On Wed, May 06, 2009 at 11:13:21PM -0700, Sukadev Bhattiprolu wrote: > Oren Laadan [orenl@xxxxxxxxxxxxxxx] wrote: <snip> > | > | diff --git a/checkpoint/memory.c b/checkpoint/memory.c > | > | index 7637c1e..5ae2b41 100644 > | > | --- a/checkpoint/memory.c > | > | +++ b/checkpoint/memory.c > | > | @@ -687,6 +687,8 @@ static int do_checkpoint_mm(struct ckpt_ctx *ctx, struct mm_struct *mm) > | > | ret = exe_objref; > | > | goto out; > | > | } > | > | + /* account for all references through vma/exe_file */ > | > | + ckpt_obj_users_inc(ctx, mm->exe_file, mm->num_exe_file_vmas); > | > > | > Do we really need to add num_exe_file_vmas here ? > | > > | > A quick look at all callers for added_exe_file_vma() seems to show that > | > those callers also do a get_file(). > | > | Each vma whose file is the same as mm->exe_file causes the refcount > | of that file to increase by 2: once by vma->vm_file, and once via > | added_exe_file_vma(). The c/r code calls ckpt_obj_checkpoint() only > | once, thus once one obj_file_grab() for that file. The code above > | accounts for the missing count. Perhaps I'm misreading Oren's explanation, but the refcount on the file should not be: 2*#vmas(vm_file==mm->exe_file) + #fds(filp==mm->exe_file) It should be: #vmas(vm_file==mm->exe_file) + #fds(filp==mm->exe_file) + 1(for mm->exe_file). because added_exe_file_vma() increments num_exe_file_vmas but does not change the file reference count. So incrementing the obj count while walking the vmas and fds should bring the count 1 short of matching. Cheers, -Matt _______________________________________________ Containers mailing list Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/containers