Minor VIA Padlock engine update: eliminate -Wunused warning when *not*
[openssl.git] / crypto / engine / eng_padlock.c
index 09f10c6..95de35d 100644 (file)
 
 #include <stdio.h>
 #include <string.h>
+#include <malloc.h>
+#ifdef _MSC_VER
+# define alloca   _alloca
+# define snprintf _snprintf
+#endif
 
 #include <openssl/crypto.h>
 #include <openssl/dso.h>
 #if !defined(I386_ONLY) && !defined(OPENSSL_NO_INLINE_ASM)
 # if defined(__i386__) || defined(__i386) || defined(_M_IX86)
 #  define COMPILE_HW_PADLOCK
+static ENGINE *ENGINE_padlock (void);
 # endif
 #endif
 
-static ENGINE *ENGINE_padlock (void);
-
 void ENGINE_load_padlock (void)
 {
 /* On non-x86 CPUs it just returns. */
@@ -246,13 +250,21 @@ struct padlock_cipher_data
        AES_KEY ks;             /* Encryption key */
 };
 
+/*
+ * Essentially this variable belongs in thread local storage.
+ * Having this variable global on the other hand can only cause
+ * few bogus key reloads [if any at all on single-CPU system],
+ * so we accept the penatly...
+ */
+static volatile struct padlock_cipher_data *padlock_saved_context;
+
 /*
  * =======================================================
  * Inline assembler section(s).
  * =======================================================
  * Order of arguments is chosen to facilitate Windows port
  * using __fastcall calling convention. If you wish to add
- * more routines, keep in mind that in __fastcall first
+ * more routines, keep in mind that first __fastcall
  * argument is passed in %ecx and second - in %edx.
  * =======================================================
  */
@@ -362,16 +374,14 @@ padlock_reload_key(void)
  * This is heuristic key context tracing. At first one
  * believes that one should use atomic swap instructions,
  * but it's not actually necessary. Point is that if
- * saved_cdata was changed by another thread after we've
- * read it and before we compare it with cdata, our key
- * *shall* be reloaded upon thread context switch and
- * we are therefore set in either case...
+ * padlock_saved_context was changed by another thread
+ * after we've read it and before we compare it with cdata,
+ * our key *shall* be reloaded upon thread context switch
+ * and we are therefore set in either case...
  */
 static inline void
 padlock_verify_context(struct padlock_cipher_data *cdata)
 {
-       static struct padlock_cipher_data *saved_cdata;
-
        asm volatile (
        "pushfl\n"
 "      bt      $30,(%%esp)\n"
@@ -382,7 +392,8 @@ padlock_verify_context(struct padlock_cipher_data *cdata)
 "      popfl\n"
 "      sub     $4,%%esp\n"
 "1:    add     $4,%%esp"
-        :"+m"(saved_cdata) : "r"(saved_cdata), "r"(cdata) : "cc");
+       :"+m"(padlock_saved_context)
+       : "r"(padlock_saved_context), "r"(cdata) : "cc");
 }
 
 /* Template for padlock_xcrypt_* modes */
@@ -427,6 +438,29 @@ padlock_xstore(void *addr, unsigned int edx_in)
        return eax_out;
 }
 
+/* Why not inline 'rep movsd'? I failed to find information on what
+ * value in Direction Flag one can expect and consequently have to
+ * apply "better-safe-than-sorry" approach and assume "undefined."
+ * I could explicitly clear it and restore the original value upon
+ * return from padlock_aes_cipher, but it's presumably too much
+ * trouble for too little gain...
+ *
+ * In case you wonder 'rep xcrypt*' instructions above are *not*
+ * affected by the Direction Flag and pointers advance toward
+ * larger addresses unconditionally.
+ */ 
+static inline unsigned char *
+padlock_memcpy(void *dst,const void *src,size_t n)
+{
+       long       *d=dst;
+       const long *s=src;
+
+       n /= sizeof(*d);
+       do { *d++ = *s++; } while (--n);
+
+       return dst;
+}
+
 #elif defined(_MSC_VER)
 /*
  * Unlike GCC these are real functions. In order to minimize impact
@@ -450,8 +484,8 @@ static void * __fastcall            \
        name (size_t cnt, void *cdata,  \
        void *outp, const void *inp)    \
 {      _asm    mov     eax,edx         \
-       _asm    lea     ebx,[eax+16]    \
-       _asm    lea     edx,[eax+32]    \
+       _asm    lea     edx,[eax+16]    \
+       _asm    lea     ebx,[eax+32]    \
        _asm    mov     edi,outp        \
        _asm    mov     esi,inp         \
        REP_XCRYPT(code)                \
@@ -474,22 +508,20 @@ padlock_reload_key(void)
 
