X-Git-Url: https://git.openssl.org/?p=openssl.git;a=blobdiff_plain;f=crypto%2Fbn%2Fbn_div.c;h=257fe0ce01b2f19fdfb10062be775fa852fd9dc1;hp=d0b68dc246e7c6d8f24242c1e0db9fa8d6b4be75;hb=d1e7d1d96cbec34a44449e19433b5fe7d748e06a;hpb=500230ee94ef5fc9d08b3cf7c92996f4366f1584 diff --git a/crypto/bn/bn_div.c b/crypto/bn/bn_div.c index d0b68dc246..257fe0ce01 100644 --- a/crypto/bn/bn_div.c +++ b/crypto/bn/bn_div.c @@ -130,7 +130,7 @@ int BN_div(BIGNUM *dv, BIGNUM *rem, const BIGNUM *m, const BIGNUM *d, #if !defined(OPENSSL_NO_ASM) && !defined(OPENSSL_NO_INLINE_ASM) \ && !defined(PEDANTIC) && !defined(BN_DIV3W) # if defined(__GNUC__) && __GNUC__>=2 -# if defined(__i386) +# if defined(__i386) || defined (__i386__) /* * There were two reasons for implementing this template: * - GNU C generates a call to a function (__udivdi3 to be exact) @@ -150,6 +150,20 @@ int BN_div(BIGNUM *dv, BIGNUM *rem, const BIGNUM *m, const BIGNUM *d, q; \ }) # define REMAINDER_IS_ALREADY_CALCULATED +# elif defined(__x86_64) && defined(SIXTY_FOUR_BIT_LONG) + /* + * Same story here, but it's 128-bit by 64-bit division. Wow! + * + */ +# define bn_div_words(n0,n1,d0) \ + ({ asm volatile ( \ + "divq %4" \ + : "=a"(q), "=d"(rem) \ + : "a"(n1), "d"(n0), "g"(d0) \ + : "cc"); \ + q; \ + }) +# define REMAINDER_IS_ALREADY_CALCULATED # endif /* __ */ # endif /* __GNUC__ */ #endif /* OPENSSL_NO_ASM */ @@ -165,12 +179,19 @@ int BN_div(BIGNUM *dv, BIGNUM *rem, const BIGNUM *m, const BIGNUM *d, int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor, BN_CTX *ctx) { - int norm_shift,i,j,loop; + int norm_shift,i,loop; BIGNUM *tmp,wnum,*snum,*sdiv,*res; BN_ULONG *resp,*wnump; BN_ULONG d0,d1; int num_n,div_n; + if ((BN_get_flags(num, BN_FLG_CONSTTIME) != 0) || (BN_get_flags(divisor, BN_FLG_CONSTTIME) != 0)) + { + return BN_div_no_branch(dv, rm, num, divisor, ctx); + } + + bn_check_top(dv); + bn_check_top(rm); bn_check_top(num); bn_check_top(divisor); @@ -196,29 +217,28 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor, res=BN_CTX_get(ctx); else res=dv; if (sdiv == NULL || res == NULL) goto err; - tmp->neg=0; /* First we normalise the numbers */ norm_shift=BN_BITS2-((BN_num_bits(divisor))%BN_BITS2); - BN_lshift(sdiv,divisor,norm_shift); + if (!(BN_lshift(sdiv,divisor,norm_shift))) goto err; sdiv->neg=0; norm_shift+=BN_BITS2; - BN_lshift(snum,num,norm_shift); + if (!(BN_lshift(snum,num,norm_shift))) goto err; snum->neg=0; div_n=sdiv->top; num_n=snum->top; loop=num_n-div_n; - /* Lets setup a 'window' into snum * This is the part that corresponds to the current * 'area' being divided */ - BN_init(&wnum); - wnum.d= &(snum->d[loop]); - wnum.top= div_n; - wnum.dmax= snum->dmax+1; /* a bit of a lie */ + wnum.neg = 0; + wnum.d = &(snum->d[loop]); + wnum.top = div_n; + /* only needed when BN_ucmp messes up the values between top and max */ + wnum.dmax = snum->dmax - loop; /* so we don't step out of bounds */ /* Get the top 2 words of sdiv */ - /* i=sdiv->top; */ + /* div_n=sdiv->top; */ d0=sdiv->d[div_n-1]; d1=(div_n == 1)?0:sdiv->d[div_n-2]; @@ -236,19 +256,28 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor, if (BN_ucmp(&wnum,sdiv) >= 0) { - if (!BN_usub(&wnum,&wnum,sdiv)) goto err; + /* If BN_DEBUG_RAND is defined BN_ucmp changes (via + * bn_pollute) the const bignum arguments => + * clean the