On Wed, Aug 1, 2018 at 2:55 PM, Hugh Dickins <hughd@xxxxxxxxxx> wrote: > On Wed, 1 Aug 2018, Kirill A. Shutemov wrote: >> On Wed, Aug 01, 2018 at 11:31:52AM -0700, Hugh Dickins wrote: >> > On Wed, 1 Aug 2018, Linus Torvalds wrote: >> > > >> > > Anyway, the upshot of all this is that I think I know what the ia64 >> > > problem was, and John sent the patch for the ashmem case, and I'm >> > > going to hold off reverting that vma_is_anonymous() false-positives >> > > commit after all. >> > >> > I'd better send deletion of zap_pmd_range()'s VM_BUG_ON_VMA(): below >> > (but I've no proprietorial interest, if you prefer to do your own). >> >> Agreed. >> >> Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx> > > Thanks. > >> >> > John's patch is good, and originally I thought it was safe from that >> > VM_BUG_ON_VMA(), because the /dev/ashmem fd exposed to the user is >> > disconnected from the vm_file in the vma, and madvise(,,MADV_REMOVE) >> > insists on VM_SHARED. But afterwards read John's earlier mail, >> > drawing attention to the vfs_fallocate() in there: I may be wrong, >> > and I don't know if Android has THP in the config anyway, but it looks >> > to me like an unmap_mapping_range() from ashmem's vfs_fallocate() >> > could hit precisely the VM_BUG_ON_VMA(), once it's vma_is_anonymous(). >> > >> > (I'm not familiar with ashmem, and I certainly don't understand the >> > role of MAP_PRIVATE ashmem mappings - hole-punch's zap_pte_range() >> > should end up leaving any anon pages in place; but the presence of >> > the BUG is requiring us all to understand too much too quickly.) >> >> Hugh, do you see any reason why ashmem shouldn't have vm_ops == >> shmem_vm_ops? > > I cannot immediately think of an absolute reason why not, but I'm > not giving it much thought; and that might turn it into a stranger > beast than it already is. > >> >> I don't understand ashmem, but I feel uncomfortable that we have this >> sneaky way to create an anonymous VMA. It feels wrong to me. > > I agree it's odd, but in this respect it's no odder than /dev/zero: > that has exactly the same pattern of shmem_zero_setup() for VM_SHARED, > else vma_set_anonymous(): which made me comfortable with John's patch, > restoring the way it worked before. > > Admittedly, the subsequent vfs_fallocate() might generate surprises; > and the business of doing a shmem_file_setup() first, and then undoing > it with a shmem_zero_setup(), looks weird - from John's old XXX comment, Yes, it is weird. :) > I think it was a quick hack to piece together some functionality they > needed in a hurry, which never got revisited (they wanted a name for > the area? maybe memfd would be good for that now). So my XXX comment is to do with a change I made to ashmem in order for it to go into staging, compared with the original Android implementation. They still carry the patch to undo it here: https://android.googlesource.com/kernel/common.git/+/ebfc8d6476a66dc91a1b30cbfc18e67d4401af09%5E%21/ But it has more to do w/ in the shared mapping case, providing a cleaner way of setting the vma->vm_ops = &shmem_vm_ops without having to use shmem_zero_setup(), and doesn't change the behavior in the private mapping case. This discussion has spurred Joel and I to chat a bit about reviving the effort to "properly" upstream ashmem. We're still in discussion but I'm thinking we might just try to add the pin/unpin functionality to memfd. We'll see how the discussion evolves. thanks -john
