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.
59 #include <ubi_uboot.h>
60 #include <dm/devres.h>
62 #include <linux/bitops.h>
65 #if defined(__FreeBSD__)
66 #include <sys/endian.h>
67 #elif defined(__APPLE__)
68 #include <machine/endian.h>
69 #include <libkern/OSByteOrder.h>
78 #if defined(__APPLE__)
79 #define cpu_to_be32 OSSwapHostToBigInt32
81 #define cpu_to_be32 htobe32
83 #define DIV_ROUND_UP(n,d) (((n) + (d) - 1) / (d))
84 #define kmalloc(size, flags) malloc(size)
85 #define kzalloc(size, flags) calloc(1, size)
87 #define ARRAY_SIZE(arr) (sizeof(arr) / sizeof((arr)[0]))
90 #include <asm/byteorder.h>
91 #include <linux/bch.h>
93 #if defined(CONFIG_BCH_CONST_PARAMS)
94 #define GF_M(_p) (CONFIG_BCH_CONST_M)
95 #define GF_T(_p) (CONFIG_BCH_CONST_T)
96 #define GF_N(_p) ((1 << (CONFIG_BCH_CONST_M))-1)
98 #define GF_M(_p) ((_p)->m)
99 #define GF_T(_p) ((_p)->t)
100 #define GF_N(_p) ((_p)->n)
103 #define BCH_ECC_WORDS(_p) DIV_ROUND_UP(GF_M(_p)*GF_T(_p), 32)
104 #define BCH_ECC_BYTES(_p) DIV_ROUND_UP(GF_M(_p)*GF_T(_p), 8)
107 #define dbg(_fmt, args...) do {} while (0)
111 * represent a polynomial over GF(2^m)
114 unsigned int deg; /* polynomial degree */
115 unsigned int c[0]; /* polynomial terms */
118 /* given its degree, compute a polynomial size in bytes */
119 #define GF_POLY_SZ(_d) (sizeof(struct gf_poly)+((_d)+1)*sizeof(unsigned int))
121 /* polynomial of degree 1 */
122 struct gf_poly_deg1 {
128 #if !defined(__DragonFly__) && !defined(__FreeBSD__) && !defined(__APPLE__)
129 static int fls(int x)
135 if (!(x & 0xffff0000u)) {
139 if (!(x & 0xff000000u)) {
143 if (!(x & 0xf0000000u)) {
147 if (!(x & 0xc0000000u)) {
151 if (!(x & 0x80000000u)) {
161 * same as encode_bch(), but process input data one byte at a time
163 static void encode_bch_unaligned(struct bch_control *bch,
164 const unsigned char *data, unsigned int len,
169 const int l = BCH_ECC_WORDS(bch)-1;
172 p = bch->mod8_tab + (l+1)*(((ecc[0] >> 24)^(*data++)) & 0xff);
174 for (i = 0; i < l; i++)
175 ecc[i] = ((ecc[i] << 8)|(ecc[i+1] >> 24))^(*p++);
177 ecc[l] = (ecc[l] << 8)^(*p);
182 * convert ecc bytes to aligned, zero-padded 32-bit ecc words
184 static void load_ecc8(struct bch_control *bch, uint32_t *dst,
187 uint8_t pad[4] = {0, 0, 0, 0};
188 unsigned int i, nwords = BCH_ECC_WORDS(bch)-1;
190 for (i = 0; i < nwords; i++, src += 4)
191 dst[i] = (src[0] << 24)|(src[1] << 16)|(src[2] << 8)|src[3];
193 memcpy(pad, src, BCH_ECC_BYTES(bch)-4*nwords);
194 dst[nwords] = (pad[0] << 24)|(pad[1] << 16)|(pad[2] << 8)|pad[3];
198 * convert 32-bit ecc words to ecc bytes
200 static void store_ecc8(struct bch_control *bch, uint8_t *dst,
204 unsigned int i, nwords = BCH_ECC_WORDS(bch)-1;
206 for (i = 0; i < nwords; i++) {
207 *dst++ = (src[i] >> 24);
208 *dst++ = (src[i] >> 16) & 0xff;
209 *dst++ = (src[i] >> 8) & 0xff;
210 *dst++ = (src[i] >> 0) & 0xff;
212 pad[0] = (src[nwords] >> 24);
213 pad[1] = (src[nwords] >> 16) & 0xff;
214 pad[2] = (src[nwords] >> 8) & 0xff;
215 pad[3] = (src[nwords] >> 0) & 0xff;
216 memcpy(dst, pad, BCH_ECC_BYTES(bch)-4*nwords);
220 * encode_bch - calculate BCH ecc parity of data
221 * @bch: BCH control structure
222 * @data: data to encode
223 * @len: data length in bytes
224 * @ecc: ecc parity data, must be initialized by caller
226 * The @ecc parity array is used both as input and output parameter, in order to
227 * allow incremental computations. It should be of the size indicated by member
228 * @ecc_bytes of @bch, and should be initialized to 0 before the first call.
