On 17.06.2020 15:00, Tamas K Lengyel wrote: > On Wed, Jun 17, 2020 at 3:59 AM Jan Beulich <jbeul...@suse.com> wrote: >> If there are code paths of both kinds, which approach to use in >> vmx_load_pdptrs() may need to be chosen based on what >> paging_locked_by_me() returns. Or perhaps an unlocked query is >> fine in either case? > > Perhaps adjusting vmx_load_pdptrs to chose the unlocked query would be > fine. But at that point what is the reason for having the lock > ordering at all? Why not just have a single recursive lock and avoid > issues like this altogether?
With just a single lock, contention problems we already know we have would be even worse. When the current locking model was introduced, there was actually a plan to make gfn_lock() more fine-grained (i.e. not simply "de-generate" to p2m_lock()), for example. Jan
