Reduction globality of tasklist_lock

tasklist_lock is used for protection of many different task links. These are place in init_task.tasks, parent-child relationship, ptrace waiting, place in PID lists, SID/PGID leadership, proibition of creation of new tasks etc. It's like gone BKL, and this badly affects on granularity of a system.

The series aims to decrease its globality and introduces a new lock for protection of task-task relationship. The lock's name is kin_lock. Firstly, it protects parent-child relationship. Children sibling and real_parent fields are protected by father's kin_lock. For transfering a child from one parent to another you should take both fathers locks.

Child's parent is protected by ptracer's kin_lock. We should take real_parent's and ptracer's locks to attach a child to ptracer. We should use parent's lock to notify it about exiting. tasklist_lock is not used in exit/wait notifications since now.

Also, real_parent's kin lock protects child's threads (is the child is multithreaded) and multithread exiting. Task's sighand is protected by it too. After all I nested tasklist_lock under kin_lock, so the lock order now is

		kin_lock
		tasklist_lock
		sighand->lock

But. sighand, task_struct::tasks, thread group is still safe under tasklist_lock. We may change __exit_signal() a little bit more, but it wants additional changing of all tasklist_lock users and makes the series bigger (plus 8-9 patches). I don't thing it's so necessary right now. We may do that in the future if we want. All new users shouldn't use tasklist_lock since now if possible (but mostly current users may be easy rewritten using RCU. In a couple of place we will have to use tasklist_lock to stop process creation).

Besides that, tasklist_lock still protects init_task.tasks list, PID lists, SID and PGID leadership.

The series is in RFC format, because I didn't add exhaustive comments to the code yet. Also, nesting of rwlock_t isn't reflected in lockdep, and arch code still uses tasklist_lock (I hadn't analyze it yet).

I'd like to hear people's opinions about that. Welcome your comments/ ideas/suggestions.

Thanks!

---

Kirill Tkhai (13):
exit: Clarify choice of new parent in forget_original_parent()
rwlock_t: Implement double_write_{,un}lock()
pid_ns: Implement rwlock_t pid_ns::cr_lock for locking child_reaper
exit: Small refactoring mm_update_next_owner()
fs: Refactoring in get_children_pid()
core: Add rwlock_t task_list::kin_lock
kin_lock: Implement helpers for kin_lock locking.
core: Use kin_lock synchronizations between parent and child and for thread group
exit: Use for_each_thread() in do_wait()
exit: Add struct wait_opts's member held_lock and use it for tasklist_lock
exit: Syncronize on kin_lock while do_notify_parent()
exit: Delete write dependence on tasklist_lock in exit_notify()
core: Nest tasklist_lock into task_struct::kin_lock

 fs/exec.c                     |   26 ++--
 fs/proc/array.c               |   28 ++--
 include/linux/init_task.h     |    1 
 include/linux/pid_namespace.h |    1 
 include/linux/sched.h         |  303 +++++++++++++++++++++++++++++++++++++++++
 include/linux/spinlock.h      |   19 +++
 kernel/exit.c                 |  178 +++++++++++++++++-------
 kernel/fork.c                 |   13 +-
 kernel/pid.c                  |   10 +
 kernel/pid_namespace.c        |    5 -
 kernel/ptrace.c               |   53 +++++--
 kernel/signal.c               |   20 +--
 kernel/sys.c                  |   19 +--
 mm/oom_kill.c                 |    9 +
 security/keys/keyctl.c        |    4 -
 security/selinux/hooks.c      |    4 -
 16 files changed, 570 insertions(+), 123 deletions(-)

--

Yes...

Kirill. I just want to let you know that I am not going to ignore this series. I'll try very much to review it this week.

But I have to admit this scares me ;) and I don't feel this all is right. We will see.

Oleg.

Hi, Oleg,

26.05.2015, 22:44, "Oleg Nesterov" <[email protected]>:

Nice to hear. I'll wait for your commentaries.
A couple of words how this all was tested.

The strace test was endless kernel compilation using standard or straced gcc/make (CC="strace -o /dev/null gcc" ) and periodical killing gcc, make or strace. No problems were observed. If there is an idea of a test, which is mostly for threads, I'll be happy to hear it.

Thanks!
Kirill

27.05.2015, 12:30, "Kirill Tkhai" <[email protected]>:

I found some problems in the patchset and going to fix them. Please, ignore it. I'll send new series next week.

Kirill