On Thu, 2017-11-30 at 18:43 +0200, Jarkko Sakkinen wrote: > On Wed, Nov 29, 2017 at 03:13:57PM -0800, Christoph Hellwig wrote: > > > > On Tue, Nov 28, 2017 at 11:57:53PM +0200, Jarkko Sakkinen wrote: > > > > > > > > > > > Yes. You still shall not play nasty games with file > > > > descriptors. > > > > > > I need to put something to file descriptors in order to have a IO > > > channels for the launch enclave hosting process. > > > > Just do it like any other program - open it from your userspace > > program using open() and related syscalls. > > In this case it would not work as the launch enclave is still part of > the kernel and it would create a dependency how the user space > defines paths. If using pipe specifically is an issue, I could easily > use shmem file as a mean for communiation. Can't you simply use sys_pipe2() sys_close() sys_dup2() To achieve the same effect as replace_fd()/create_pipe_files()? The point Christoph is making is that you can call sys_ interfaces from within the kernel (carefully) and have them operate like direct invocations. Look at main.c:kernel_init_freeable() it's doing something similar to what you want, except with the console, not a pipe and it begins with the file table empty. James
