On Wed, 02 May 2018 14:21:35 -0500 ebied...@xmission.com (Eric W. Biederman)
wrote:
> Recently it was reported that mm_update_next_owner could get into
> cases where it was executing it's fallback for_each_process part of
> the loop and thus taking up a lot of time.
>
> To deal with this replace mm->owner with mm->memcg. This just reduces
> the complexity of everything. As much as possible I have maintained
> the current semantics. There are two siginificant exceptions. During
> fork the memcg of the process calling fork is charged rather than
> init_css_set. During memory cgroup migration the charges are migrated
> not if the process is the owner of the mm, but if the process being
> migrated has the same memory cgroup as the mm.
>
> I believe it was a bug if init_css_set is charged for memory activity
> during fork, and the old behavior was simply a consequence of the new
> task not having tsk->cgroup not initialized to it's proper cgroup.
>
> Durhing cgroup migration only thread group leaders are allowed to
> migrate. Which means in practice there should only be one. Linux
> tasks created with CLONE_VM are the only exception, but the common
> cases are already ruled out. Processes created with vfork have a
> suspended parent and can do nothing but call exec so they should never
> show up. Threads of the same cgroup are not the thread group leader
> so also should not show up. That leaves the old LinuxThreads library
> which is probably out of use by now, and someone doing something very
> creative with cgroups, and rolling their own threads with CLONE_VM.
> So in practice I don't think the difference charge migration will
> affect anyone.
>
> To ensure that mm->memcg is updated appropriately I have implemented
> cgroup "attach" and "fork" methods. This ensures that at those
> points the mm pointed to the task has the appropriate memory cgroup.
>
> For simplicity instead of introducing a new mm lock I simply use
> exchange on the pointer where the mm->memcg is updated to get
> atomic updates.
>
> Looking at the history effectively this change is a revert. The
> reason given for adding mm->owner is so that multiple cgroups can be
> attached to the same mm. In the last 8 years a second user of
> mm->owner has not appeared. A feature that has never used, makes the
> code more complicated and has horrible worst case performance should
> go.
Cleanliness nit: I'm not sure that the removal and open-coding of
mem_cgroup_from_task() actually improved things. Should we restore it?
--- a/mm/memcontrol.c~memcg-replace-mm-owner-with-mm-memcg-fix
+++ a/mm/memcontrol.c
@@ -664,6 +664,11 @@ static void memcg_check_events(struct me
}
}
+static inline struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
+{
+ return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
+}
+
struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
{
struct mem_cgroup *memcg = NULL;
@@ -1011,7 +1016,7 @@ bool task_in_mem_cgroup(struct task_stru
* killed to prevent needlessly killing additional tasks.
*/
rcu_read_lock();
- task_memcg = mem_cgroup_from_css(task_css(task,
memory_cgrp_id));
+ task_memcg = mem_cgroup_from_task(task);
css_get(&task_memcg->css);
rcu_read_unlock();
}
@@ -4829,7 +4834,7 @@ static int mem_cgroup_can_attach(struct
if (!move_flags)
return 0;
- from = mem_cgroup_from_css(task_css(p, memory_cgrp_id));
+ from = mem_cgroup_from_task(p);
VM_BUG_ON(from == memcg);
@@ -5887,7 +5892,7 @@ void mem_cgroup_sk_alloc(struct sock *sk
}
rcu_read_lock();
- memcg = mem_cgroup_from_css(task_css(current, memory_cgrp_id));
+ memcg = mem_cgroup_from_task(current);
if (memcg == root_mem_cgroup)
goto out;
if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
_