230 * The exact number of computed ecc parity bits is given by member @ecc_bits of
231 * @bch; it may be less than m*t for large values of t.
233 void encode_bch(struct bch_control *bch, const uint8_t *data,
234 unsigned int len, uint8_t *ecc)
236 const unsigned int l = BCH_ECC_WORDS(bch)-1;
237 unsigned int i, mlen;
240 const uint32_t * const tab0 = bch->mod8_tab;
241 const uint32_t * const tab1 = tab0 + 256*(l+1);
242 const uint32_t * const tab2 = tab1 + 256*(l+1);
243 const uint32_t * const tab3 = tab2 + 256*(l+1);
244 const uint32_t *pdata, *p0, *p1, *p2, *p3;
247 /* load ecc parity bytes into internal 32-bit buffer */
248 load_ecc8(bch, bch->ecc_buf, ecc);
250 memset(bch->ecc_buf, 0, sizeof(r));
253 /* process first unaligned data bytes */
254 m = ((unsigned long)data) & 3;
256 mlen = (len < (4-m)) ? len : 4-m;
257 encode_bch_unaligned(bch, data, mlen, bch->ecc_buf);
262 /* process 32-bit aligned data words */
263 pdata = (uint32_t *)data;
267 memcpy(r, bch->ecc_buf, sizeof(r));
270 * split each 32-bit word into 4 polynomials of weight 8 as follows:
272 * 31 ...24 23 ...16 15 ... 8 7 ... 0
273 * xxxxxxxx yyyyyyyy zzzzzzzz tttttttt
274 * tttttttt mod g = r0 (precomputed)
275 * zzzzzzzz 00000000 mod g = r1 (precomputed)
276 * yyyyyyyy 00000000 00000000 mod g = r2 (precomputed)
277 * xxxxxxxx 00000000 00000000 00000000 mod g = r3 (precomputed)
278 * xxxxxxxx yyyyyyyy zzzzzzzz tttttttt mod g = r0^r1^r2^r3
281 /* input data is read in big-endian format */
282 w = r[0]^cpu_to_be32(*pdata++);
283 p0 = tab0 + (l+1)*((w >> 0) & 0xff);
284 p1 = tab1 + (l+1)*((w >> 8) & 0xff);
285 p2 = tab2 + (l+1)*((w >> 16) & 0xff);
286 p3 = tab3 + (l+1)*((w >> 24) & 0xff);
288 for (i = 0; i < l; i++)
289 r[i] = r[i+1]^p0[i]^p1[i]^p2[i]^p3[i];
291 r[l] = p0[l]^p1[l]^p2[l]^p3[l];
293 memcpy(bch->ecc_buf, r, sizeof(r));
295 /* process last unaligned bytes */
297 encode_bch_unaligned(bch, data, len, bch->ecc_buf);
299 /* store ecc parity bytes into original parity buffer */
301 store_ecc8(bch, ecc, bch->ecc_buf);
304 static inline int modulo(struct bch_control *bch, unsigned int v)
306 const unsigned int n = GF_N(bch);
309 v = (v & n) + (v >> GF_M(bch));
315 * shorter and faster modulo function, only works when v < 2N.
317 static inline int mod_s(struct bch_control *bch, unsigned int v)
319 const unsigned int n = GF_N(bch);
320 return (v < n) ? v : v-n;
323 static inline int deg(unsigned int poly)
325 /* polynomial degree is the most-significant bit index */
329 static inline int parity(unsigned int x)
332 * public domain code snippet, lifted from
333 * http://www-graphics.stanford.edu/~seander/bithacks.html
337 x = (x & 0x11111111U) * 0x11111111U;
338 return (x >> 28) & 1;
341 /* Galois field basic operations: multiply, divide, inverse, etc. */
343 static inline unsigned int gf_mul(struct bch_control *bch, unsigned int a,
346 return (a && b) ? bch->a_pow_tab[mod_s(bch, bch->a_log_tab[a]+
347 bch->a_log_tab[b])] : 0;
350 static inline unsigned int gf_sqr(struct bch_control *bch, unsigned int a)
352 return a ? bch->a_pow_tab[mod_s(bch, 2*bch->a_log_tab[a])] : 0;
355 static inline unsigned int gf_div(struct bch_control *bch, unsigned int a,
358 return a ? bch->a_pow_tab[mod_s(bch, bch->a_log_tab[a]+
359 GF_N(bch)-bch->a_log_tab[b])] : 0;
362 static inline unsigned int gf_inv(struct bch_control *bch, unsigned int a)
364 return bch->a_pow_tab[GF_N(bch)-bch->a_log_tab[a]];
367 static inline unsigned int a_pow(struct bch_control *bch, int i)
