On Thu, Oct 13, 2022 at 09:48:29AM -0400, Dennis Dalessandro wrote: > On 10/13/22 9:21 AM, Jason Gunthorpe wrote: > > On Thu, Oct 13, 2022 at 07:59:52AM -0500, Dean Luick wrote: > >> Hi Jason, > >> > >> I am looking at the FIXME you left in hfi1/user_exp_rcv.c with git commit 3889551db212 > >> > >> Link: https://lore.kernel.org/r/20191112202231.3856-7-jgg@xxxxxxxx > >> > >> Can you please explain in more detail what made you add the FIXME and what may be "racy"? > > > > The comment seems self explanatory, the ordering is upposed to have > > mmu_interval_read_begin() done before the page tables are read, not > > after - since we already have a page list at this point it can't be > > right. > > Is the race you are worried about here when a user proc tries to free the memory > before we get done handling the IOCTL to set up the recvs? The mechanism of mmu notifiers allows coherently reading the page table with a notification when the read becomes invalid. If you don't order things properly then it just doesn't work as advertised. Jason
