn, @errloc)
1013 * Once decode_bch() has successfully returned with a positive value, error
1014 * locations returned in array @errloc should be interpreted as follows -
1016 * if (errloc[n] >= 8*len), then n-th error is located in ecc (no need for
1019 * if (errloc[n] < 8*len), then n-th error is located in data and can be
1020 * corrected with statement data[errloc[n]/8] ^= 1 << (errloc[n] % 8);
1022 * Note that this function does not perform any data correction by itself, it
1023 * merely indicates error locations.
1025 int decode_bch(struct bch_control *bch, const uint8_t *data, unsigned int len,
1026 const uint8_t *recv_ecc, const uint8_t *calc_ecc,
1027 const unsigned int *syn, unsigned int *errloc)
1029 const unsigned int ecc_words = BCH_ECC_WORDS(bch);
1034 /* sanity check: make sure data length can be handled */
1035 if (8*len > (bch->n-bch->ecc_bits))
1038 /* if caller does not provide syndromes, compute them */
1041 /* compute received data ecc into an internal buffer */
1042 if (!data || !recv_ecc)
1044 encode_bch(bch, data, len, NULL);
1046 /* load provided calculated ecc */
1047 load_ecc8(bch, bch->ecc_buf, calc_ecc);
1049 /* load received ecc or assume it was XORed in calc_ecc */
1051 load_ecc8(bch, bch->ecc_buf2, recv_ecc);
1052 /* XOR received and calculated ecc */
1053 for (i = 0, sum = 0; i < (int)ecc_words; i++) {
1054 bch->ecc_buf[i] ^= bch->ecc_buf2[i];
1055 sum |= bch->ecc_buf[i];
1058 /* no error found */
1061 compute_syndromes(bch, bch->ecc_buf, bch->syn);
1065 err = compute_error_locator_polynomial(bch, syn);
1067 nroots = find_poly_roots(bch, 1, bch->elp, errloc);
1072 /* post-process raw error locations for easier correction */
1073 nbits = (len*8)+bch->ecc_bits;
1074 for (i = 0; i < err; i++) {
1075 if (errloc[i] >= nbits) {
1079 errloc[i] = nbits-1-errloc[i];
1080 errloc[i] = (errloc[i] & ~7)|(7-(errloc[i] & 7));
1083 return (err >= 0) ? err : -EBADMSG;
1087 * generate Galois field lookup tables
1089 static int build_gf_tables(struct bch_control *bch, unsigned int poly)
1091 unsigned int i, x = 1;
1092 const unsigned int k = 1 << deg(poly);
1094 /* primitive polynomial must be of degree m */
1095 if (k != (1u << GF_M(bch)))
1098 for (i = 0; i < GF_N(bch); i++) {
1099 bch->a_pow_tab[i] = x;
1100 bch->a_log_tab[x] = i;
1102 /* polynomial is not primitive (a^i=1 with 0<i<2^m-1) */
1108 bch->a_pow_tab[GF_N(bch)] = 1;
1109 bch->a_log_tab[0] = 0;
1115 * compute generator polynomial remainder tables for fast encoding
1117 static void build_mod8_tables(struct bch_control *bch, const uint32_t *g)
1120 uint32_t data, hi, lo, *tab;
1121 const int l = BCH_ECC_WORDS(bch);
1122 const int plen = DIV_ROUND_UP(bch->ecc_bits+1, 32);
1123 const int ecclen = DIV_ROUND_UP(bch->ecc_bits, 32);
1125 memset(bch->mod8_tab, 0, 4*256*l*sizeof(*bch->mod8_tab));
1127 for (i = 0; i < 256; i++) {
1128 /* p(X)=i is a small polynomial of weight <= 8 */
1129 for (b = 0; b < 4; b++) {
1130 /* we want to compute (p(X).X^(8*b+deg(g))) mod g(X) */
1131 tab = bch->mod8_tab + (b*256+i)*l;
1135 /* subtract X^d.g(X) from p(X).X^(8*b+deg(g)) */
1136 data ^= g[0] >> (31-d);
1137 for (j = 0; j < ecclen; j++) {
1138 hi = (d < 31) ? g[j] << (d+1) : 0;
1140 g[j+1] >> (31-d) : 0;
1149 * build a base for factoring degree 2 polynomials
1151 static int build_deg2_base(struct bch_control *bch)
1153 const int m = GF_M(bch);
1155 unsigned int sum, x, y, remaining, ak = 0, xi[m];
1157 /* find k s.t. Tr(a^k) = 1 and 0 <= k < m */
1158 for (i = 0; i < m; i++) {
1159 for (j = 0, sum = 0; j < m; j++)
1160 sum ^= a_pow(bch, i*(1 << j));
1163 ak = bch->a_pow_tab[i];
1167 /* find xi, i=0..m-1 such that xi^2+xi = a^i+Tr(a^i).a^k */
1169 memset(xi, 0, sizeof(xi));
1171 for (x = 0; (x <= GF_N(bch)) && remaining; x++) {
1172 y = gf_sqr(bch, x)^x;
1173 for (i = 0; i < 2; i++) {
1175 if (y && (r < m) && !xi[r]) {
1179 dbg("x%d = %x\n", r, x);
1185 /* should not happen but check anyway */
1186 return remaining ? -1 : 0;
1189 static void *bch_alloc(size_t size, int *err)
1193 ptr = kmalloc(size, GFP_KERNEL);
1200 * compute generator polynomial for given (m,t) parameters.
