Gracefully handle an asynchronous write disconnect when an SSL read is pending

net/base/nss_memio.c

Index: net/base/nss_memio.c
diff --git a/net/base/nss_memio.c b/net/base/nss_memio.c
index bd0471067f4e57e31b39d09889f0caa1bdeef378..9e6e84611cfb257a9a944cd9ff2072770616896a 100644
--- a/net/base/nss_memio.c
+++ b/net/base/nss_memio.c
@@ -228,8 +228,17 @@ static int PR_CALLBACK memio_Recv(PRFileDesc *fd, void *buf, PRInt32 len,
     rv = memio_buffer_get(mb, buf, len);
     if (rv == 0 && !secret->eof) {
         secret->read_requested = len;
+        /* If there is no more data in the buffer, report any pending errors
+         * that were previously observed. Note that both the readbuf and the
+         * writebuf are checked for errors, since the application may have
+         * encountered a socket error while writing that would otherwise not
+         * be reported until the application attempted to write again - which
+         * it may never do.
+         */
         if (mb->last_err)
             PR_SetError(mb->last_err, 0);
+        else if (secret->writebuf.last_err)
+            PR_SetError(secret->writebuf.last_err, 0);
         else
             PR_SetError(PR_WOULD_BLOCK_ERROR, 0);
         return -1;
@@ -256,6 +265,11 @@ static int PR_CALLBACK memio_Send(PRFileDesc *fd, const void *buf, PRInt32 len,
     mb = &secret->writebuf;
     PR_ASSERT(mb->bufsize);
 
+    /* Note that the read error state is not reported, because it cannot be
+     * reported until all buffered data has been read. If there is an error
+     * with the next layer, attempting to call Send again will report the
+     * error appropriately.
+     */
     if (mb->last_err) {
         PR_SetError(mb->last_err, 0);
         return -1;