values between top and max again */ + bn_clear_top2max(&wnum); + bn_sub_words(wnum.d, wnum.d, sdiv->d, div_n); *resp=1; - res->d[res->top-1]=1; } else res->top--; + /* if res->top == 0 then clear the neg value otherwise decrease + * the resp pointer */ if (res->top == 0) res->neg = 0; - resp--; + else + resp--; - for (i=0; i 0x%08X\n", + n0, n1, d0, q); +#endif #endif #ifndef REMAINDER_IS_ALREADY_CALCULATED @@ -292,11 +326,18 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor, BN_ULONG t2l,t2h,ql,qh; q=bn_div_words(n0,n1,d0); +#ifdef BN_DEBUG_LEVITTE + fprintf(stderr,"DEBUG: bn_div_words(0x%08X,0x%08X,0x%08\ +X) -> 0x%08X\n", + n0, n1, d0, q); +#endif #ifndef REMAINDER_IS_ALREADY_CALCULATED rem=(n1-q*d0)&BN_MASK2; #endif -#ifdef BN_UMULT_HIGH +#if defined(BN_UMULT_LOHI) + BN_UMULT_LOHI(t2l,t2h,d1,q); +#elif defined(BN_UMULT_HIGH) t2l = d1 * q; t2h = BN_UMULT_HIGH(d1,q); #else @@ -320,27 +361,252 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor, #endif /* !BN_DIV3W */ l0=bn_mul_words(tmp->d,sdiv->d,div_n,q); - wnum.d--; wnum.top++; tmp->d[div_n]=l0; - for (j=div_n+1; j>0; j--) - if (tmp->d[j-1]) break; - tmp->top=j; + wnum.d--; + /* ingore top values of the bignums just sub the two + * BN_ULONG arrays with bn_sub_words */ + if (bn_sub_words(wnum.d, wnum.d, tmp->d, div_n+1)) + { + /* Note: As we have considered only the leading + * two BN_ULONGs in the calculation of q, sdiv * q + * might be greater than wnum (but then (q-1) * sdiv + * is less or equal than wnum) + */ + q--; + if (bn_add_words(wnum.d, wnum.d, sdiv->d, div_n)) + /* we can't have an overflow here (assuming + * that q != 0, but if q == 0 then tmp is + * zero anyway) */ + (*wnump)++; + } + /* store part of the result */ + *resp = q; + } + bn_correct_top(snum); + if (rm != NULL) + { + /* Keep a copy of the neg flag in num because if rm==num + * BN_rshift() will overwrite it. + */ + int neg = num->neg; + BN_rshift(rm,snum,norm_shift); + if (!BN_is_zero(rm)) + rm->neg = neg; + bn_check_top(rm); + } + BN_CTX_end(ctx); + return(1); +err: + bn_check_top(rm); + BN_CTX_end(ctx); + return(0); + } - j=wnum.top; - BN_sub(&wnum,&wnum,tmp); - snum->top=snum->top+wnum.top-j; +/* BN_div_no_branch is a special version of BN_div. It does not contain + * branches that may leak sensitive information. + */ +int BN_div_no_branch(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, + const BIGNUM *divisor, BN_CTX *ctx) + { + int norm_shift,i,loop; + BIGNUM *tmp,wnum,*snum,*sdiv,*res; + BN_ULONG *resp,*wnump; + BN_ULONG d0,d1; + int num_n,div_n; + + bn_check_top(dv); + bn_check_top(rm); + bn_check_top(num); + bn_check_top(divisor); + + if (BN_is_zero(divisor)) + { + BNerr(BN_F_BN_DIV,BN_R_DIV_BY_ZERO); + return(0); + } + + BN_CTX_start(ctx); + tmp=BN_CTX_get(ctx); + snum=BN_CTX_get(ctx); + sdiv=BN_CTX_get(ctx); + if (dv == NULL) + res=BN_CTX_get(ctx); + else res=dv; + if (sdiv == NULL || res == NULL) goto err; + + /* First we normalise the numbers */ + norm_shift=BN_BITS2-((BN_num_bits(divisor))%BN_BITS2); + if (!(BN_lshift(sdiv,divisor,norm_shift))) goto err; + sdiv->neg=0; + norm_shift+=BN_BITS2; + if (!