> We've all seen the arguments about the in-kernel HTTP daemons and whatnot. I tend to be on the side that says that more things should be pulled out.
I always wondered if a lot of components like this could be either sped up, or moved out of the kernel, using a zero-copy messaging system based on splice() / vmsplice() in the Linux kernel (which works by page table remapping as far as I understand). I guess the prime candidate would be device drivers.
That's one approach. Another one attempted sometimes is to go the other way and move the networking stack into userland as well. Depends on your bottlenecks etc
I always wondered if a lot of components like this could be either sped up, or moved out of the kernel, using a zero-copy messaging system based on splice() / vmsplice() in the Linux kernel (which works by page table remapping as far as I understand). I guess the prime candidate would be device drivers.