369 return bch->a_pow_tab[modulo(bch, i)];
372 static inline int a_log(struct bch_control *bch, unsigned int x)
374 return bch->a_log_tab[x];
377 static inline int a_ilog(struct bch_control *bch, unsigned int x)
379 return mod_s(bch, GF_N(bch)-bch->a_log_tab[x]);
383 * compute 2t syndromes of ecc polynomial, i.e. ecc(a^j) for j=1..2t
385 static void compute_syndromes(struct bch_control *bch, uint32_t *ecc,
391 const int t = GF_T(bch);
395 /* make sure extra bits in last ecc word are cleared */
396 m = ((unsigned int)s) & 31;
398 ecc[s/32] &= ~((1u << (32-m))-1);
399 memset(syn, 0, 2*t*sizeof(*syn));
401 /* compute v(a^j) for j=1 .. 2t-1 */
407 for (j = 0; j < 2*t; j += 2)
408 syn[j] ^= a_pow(bch, (j+1)*(i+s));
414 /* v(a^(2j)) = v(a^j)^2 */
415 for (j = 0; j < t; j++)
416 syn[2*j+1] = gf_sqr(bch, syn[j]);
419 static void gf_poly_copy(struct gf_poly *dst, struct gf_poly *src)
421 memcpy(dst, src, GF_POLY_SZ(src->deg));
424 static int compute_error_locator_polynomial(struct bch_control *bch,
425 const unsigned int *syn)
427 const unsigned int t = GF_T(bch);
428 const unsigned int n = GF_N(bch);
429 unsigned int i, j, tmp, l, pd = 1, d = syn[0];
430 struct gf_poly *elp = bch->elp;
431 struct gf_poly *pelp = bch->poly_2t[0];
432 struct gf_poly *elp_copy = bch->poly_2t[1];
435 memset(pelp, 0, GF_POLY_SZ(2*t));
436 memset(elp, 0, GF_POLY_SZ(2*t));
443 /* use simplified binary Berlekamp-Massey algorithm */
444 for (i = 0; (i < t) && (elp->deg <= t); i++) {
447 gf_poly_copy(elp_copy, elp);
448 /* e[i+1](X) = e[i](X)+di*dp^-1*X^2(i-p)*e[p](X) */
449 tmp = a_log(bch, d)+n-a_log(bch, pd);
450 for (j = 0; j <= pelp->deg; j++) {
452 l = a_log(bch, pelp->c[j]);
453 elp->c[j+k] ^= a_pow(bch, tmp+l);
456 /* compute l[i+1] = max(l[i]->c[l[p]+2*(i-p]) */
458 if (tmp > elp->deg) {
460 gf_poly_copy(pelp, elp_copy);
465 /* di+1 = S(2i+3)+elp[i+1].1*S(2i+2)+...+elp[i+1].lS(2i+3-l) */
468 for (j = 1; j <= elp->deg; j++)
469 d ^= gf_mul(bch, elp->c[j], syn[2*i+2-j]);
472 dbg("elp=%s\n", gf_poly_str(elp));
473 return (elp->deg > t) ? -1 : (int)elp->deg;
477 * solve a m x m linear system in GF(2) with an expected number of solutions,
478 * and return the number of found solutions
480 static int solve_linear_system(struct bch_control *bch, unsigned int *rows,
481 unsigned int *sol, int nsol)
483 const int m = GF_M(bch);
484 unsigned int tmp, mask;
485 int rem, c, r, p, k, param[m];
490 /* Gaussian elimination */
491 for (c = 0; c < m; c++) {
494 /* find suitable row for elimination */
495 for (r = p; r < m; r++) {
496 if (rows[r] & mask) {
507 /* perform elimination on remaining rows */
509 for (r = rem; r < m; r++) {
514 /* elimination not needed, store defective row index */
519 /* rewrite system, inserting fake parameter rows */
522 for (r = m-1; r >= 0; r--) {
523 if ((r > m-1-k) && rows[r])
524 /* system has no solution */
527 rows[r] = (p && (r == param[p-1])) ?
528 p--, 1u << (m-r) : rows[r-p];
532 if (nsol != (1 << k))
533 /* unexpected number of solutions */
536 for (p = 0; p < nsol; p++) {
537 /* set parameters for p-th solution */
538 for (c = 0; c < k; c++)
539 rows[param[c]] = (rows[param[c]] & ~1)|((p >> c) & 1);
541 /* compute unique solution */
543 for (r = m-1; r >= 0; r--) {
544 mask = rows[r] & (tmp|1);
545 tmp |= parity(mask) << (m-r);
553 * this function builds and solves a linear system for finding roots of a degree
554 * 4 affine monic polynomial X^4+aX^2+bX+c over GF(2^m).
