Batch up uevent notifications

Uevent notifications come in bursts especially at boot when we're
iterating over all devices. This batches the reports to Erlang up so
that there are significantly fewer context switches.

On a simple RPi0 W application, this trims 1.5 seconds off the boot time.

Author: fhunleth

src/uevent.c

@@ -130,8 +130,12 @@ static int nl_uevent_process_one(struct mnl_socket *nl_uevent, char *resp)
 {
     char nlbuf[8192]; // See MNL_SOCKET_BUFFER_SIZE
     int bytecount = mnl_socket_recvfrom(nl_uevent, nlbuf, sizeof(nlbuf));
-    if (bytecount <= 0)
-        err(EXIT_FAILURE, "mnl_socket_recvfrom");
+    if (bytecount <= 0) {
+        if (errno == EAGAIN)
+            return -1;
+        else
+            err(EXIT_FAILURE, "mnl_socket_recvfrom");
+    }
 
     char *str = nlbuf;
     char *str_end = str + bytecount;
@@ -207,15 +211,23 @@ static int nl_uevent_process_one(struct mnl_socket *nl_uevent, char *resp)
 
 static void nl_uevent_process_all(struct mnl_socket *nl_uevent)
 {
-    // Erlang response processing
+    // Erlang response buffer
     char resp[8192];
-    int resp_index;
+    size_t resp_index;
+
+    // Process uevents until there aren't any more or we're
+    // within 1K of the end. This is pretty conservative since
+    // the Erlang reports looks like they're nearly always < 200 bytes.
+    for (resp_index = 0; resp_index < sizeof(resp) - 1024;) {
+        int bytes_added = nl_uevent_process_one(nl_uevent, &resp[resp_index]);
+        if (bytes_added < 0)
+            break;
 
-    resp_index = nl_uevent_process_one(nl_uevent, resp);
-    if (resp_index <= 0)
-        return;
+        resp_index += bytes_added;
+    }
 
-    write_all(resp, resp_index);
+    if (resp_index > 0)
+        write_all(resp, resp_index);
 }
 
 static int filter(const struct dirent *dirp)