On Tue 12-07-22 10:40:11, Yang Shi wrote: > On Fri, Jul 8, 2022 at 1:29 AM Mike Rapoport <rppt@xxxxxxxxxx> wrote: > > > > On Thu, Jul 07, 2022 at 03:09:32PM -0700, Yang Shi wrote: > > > On Thu, Jul 7, 2022 at 1:55 PM Darrick J. Wong <djwong@xxxxxxxxxx> wrote: > > > > > > > > On Thu, Jul 07, 2022 at 10:48:00AM -0700, Yang Shi wrote: > > > > > On Thu, Jul 7, 2022 at 9:57 AM Mike Rapoport <rppt@xxxxxxxxxx> wrote: > > > > > > > > > > > > Eric Biggers suggested that this happens when > > > > > > secretmem_setattr()->simple_setattr() races with secretmem_fault() so > > > > > > that a page that is faulted in by secretmem_fault() (and thus removed > > > > > > from the direct map) is zeroed by inode truncation right afterwards. > > > > > > > > > > > > Since do_truncate() takes inode_lock(), adding inode_lock_shared() to > > > > > > secretmem_fault() prevents the race. > > > > > > > > > > Should invalidate_lock be used to serialize between page fault and truncate? > > > > > > > > I would have thought so, given Documentation/filesystems/locking.rst: > > > > > > > > "->fault() is called when a previously not present pte is about to be > > > > faulted in. The filesystem must find and return the page associated with > > > > the passed in "pgoff" in the vm_fault structure. If it is possible that > > > > the page may be truncated and/or invalidated, then the filesystem must > > > > lock invalidate_lock, then ensure the page is not already truncated > > > > (invalidate_lock will block subsequent truncate), and then return with > > > > VM_FAULT_LOCKED, and the page locked. The VM will unlock the page." > > > > > > > > IIRC page faults aren't supposed to take i_rwsem because the fault could > > > > be in response to someone mmaping a file into memory and then write()ing > > > > to the same file using the mmapped region. The write() takes > > > > inode_lock and faults on the buffer, so the fault cannot take inode_lock > > > > again. > > > > > > Do you mean writing from one part of the file to the other part of the > > > file so the "from" buffer used by copy_from_user() is part of the > > > mmaped region? > > > > > > Another possible deadlock issue by using inode_lock in page faults is > > > mmap_lock is acquired before inode_lock, but write may acquire > > > inode_lock before mmap_lock, it is a AB-BA lock pattern, but it should > > > not cause real deadlock since mmap_lock is not exclusive for page > > > faults. But such pattern should be avoided IMHO. > > > > > > > That said... I don't think memfd_secret files /can/ be written to? > > > > memfd_secret files cannot be written to, they can only be mmap()ed. > > Synchronization is only required between > > do_truncate()->...->simple_setatt() and secretmem->fault() and I don't see > > how that can deadlock. > > Sure, there is no deadlock. > > > > > I'm not an fs expert though, so if you think that invalidate_lock() is > > safer, I don't mind s/inode_lock/invalidate_lock/ in the patch. > > IIUC invalidate_lock should be preferred per the filesystem's locking > document. And I found Jan Kara's email of the invalidate_lock > patchset, please refer to > https://lore.kernel.org/linux-mm/20210715133202.5975-1-jack@xxxxxxx/. Yeah, so using invalidate_lock for such synchronization would be certainly more standard than using inode_lock. Although I agree that for filesystems that do not support read(2) and write(2) there does not seem to be an immediate risk of a deadlock when inode_lock is used inside a page fault. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