 static void __fastcall
 padlock_verify_context(void *cdata)
-{      static void *saved_cdata;
-
-       _asm    {
+{      _asm    {
                pushfd
                bt      DWORD PTR[esp],30
                jnc     skip
-               cmp     ecx,saved_cdata
+               cmp     ecx,padlock_saved_context
                je      skip
-               mov     saved_cdata,ecx
+               mov     padlock_saved_context,ecx
                popfd
                sub     esp,4
        skip:   add     esp,4
                }
 }
 
-sttic int __fastcall
+static int
 padlock_available(void)
 {      _asm    {
                pushfd
@@ -546,14 +578,18 @@ padlock_bswapl(void *key)
                mov     esi,ecx
                mov     edi,ecx
                mov     ecx,60
-       up:
-               lodsd
+       up:     lodsd
                bswap   eax
                stosd
                loop    up
                popfd
                }
 }
+
+/* MS actually specifies status of Direction Flag and compiler even
+ * manages to compile following as 'rep movsd' all by itself...
+ */
+#define padlock_memcpy(o,i,n) ((unsigned char *)memcpy((o),(i),(n)&~3U))
 #endif
 
 /* ===== AES encryption/decryption ===== */
@@ -591,13 +627,17 @@ static int padlock_cipher_nids[] = {
 
        NID_aes_192_ecb,
        NID_aes_192_cbc,
-//     NID_aes_192_cfb,        /* FIXME: AES192/256 CFB/OFB don't work. */
-//     NID_aes_192_ofb,
+#if 0
+       NID_aes_192_cfb,        /* FIXME: AES192/256 CFB/OFB don't work. */
+       NID_aes_192_ofb,
+#endif
 
        NID_aes_256_ecb,
        NID_aes_256_cbc,
-//     NID_aes_256_cfb,
-//     NID_aes_256_ofb,
+#if 0
+       NID_aes_256_cfb,
+       NID_aes_256_ofb,
+#endif
 };
 static int padlock_cipher_nids_num = (sizeof(padlock_cipher_nids)/
                                      sizeof(padlock_cipher_nids[0]));
@@ -608,7 +648,10 @@ static int padlock_aes_init_key(EVP_CIPHER_CTX *ctx, const unsigned char *key,
 static int padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
                              const unsigned char *in, unsigned int nbytes);
 
-#define ALIGNED_CIPHER_DATA(ctx) ((struct padlock_cipher_data *)(ctx->cipher_data + ((0x10 - ((size_t)(ctx->cipher_data) & 0x0F)) & 0x0F)))
+#define NEAREST_ALIGNED(ptr) ( (char *)(ptr) +         \
+       ( (0x10 - ((size_t)(ptr) & 0x0F)) & 0x0F )      )
+#define ALIGNED_CIPHER_DATA(ctx) ((struct padlock_cipher_data *)\
+       NEAREST_ALIGNED(ctx->cipher_data))
 
 /* Declaring so many ciphers by hand would be a pain.
    Instead introduce a bit of preprocessor magic :-) */
@@ -733,6 +776,10 @@ padlock_aes_init_key (EVP_CIPHER_CTX *ctx, const unsigned char *key,
                case 256:
                        /* Generate an extended AES key in software.
                           Needed for AES192/AES256 */
+                       /* Well, the above applies to Stepping 8 CPUs
+                          and is listed as hardware errata. They most
+                          likely will fix it at some point and then
+                          a check for stepping would be due here. */
                        if (enc)
                                AES_set_encrypt_key(key, key_len, &cdata->ks);
                        else
@@ -822,21 +869,21 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
 {
        struct padlock_cipher_data *cdata;
        const  void *inp;
-       void  *out, *iv;
+       char  *out;
+       void  *iv;
        int    inp_misaligned, out_misaligned, realign_in_loop;
-       size_t chunk, allocated;
+       size_t chunk, allocated=0;
 
        if (nbytes == 0)
                return 1;
        if (nbytes % AES_BLOCK_SIZE)
                return 0; /* are we expected to do tail processing? */
 
