06mmap.txt

# Copyright (C) 2005-2020 Junjiro R. Okajima
# 
# This program is free software; you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation; either version 2 of the License, or
# (at your option) any later version.
# 
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
# 
# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <http://www.gnu.org/licenses/>.

mmap(2) -- File Memory Mapping
----------------------------------------------------------------------
In aufs, the file-mapped pages are handled by a branch fs directly, no
interaction with aufs. It means aufs_mmap() calls the branch fs's
->mmap().
This approach is simple and good, but there is one problem.
Under /proc, several entries show the mmapped files by its path (with
device and inode number), and the printed path will be the path on the
branch fs's instead of virtual aufs's.
This is not a problem in most cases, but some utilities lsof(1) (and its
user) may expect the path on aufs.

To address this issue, aufs adds a new member called vm_prfile in struct
vm_area_struct (and struct vm_region). The original vm_file points to
the file on the branch fs in order to handle everything correctly as
usual. The new vm_prfile points to a virtual file in aufs, and the
show-functions in procfs refers to vm_prfile if it is set.
Also we need to maintain several other places where touching vm_file
such like
- fork()/clone() copies vma and the reference count of vm_file is
  incremented.
- merging vma maintains the ref count too.

This is not a good approach. It just fakes the printed path. But it
leaves all behaviour around f_mapping unchanged. This is surely an
advantage.
Actually aufs had adopted another complicated approach which calls
generic_file_mmap() and handles struct vm_operations_struct. In this
approach, aufs met a hard problem and I could not solve it without
switching the approach.

There may be one more another approach which is
- bind-mount the branch-root onto the aufs-root internally
- grab the new vfsmount (ie. struct mount)
- lazy-umount the branch-root internally
- in open(2) the aufs-file, open the branch-file with the hidden
  vfsmount (instead of the original branch's vfsmount)
- ideally this "bind-mount and lazy-umount" should be done atomically,
  but it may be possible from userspace by the mount helper.

Adding the internal hidden vfsmount and using it in opening a file, the
file path under /proc will be printed correctly. This approach looks
smarter, but is not possible I am afraid.
- aufs-root may be bind-mount later. when it happens, another hidden
  vfsmount will be required.
- it is hard to get the chance to bind-mount and lazy-umount
  + in kernel-space, FS can have vfsmount in open(2) via
    file->f_path, and aufs can know its vfsmount. But several locks are
    already acquired, and if aufs tries to bind-mount and lazy-umount
    here, then it may cause a deadlock.
  + in user-space, bind-mount doesn't invoke the mount helper.
- since /proc shows dev and ino, aufs has to give vma these info. it
  means a new member vm_prinode will be necessary. this is essentially
  equivalent to vm_prfile described above.

I have to give up this "looks-smater" approach.