Hi Vladimir,
On Mon, Dec 14, 2015 at 08:14:55PM +0300, Vladimir Davydov wrote:
> On Fri, Dec 11, 2015 at 02:54:13PM -0500, Johannes Weiner wrote:
> ...
> > -static int
> > -mem_cgroup_css_online(struct cgroup_subsys_state *css)
> > +static struct cgroup_subsys_state * __ref
> > +mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
> > {
> > - struct mem_cgroup *memcg = mem_cgroup_from_css(css);
> > - struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
> > - int ret;
> > -
> > - if (css->id > MEM_CGROUP_ID_MAX)
> > - return -ENOSPC;
> > + struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
> > + struct mem_cgroup *memcg;
> > + long error = -ENOMEM;
> >
> > - if (!parent)
> > - return 0;
> > + memcg = mem_cgroup_alloc();
> > + if (!memcg)
> > + return ERR_PTR(error);
> >
> > mutex_lock(&memcg_create_mutex);
>
> It is pointless to take memcg_create_mutex in ->css_alloc. It won't
> prevent setting use_hierarchy for parent after a new child was
> allocated, but before it was added to the list of children (see
> create_css()). Taking the mutex in ->css_online renders this race
> impossible. That is, your cleanup breaks use_hierarchy consistency
> check.
>
> Can we drop this use_hierarchy consistency check at all and allow
> children of a cgroup with use_hierarchy=1 have use_hierarchy=0? Yeah,
> that might result in some strangeness if cgroups are created in parallel
> with use_hierarchy flipped, but is it a valid use case? I surmise, one
> just sets use_hierarchy for a cgroup once and for good before starting
> to create sub-cgroups.
I don't think we have to support airtight exclusion between somebody
changing the parent attribute and creating new children that inherit
these attributes. Everything will still work if this race happens.
Does that mean we have to remove the restriction altogether? I'm not
convinced. We can just keep it for historical purposes so that we do
not *encourage* this weird setting.
I think that's good enough. Let's just remove the memcg_create_mutex.
> > - memcg->use_hierarchy = parent->use_hierarchy;
> > - memcg->oom_kill_disable = parent->oom_kill_disable;
> > - memcg->swappiness = mem_cgroup_swappiness(parent);
> > -
> > - if (parent->use_hierarchy) {
> > + memcg->high = PAGE_COUNTER_MAX;
> > + memcg->soft_limit = PAGE_COUNTER_MAX;
> > + if (parent)
> > + memcg->swappiness = mem_cgroup_swappiness(parent);
> > + if (parent && parent->use_hierarchy) {
> > + memcg->use_hierarchy = true;
> > + memcg->oom_kill_disable = parent->oom_kill_disable;
>
> oom_kill_disable was propagated to child cgroup despite use_hierarchy
> configuration. I don't see any reason to change this.
Just a weird quirk of !use_hierarchy, I guess. Technically, if you're
not in hierarchy mode then your parent is the root, not whatever is in
the css tree above you. And root can't disable OOM killing. So I don't
really care either way. Restored to previous behavior.
> > @@ -4296,19 +4211,30 @@ mem_cgroup_css_online(struct cgroup_subsys_state *css)
> > }
> > mutex_unlock(&memcg_create_mutex);
> >
> > - ret = memcg_propagate_kmem(memcg);
> > - if (ret)
> > - return ret;
> > + /* The following stuff does not apply to the root */
> > + if (!parent) {
> > + root_mem_cgroup = memcg;
> > + return &memcg->css;
> > + }
> > +
> > + error = memcg_propagate_kmem(parent, memcg);
>
> I don't think ->css_alloc is the right place for this function: if
> create_css() fails after ->css_alloc and before ->css_online, it'll call
> ->css_free, which won't cleanup kmem properly.
Hm, on the other hand, the static branch maintenance is done from
css_free. If online doesn't run we'll have an underflow there.
I think the easiest way to deal with this is to have memcg_free_kmem()
check if offlining has happened, and if not call memcg_offline_kmem().
> > + if (error)
> > + goto fail;
> >
> > if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
> > static_branch_inc(&memcg_sockets_enabled_key);
>
> Frankly, I don't get why this should live here either. This has nothing
> to do with memcg allocation and looks rather like a preparation for
> online.