-#if 0
-       /* There is more work to support CPUs that don't require alignment.
-          Therefore disabled completely for now... */
+       /* VIA promises CPUs that won't require alignment in the future.
+          For now padlock_aes_align_required is initialized to 1 and
+          the condition is never met... */
        if (!padlock_aes_align_required)
                return padlock_aes_cipher_omnivorous(ctx, out_arg, in_arg, nbytes);
-#endif
 
        inp_misaligned = (((size_t)in_arg) & 0x0F);
        out_misaligned = (((size_t)out_arg) & 0x0F);
@@ -858,12 +905,8 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
        if (out_misaligned) {
                /* optmize for small input */
                allocated = (chunk<nbytes?PADLOCK_CHUNK:nbytes);
-#ifdef _MSC_VER
-               out  = _alloca(0x10 + allocated);
-#else
-               out  =  alloca(0x10 + allocated);
-#endif
-               out += (0x10 - ((size_t)out & 0x0F)) & 0x0F;
+               out = alloca(0x10 + allocated);
+               out = NEAREST_ALIGNED(out);
        }
        else
                out = out_arg;
@@ -875,7 +918,7 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
        case EVP_CIPH_ECB_MODE:
                do      {
                        if (inp_misaligned)
-                               inp = memcpy(out, in_arg, chunk);
+                               inp = padlock_memcpy(out, in_arg, chunk);
                        else
                                inp = in_arg;
                        in_arg += chunk;
@@ -883,7 +926,7 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
                        padlock_xcrypt_ecb(chunk/AES_BLOCK_SIZE, cdata, out, inp);
 
                        if (out_misaligned)
-                               out_arg = memcpy(out_arg, out, chunk) + chunk;
+                               out_arg = padlock_memcpy(out_arg, out, chunk) + chunk;
                        else
                                out     = out_arg+=chunk;
 
@@ -901,7 +944,7 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
                        chunk = PADLOCK_CHUNK;
                cbc_shortcut: /* optimize for small input */
                        if (inp_misaligned)
-                               inp = memcpy(out, in_arg, chunk);
+                               inp = padlock_memcpy(out, in_arg, chunk);
                        else
                                inp = in_arg;
                        in_arg += chunk;
@@ -909,7 +952,7 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
                        iv = padlock_xcrypt_cbc(chunk/AES_BLOCK_SIZE, cdata, out, inp);
 
                        if (out_misaligned)
-                               out_arg = memcpy(out_arg, out, chunk) + chunk;
+                               out_arg = padlock_memcpy(out_arg, out, chunk) + chunk;
                        else
                                out     = out_arg+=chunk;
 
@@ -926,7 +969,7 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
                        chunk = PADLOCK_CHUNK;
                cfb_shortcut: /* optimize for small input */
                        if (inp_misaligned)
-                               inp = memcpy(out, in_arg, chunk);
+                               inp = padlock_memcpy(out, in_arg, chunk);
                        else
                                inp = in_arg;
                        in_arg += chunk;
@@ -934,7 +977,7 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
                        iv = padlock_xcrypt_cfb(chunk/AES_BLOCK_SIZE, cdata, out, inp);
 
                        if (out_misaligned)
-                               out_arg = memcpy(out_arg, out, chunk) + chunk;
+                               out_arg = padlock_memcpy(out_arg, out, chunk) + chunk;
                        else
                                out     = out_arg+=chunk;
 
@@ -946,7 +989,7 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
                memcpy(cdata->iv, ctx->iv, AES_BLOCK_SIZE);
                do      {
                        if (inp_misaligned)
-                               inp = memcpy(out, in_arg, chunk);
+                               inp = padlock_memcpy(out, in_arg, chunk);
                        else
                                inp = in_arg;
                        in_arg += chunk;
@@ -954,7 +997,7 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
                        padlock_xcrypt_ofb(chunk/AES_BLOCK_SIZE, cdata, out, inp);
 
                        if (out_misaligned)
-                               out_arg = memcpy(out_arg, out, chunk) + chunk;
+                               out_arg = padlock_memcpy(out_arg, out, chunk) + chunk;
                        else
                                out     = out_arg+=chunk;
 
@@ -970,7 +1013,7 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
 
        /* Clean the realign buffer if it was used */
        if (out_misaligned) {
-               volatile unsigned long *p=out;
+               volatile unsigned long *p=(void *)out;
                size_t   n = allocated/sizeof(*p);
                while (n--) *p++=0;
        }