On Tue, Sep 15, 2020 at 03:13:46PM -0400, Peter Xu wrote: > On Tue, Sep 15, 2020 at 03:29:33PM -0300, Jason Gunthorpe wrote: > > On Tue, Sep 15, 2020 at 01:05:53PM -0300, Jason Gunthorpe wrote: > > > On Tue, Sep 15, 2020 at 10:50:40AM -0400, Peter Xu wrote: > > > > On Mon, Sep 14, 2020 at 08:28:51PM -0300, Jason Gunthorpe wrote: > > > > > Yes, this stuff does pin_user_pages_fast() and MADV_DONTFORK > > > > > together. It sets FOLL_FORCE and FOLL_WRITE to get an exclusive copy > > > > > of the page and MADV_DONTFORK was needed to ensure that a future fork > > > > > doesn't establish a COW that would break the DMA by moving the > > > > > physical page over to the fork. DMA should stay with the process that > > > > > called pin_user_pages_fast() (Is MADV_DONTFORK still needed with > > > > > recent years work to GUP/etc? It is a pretty terrible ancient thing) > > > > > > > > ... Now I'm more confused on what has happened. > > > > > > I'm going to try to confirm that the MADV_DONTFORK is actually being > > > done by userspace properly, more later. > > > > It turns out the test is broken and does not call MADV_DONTFORK when > > doing forks - it is an opt-in it didn't do. > > > > It looks to me like this patch makes it much more likely that the COW > > break after page pinning will end up moving the pinned physical page > > to the fork while before it was not very common. Does that make sense? > > My understanding is that the fix should not matter much with current failing > test case, as long as it's with FOLL_FORCE & FOLL_WRITE. However what I'm not > sure is what if the RDMA/DMA buffers are designed for pure read from > userspace.
No, they are write. Always FOLL_WRITE. > E.g. for vfio I'm looking at vaddr_get_pfn() where I believe such pure read > buffers will be a GUP with FOLL_PIN and !FOLL_WRITE which will finally pass to > pin_user_pages_remote(). So what I'm worrying is something like this: I think the !(prot & IOMMU_WRITE) case is probably very rare for VFIO. I'm also not sure it will work reliably, in RDMA we had this as a more common case and long ago found bugs. The COW had to be broken for the pin anyhow. > 1. Proc A gets a private anon page X for DMA, mapcount==refcount==1. > > 2. Proc A fork()s and gives birth to proc B, page X will now have > mapcount==refcount==2, write-protected. proc B quits. Page X goes back > to mapcount==refcount==1 (note! without WRITE bits set in the PTE). > 3. pin_user_pages(write=false) for page X. Since it's with !FORCE & !WRITE, > no COW needed. Refcount==2 after that. > > 4. Pass these pages to device. We either setup IOMMU page table or just use > the PFNs, which is not important imho - the most important thing is the > device will DMA into page X no matter what. > > 5. Some thread of proc A writes to page X, trigger COW since it's > write-protected with mapcount==1 && refcount==2. The HVA that pointing > to > page X will be changed to point to another page Y after the COW. > > 6. Device DMA happens, data resides on X. Proc A can never get the data, > though, because it's looking at page Y now. RDMA doesn't ever use !WRITE I'm guessing #5 is the issue, just with a different ordering. If the #3 pin_user_pages() preceeds the #2 fork, don't we get to the same #5? > If this is a problem, we may still need the fix patch (maybe not as urgent as > before at least). But I'd like to double confirm, just in case I miss some > obvious facts above. I'm worred that the sudden need to have MAD_DONTFORK is going to be a turn into a huge regression. It already blew up our first level of synthetic test cases. I'm worried what we will see when the application suite is run in a few months :\ > > Given that the tests are wrong it seems like broken userspace, > > however, it also worked reliably for a fairly long time. > > IMHO it worked because the page to do RDMA has mapcount==1, so it was reused > previously just as-is even after the fork without MADV_DONTFORK and after the > child quits. That would match the results we see.. So this patch changes things so it is not re-used as-is, but replaced with Y? This basically means any driver using pin_user_pages() can no longer have fork() in userspace, when before fork() only failed in fairly narrow cases. Unfortunately I think this will break things broadly beyond RDMA. Jason