(BN_lshift(snum,num,norm_shift))) goto err; + snum->neg=0; + + /* Since we don't know whether snum is larger than sdiv, + * we pad snum with enough zeroes without changing its + * value. + */ + if (snum->top <= sdiv->top+1) + { + if (bn_wexpand(snum, sdiv->top + 2) == NULL) goto err; + for (i = snum->top; i < sdiv->top + 2; i++) snum->d[i] = 0; + snum->top = sdiv->top + 2; + } + else + { + if (bn_wexpand(snum, snum->top + 1) == NULL) goto err; + snum->d[snum->top] = 0; + snum->top ++; + } + + div_n=sdiv->top; + num_n=snum->top; + loop=num_n-div_n; + /* Lets setup a 'window' into snum + * This is the part that corresponds to the current + * 'area' being divided */ + wnum.neg = 0; + wnum.d = &(snum->d[loop]); + wnum.top = div_n; + /* only needed when BN_ucmp messes up the values between top and max */ + wnum.dmax = snum->dmax - loop; /* so we don't step out of bounds */ + + /* Get the top 2 words of sdiv */ + /* div_n=sdiv->top; */ + d0=sdiv->d[div_n-1]; + d1=(div_n == 1)?0:sdiv->d[div_n-2]; + + /* pointer to the 'top' of snum */ + wnump= &(snum->d[num_n-1]); - if (wnum.neg) + /* Setup to 'res' */ + res->neg= (num->neg^divisor->neg); + if (!bn_wexpand(res,(loop+1))) goto err; + res->top=loop-1; + resp= &(res->d[loop-1]); + + /* space for temp */ + if (!bn_wexpand(tmp,(div_n+1))) goto err; + + /* if res->top == 0 then clear the neg value otherwise decrease + * the resp pointer */ + if (res->top == 0) + res->neg = 0; + else + resp--; + + for (i=0; i 0x%08X\n", + n0, n1, d0, q); +#endif +#endif + +#ifndef REMAINDER_IS_ALREADY_CALCULATED + /* + * rem doesn't have to be BN_ULLONG. The least we + * know it's less that d0, isn't it? + */ + rem=(n1-q*d0)&BN_MASK2; +#endif + t2=(BN_ULLONG)d1*q; + + for (;;) + { + if (t2 <= ((((BN_ULLONG)rem)< 0x%08X\n", + n0, n1, d0, q); +#endif +#ifndef REMAINDER_IS_ALREADY_CALCULATED + rem=(n1-q*d0)&BN_MASK2; +#endif + +#if defined(BN_UMULT_LOHI) + BN_UMULT_LOHI(t2l,t2h,d1,q); +#elif defined(BN_UMULT_HIGH) + t2l = d1 * q; + t2h = BN_UMULT_HIGH(d1,q); +#else + t2l=LBITS(d1); t2h=HBITS(d1); + ql =LBITS(q); qh =HBITS(q); + mul64(t2l,t2h,ql,qh); /* t2=(BN_ULLONG)d1*q; */ +#endif + + for (;;) + { + if ((t2h < rem) || + ((t2h == rem) && (t2l <= wnump[-2]))) + break; + q--; + rem += d0; + if (rem < d0) break; /* don't let rem overflow */ + if (t2l < d1) t2h--; t2l -= d1; + } +#endif /* !BN_LLONG */ + } +#endif /* !BN_DIV3W */ + + l0=bn_mul_words(tmp->d,sdiv->d,div_n,q); + tmp->d[div_n]=l0; + wnum.d--; + /* ingore top values of the bignums just sub the two + * BN_ULONG arrays with bn_sub_words */ + if (bn_sub_words(wnum.d, wnum.d, tmp->d, div_n+1)) { + /* Note: As we have considered only the leading + * two BN_ULONGs in the calculation of q, sdiv * q + * might be greater than wnum (but then (q-1) * sdiv + * is less or equal than wnum) + */ q--; - j=wnum.top; - BN_add(&wnum,&wnum,sdiv); - snum->top+=wnum.top-j; + if (bn_add_words(wnum.d, wnum.d, sdiv->d, div_n)) + /* we can't have an overflow here (assuming + * that q != 0, but if q == 0 then tmp is + * zero anyway) */ + (*wnump)++; } - *(resp--)=q; - wnump--; + /* store part of the result */ + *resp = q; } + bn_correct_top(snum); if (rm != NULL) { /* Keep a copy of the neg flag in num because if rm==num @@ -350,10 +616,13 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor, BN_rshift(rm,snum,norm_shift); if (!BN_is_zero(rm)) rm->neg = neg; + bn_check_top(rm); } + bn_correct_top(res); BN_CTX_end(ctx); return(1); err: + bn_check_top(rm); BN_CTX_end(ctx); return(0); }