556 static int find_affine4_roots(struct bch_control *bch, unsigned int a,
557 unsigned int b, unsigned int c,
561 const int m = GF_M(bch);
562 unsigned int mask = 0xff, t, rows[16] = {0,};
568 /* buid linear system to solve X^4+aX^2+bX+c = 0 */
569 for (i = 0; i < m; i++) {
570 rows[i+1] = bch->a_pow_tab[4*i]^
571 (a ? bch->a_pow_tab[mod_s(bch, k)] : 0)^
572 (b ? bch->a_pow_tab[mod_s(bch, j)] : 0);
577 * transpose 16x16 matrix before passing it to linear solver
578 * warning: this code assumes m < 16
580 for (j = 8; j != 0; j >>= 1, mask ^= (mask << j)) {
581 for (k = 0; k < 16; k = (k+j+1) & ~j) {
582 t = ((rows[k] >> j)^rows[k+j]) & mask;
587 return solve_linear_system(bch, rows, roots, 4);
591 * compute root r of a degree 1 polynomial over GF(2^m) (returned as log(1/r))
593 static int find_poly_deg1_roots(struct bch_control *bch, struct gf_poly *poly,
599 /* poly[X] = bX+c with c!=0, root=c/b */
600 roots[n++] = mod_s(bch, GF_N(bch)-bch->a_log_tab[poly->c[0]]+
601 bch->a_log_tab[poly->c[1]]);
606 * compute roots of a degree 2 polynomial over GF(2^m)
608 static int find_poly_deg2_roots(struct bch_control *bch, struct gf_poly *poly,
611 int n = 0, i, l0, l1, l2;
612 unsigned int u, v, r;
614 if (poly->c[0] && poly->c[1]) {
616 l0 = bch->a_log_tab[poly->c[0]];
617 l1 = bch->a_log_tab[poly->c[1]];
618 l2 = bch->a_log_tab[poly->c[2]];
620 /* using z=a/bX, transform aX^2+bX+c into z^2+z+u (u=ac/b^2) */
621 u = a_pow(bch, l0+l2+2*(GF_N(bch)-l1));
623 * let u = sum(li.a^i) i=0..m-1; then compute r = sum(li.xi):
624 * r^2+r = sum(li.(xi^2+xi)) = sum(li.(a^i+Tr(a^i).a^k)) =
625 * u + sum(li.Tr(a^i).a^k) = u+a^k.Tr(sum(li.a^i)) = u+a^k.Tr(u)
626 * i.e. r and r+1 are roots iff Tr(u)=0
636 if ((gf_sqr(bch, r)^r) == u) {
637 /* reverse z=a/bX transformation and compute log(1/r) */
638 roots[n++] = modulo(bch, 2*GF_N(bch)-l1-
639 bch->a_log_tab[r]+l2);
640 roots[n++] = modulo(bch, 2*GF_N(bch)-l1-
641 bch->a_log_tab[r^1]+l2);
648 * compute roots of a degree 3 polynomial over GF(2^m)
650 static int find_poly_deg3_roots(struct bch_control *bch, struct gf_poly *poly,
654 unsigned int a, b, c, a2, b2, c2, e3, tmp[4];
657 /* transform polynomial into monic X^3 + a2X^2 + b2X + c2 */
659 c2 = gf_div(bch, poly->c[0], e3);
660 b2 = gf_div(bch, poly->c[1], e3);
661 a2 = gf_div(bch, poly->c[2], e3);
663 /* (X+a2)(X^3+a2X^2+b2X+c2) = X^4+aX^2+bX+c (affine) */
664 c = gf_mul(bch, a2, c2); /* c = a2c2 */
665 b = gf_mul(bch, a2, b2)^c2; /* b = a2b2 + c2 */
666 a = gf_sqr(bch, a2)^b2; /* a = a2^2 + b2 */
668 /* find the 4 roots of this affine polynomial */
669 if (find_affine4_roots(bch, a, b, c, tmp) == 4) {
670 /* remove a2 from final list of roots */
671 for (i = 0; i < 4; i++) {
673 roots[n++] = a_ilog(bch, tmp[i]);
681 * compute roots of a degree 4 polynomial over GF(2^m)
683 static int find_poly_deg4_roots(struct bch_control *bch, struct gf_poly *poly,
687 unsigned int a, b, c, d, e = 0, f, a2, b2, c2, e4;
692 /* transform polynomial into monic X^4 + aX^3 + bX^2 + cX + d */
694 d = gf_div(bch, poly->c[0], e4);
695 c = gf_div(bch, poly->c[1], e4);
696 b = gf_div(bch, poly->c[2], e4);
697 a = gf_div(bch, poly->c[3], e4);
699 /* use Y=1/X transformation to get an affine polynomial */
701 /* first, eliminate cX by using z=X+e with ae^2+c=0 */
703 /* compute e such that e^2 = c/a */
704 f = gf_div(bch, c, a);
706 l += (l & 1) ? GF_N(bch) : 0;
709 * use transformation z=X+e:
710 * z^4+e^4 + a(z^3+ez^2+e^2z+e^3) + b(z^2+e^2) +cz+ce+d
711 * z^4 + az^3 + (ae+b)z^2 + (ae^2+c)z+e^4+be^2+ae^3+ce+d
712 * z^4 + az^3 + (ae+b)z^2 + e^4+be^2+d
713 * z^4 + az^3 + b'z^2 + d'
715 d = a_pow(bch, 2*l)^gf_mul(bch, b, f)^d;
