On 4/26/23 00:20, Lorenzo Stoakes wrote: ... >> Could you elaborate? GUP calls handle_mm_fault() that invokes the write >> notify the pte is made first writable. Of course, virt(pinned_page)[0] = 'x' >> is not supposed to fault while using the kernel mapping. >> > > The issue is how dirtying works. Typically for a dirty-tracking mapping the > kernel makes the mapping read-only, then when a write fault occurs, > writenotify is called and the folio is marked dirty. This way the file > system knows which files to writeback, then after writeback it 'cleans' > them, restoring the read-only mapping and relying on the NEXT write marking > write notifying and marking the folio dirty again. > > If we GUP, _especially_ if it's long term, we run the risk of a write to > the folio _after_ it has been cleaned and if the caller tries to do the > 'right thing' and mark the folio dirty, it being marked dirty at an > unexpected time which might race with other things and thus oops. > > The issue is that this dirtying mechanism implicitly relies on the memory > _only_ being accessed via user mappings, but we're providing another way to > access that memory bypassing all that. > > It's been a fundamental flaw in GUP for some time. This change tries to > make things a little better by precluding the riskiest version of this > which is the caller indicating that the pin is longterm (via > FOLL_LONGTERM). > > For an example of a user trying to sensibly avoid this, see io_pin_pages() > in io_uring/rsrc.c. This is where io_uring tries to explicitly avoid this > themselves, something that GUP should clearly be doing. > > After this change, that code can be removed and we will live in a world > where linux has a saner GUP :) Hi Lorenzo, This is the "missing writeup" that really helps people visualize the problem, very nice. I think if you include the above, plus maybe a link to Jan Kara's original report [1], in the commit description, that will help a lot. > > Of course we need to make more fundamental changes moving forward, the idea > is this improves the situation and eliminates the need for the open-coded > solution for io_uring which unblocks my other patch series which is also > trying to make GUP more sane. It is true that solving the problem has taken "a few"...years, and is still not done, yes. :) I'll respond to the patch with a review, as well. [1] https://lore.kernel.org/linux-mm/20180103100430.GE4911@xxxxxxxxxxxxxx/ thanks, -- John Hubbard NVIDIA
