diff --git a/lib/curl_setup.h b/lib/curl_setup.h
index da26e48cb..d78873fe5 100644
--- a/lib/curl_setup.h
+++ b/lib/curl_setup.h
@@ -722,4 +722,19 @@ endings either CRLF or LF so 't' is appropriate.
 #define FOPEN_WRITETEXT "w"
 #endif
 
+/* WinSock destroys recv() buffer when send() failed.
+ * Enabled automatically for Windows and for Cygwin as Cygwin sockets are
+ * wrappers for WinSock sockets. https://github.com/curl/curl/issues/657
+ * Define DONT_USE_RECV_BEFORE_SEND_WORKAROUND to force disable workaround.
+ */
+#if !defined(DONT_USE_RECV_BEFORE_SEND_WORKAROUND)
+#  if defined(WIN32) || defined(__CYGWIN__)
+#    define USE_RECV_BEFORE_SEND_WORKAROUND
+#  endif
+#else  /* DONT_USE_RECV_BEFORE_SEND_WORKAROUNDS */
+#  ifdef USE_RECV_BEFORE_SEND_WORKAROUND
+#    undef USE_RECV_BEFORE_SEND_WORKAROUND
+#  endif
+#endif /* DONT_USE_RECV_BEFORE_SEND_WORKAROUNDS */
+
 #endif /* HEADER_CURL_SETUP_H */
diff --git a/lib/sendf.c b/lib/sendf.c
index a75c5c743..23bcfa2c9 100644
--- a/lib/sendf.c
+++ b/lib/sendf.c
@@ -33,6 +33,7 @@
 #include "non-ascii.h"
 #include "curl_printf.h"
 #include "strerror.h"
+#include "select.h"
 
 /* The last #include files should be: */
 #include "curl_memory.h"
@@ -120,6 +121,90 @@ static size_t convert_lineends(struct SessionHandle *data,
 }
 #endif /* CURL_DO_LINEEND_CONV */
 
+#ifdef USE_RECV_BEFORE_SEND_WORKAROUND
+static void pre_receive_plain(struct connectdata *conn, int num)
+{
+  const curl_socket_t sockfd = conn->sock[num];
+  struct postponed_data * const psnd = &(conn->postponed[num]);
+  size_t bytestorecv = psnd->allocated_size - psnd->recv_size;
+  /* WinSock will destroy unread received data if send() is
+     failed.
+     To avoid lossage of received data, recv() must be
+     performed before every send() if any incoming data is
+     available. However, skip this, if buffer is already full. */
+  if((conn->handler->protocol&PROTO_FAMILY_HTTP) != 0 &&
+     conn->recv[num] == Curl_recv_plain &&
+     (!psnd->buffer || bytestorecv)) {
+    const int readymask = Curl_socket_check(sockfd, CURL_SOCKET_BAD,
+                                            CURL_SOCKET_BAD, 0);
+    if(readymask != -1 && (readymask & CURL_CSELECT_IN) != 0) {
+      /* Have some incoming data */
+      if(!psnd->buffer) {
+        /* Use buffer double default size for intermediate buffer */
+        psnd->allocated_size = 2 * BUFSIZE;
+        psnd->buffer = malloc(psnd->allocated_size);
+        psnd->recv_size = 0;
+        psnd->recv_processed = 0;
+#ifdef DEBUGBUILD
+        psnd->bindsock = sockfd; /* Used only for DEBUGASSERT */
+#endif /* DEBUGBUILD */
+        bytestorecv = psnd->allocated_size;
+      }
+      if(psnd->buffer) {
+        ssize_t recvedbytes;
+        DEBUGASSERT(psnd->bindsock == sockfd);
+        recvedbytes = sread(sockfd, psnd->buffer + psnd->recv_size,
+                            bytestorecv);
+        if(recvedbytes > 0)
+          psnd->recv_size += recvedbytes;
+      }
+      else
+        psnd->allocated_size = 0;
+    }
+  }
+}
+
+static ssize_t get_pre_recved(struct connectdata *conn, int num, char *buf,
+                              size_t len)
+{
+  struct postponed_data * const psnd = &(conn->postponed[num]);
+  size_t copysize;
+  if(!psnd->buffer)
+    return 0;
+
+  DEBUGASSERT(psnd->allocated_size > 0);
+  DEBUGASSERT(psnd->recv_size <= psnd->allocated_size);
+  DEBUGASSERT(psnd->recv_processed <= psnd->recv_size);
+  /* Check and process data that already received and storied in internal
+     intermediate buffer */
+  if(psnd->recv_size > psnd->recv_processed) {
+    DEBUGASSERT(psnd->bindsock == conn->sock[num]);
+    copysize = CURLMIN(len, psnd->recv_size - psnd->recv_processed);
+    memcpy(buf, psnd->buffer + psnd->recv_processed, copysize);
+    psnd->recv_processed += copysize;
+  }
+  else
+    copysize = 0; /* buffer was allocated, but nothing was received */
+
+  /* Free intermediate buffer if it has no unprocessed data */
+  if(psnd->recv_processed == psnd->recv_size) {
+    free(psnd->buffer);
+    psnd->buffer = NULL;
+    psnd->allocated_size = 0;
+    psnd->recv_size = 0;
+    psnd->recv_processed = 0;
+#ifdef DEBUGBUILD
+    psnd->bindsock = CURL_SOCKET_BAD;
+#endif /* DEBUGBUILD */
+  }
+  return (ssize_t)copysize;
+}
+#else  /* ! USE_RECV_BEFORE_SEND_WORKAROUND */
+/* Use "do-nothing" macros instead of functions when workaround not used */
+#define pre_receive_plain(c,n) do {} WHILE_FALSE
+#define get_pre_recved(c,n,b,l) 0
+#endif /* ! USE_RECV_BEFORE_SEND_WORKAROUND */
+
 /* Curl_infof() is for info message along the way */
 
