On Wed 30-01-19 18:01:33, Weiny, Ira wrote: > > > > On Tue, Jan 29, 2019 at 10:50:05AM -0800, Ira Weiny wrote: > > > > .. and I'm looking at some of the other conversions here.. *most > > > > likely* any caller that is manipulating rlimit for get_user_pages > > > > should really be calling get_user_pages_longterm, so they should not > > > > be converted to use _fast? > > > > > > Is this a question? I'm not sure I understand the meaning here? > > > > More an invitation to disprove the statement > > Generally I agree. But would be best if we could get fast GUP for > performance. I have not worked out if that will be possible with the > final "longterm" solutions. Initially probably not, longer-term it might be added if there are performance data supporting that (i.e., showing real workload that would benefit). In principle there's nothing that would prevent gup_fast like functionality for long-term pins but I expect there will be always additional overhead (compared to plain gup_fast()) of establishing something like leases to identify long-term pins. But we haven't figured out the details yet. For now we concentrate on fixing short-term pins and issues with those. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
