On Tue, Feb 18, 2020 at 10:00:35AM -0500, Qian Cai wrote: > On Tue, 2020-02-18 at 15:09 +0100, Marco Elver wrote: > > On Tue, 18 Feb 2020 at 13:40, Qian Cai <cai@xxxxxx> wrote: > > > > > > > > > > > > > On Feb 18, 2020, at 5:29 AM, Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote: > > > > > > > > I think I've got this: > > > > > > > > vm_area_dup() blindly copies all fields of orignal VMA to the new one. > > > > This includes coping vm_area_struct::shared.rb which is normally protected > > > > by i_mmap_lock. But this is fine because the read value will be > > > > overwritten on the following __vma_link_file() under proper protectection. > > > > > > Right, multiple processes could share the same file-based address space where those vma have been linked into address_space::i_mmap via vm_area_struct::shared.rb. Thus, the reader could see its shared.rb linkage pointers got updated by other processes. > > > > > > > > > > > So the fix is correct, but justificaiton is lacking. > > > > > > > > Also, I would like to more fine-grained annotation: marking with > > > > data_race() 200 bytes copy may hide other issues. > > > > > > That is the harder part where I don’t think we have anything for that today. Macro, any suggestions? ASSERT_IGNORE_FIELD()? > > > > There is no nice interface I can think of. All options will just cause > > more problems, inconsistencies, or annoyances. > > > > Ideally, to not introduce more types of macros and keep it consistent, > > ASSERT_EXCLUSIVE_FIELDS_EXCEPT(var, ...) maybe what you're after: > > "Check no concurrent writers to struct, except ignore the provided > > fields". > > > > This option doesn't quite work, unless you just restrict it to 1 field > > (we can't use ranges, because e.g. vm_area_struct has > > __randomize_layout). The next time around, you'll want 2 fields, and > > it won't work. Also, do we know that 'shared.rb' is the only thing we > > want to ignore? > > > > If you wanted something similar to ASSERT_EXCLUSIVE_BITS, it'd have to > > be ASSERT_EXCLUSIVE_FIELDS(var, ...), however, this is quite annoying > > for structs with many fields as you'd have to list all of them. It's > > similar to what you can already do currently (but I also don't > > recommend because it's unmaintainable): > > > > ASSERT_EXCLUSIVE_WRITER(orig->vm_start); > > ASSERT_EXCLUSIVE_WRITER(orig->vm_end); > > ASSERT_EXCLUSIVE_WRITER(orig->vm_next); > > ASSERT_EXCLUSIVE_WRITER(orig->vm_prev); > > ... and so on ... > > *new = data_race(*orig); > > > > Also note that vm_area_struct has __randomize_layout, which makes > > using ranges impossible. All in all, I don't see a terribly nice > > option. > > > > If, however, you knew that there are 1 or 2 fields only that you want > > to make sure are not modified concurrently, ASSERT_EXCLUSIVE_WRITER + > > data_race() would probably work well (or even ASSERT_EXCLUSIVE_ACCESS > > if you want to make sure there are no writers nor _readers_). > > I am testing an idea that just do, > > lockdep_assert_held_write(&orig->vm_mm->mmap_sem); > *new = data_race(*orig); > > The idea is that as long as we have the exclusive mmap_sem held in all paths > (auditing indicated so), no writer should be able to mess up our vm_area_struct > except the "shared.rb" field which has no harm. Well, some fields protected by page_table_lock and can be written to without exclusive mmap_sem. Probably even without any mmap_sem: pin mm_struct + page_table_lock should be enough. -- Kirill A. Shutemov
