On Sat, Apr 22, 2023 at 02:37:05PM +0100, Lorenzo Stoakes wrote: > It isn't safe to write to file-backed mappings as GUP does not ensure that > the semantics associated with such a write are performed correctly, for > instance filesystems which rely upon write-notify will not be correctly > notified. > > There are exceptions to this - shmem and hugetlb mappings are (in effect) > anonymous mappings by other names so we do permit this operation in these > cases. > > In addition, if no pinning takes place (neither FOLL_GET nor FOLL_PIN is > specified and neither flags gets implicitly set) then no writing can occur > so we do not perform the check in this instance. > > This is an important exception, as populate_vma_page_range() invokes > __get_user_pages() in this way (and thus so does __mm_populate(), used by > MAP_POPULATE mmap() and mlock() invocations). > > There are GUP users within the kernel that do nevertheless rely upon this > behaviour, so we introduce the FOLL_ALLOW_BROKEN_FILE_MAPPING flag to > explicitly permit this kind of GUP access. > > This is required in order to not break userspace in instances where the > uAPI might permit file-mapped addresses - a number of RDMA users require > this for instance, as do the process_vm_[read/write]v() system calls, > /proc/$pid/mem, ptrace and SDT uprobes. Each of these callers have been > updated to use this flag. > > Making this change is an important step towards a more reliable GUP, and > explicitly indicates which callers might encouter issues moving forward. nit: s/encouter/encounter/
