On 04/21/2018 02:07 AM, Christoph Hellwig wrote: > On Fri, Apr 20, 2018 at 08:55:38AM -0700, Eric Dumazet wrote: >> This patch series provide a new mmap_hook to fs willing to grab >> a mutex before mm->mmap_sem is taken, to ensure lockdep sanity. >> >> This hook allows us to shorten tcp_mmap() execution time (while mmap_sem >> is held), and improve multi-threading scalability. > > Missing CC to linu-fsdevel and linux-mm that will have to decide. > > We've rejected this approach multiple times before, so you better > make a really good argument for it. > Well, tcp code needs to hold socket lock before mm->mmap_sem, so current mmap hook can not fit. Or we need to revisit all code doing copyin/copyout while holding a socket lock. (Not feasible really) > introducing a multiplexer that overloads a single method certainly > doesn't help making that case. Well, if you refer to multiple hooks instead of a single one, I basically thought that since only TCP needs this hook at the moment, it was not worth adding extra 8-bytes loads for all other mmap() users. I have no issue adding more hooks and more memory pressure if this is the blocking factor. We need two actions at this moment, (to lock the socket or release it) and a third one would allow us to build the array of pages before grabbing mmap_sem (as I mentioned in the last patch changelog)
