Optimize payload exhaustion

https://github.com/warmcat/libwebsockets/pull/462

AG refactor and do loop unrolling

diff --git a/lib/parsers.c b/lib/parsers.c
index 360e261..963804e 100644 (file)
--- a/lib/parsers.c
+++ b/lib/parsers.c
@@ -1415,3 +1415,63 @@ lws_remaining_packet_payload(struct lws *wsi)
 {
        return wsi->u.ws.rx_packet_length;
 }
+
+/* Once we reach LWS_RXPS_PAYLOAD_UNTIL_LENGTH_EXHAUSTED, we know how much
+ * to expect in that state and can deal with it in bulk more efficiently.
+ */
+
+void
+lws_payload_until_length_exhausted(struct lws *wsi, unsigned char **buf,
+                                  size_t *len)
+{
+       unsigned char *buffer = *buf, mask[4];
+       int buffer_size, avail, n;
+       char *rx_ubuf;
+
+       if (wsi->protocol->rx_buffer_size)
+               buffer_size = wsi->protocol->rx_buffer_size;
+       else
+               buffer_size = LWS_MAX_SOCKET_IO_BUF;
+       avail = buffer_size - wsi->u.ws.rx_ubuf_head;
+
+       /* do not consume more than we should */
+       if (avail > wsi->u.ws.rx_packet_length)
+               avail = wsi->u.ws.rx_packet_length;
+
+       /* do not consume more than what is in the buffer */
+       if (avail > *len)
+               avail = *len;
+
+       /* we want to leave 1 byte for the parser to handle properly */
+       if (avail <= 1)
+               return;
+
+       avail--;
+       rx_ubuf = wsi->u.ws.rx_ubuf + LWS_PRE + wsi->u.ws.rx_ubuf_head;
+       if (wsi->u.ws.all_zero_nonce)
+               memcpy(rx_ubuf, buffer, avail);
+       else {
+
+               for (n = 0; n < 4; n++)
+                       mask[n] = wsi->u.ws.mask[(wsi->u.ws.mask_idx + n) & 3];
+
+               /* deal with 4-byte chunks using unwrapped loop */
+               n = avail >> 2;
+               while (n--) {
+                       *(rx_ubuf++) = *(buffer++) ^ mask[0];
+                       *(rx_ubuf++) = *(buffer++) ^ mask[1];
+                       *(rx_ubuf++) = *(buffer++) ^ mask[2];
+                       *(rx_ubuf++) = *(buffer++) ^ mask[3];
+               }
+               /* and the remaining bytes bytewise */
+               for (n = 0; n < (avail & 3); n++)
+                       *(rx_ubuf++) = *(buffer++) ^ mask[n];
+
+               wsi->u.ws.mask_idx = (wsi->u.ws.mask_idx + avail) & 3;
+       }
+
+       (*buf) += avail;
+       wsi->u.ws.rx_ubuf_head += avail;
+       wsi->u.ws.rx_packet_length -= avail;
+       *len -= avail;
+}

diff --git a/lib/private-libwebsockets.h b/lib/private-libwebsockets.h
index 4c10b68..3237798 100644 (file)
--- a/lib/private-libwebsockets.h
+++ b/lib/private-libwebsockets.h
@@ -1217,6 +1217,9 @@ lws_client_interpret_server_handshake(struct lws *wsi);
 LWS_EXTERN int LWS_WARN_UNUSED_RESULT
 lws_rx_sm(struct lws *wsi, unsigned char c);
 
+LWS_EXTERN void
+lws_payload_until_length_exhausted(struct lws *wsi, unsigned char **buf, size_t *len);
+
 LWS_EXTERN int LWS_WARN_UNUSED_RESULT
 lws_issue_raw_ext_access(struct lws *wsi, unsigned char *buf, size_t len);
 

diff --git a/lib/server.c b/lib/server.c
index 5ab695e..0b6e4c1 100644 (file)
--- a/lib/server.c
+++ b/lib/server.c
@@ -1174,6 +1174,11 @@ lws_interpret_incoming_packet(struct lws *wsi, unsigned char **buf, size_t len)
                if (wsi->rxflow_buffer)
                        wsi->rxflow_pos++;
 
+               /* consume payload bytes efficiently */
+               if (wsi->lws_rx_parse_state ==
+                   LWS_RXPS_PAYLOAD_UNTIL_LENGTH_EXHAUSTED)
+                       lws_payload_until_length_exhausted(wsi, buf, &len);
+
                /* process the byte */
                m = lws_rx_sm(wsi, *(*buf)++);
                if (m < 0)