Alexandre Julliard wrote: 
Robert Shearman <rob@codeweavers.com> writes:

  
Andrey Turkin wrote:
    
Robert Shearman wrote:
  
      
Andrey Turkin wrote:
    
        
 /******************************************************************************

  *        BindIoCompletionCallback (KERNEL32.@)
  */
+extern NTSTATUS WINAPI
RtlSetIoCompletionCallback(HANDLE,LPOVERLAPPED_COMPLETION_ROUTINE,ULONG);

  
      
          
This should go in winternl.h.
    
        
winternl.h may not depend on winbase.h for some reason, so this would
mean either changing prototype to "(HANDLE,LPVOID,ULONG)", which seems
wrong, or redefining LPOVERLAPPED_COMPLETION_ROUTINE and
OVERLAPPED/LPOVERLAPPED, which seems to be too much duplication.
  
      
I think you should guard the function with an #ifdef __WINE_WINBASE_H 
statement, but I would wait for some other developers to chime in to get 
a consensus before accepting this though.
    

It doesn't seem right for an ntdll function to use an OVERLAPPED
pointer. Shouldn't it take an IO_STATUS_BLOCK instead?

It is not used in the function itself, it is only passed to user-provided callback. So, I see several possibilities:
1) change RtlSetIoCompletionCallback to use another callback prototype, with LPVOID or ULONG_PTR or something instead of LPOVERLAPPED
2) leave callback prototype as in BindIoCompletionCallback, and protect RtlSetIoCompletionCallback with #ifdefs as Robert suggested
3) do not implement RtlSetIoCompletionCallback at all and move all code in kernel32