On Mon, Feb 17, 2020 at 10:59:47PM -0500, Qian Cai wrote: > > > > On Feb 17, 2020, at 5:31 PM, Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote: > > > > I'm confused. AFAICS both sides hold mmap_sem on write: > > > > - vm_mmap_pgoff() takes mmap_sem for the write on the write side > > > > - do_mprotect_pkey() takes mmap_sem for the write on the read side > > > > > > What do I miss? > > Ah, good catch. I missed the locking for the read there. This is interesting because Marco > did confirmed that the concurrency could happen, > > https://lore.kernel.org/lkml/20191025173511.181416-1-elver@xxxxxxxxxx/ > > If that means KCSAN is not at fault, then I could think of two things, > > 1) someone downgrades the lock. > > I don’t think that a case here. Only __do_munmap() will do that but I did not see how > it will affect us here. > > 2) the reader and writer are two different processes. > > So, they held a different mmap_sem, but I can’t see how could two processes shared > the same vm_area_struct. Also, file->f_mapping->i_mmap was also stored in the > writer, but I can’t see how it was also loaded in the reader. > > Any ideas? 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. 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. -- Kirill A. Shutemov
