On Mon, Apr 24, 2023 at 10:39:25AM -0300, Jason Gunthorpe wrote: > On Mon, Apr 24, 2023 at 01:38:49PM +0100, Lorenzo Stoakes wrote: > > > I was being fairly conservative in that list, though we certainly need to > > set the flag for /proc/$pid/mem and ptrace to avoid breaking this > > functionality (I observed breakpoints breaking without it which obviously > > is a no go :). I'm not sure if there's a more general way we could check > > for this though? > > More broadly we should make sure these usages of GUP safe somehow so > that it can reliably write to those types of pages without breaking > the current FS contract.. > > I forget exactly, but IIRC, don't you have to hold some kind of page > spinlock while writing to the page memory? > I think perhaps you're thinking of the mm->mmap_lock? Which will be held for the FOLL_GET cases and simply prevent the VMA from disappearing below us but not do much else. > So, users that do this, or can be fixed to do this, can get file > backed pages. It suggests that a flag name is more like > FOLL_CALLER_USES_FILE_WRITE_LOCKING > As stated above, I'm not sure what locking you're referring to, but seems to me that FOLL_GET already implies what you're thinking? I wonder whether we should do this check purely for FOLL_PIN to be honest? As this indicates medium to long-term access without mmap_lock held. This would exclude the /proc/$pid/mem and ptrace paths which use gup_remote(). That and a very specific use of uprobes are the only places that use FOLL_GET in this instance and each of them are careful in any case to handle setting the dirty page flag. All PUP cases that do not specify FOLL_LONGTERM also do this, so we could atually go so far as to reduce the patch to simply performing the vma_wants_writenotify() check if (FOLL_PIN | FOLL_LONGTERM) is specified, which covers the io_uring case. Alternatively if we wanted to be safer, we could add a FOLL_ALLOW_FILE_PIN that is checked on FOLL_PIN and ignored on FOLL_LONGTERM? > > I wouldn't be totally opposed to dropping it for RDMA too, because I > > suspect accessing file-backed mappings for that is pretty iffy. > > > > Do you have a sense of which in the list you feel could be pared back? > > Anything using FOLL_LONGTERM should not set the flag, GUP should even > block the combination. OK > > And we need to have in mind that the flag indicates the code is > buggy, so if you set it then we should understand how is that caller > expected to be fixed. > > Jason I think we are working towards a much simpler solution in any case!
