Matthew Wilcox <willy@xxxxxxxxxxxxx> 于2024年6月19日周三 00:51写道： > > On Tue, Jun 18, 2024 at 12:40:23PM -0400, JunChao Sun wrote: > > Darrick J. Wong <djwong@xxxxxxxxxx> 于2024年6月18日周二 12:23写道： > > > > > > On Tue, Jun 18, 2024 at 07:35:04PM +0800, Junchao Sun wrote: > > > > By reordering the elements in the xfs_inode structure, we can > > > > reduce the padding needed on an x86_64 system by 8 bytes. > > > > > > > > > > Does this result in denser packing of xfs_inode objects in the slab > > > > page? > > > > No. Before applying the patch, the size of xfs_inode is 1800 bytes > > with my config, and after applying the patch, the size is 1792 bytes. > > This slight reduction does not result in a denser packing of xfs_inode > > objects within a single page. > > > > The "config dependent" part of this is important though. On my > > laptop running Debian 6.6.15-amd64, xfs_inode is exactly 1024 bytes, > > and slab chooses to allocate 32 of them from an order-3 slab. > > > > Your config gets you 18 from an order-3 slab, and you'd need to get > > it down to 1724 (probably 1720 bytes due to alignment) to get 19 > > from an order-3 slab. I bet you have lockdep or something on. Hi, I couldn't find the exact 6.6.15-amd64 kernel, but I installed the Debian 6.8.12-amd64 and 6.1.0-21-amd64 kernels, along with their corresponding debug packages. In both cases, the size of xfs_inode is 1000 bytes. By eliminating the padding bytes, the number of xfs_inode objects allocated from an order-3 slab increased from 32 to 33. I'm not sure what specific differences there are between our Debian kernels, but I have submitted the v2 version of the patch. If there are any issues, please feel free to let me know. Thank you! Best regards -- Junchao Sun <sunjunchao2870@xxxxxxxxx>