716 b = gf_mul(bch, a, e)^b;
718 /* now, use Y=1/X to get Y^4 + b/dY^2 + a/dY + 1/d */
720 /* assume all roots have multiplicity 1 */
724 b2 = gf_div(bch, a, d);
725 a2 = gf_div(bch, b, d);
727 /* polynomial is already affine */
732 /* find the 4 roots of this affine polynomial */
733 if (find_affine4_roots(bch, a2, b2, c2, roots) == 4) {
734 for (i = 0; i < 4; i++) {
735 /* post-process roots (reverse transformations) */
736 f = a ? gf_inv(bch, roots[i]) : roots[i];
737 roots[i] = a_ilog(bch, f^e);
745 * build monic, log-based representation of a polynomial
747 static void gf_poly_logrep(struct bch_control *bch,
748 const struct gf_poly *a, int *rep)
750 int i, d = a->deg, l = GF_N(bch)-a_log(bch, a->c[a->deg]);
752 /* represent 0 values with -1; warning, rep[d] is not set to 1 */
753 for (i = 0; i < d; i++)
754 rep[i] = a->c[i] ? mod_s(bch, a_log(bch, a->c[i])+l) : -1;
758 * compute polynomial Euclidean division remainder in GF(2^m)[X]
760 static void gf_poly_mod(struct bch_control *bch, struct gf_poly *a,
761 const struct gf_poly *b, int *rep)
764 unsigned int i, j, *c = a->c;
765 const unsigned int d = b->deg;
770 /* reuse or compute log representation of denominator */
773 gf_poly_logrep(bch, b, rep);
776 for (j = a->deg; j >= d; j--) {
778 la = a_log(bch, c[j]);
780 for (i = 0; i < d; i++, p++) {
783 c[p] ^= bch->a_pow_tab[mod_s(bch,
789 while (!c[a->deg] && a->deg)
794 * compute polynomial Euclidean division quotient in GF(2^m)[X]
796 static void gf_poly_div(struct bch_control *bch, struct gf_poly *a,
797 const struct gf_poly *b, struct gf_poly *q)
799 if (a->deg >= b->deg) {
800 q->deg = a->deg-b->deg;
801 /* compute a mod b (modifies a) */
802 gf_poly_mod(bch, a, b, NULL);
803 /* quotient is stored in upper part of polynomial a */
804 memcpy(q->c, &a->c[b->deg], (1+q->deg)*sizeof(unsigned int));
812 * compute polynomial GCD (Greatest Common Divisor) in GF(2^m)[X]
814 static struct gf_poly *gf_poly_gcd(struct bch_control *bch, struct gf_poly *a,
819 dbg("gcd(%s,%s)=", gf_poly_str(a), gf_poly_str(b));
821 if (a->deg < b->deg) {
828 gf_poly_mod(bch, a, b, NULL);
834 dbg("%s\n", gf_poly_str(a));
840 * Given a polynomial f and an integer k, compute Tr(a^kX) mod f
841 * This is used in Berlekamp Trace algorithm for splitting polynomials
843 static void compute_trace_bk_mod(struct bch_control *bch, int k,
844 const struct gf_poly *f, struct gf_poly *z,
847 const int m = GF_M(bch);
850 /* z contains z^2j mod f */
853 z->c[1] = bch->a_pow_tab[k];
856 memset(out, 0, GF_POLY_SZ(f->deg));
858 /* compute f log representation only once */
859 gf_poly_logrep(bch, f, bch->cache);
861 for (i = 0; i < m; i++) {
862 /* add a^(k*2^i)(z^(2^i) mod f) and compute (z^(2^i) mod f)^2 */
863 for (j = z->deg; j >= 0; j--) {
864 out->c[j] ^= z->c[j];
865 z->c[2*j] = gf_sqr(bch, z->c[j]);
868 if (z->deg > out->deg)
873 /* z^(2(i+1)) mod f = (z^(2^i) mod f)^2 mod f */
874 gf_poly_mod(bch, z, f, bch->cache);
877 while (!out->c[out->deg] && out->deg)
880 dbg("Tr(a^%d.X) mod f = %s\n", k, gf_poly_str(out));
884 * factor a polynomial using Berlekamp Trace algorithm (BTA)
886 static void factor_polynomial(struct bch_control *bch, int k, struct gf_poly *f,
887 struct gf_poly **g, struct gf_poly **h)
889 struct gf_poly *f2 = bch->poly_2t[0];
890 struct gf_poly *q = bch->poly_2t[1];
891 struct gf_poly *tk = bch->poly_2t[2];
892 struct gf_poly *z = bch->poly_2t[3];
895 dbg("factoring %s...\n", gf_poly_str(f));
900 /* tk = Tr(a^k.X) mod f */
901 compute_trace_bk_mod(bch, k, f, z, tk);
904 /* compute g = gcd(f, tk) (destructive operation) */
906 gcd = gf_poly_gcd(bch, f2, tk);
907 if (gcd->deg < f->deg) {