 void Curl_infof(struct SessionHandle *data, const char *fmt, ...)
@@ -255,6 +340,12 @@ ssize_t Curl_send_plain(struct connectdata *conn, int num,
 {
   curl_socket_t sockfd = conn->sock[num];
   ssize_t bytes_written;
+  /* WinSock will destroy unread received data if send() is
+     failed.
+     To avoid lossage of received data, recv() must be
+     performed before every send() if any incoming data is
+     available. */
+  pre_receive_plain(conn, num);
 
 #ifdef MSG_FASTOPEN /* Linux */
   if(conn->bits.tcp_fastopen) {
@@ -322,7 +413,16 @@ ssize_t Curl_recv_plain(struct connectdata *conn, int num, char *buf,
                         size_t len, CURLcode *code)
 {
   curl_socket_t sockfd = conn->sock[num];
-  ssize_t nread = sread(sockfd, buf, len);
+  ssize_t nread;
+  /* Check and return data that already received and storied in internal
+     intermediate buffer */
+  nread = get_pre_recved(conn, num, buf, len);
+  if(nread > 0) {
+    *code = CURLE_OK;
+    return nread;
+  }
+
+  nread = sread(sockfd, buf, len);
 
   *code = CURLE_OK;
   if(-1 == nread) {
diff --git a/lib/url.c b/lib/url.c
index f1028f34c..184e27049 100644
--- a/lib/url.c
+++ b/lib/url.c
@@ -2694,6 +2694,43 @@ CURLcode Curl_setopt(struct SessionHandle *data, CURLoption option,
   return result;
 }
 
