[or-cvs] r12002: oprofile was telling me that a fair bit of our time in opens (in tor/trunk: . src/common src/or)

nickm at seul.org nickm at seul.org
Wed Oct 17 19:23:57 UTC 2007


Author: nickm
Date: 2007-10-17 15:23:56 -0400 (Wed, 17 Oct 2007)
New Revision: 12002

Modified:
   tor/trunk/
   tor/trunk/ChangeLog
   tor/trunk/src/common/crypto.c
   tor/trunk/src/or/test.c
Log:
 r15882 at catbus:  nickm | 2007-10-17 15:23:05 -0400
 oprofile was telling me that a fair bit of our time in openssl was spent in base64_decode, so replace base64_decode with an all-at-once fairly optimized implementation.  For decoding keys and digests, it seems 3-3.5x faster than calling out to openssl.  (Yes, I wrote it from scratch.)



Property changes on: tor/trunk
___________________________________________________________________
 svk:merge ticket from /tor/trunk [r15882] on 8246c3cf-6607-4228-993b-4d95d33730f1

Modified: tor/trunk/ChangeLog
===================================================================
--- tor/trunk/ChangeLog	2007-10-17 19:23:52 UTC (rev 12001)
+++ tor/trunk/ChangeLog	2007-10-17 19:23:56 UTC (rev 12002)
@@ -49,6 +49,12 @@
     - Make base32_decode() accept upper-case letters.  Bugfix on
       0.2.0.7-alpha.
 
+  o Minor bugfixes (performance):
+    - Base64 decoding was actually showing up on our profile when parsing
+      the initial descriptor file; switch to an in-process all-at-once
+      implementation that's about 3.5x times faster than calling out to
+      OpenSSL.
+
   o Code simplifications and refactoring:
     - Remove support for the old bw_accounting file: we've been storing
       bandwidth accounting information in the state file since 0.1.2.5-alpha.

Modified: tor/trunk/src/common/crypto.c
===================================================================
--- tor/trunk/src/common/crypto.c	2007-10-17 19:23:52 UTC (rev 12001)
+++ tor/trunk/src/common/crypto.c	2007-10-17 19:23:56 UTC (rev 12002)
@@ -1768,6 +1768,8 @@
 int
 base64_encode(char *dest, size_t destlen, const char *src, size_t srclen)
 {
+  /* XXXX we might want to rewrite this along the lines of base64_decode, if
+   * it ever shows up in the profile. */
   EVP_ENCODE_CTX ctx;
   int len, ret;
 
@@ -1787,18 +1789,48 @@
   return ret;
 }
 
+#define X 255
+#define SP 64
+#define PAD 65
+/** Internal table mapping byte values to what they represent in base64.
+ * Numbers 0..63 are 6-bit integers.  SPs are spaces, and should be
+ * skipped.  Xs are invalid and must not appear in base64. PAD indicates
+ * end-of-string. */
+static const uint8_t base64_decode_table[256] = {
+  X, X, X, X, X, X, X, X, X, SP, SP, SP, X, SP, X, X, /* */
+  X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X,
+  SP, X, X, X, X, X, X, X, X, X, X, 62, X, X, X, 63,
+  52, 53, 54, 55, 56, 57, 58, 59, 60, 61, X, X, X, PAD, X, X,
+  X, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
+  15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, X, X, X, X, X,
+  X, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40,
+  41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, X, X, X, X, X,
+  X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X,
+  X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X,
+  X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X,
+  X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X,
+  X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X,
+  X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X,
+  X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X,
+  X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X,
+};
+
 /** Base-64 decode <b>srclen</b> bytes of data from <b>src</b>.  Write
  * the result into <b>dest</b>, if it will fit within <b>destlen</b>
  * bytes.  Return the number of bytes written on success; -1 if
  * destlen is too short, or other failure.
  *
- * NOTE: destlen should be a little longer than the amount of data it
- * will contain, since we check for sufficient space conservatively.
- * Here, "a little" is around 64-ish bytes.
+ * NOTE 1: destlen is checked conservatively, as though srclen contained no
+ * spaces or padding.
+ *
+ * NOTE 2: This implementation does not check for the correct number of
+ * padding "=" characters at the end of the string, and does not check
+ * for internal padding characters.
  */
 int
 base64_decode(char *dest, size_t destlen, const char *src, size_t srclen)
 {
+#ifdef USE_OPENSSL_BASE64
   EVP_ENCODE_CTX ctx;
   int len, ret;
   /* 64 bytes of input -> *up to* 48 bytes of output.
@@ -1815,7 +1847,80 @@
   EVP_DecodeFinal(&ctx, (unsigned char*)dest, &ret);
   ret += len;
   return ret;
+#else
+  #define ACC32
+  const char *eos = src+srclen;
+  uint32_t n=0;
+  int n_idx=0;
+  char *dest_orig = dest;
+
+  /* Max number of bits == srclen*6.
+   * Number of bytes required to hold all bits == (srclen*6)/8.
+   * Yes, we want to round down: anything that hangs over the end of a
+   * byte is padding. */
+  if (destlen < (srclen*3)/4)
+    return -1;
+  if (destlen > SIZE_T_CEILING)
+    return -1;
+
+  /* Iterate over all the bytes in src.  Each one will add 0 or 6 bits to the
+   * value we're decoding.  Accumulate bits in <b>n</b>, and whenever we have
+   * 24 bits, batch them into 3 bytes and flush those bytes to dest.
+   */
+  for ( ; src < eos; ++src) {
+    unsigned char c = (unsigned char) *src;
+    uint8_t v = base64_decode_table[c];
+    switch (v) {
+      case X:
+        /* This character isn't allowed in base64. */
+        return -1;
+      case SP:
+        /* This character is whitespace, and has no effect. */
+        continue;
+      case PAD:
+        /* We've hit an = character: the data is over. */
+        goto end_of_loop;
+      default:
+        /* We have an actual 6-bit value.  Append it to the bits in n. */
+        n = (n<<6) | v;
+        if ((++n_idx) == 4) {
+          /* We've accumulated 24 bits in n. Flush them. */
+          *dest++ = (n>>16);
+          *dest++ = (n>>8) & 0xff;
+          *dest++ = (n) & 0xff;
+          n_idx = 0;
+          n = 0;
+        }
+    }
+  }
+ end_of_loop:
+  /* If we have leftover bits, we need to cope. */
+  switch (n_idx) {
+    case 0:
+    default:
+      /* No leftover bits.  We win. */
+      break;
+    case 1:
+      /* 6 leftover bits. That's invalid; we can't form a byte out of that. */
+      return -1;
+    case 2:
+      /* 12 leftover bits: The last 4 are padding and the first 8 are data. */
+      *dest++ = n >> 4;
+      break;
+    case 3:
+      /* 18 leftover bits: The last 2 are padding and the first 16 are data. */
+      *dest++ = n >> 10;
+      *dest++ = n >> 2;
+  }
+
+  tor_assert((dest-dest_orig) <= (ssize_t)destlen);
+
+  return dest-dest_orig;
+#endif
 }
+#undef X
+#undef SP
+#undef NIL
 