908 /* compute h=f/gcd(f,tk); this will modify f and q */
909 gf_poly_div(bch, f, gcd, q);
910 /* store g and h in-place (clobbering f) */
911 *h = &((struct gf_poly_deg1 *)f)[gcd->deg].poly;
912 gf_poly_copy(*g, gcd);
919 * find roots of a polynomial, using BTZ algorithm; see the beginning of this
922 static int find_poly_roots(struct bch_control *bch, unsigned int k,
923 struct gf_poly *poly, unsigned int *roots)
926 struct gf_poly *f1, *f2;
929 /* handle low degree polynomials with ad hoc techniques */
931 cnt = find_poly_deg1_roots(bch, poly, roots);
934 cnt = find_poly_deg2_roots(bch, poly, roots);
937 cnt = find_poly_deg3_roots(bch, poly, roots);
940 cnt = find_poly_deg4_roots(bch, poly, roots);
943 /* factor polynomial using Berlekamp Trace Algorithm (BTA) */
945 if (poly->deg && (k <= GF_M(bch))) {
946 factor_polynomial(bch, k, poly, &f1, &f2);
948 cnt += find_poly_roots(bch, k+1, f1, roots);
950 cnt += find_poly_roots(bch, k+1, f2, roots+cnt);
957 #if defined(USE_CHIEN_SEARCH)
959 * exhaustive root search (Chien) implementation - not used, included only for
960 * reference/comparison tests
962 static int chien_search(struct bch_control *bch, unsigned int len,
963 struct gf_poly *p, unsigned int *roots)
966 unsigned int i, j, syn, syn0, count = 0;
967 const unsigned int k = 8*len+bch->ecc_bits;
969 /* use a log-based representation of polynomial */
970 gf_poly_logrep(bch, p, bch->cache);
971 bch->cache[p->deg] = 0;
972 syn0 = gf_div(bch, p->c[0], p->c[p->deg]);
974 for (i = GF_N(bch)-k+1; i <= GF_N(bch); i++) {
975 /* compute elp(a^i) */
976 for (j = 1, syn = syn0; j <= p->deg; j++) {
979 syn ^= a_pow(bch, m+j*i);
982 roots[count++] = GF_N(bch)-i;
987 return (count == p->deg) ? count : 0;
989 #define find_poly_roots(_p, _k, _elp, _loc) chien_search(_p, len, _elp, _loc)
990 #endif /* USE_CHIEN_SEARCH */
993 * decode_bch - decode received codeword and find bit error locations
994 * @bch: BCH control structure
995 * @data: received data, ignored if @calc_ecc is provided
996 * @len: data length in bytes, must always be provided
997 * @recv_ecc: received ecc, if NULL then assume it was XORed in @calc_ecc
998 * @calc_ecc: calculated ecc, if NULL then calc_ecc is computed from @data
999 * @syn: hw computed syndrome data (if NULL, syndrome is calculated)
1000 * @errloc: output array of error locations
1003 * The number of errors found, or -EBADMSG if decoding failed, or -EINVAL if
1004 * invalid parameters were provided
1006 * Depending on the available hw BCH support and the need to compute @calc_ecc
1007 * separately (using encode_bch()), this function should be called with one of
1008 * the following parameter configurations -
1010 * by providing @data and @recv_ecc only:
1011 * decode_bch(@bch, @data, @len, @recv_ecc, NULL, NULL, @errloc)
1013 * by providing @recv_ecc and @calc_ecc:
1014 * decode_bch(@bch, NULL, @len, @recv_ecc, @calc_ecc, NULL, @errloc)
1016 * by providing ecc = recv_ecc XOR calc_ecc:
1017 * decode_bch(@bch, NULL, @len, NULL, ecc, NULL, @errloc)
1019 * by providing syndrome results @syn:
1020 * decode_bch(@bch, NULL, @len, NULL, NULL, @syn, @errloc)
1022 * Once decode_bch() has successfully returned with a positive value, error
1023 * locations returned in array @errloc should be interpreted as follows -
1025 * if (errloc[n] >= 8*len), then n-th error is located in ecc (no need for
1028 * if (errloc[n] < 8*len), then n-th error is located in data and can be
1029 * corrected with statement data[errloc[n]/8] ^= 1 << (errloc[n] % 8);
1031 * Note that this function does not perform any data correction by itself, it
1032 * merely indicates error locations.
