Decrease CPU usage caused by tls:send with large data.

Sending one large chunk of data with tls:send eats lots of
CPU power and blocks whole Erlang emulator. This is caused by the
fact that encrypted output is read from memory BIO in 1k chunks.
Memory BIO, after reading data, shifts the remaining part.
If large chunks of data (few MB) is sent and then read in 1k
chunks, then a _lot_ of shifting is performed eating CPU.

The solution is to simply allocate binary of the needed size
(amount of data in memory BIO can be retrieved with
BIO_ctrl_pending) and then issue only one read that reads the
whole data.

diff --git a/src/tls/tls_drv.c b/src/tls/tls_drv.c
index 36cd8d9b926836bcae1550cc7710c05c8cef4619..bdb5446f2235acde953d022cbe70f84d2a252c14 100644 (file)
--- a/src/tls/tls_drv.c
+++ b/src/tls/tls_drv.c
@@ -407,22 +407,12 @@ static int tls_drv_control(ErlDrvData handle,
         break;
       case GET_ENCRYPTED_OUTPUT:
         die_unless(d->ssl, "SSL not initialized");
-        size = BUF_SIZE + 1;
-        rlen = 1;
+        size = BIO_ctrl_pending(d->bio_write) + 1;
         b = driver_alloc_binary(size);
         b->orig_bytes[0] = 0;
-        while ((res = BIO_read(d->bio_write,
-                               b->orig_bytes + rlen, BUF_SIZE)) > 0)
-        {
-           //printf("%d bytes of encrypted data read from state machine\r\n", res);
-
-           rlen += res;
-           size += BUF_SIZE;
-           b = driver_realloc_binary(b, size);
-        }
-        b = driver_realloc_binary(b, rlen);
+        BIO_read(d->bio_write, b->orig_bytes + 1, size - 1);
         *rbuf = (char *)b;
-        return rlen;
+        return size;
       case GET_DECRYPTED_INPUT:
         if (!SSL_is_init_finished(d->ssl))
         {