 /** Base-64 encode DIGEST_LINE bytes from <b>digest</b>, remove the trailing =
  * and newline characters, and store the nul-terminated result in the first
@@ -1836,6 +1941,7 @@
 int
 digest_from_base64(char *digest, const char *d64)
 {
+#ifdef USE_OPENSSL_BASE64
   char buf_in[BASE64_DIGEST_LEN+3];
   char buf[256];
   if (strlen(d64) != BASE64_DIGEST_LEN)
@@ -1846,6 +1952,12 @@
     return -1;
   memcpy(digest, buf, DIGEST_LEN);
   return 0;
+#else
+  if (base64_decode(digest, DIGEST_LEN, d64, strlen(d64)) == DIGEST_LEN)
+    return 0;
+  else
+    return -1;
+#endif
 }
 
 /** Implements base32 encoding as in rfc3548.  Limitation: Requires
@@ -1878,6 +1990,8 @@
 int
 base32_decode(char *dest, size_t destlen, const char *src, size_t srclen)
 {
+  /* XXXX we might want to rewrite this along the lines of base64_decode, if
+   * it ever shows up in the profile. */
   unsigned int nbits, i, j, bit;
   char *tmp;
   nbits = srclen * 5;

Modified: tor/trunk/src/or/test.c
===================================================================
--- tor/trunk/src/or/test.c	2007-10-17 19:23:52 UTC (rev 12001)
+++ tor/trunk/src/or/test.c	2007-10-17 19:23:56 UTC (rev 12002)
@@ -363,7 +363,7 @@
   crypto_cipher_env_t *env1, *env2;
   crypto_pk_env_t *pk1, *pk2;
   char *data1, *data2, *data3, *cp;
-  int i, j, p, len;
+  int i, j, p, len, idx;
   size_t size;
 
   data1 = tor_malloc(1024);
@@ -528,13 +528,21 @@
   crypto_free_pk_env(pk2);
 
   /* Base64 tests */
+  memset(data1, 6, 1024);
+  for (idx = 0; idx < 10; ++idx) {
+    i = base64_encode(data2, 1024, data1, idx);
+    j = base64_decode(data3, 1024, data2, i);
+    test_eq(j,idx);
+    test_memeq(data3, data1, idx);
+  }
+
   strlcpy(data1, "Test string that contains 35 chars.", 1024);
   strlcat(data1, " 2nd string that contains 35 chars.", 1024);
 
   i = base64_encode(data2, 1024, data1, 71);
   j = base64_decode(data3, 1024, data2, i);
+  test_eq(j, 71);
   test_streq(data3, data1);
-  test_eq(j, 71);
   test_assert(data2[i] == '\0');
 
   crypto_rand(data1, DIGEST_LEN);
@@ -543,7 +551,7 @@
   test_eq(BASE64_DIGEST_LEN, strlen(data2));
   test_eq(100, data2[BASE64_DIGEST_LEN+2]);
   memset(data3, 99, 1024);
-  digest_from_base64(data3, data2);
+  test_eq(digest_from_base64(data3, data2), 0);
   test_memeq(data1, data3, DIGEST_LEN);
   test_eq(99, data3[DIGEST_LEN+1]);
 
@@ -2639,6 +2647,7 @@
 
   v3_text = format_networkstatus_vote(sign_skey_3, vote);
   test_assert(v3_text);
+
   v3 = networkstatus_parse_vote_from_string(v3_text, NULL, 1);
   test_assert(v3);
 



More information about the tor-commits mailing list