1034 int decode_bch(struct bch_control *bch, const uint8_t *data, unsigned int len,
1035 const uint8_t *recv_ecc, const uint8_t *calc_ecc,
1036 const unsigned int *syn, unsigned int *errloc)
1038 const unsigned int ecc_words = BCH_ECC_WORDS(bch);
1043 /* sanity check: make sure data length can be handled */
1044 if (8*len > (bch->n-bch->ecc_bits))
1047 /* if caller does not provide syndromes, compute them */
1050 /* compute received data ecc into an internal buffer */
1051 if (!data || !recv_ecc)
1053 encode_bch(bch, data, len, NULL);
1055 /* load provided calculated ecc */
1056 load_ecc8(bch, bch->ecc_buf, calc_ecc);
1058 /* load received ecc or assume it was XORed in calc_ecc */
1060 load_ecc8(bch, bch->ecc_buf2, recv_ecc);
1061 /* XOR received and calculated ecc */
1062 for (i = 0, sum = 0; i < (int)ecc_words; i++) {
1063 bch->ecc_buf[i] ^= bch->ecc_buf2[i];
1064 sum |= bch->ecc_buf[i];
1067 /* no error found */
1070 compute_syndromes(bch, bch->ecc_buf, bch->syn);
1074 err = compute_error_locator_polynomial(bch, syn);
1076 nroots = find_poly_roots(bch, 1, bch->elp, errloc);
1081 /* post-process raw error locations for easier correction */
1082 nbits = (len*8)+bch->ecc_bits;
1083 for (i = 0; i < err; i++) {
1084 if (errloc[i] >= nbits) {
1088 errloc[i] = nbits-1-errloc[i];
1089 errloc[i] = (errloc[i] & ~7)|(7-(errloc[i] & 7));
1092 return (err >= 0) ? err : -EBADMSG;
1096 * generate Galois field lookup tables
1098 static int build_gf_tables(struct bch_control *bch, unsigned int poly)
1100 unsigned int i, x = 1;
1101 const unsigned int k = 1 << deg(poly);
1103 /* primitive polynomial must be of degree m */
1104 if (k != (1u << GF_M(bch)))
1107 for (i = 0; i < GF_N(bch); i++) {
1108 bch->a_pow_tab[i] = x;
1109 bch->a_log_tab[x] = i;
1111 /* polynomial is not primitive (a^i=1 with 0<i<2^m-1) */
1117 bch->a_pow_tab[GF_N(bch)] = 1;
1118 bch->a_log_tab[0] = 0;
1124 * compute generator polynomial remainder tables for fast encoding
1126 static void build_mod8_tables(struct bch_control *bch, const uint32_t *g)
1129 uint32_t data, hi, lo, *tab;
1130 const int l = BCH_ECC_WORDS(bch);
1131 const int plen = DIV_ROUND_UP(bch->ecc_bits+1, 32);
1132 const int ecclen = DIV_ROUND_UP(bch->ecc_bits, 32);
1134 memset(bch->mod8_tab, 0, 4*256*l*sizeof(*bch->mod8_tab));
1136 for (i = 0; i < 256; i++) {
1137 /* p(X)=i is a small polynomial of weight <= 8 */
1138 for (b = 0; b < 4; b++) {
1139 /* we want to compute (p(X).X^(8*b+deg(g))) mod g(X) */
1140 tab = bch->mod8_tab + (b*256+i)*l;
1144 /* subtract X^d.g(X) from p(X).X^(8*b+deg(g)) */
1145 data ^= g[0] >> (31-d);
1146 for (j = 0; j < ecclen; j++) {
1147 hi = (d < 31) ? g[j] << (d+1) : 0;
1149 g[j+1] >> (31-d) : 0;
1158 * build a base for factoring degree 2 polynomials
1160 static int build_deg2_base(struct bch_control *bch)
1162 const int m = GF_M(bch);
1164 unsigned int sum, x, y, remaining, ak = 0, xi[m];
1166 /* find k s.t. Tr(a^k) = 1 and 0 <= k < m */
1167 for (i = 0; i < m; i++) {
1168 for (j = 0, sum = 0; j < m; j++)
1169 sum ^= a_pow(bch, i*(1 << j));
1172 ak = bch->a_pow_tab[i];
1176 /* find xi, i=0..m-1 such that xi^2+xi = a^i+Tr(a^i).a^k */
1178 memset(xi, 0, sizeof(xi));
1180 for (x = 0; (x <= GF_N(bch)) && remaining; x++) {
1181 y = gf_sqr(bch, x)^x;
1182 for (i = 0; i < 2; i++) {
1184 if (y && (r < m) && !xi[r]) {
1188 dbg("x%d = %x\n", r, x);
1194 /* should not happen but check anyway */
1195 return remaining ? -1 : 0;
1198 static void *bch_alloc(size_t size, int *err)
1202 ptr = kmalloc(size, GFP_KERNEL);
1209 * compute generator polynomial for given (m,t) parameters.
