Hi, There are a number of 'global' locks taken in epoll. The first two patches remove taking these in the poll() and wakeup paths, since we're already preventing loops, and excessive wakeup paths during EPOLL_CTL_ADD. The final 3 introduce the idea of breaking up the current global 'epmutex' by keeping track of the files that make up a connected epoll set. In this way we can limit locking to be within in a connected component as opposed to the current design which limits things globally. This will mostly help workloads that are doing deeper than 1 level epoll nesting (which may that be that common), since we currently don't take the 'epmutex' for EPOLL_CTL_ADD when there is only 1 level of nesting. However, we do eliminate the global 'epmutex' from the close() path. There is more detail on this in the patch descriptions. One aspect that I'd like to improve is that these patches add a 'struct list_head' and a pointer to the 'struct file', so essentially 3 pointers. I think there are ways to reduce this (for example, by using single-link lists where approriate), but I wanted to get general feedback on the approach before attempting this. Previous changes in this area showed really good improvements on SPECjbb, see commit: 67347fe4e6326338ee217d7eb826bedf30b2e155. However, I don't think that this patch series will improve things that much, as its mostly going to help deeper epoll nesting setups. The motivation for these patches was that it was bothering me that we were still taking some global locks in some fairly hot paths (namely close()). Patches are fairly lightly tested at this point (and need a lot more testing), but I'm not aware of any outstanding issues. Finally, I'd also like to potentially co-ordinate this series with the recent syscall enhancements from Fam Zheng: http://lwn.net/Articles/628828/ since these patches are somewhat invasive. Thanks, -Jason Jason Baron (5): epoll: Remove ep_call_nested() from ep_eventpoll_poll() epoll: Remove ep_call_nested() from ep_poll_safewake() epoll: Add ep_call_nested_nolock() epoll: Allow topology checks to be parallelized epoll: Introduce epoll connected components (remove the epmutex) fs/eventpoll.c | 596 +++++++++++++++++++++++++++++++--------------- include/linux/eventpoll.h | 52 ++-- include/linux/fs.h | 3 + 3 files changed, 442 insertions(+), 209 deletions(-) -- 1.8.2.rc2 -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
