On Sat, May 09, 2020 at 11:25:16AM +0200, Peter Zijlstra wrote: > On Fri, May 08, 2020 at 11:34:07PM +0200, Joerg Roedel wrote: > > On Fri, May 08, 2020 at 09:20:00PM +0200, Peter Zijlstra wrote: > > > The only concern I have is the pgd_lock lock hold times. > > > > > > By not doing on-demand faults anymore, and consistently calling > > > sync_global_*(), we iterate that pgd_list thing much more often than > > > before if I'm not mistaken. > > > > Should not be a problem, from what I have seen this function is not > > called often on x86-64. The vmalloc area needs to be synchronized at > > the top-level there, which is by now P4D (with 4-level paging). And the > > vmalloc area takes 64 entries, when all of them are populated the > > function will not be called again. > > Right; it's just that the moment you do trigger it, it'll iterate that > pgd_list and that is potentially huge. Then again, that's not a new > problem. > > I suppose we can deal with it if/when it becomes an actual problem. > > I had a quick look and I think it might be possible to make it an RCU > managed list. We'd have to remove the pgd_list entry in > put_task_struct_rcu_user(). Then we can play games in sync_global_*() > and use RCU iteration. New tasks (which can be missed in the RCU > iteration) will already have a full clone of the PGD anyway.
One of the things on my long-term todo list is to replace mm_struct.mmlist with an XArray containing all mm_structs. Then we can use one mark to indicate maybe-swapped and another mark to indicate ... whatever it is pgd_list indicates. Iterating an XArray (whether the entire thing or with marks) is RCU-safe and faster than iterating a linked list, so this should solve the problem?