1211 static uint32_t *compute_generator_polynomial(struct bch_control *bch)
1213 const unsigned int m = GF_M(bch);
1214 const unsigned int t = GF_T(bch);
1216 unsigned int i, j, nbits, r, word, *roots;
1220 g = bch_alloc(GF_POLY_SZ(m*t), &err);
1221 roots = bch_alloc((bch->n+1)*sizeof(*roots), &err);
1222 genpoly = bch_alloc(DIV_ROUND_UP(m*t+1, 32)*sizeof(*genpoly), &err);
1230 /* enumerate all roots of g(X) */
1231 memset(roots , 0, (bch->n+1)*sizeof(*roots));
1232 for (i = 0; i < t; i++) {
1233 for (j = 0, r = 2*i+1; j < m; j++) {
1235 r = mod_s(bch, 2*r);
1238 /* build generator polynomial g(X) */
1241 for (i = 0; i < GF_N(bch); i++) {
1243 /* multiply g(X) by (X+root) */
1244 r = bch->a_pow_tab[i];
1246 for (j = g->deg; j > 0; j--)
1247 g->c[j] = gf_mul(bch, g->c[j], r)^g->c[j-1];
1249 g->c[0] = gf_mul(bch, g->c[0], r);
1253 /* store left-justified binary representation of g(X) */
1258 nbits = (n > 32) ? 32 : n;
1259 for (j = 0, word = 0; j < nbits; j++) {
1261 word |= 1u << (31-j);
1263 genpoly[i++] = word;
1266 bch->ecc_bits = g->deg;
1276 * init_bch - initialize a BCH encoder/decoder
1277 * @m: Galois field order, should be in the range 5-15
1278 * @t: maximum error correction capability, in bits
1279 * @prim_poly: user-provided primitive polynomial (or 0 to use default)
1282 * a newly allocated BCH control structure if successful, NULL otherwise
1284 * This initialization can take some time, as lookup tables are built for fast
1285 * encoding/decoding; make sure not to call this function from a time critical
1286 * path. Usually, init_bch() should be called on module/driver init and
1287 * free_bch() should be called to release memory on exit.
1289 * You may provide your own primitive polynomial of degree @m in argument
1290 * @prim_poly, or let init_bch() use its default polynomial.
1292 * Once init_bch() has successfully returned a pointer to a newly allocated
1293 * BCH control structure, ecc length in bytes is given by member @ecc_bytes of
1296 struct bch_control *init_bch(int m, int t, unsigned int prim_poly)
1299 unsigned int i, words;
1301 struct bch_control *bch = NULL;
1303 const int min_m = 5;
1304 const int max_m = 15;
1306 /* default primitive polynomials */
1307 static const unsigned int prim_poly_tab[] = {
1308 0x25, 0x43, 0x83, 0x11d, 0x211, 0x409, 0x805, 0x1053, 0x201b,
1312 #if defined(CONFIG_BCH_CONST_PARAMS)
1313 if ((m != (CONFIG_BCH_CONST_M)) || (t != (CONFIG_BCH_CONST_T))) {
1314 printk(KERN_ERR "bch encoder/decoder was configured to support "
1315 "parameters m=%d, t=%d only!\n",
1316 CONFIG_BCH_CONST_M, CONFIG_BCH_CONST_T);
1320 if ((m < min_m) || (m > max_m))
1322 * values of m greater than 15 are not currently supported;
1323 * supporting m > 15 would require changing table base type
1324 * (uint16_t) and a small patch in matrix transposition
1329 if ((t < 1) || (m*t >= ((1 << m)-1)))
1330 /* invalid t value */
1333 /* select a primitive polynomial for generating GF(2^m) */
1335 prim_poly = prim_poly_tab[m-min_m];
1337 bch = kzalloc(sizeof(*bch), GFP_KERNEL);
1343 bch->n = (1 << m)-1;
1344 words = DIV_ROUND_UP(m*t, 32);
1345 bch->ecc_bytes = DIV_ROUND_UP(m*t, 8);
1346 bch->a_pow_tab = bch_alloc((1+bch->n)*sizeof(*bch->a_pow_tab), &err);
1347 bch->a_log_tab = bch_alloc((1+bch->n)*sizeof(*bch->a_log_tab), &err);
1348 bch->mod8_tab = bch_alloc(words*1024*sizeof(*bch->mod8_tab), &err);
1349 bch->ecc_buf = bch_alloc(words*sizeof(*bch->ecc_buf), &err);
1350 bch->ecc_buf2 = bch_alloc(words*sizeof(*bch->ecc_buf2), &err);
1351 bch->xi_tab = bch_alloc(m*sizeof(*bch->xi_tab), &err);
1352 bch->syn = bch_alloc(2*t*sizeof(*bch->syn), &err);
1353 bch->cache = bch_alloc(2*t*sizeof(*bch->cache), &err);
1354 bch->elp = bch_alloc((t+1)*sizeof(struct gf_poly_deg1), &err);
1356 for (i = 0; i < ARRAY_SIZE(bch->poly_2t); i++)
1357 bch->poly_2t[i] = bch_alloc(GF_POLY_SZ(2*t), &err);
1362 err = build_gf_tables(bch, prim_poly);
1366 /* use generator polynomial for computing encoding tables */
1367 genpoly = compute_generator_polynomial(bch);
1368 if (genpoly == NULL)
1371 build_mod8_tables(bch, genpoly);
1374 err = build_deg2_base(bch);
1386 * free_bch - free the BCH control structure
1387 * @bch: BCH control structure to release
1389 void free_bch(struct bch_control *bch)
1394 kfree(bch->a_pow_tab);
1395 kfree(bch->a_log_tab);
1396 kfree(bch->mod8_tab);
1397 kfree(bch->ecc_buf);
1398 kfree(bch->ecc_buf2);
1404 for (i = 0; i < ARRAY_SIZE(bch->poly_2t); i++)
1405 kfree(bch->poly_2t[i]);