[patch] Re: Bug#448775: Uses too much entropy (Debian Bug #343085)

[Werner Koch]

On Tue,  8 Jan 2008 11:59, wk at gnupg.org said:

> Anyway there 3000 calls to /dev/urandom are far too many for an initial
> pool filling.  I need to check this.

Found it.  The bug was introduced with libgcrypt 1.3.1.  Here is a patch:

2008-01-08  Werner Koch  <wk at g10code.com>

	* random.c (add_randomness): Do not just increment
	POOL_FILLED_COUNTER but update it by the actual amount of data.

Index: cipher/random.c
===================================================================
--- cipher/random.c	(revision 1277)
+++ cipher/random.c	(working copy)
@@ -1115,6 +1115,7 @@
 add_randomness (const void *buffer, size_t length, enum random_origins origin)
 {
   const unsigned char *p = buffer;
+  size_t count = 0;
 
   assert (pool_is_locked);
 
@@ -1123,6 +1124,7 @@
   while (length-- )
     {
       rndpool[pool_writepos++] ^= *p++;
+      count++;
       if (pool_writepos >= POOLSIZE )
         {
           /* It is possible that we are invoked before the pool is
@@ -1132,7 +1134,9 @@
              separately.  See also the remarks about the seed file. */
           if (origin >= RANDOM_ORIGIN_SLOWPOLL && !pool_filled)
             {
-              if (++pool_filled_counter >= POOLSIZE)
+              pool_filled_counter += count;
+              count = 0;
+              if (pool_filled_counter >= POOLSIZE)
                 pool_filled = 1;
             }
           pool_writepos = 0;


Also commited to SVN.  Old and new stats:

$ LD_PRELOAD=/usr/local/lib/libgcrypt.so ./benchmark --verbose random
random       130ms    30ms
random usage: poolsize=600 mixed=972 polls=3000/200 added=4200/378400
              outmix=200 getlvl1=200/13600 getlvl2=0/0

$ ./benchmark --verbose random
random        40ms    30ms
random usage: poolsize=600 mixed=377 polls=25/200 added=1225/21400
              outmix=200 getlvl1=200/13600 getlvl2=0/0



Shalom-Salam,

   Werner



-- 
Die Gedanken sind frei.  Auschnahme regelt ein Bundeschgesetz.