It matches css_free() as per above. Unless we need synchronization
with showing up as a child in the parent, there is no need to use
online and fragment the initialization. So I'd rather consolidate.
> > @@ -4347,9 +4270,11 @@ static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
> > if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
> > static_branch_dec(&memcg_sockets_enabled_key);
> >
> > + vmpressure_cleanup(&memcg->vmpressure);
>
> vmpressure->work can be scheduled after offline, so ->css_free is
> definitely the right place for vmpressure_cleanup. Looks like you've
> just fixed a potential use-after-free bug.
Yay! Since nobody reported this problem, it should be okay to keep it
in this patch.
So, how about the following fixlets on top to address your comments?
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index af8714a..124a802 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -250,13 +250,6 @@ enum res_type {
/* Used for OOM nofiier */
#define OOM_CONTROL (0)
-/*
- * The memcg_create_mutex will be held whenever a new cgroup is created.
- * As a consequence, any change that needs to protect against new child cgroups
- * appearing has to hold it as well.
- */
-static DEFINE_MUTEX(memcg_create_mutex);
-
/* Some nice accessors for the vmpressure. */
struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
{
@@ -2660,14 +2653,6 @@ static inline bool memcg_has_children(struct mem_cgroup *memcg)
{
bool ret;
- /*
- * The lock does not prevent addition or deletion of children, but
- * it prevents a new child from being initialized based on this
- * parent in css_online(), so it's enough to decide whether
- * hierarchically inherited attributes can still be changed or not.
- */
- lockdep_assert_held(&memcg_create_mutex);
-
rcu_read_lock();
ret = css_next_child(NULL, &memcg->css);
rcu_read_unlock();
@@ -2730,10 +2715,8 @@ static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
struct mem_cgroup *memcg = mem_cgroup_from_css(css);
struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
- mutex_lock(&memcg_create_mutex);
-
if (memcg->use_hierarchy == val)
- goto out;
+ return 0;
/*
* If parent's use_hierarchy is set, we can't make any modifications
@@ -2752,9 +2735,6 @@ static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
} else
retval = -EINVAL;
-out:
- mutex_unlock(&memcg_create_mutex);
-
return retval;
}
@@ -2929,6 +2909,10 @@ static void memcg_offline_kmem(struct mem_cgroup *memcg)
static void memcg_free_kmem(struct mem_cgroup *memcg)
{
+ /* css_alloc() failed, offlining didn't happen */
+ if (unlikely(memcg->kmem_state == KMEM_ONLINE))
+ memcg_offline_kmem(memcg);
+
if (memcg->kmem_state == KMEM_ALLOCATED) {
memcg_destroy_kmem_caches(memcg);
static_branch_dec(&memcg_kmem_enabled_key);
@@ -2956,11 +2940,9 @@ static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
mutex_lock(&memcg_limit_mutex);
/* Top-level cgroup doesn't propagate from root */
if (!memcg_kmem_online(memcg)) {
- mutex_lock(&memcg_create_mutex);
if (cgroup_is_populated(memcg->css.cgroup) ||
(memcg->use_hierarchy && memcg_has_children(memcg)))
ret = -EBUSY;
- mutex_unlock(&memcg_create_mutex);
if (ret)
goto out;
ret = memcg_online_kmem(memcg);
@@ -4184,14 +4166,14 @@ mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
if (!memcg)
return ERR_PTR(error);
- mutex_lock(&memcg_create_mutex);
memcg->high = PAGE_COUNTER_MAX;
memcg->soft_limit = PAGE_COUNTER_MAX;
- if (parent)
+ if (parent) {
memcg->swappiness = mem_cgroup_swappiness(parent);
+ memcg->oom_kill_disable = parent->oom_kill_disable;
+ }
if (parent && parent->use_hierarchy) {
memcg->use_hierarchy = true;
- memcg->oom_kill_disable = parent->oom_kill_disable;
page_counter_init(&memcg->memory, &parent->memory);
page_counter_init(&memcg->memsw, &parent->memsw);
page_counter_init(&memcg->kmem, &parent->kmem);
@@ -4209,7 +4191,6 @@ mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
if (parent != root_mem_cgroup)
memory_cgrp_subsys.broken_hierarchy = true;
}
- mutex_unlock(&memcg_create_mutex);
/* The following stuff does not apply to the root */
if (!parent) {
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>