On Mon, Mar 29, 2021 at 04:24:19PM -0700, Dan Williams wrote: > On Thu, Mar 25, 2021 at 7:34 AM Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > > > On Thu, Mar 18, 2021 at 10:03:06AM -0700, Dan Williams wrote: > > > Yes. I still need to answer the question of whether mapping > > > invalidation triggers longterm pin holders to relinquish their hold, > > > but that's a problem regardless of whether gup-fast is supported or > > > not. > > > > It does not, GUP users do not interact with addres_space or mmu > > notifiers > > Ok, but the SIGKILL from the memory_failure() will drop the pin? In the general case I doubt it.. I think this is fine, we shouldn't expect unplugging a driver to not block if there are open users. Demading DMA users to be able to shoot down access to system memory is a really big kernel change. Jason
