On Fri, Jun 21, 2024 at 09:07:19PM +0200, Sebastian Andrzej Siewior wrote: > On 2024-06-21 14:49:23 [-0400], Kent Overstreet wrote: > > On Fri, Jun 21, 2024 at 08:29:15PM +0200, Sebastian Andrzej Siewior wrote: > > > On 2024-06-21 10:20:58 [-0400], Kent Overstreet wrote: > > > > On Fri, Jun 21, 2024 at 12:27:50PM +0200, Sebastian Andrzej Siewior wrote: > > > > > The alloc_tag member has been added to task_struct at the very > > > > > beginning. This is a pointer and on 64bit architectures it forces 4 byte > > > > > padding after `ptrace' and then forcing another another 4 byte padding > > > > > after `on_cpu'. A few members later, `se' requires a cacheline aligned > > > > > due to struct sched_avg resulting in 52 hole before `se'. > > > > > > > > > > This is the case on 64bit-SMP architectures. > > > > > The 52 byte hole can be avoided by moving alloc_tag away where it > > > > > currently resides. > > > > > > > > > > Move alloc_tag to the end of task_struct. There is likely a hole before > > > > > `thread' due to its alignment requirement and the previous members are > > > > > likely to be already pointer-aligned. > > > > > > > > We sure we want it at the end? we do want it on a hot cacheline > > > > > > Well, the front is bad. > > > Looking at pgalloc_tag_add() and its callers, there is no task_struct > > > touching. alloc_tag_save()/restore might be the critical one. This is > > > random code… Puh. So if the end is too cold, what about around the mm > > > pointer? > > > > Not there, that's not actually that hot. It needs to be by > > task_struct->flags; that's used in the same paths. > > But there is no space without the additional 52 bytes. Was this by any > chance on purpose? No, that wasn't, and it doesn't have to be the exact same cacheline, but we do want it near current->flags; we store PF_MEMALLOC flags there that are converted to gfp flags and used exactly where we're using current->alloc_tag.
