On Wed, Nov 24, 2004 at 10:06:02PM +0000, jra@xxxxxxxxx wrote:
> Fix the problem we get on Linux where sendfile fails, but we've already sent
> header using send(). As our implementation of sendfile can't return EINTR (it
> restarts in that case) use an errno of EINTR to signal the linux sendfile fail
> after header case. When that happens send the rest of the data and then turn
> off sendfile. Sendfile should be safe to enable on all systems now (even
> it may not help in all performance cases).
How much does sendfile gain us at all? I've had to debug really nasty problems,
I scratched my head why samba simply stops responding, making XP reconnect.
sendfile = no was the very simple answer after hours of staring at sniffs,
trying to reproduce the errors. If we don't see a huge performance improvement
I'd VERY strongly argue that we dump sendfile completely. It is just way too
buggy in common kernels.