On Tue 10-10-17 00:43:31, Yang Shi wrote: > > > On 10/8/17 11:48 PM, Michal Hocko wrote: > > On Sun 08-10-17 15:56:51, Kirill A. Shutemov wrote: > > > On Sat, Oct 07, 2017 at 04:22:10AM +0800, Yang Shi wrote: > > > > When passing "huge=always" option for mounting tmpfs, THP is supposed to > > > > be allocated all the time when it can fit, but when the available space > > > > is > > > > smaller than the size of THP (2MB on x86), shmem fault handler still > > > > tries > > > > to allocate huge page every time, then fallback to regular 4K page > > > > allocation, i.e.: > > > > > > > > # mount -t tmpfs -o huge,size=3000k tmpfs /tmp > > > > # dd if=/dev/zero of=/tmp/test bs=1k count=2048 > > > > # dd if=/dev/zero of=/tmp/test1 bs=1k count=2048 > > > > > > > > The last dd command will handle 952 times page fault handler, then exit > > > > with -ENOSPC. > > > > > > > > Rounding up tmpfs size to THP size in order to use THP with "always" > > > > more efficiently. And, it will not wast too much memory (just allocate > > > > 511 extra pages in worst case). > > > > > > Hm. I don't think it's good idea to silently increase size of fs. > > > > Agreed! > > > > > Maybe better just refuse to mount with huge=always for too small fs? > > > > We cannot we simply have the remaining page !THP? What is the actual
ups s@We@Why@ > > problem? > > The remaining pages can be !THP, it will fall back to regular 4k pages when > the available space is less than THP size. > > I just wonder it sounds not make sense to *not* mount tmpfs with THP size > alignment when "huge=always" is passed. yes failure seems overly excessive reaction to me. > I guess someone would like to assume all allocation in tmpfs with > "huge=always" should be THP. Nobody can assume that because THP pages can be broken up at any point in time. We have hugetlb to provide a guarantee > But, they might not be fully aware of in some > corner cases THP might be not used, for example, the remaining space is less > then THP size, then some unexpected performance degrade might be perceived. > > So, why not we do the mount correctly at the first place. It could be > delegated to the administrator, but it should be better to give some hint > from kernel side. Because we are not trying to be more clever than the user. I still do not see what is the actual problem you are trying to fix to be honest. -- Michal Hocko SUSE Labs