1202 static uint32_t *compute_generator_polynomial(struct bch_control *bch)
1204 const unsigned int m = GF_M(bch);
1205 const unsigned int t = GF_T(bch);
1207 unsigned int i, j, nbits, r, word, *roots;
1211 g = bch_alloc(GF_POLY_SZ(m*t), &err);
1212 roots = bch_alloc((bch->n+1)*sizeof(*roots), &err);
1213 genpoly = bch_alloc(DIV_ROUND_UP(m*t+1, 32)*sizeof(*genpoly), &err);
1221 /* enumerate all roots of g(X) */
1222 memset(roots , 0, (bch->n+1)*sizeof(*roots));
1223 for (i = 0; i < t; i++) {
1224 for (j = 0, r = 2*i+1; j < m; j++) {
1226 r = mod_s(bch, 2*r);
1229 /* build generator polynomial g(X) */
1232 for (i = 0; i < GF_N(bch); i++) {
1234 /* multiply g(X) by (X+root) */
1235 r = bch->a_pow_tab[i];
1237 for (j = g->deg; j > 0; j--)
1238 g->c[j] = gf_mul(bch, g->c[j], r)^g->c[j-1];
1240 g->c[0] = gf_mul(bch, g->c[0], r);
1244 /* store left-justified binary representation of g(X) */
1249 nbits = (n > 32) ? 32 : n;
1250 for (j = 0, word = 0; j < nbits; j++) {
1252 word |= 1u << (31-j);
1254 genpoly[i++] = word;
1257 bch->ecc_bits = g->deg;
1267 * init_bch - initialize a BCH encoder/decoder
1268 * @m: Galois field order, should be in the range 5-15
1269 * @t: maximum error correction capability, in bits
1270 * @prim_poly: user-provided primitive polynomial (or 0 to use default)
1273 * a newly allocated BCH control structure if successful, NULL otherwise
1275 * This initialization can take some time, as lookup tables are built for fast
1276 * encoding/decoding; make sure not to call this function from a time critical
1277 * path. Usually, init_bch() should be called on module/driver init and
1278 * free_bch() should be called to release memory on exit.
1280 * You may provide your own primitive polynomial of degree @m in argument
1281 * @prim_poly, or let init_bch() use its default polynomial.
1283 * Once init_bch() has successfully returned a pointer to a newly allocated
1284 * BCH control structure, ecc length in bytes is given by member @ecc_bytes of
1287 struct bch_control *init_bch(int m, int t, unsigned int prim_poly)
1290 unsigned int i, words;
1292 struct bch_control *bch = NULL;
1294 const int min_m = 5;
1295 const int max_m = 15;
1297 /* default primitive polynomials */
1298 static const unsigned int prim_poly_tab[] = {
1299 0x25, 0x43, 0x83, 0x11d, 0x211, 0x409, 0x805, 0x1053, 0x201b,
1303 #if defined(CONFIG_BCH_CONST_PARAMS)
1304 if ((m != (CONFIG_BCH_CONST_M)) || (t != (CONFIG_BCH_CONST_T))) {
1305 printk(KERN_ERR "bch encoder/decoder was configured to support "
1306 "parameters m=%d, t=%d only!\n",
1307 CONFIG_BCH_CONST_M, CONFIG_BCH_CONST_T);
1311 if ((m < min_m) || (m > max_m))
1313 * values of m greater than 15 are not currently supported;
1314 * supporting m > 15 would require changing table base type
1315 * (uint16_t) and a small patch in matrix transposition
1320 if ((t < 1) || (m*t >= ((1 << m)-1)))
1321 /* invalid t value */
1324 /* select a primitive polynomial for generating GF(2^m) */
1326 prim_poly = prim_poly_tab[m-min_m];
1328 bch = kzalloc(sizeof(*bch), GFP_KERNEL);
1334 bch->n = (1 << m)-1;
1335 words = DIV_ROUND_UP(m*t, 32);
1336 bch->ecc_bytes = DIV_ROUND_UP(m*t, 8);
1337 bch->a_pow_tab = bch_alloc((1+bch->n)*sizeof(*bch->a_pow_tab), &err);
1338 bch->a_log_tab = bch_alloc((1+bch->n)*sizeof(*bch->a_log_tab), &err);
1339 bch->mod8_tab = bch_alloc(words*1024*sizeof(*bch->mod8_tab), &err);
1340 bch->ecc_buf = bch_alloc(words*sizeof(*bch->ecc_buf), &err);
1341 bch->ecc_buf2 = bch_alloc(words*sizeof(*bch->ecc_buf2), &err);
1342 bch->xi_tab = bch_alloc(m*sizeof(*bch->xi_tab), &err);
1343 bch->syn = bch_alloc(2*t*sizeof(*bch->syn), &err);
1344 bch->cache = bch_alloc(2*t*sizeof(*bch->cache), &err);
1345 bch->elp = bch_alloc((t+1)*sizeof(struct gf_poly_deg1), &err);
1347 for (i = 0; i < ARRAY_SIZE(bch->poly_2t); i++)
1348 bch->poly_2t[i] = bch_alloc(GF_POLY_SZ(2*t), &err);
1353 err = build_gf_tables(bch, prim_poly);
1357 /* use generator polynomial for computing encoding tables */
1358 genpoly = compute_generator_polynomial(bch);
1359 if (genpoly == NULL)
1362 build_mod8_tables(bch, genpoly);
1365 err = build_deg2_base(bch);
1377 * free_bch - free the BCH control structure
1378 * @bch: BCH control structure to release
1380 void free_bch(struct bch_control *bch)
1385 kfree(bch->a_pow_tab);
1386 kfree(bch->a_log_tab);
1387 kfree(bch->mod8_tab);
1388 kfree(bch->ecc_buf);
1389 kfree(bch->ecc_buf2);
1395 for (i = 0; i < ARRAY_SIZE(bch->poly_2t); i++)
1396 kfree(bch->poly_2t[i]);