On Mon, Apr 17, 2023 at 04:24:04PM -0300, Jason Gunthorpe wrote: > On Mon, Apr 17, 2023 at 08:00:48PM +0100, Lorenzo Stoakes wrote: > > > So I don't think this route is plausible unless you were thinking of > > somehow offloading to a thread? > > ah, fair enough > > > In any case, if we institute the FOLL_ALLOW_BROKEN_FILE_MAPPINGS flag we > > can just drop FOLL_ANON altogether right, as this will be implied and > > hugetlb should work here too? > > Well.. no, as I said read-only access to the pages works fine, so GUP > should not block that. It is only write that has issues > > > Separately, I find the semantics of access_remote_vm() kind of weird, and > > with a possible mmap_lock-free future it does make me wonder whether > > something better could be done there. > > Yes, it is very weird, kthread_use_mm is much nicer. > > > (Section where I sound like I might be going mad) Perhaps having some means > > of context switching into the kernel portion of the remote process as if > > were a system call or soft interrupt handler and having that actually do > > the uaccess operation could be useful here? > > This is the kthread_use_mm() approach, that is basically what it > does. You are suggesting to extend it to kthreads that already have a > process attached... Yeah, I wonder how plausible this is as we could in theory simply eliminate these remote cases altogether which could be relatively efficient if we could find a way to batch up operations. > > access_remote_vm is basically copy_to/from_user built using kmap and > GUP. > > even a simple step of localizing FOLL_ANON to __access_remote_vm, > since it must have the VMA nyhow, would be an improvement. This is used from places where this flag might not be set though, e.g. acess_process_vm() and ptrace. However, access_remote_vm() is only used by the proc stuff, so I will spin up a patch to move this function and treat it as a helper which sets FOLL_ANON. > > Jason
