Fix float point overflow when doing pass up

The old way of rounding floor(x/((float)sizeof(int)))*sizeof(int)) will
cause overflow when x is greater than 203672784. For example, if x is
203672792, the above expression will give 203672800. When this bigger
value is minus by parent_sent_pos, negative value can be generated
which is later assigned to an unsigned integer. This will cause the
whole learning fail. This patch fixes the above expression and a
potential network problem.

1 files changed, 5 insertions, 4 deletions

diff --git a/vowpalwabbit/allreduce.h b/vowpalwabbit/allreduce.h
index 07ecafc7..ecac009b 100644
--- a/vowpalwabbit/allreduce.h
+++ b/vowpalwabbit/allreduce.h
@@ -62,15 +62,16 @@ template <class T, void (*f)(T&, const T&)> void addbufs(T* buf1, const T* buf2,
 void all_reduce_init(const string master_location, const size_t unique_id, const size_t total, const size_t node, node_socks& socks);
 
 template <class T> void pass_up(char* buffer, size_t left_read_pos, size_t right_read_pos, size_t& parent_sent_pos, socket_t parent_sock, size_t n) {
-  size_t my_bufsize = min(ar_buf_size, ((int)(floor(left_read_pos/((float)sizeof(T)))*sizeof(T)) - parent_sent_pos));
-  my_bufsize = min(my_bufsize, ((int)(floor(right_read_pos/((float)sizeof(T)))*sizeof(T)) - parent_sent_pos));
+  size_t my_bufsize = min(ar_buf_size, min(left_read_pos, right_read_pos) / sizeof(T) * sizeof(T) - parent_sent_pos);
 
   if(my_bufsize > 0) {
     //going to pass up this chunk of data to the parent
     int write_size = send(parent_sock, buffer+parent_sent_pos, (int)my_bufsize, 0);
-    if(write_size < (int)my_bufsize)
+    if(write_size < 0) {
       cerr<<"Write to parent failed "<<my_bufsize<<" "<<write_size<<" "<<parent_sent_pos<<" "<<left_read_pos<<" "<<right_read_pos<<endl ;
-    parent_sent_pos += my_bufsize;
+      throw exception();
+    }
+    parent_sent_pos += write_size;
   }
 
 }