cgroup filesystem code was derived from the original sysfs
implementation which was heavily intertwined with vfs objects and
locking with the goal of re-using the existing vfs infrastructure.
That experiment turned out rather disastrous and sysfs switched, a
long time ago, to distributed filesystem model where a separate
representation is maintained which is queried by vfs. Unfortunately,
cgroup stuck with the failed experiment all these years and
accumulated even more problems over time.
Locking and object lifetime management being entangled with vfs is
probably the most egregious. vfs is never designed to be misused like
this and cgroup ends up jumping through various convoluted dancing to
make things work. Even then, operations across multiple cgroups can't
be done safely as it'll deadlock with rename locking.
Recently, kernfs is separated out from sysfs so that it can be used by
users other than sysfs. This patch converts cgroup to use kernfs,
which will bring the following benefits.
* Separation from vfs internals. Locking and object lifetime
management is contained in cgroup proper making things a lot
simpler. This removes significant amount of locking convolutions,
hairy object lifetime rules and the restriction on multi-cgroup
operations.
* Can drop a lot of code to implement filesystem interface as most are
provided by kernfs.
* Proper "severing" semantics, which allows controllers to not worry
about lingering file accesses after offline.
While the preceding patches did as much as possible to make the
transition less painful, large part of the conversion has to be one
discrete step making this patch rather large. The rest of the commit
message lists notable changes in different areas.
Overall
-------
* vfs constructs replaced with kernfs ones. cgroup->dentry w/ ->kn,
cgroupfs_root->sb w/ ->kf_root.
* All dentry accessors are removed. Helpers to map from kernfs
constructs are added.
* All vfs plumbing around dentry, inode and bdi removed.
* cgroup_mount() now directly looks for matching root and then
proceeds to create a new one if not found.
Synchronization and object lifetime
-----------------------------------
* vfs inode locking removed. Among other things, this removes the
need for the convolution in cgroup_cfts_commit(). Future patches
will further simplify it.
* vfs refcnting replaced with cgroup internal ones. cgroup->refcnt,
cgroupfs_root->refcnt added. cgroup_put_root() now directly puts
root->refcnt and when it reaches zero proceeds to destroy it thus
merging cgroup_put_root() and the former cgroup_kill_sb().
Simliarly, cgroup_put() now directly schedules cgroup_free_rcu()
when refcnt reaches zero.
* Unlike before, kernfs objects don't hold onto cgroup objects. When
cgroup destroys a kernfs node, all existing operations are drained
and the association is broken immediately. The same for
cgroupfs_roots and mounts.
* All operations which come through kernfs guarantee that the
associated cgroup is and stays valid for the duration of operation;
however, there are two paths which need to find out the associated
cgroup from dentry without going through kernfs -
css_tryget_from_dir() and cgroupstats_build(). For these two,
kernfs_node->priv is RCU managed so that they can dereference it
under RCU read lock.
File and directory handling
---------------------------
* File and directory operations converted to kernfs_ops and
kernfs_syscall_ops.
* xattrs is implicitly supported by kernfs. No need to worry about it
from cgroup. This means that "xattr" mount option is no longer
necessary. A future patch will add a deprecated warning message
when sane_behavior.
* When cftype->max_write_len > PAGE_SIZE, it's necessary to make a
private copy of one of the kernfs_ops to set its atomic_write_len.
cftype->kf_ops is added and cgroup_init/exit_cftypes() are updated
to handle it.
* cftype->lockdep_key added so that kernfs lockdep annotation can be
per cftype.
* Inidividual file entries and open states are now managed by kernfs.
No need to worry about them from cgroup. cfent, cgroup_open_file
and their friends are removed.
* kernfs_nodes are created deactivated and kernfs_activate()
invocations added to places where creation of new nodes are
committed.
* cgroup_rmdir() uses kernfs_[un]break_active_protection() for
self-removal.
v2: - Li pointed out in an earlier patch that specifying "name="
during mount without subsystem specification should succeed if
there's an existing hierarchy with a matching name although it
should fail with -EINVAL if a new hierarchy should be created.
Prior to the conversion, this used by handled by deferring
failure from NULL return from cgroup_root_from_opts(), which was
necessary because root was being created before checking for
existing ones. Note that cgroup_root_from_opts() returned an
ERR_PTR() value for error conditions which require immediate
mount failure.
As we now have separate search and creation steps, deferring
failure from cgroup_root_from_opts() is no longer necessary.
cgroup_root_from_opts() is updated to always return ERR_PTR()
value on failure.
- The logic to match existing roots is updated so that a mount
attempt with a matching name but different subsys_mask are
rejected. This was handled by a separate matching loop under
the comment "Check for name clashes with existing mounts" but
got lost during conversion. Merge the check into the main
search loop.
- Add __rcu __force casting in RCU_INIT_POINTER() in
cgroup_destroy_locked() to avoid the sparse address space
warning reported by kbuild test bot. Maybe we want an explicit
interface to use kn->priv as RCU protected pointer?
v3: Make CONFIG_CGROUPS select CONFIG_KERNFS.
v4: Rebased on top of 0ab02ca8f8 ("cgroup: protect modifications to
cgroup_idr with cgroup_mutex").
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Li Zefan <lizefan@huawei.com>
Cc: kbuild test robot fengguang.wu@intel.com>
2bd59d48eb