+#ifdef USE_RECV_BEFORE_SEND_WORKAROUND
+static void conn_reset_postponed_data(struct connectdata *conn, int num)
+{
+  struct postponed_data * const psnd = &(conn->postponed[num]);
+  if(psnd->buffer) {
+    DEBUGASSERT(psnd->allocated_size > 0);
+    DEBUGASSERT(psnd->recv_size <= psnd->allocated_size);
+    DEBUGASSERT(psnd->recv_processed < psnd->recv_size);
+    DEBUGASSERT(psnd->bindsock != CURL_SOCKET_BAD);
+    free(psnd->buffer);
+    psnd->buffer = NULL;
+    psnd->allocated_size = 0;
+    psnd->recv_size = 0;
+    psnd->recv_processed = 0;
+#ifdef DEBUGBUILD
+    psnd->bindsock = CURL_SOCKET_BAD; /* used only for DEBUGASSERT */
+#endif /* DEBUGBUILD */
+  }
+  else {
+    DEBUGASSERT (psnd->allocated_size == 0);
+    DEBUGASSERT (psnd->recv_size == 0);
+    DEBUGASSERT (psnd->recv_processed == 0);
+    DEBUGASSERT (psnd->bindsock == CURL_SOCKET_BAD);
+  }
+}
+
+static void conn_reset_all_postponed_data(struct connectdata *conn)
+{
+  conn_reset_postponed_data(conn, 0);
+  conn_reset_postponed_data(conn, 1);
+}
+#else  /* ! USE_RECV_BEFORE_SEND_WORKAROUND */
+/* Use "do-nothing" macros instead of functions when workaround not used */
+#define conn_reset_postponed_data(c,n) do {} WHILE_FALSE
+#define conn_reset_all_postponed_data(c) do {} WHILE_FALSE
+#endif /* ! USE_RECV_BEFORE_SEND_WORKAROUND */
+
 static void conn_free(struct connectdata *conn)
 {
   if(!conn)
@@ -2744,6 +2781,8 @@ static void conn_free(struct connectdata *conn)
   Curl_safefree(conn->proxy.rawalloc); /* proxy name buffer */
   Curl_safefree(conn->master_buffer);
 
+  conn_reset_all_postponed_data(conn);
+
   Curl_llist_destroy(conn->send_pipe, NULL);
   Curl_llist_destroy(conn->recv_pipe, NULL);
 
@@ -3851,6 +3890,10 @@ static struct connectdata *allocate_conn(struct SessionHandle *data)
   conn->connection_id = -1;    /* no ID */
   conn->port = -1; /* unknown at this point */
   conn->remote_port = -1; /* unknown */
+#if defined(USE_RECV_BEFORE_SEND_WORKAROUND) && defined(DEBUGBUILD)
+  conn->postponed[0].bindsock = CURL_SOCKET_BAD; /* no file descriptor */
+  conn->postponed[1].bindsock = CURL_SOCKET_BAD; /* no file descriptor */
+#endif /* USE_RECV_BEFORE_SEND_WORKAROUND && DEBUGBUILD */
 
   /* Default protocol-independent behavior doesn't support persistent
      connections, so we set this to force-close. Protocols that support
@@ -5625,6 +5668,9 @@ static void reuse_conn(struct connectdata *old_conn,
   /* persist connection info in session handle */
   Curl_persistconninfo(conn);
 
+  conn_reset_all_postponed_data(old_conn); /* free buffers */
+  conn_reset_all_postponed_data(conn);     /* reset unprocessed data */
+
   /* re-use init */
   conn->bits.reuse = TRUE; /* yes, we're re-using here */
 
diff --git a/lib/urldata.h b/lib/urldata.h
index 016761929..f190fbdd6 100644
--- a/lib/urldata.h
+++ b/lib/urldata.h
@@ -833,6 +833,20 @@ typedef ssize_t (Curl_recv)(struct connectdata *conn, /* connection data */
                             size_t len,               /* max amount to read */
                             CURLcode *err);           /* error to return */
 
+#ifdef USE_RECV_BEFORE_SEND_WORKAROUND
+struct postponed_data {
+  char *buffer;          /* Temporal store for received data during
+                            sending, must be freed */
+  size_t allocated_size; /* Size of temporal store */
+  size_t recv_size;      /* Size of received data during sending */
+  size_t recv_processed; /* Size of processed part of postponed data */
+#ifdef DEBUGBUILD
+  curl_socket_t bindsock;/* Structure must be bound to specific socket,
+                            used only for DEBUGASSERT */
+#endif /* DEBUGBUILD */
+};
+#endif /* USE_RECV_BEFORE_SEND_WORKAROUND */
+
 /*
  * The connectdata struct contains all fields and variables that should be
  * unique for an entire connection.
@@ -931,6 +945,9 @@ struct connectdata {
   Curl_recv *recv[2];
   Curl_send *send[2];
 
+#ifdef USE_RECV_BEFORE_SEND_WORKAROUND
+  struct postponed_data postponed[2]; /* two buffers for two sockets */
+#endif /* USE_RECV_BEFORE_SEND_WORKAROUND */
   struct ssl_connect_data ssl[2]; /* this is for ssl-stuff */
   struct ssl_config_data ssl_config;
   bool tls_upgraded;