On 11/23/2012 01:20 PM, Michal Hocko wrote:
> On Thu 22-11-12 14:29:50, Glauber Costa wrote:
> [...]
>> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
>> index 05b87aa..46f7cfb 100644
>> --- a/mm/memcontrol.c
>> +++ b/mm/memcontrol.c
> [...]
>> @@ -349,6 +366,33 @@ struct mem_cgroup {
>> #endif
>> };
>>
>> +#if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MEMCG_SWAP)
>
> Can we have a common config for this something like CONFIG_MEMCG_ASYNC_DESTROY
> which would be selected if either of the two (or potentially others)
> would be selected.
> Also you are saying that the feature is only for debugging purposes so
> it shouldn't be on by default probably.
>
I personally wouldn't mind. But the big value I see from it is basically
being able to turn it off. For all the rest, we would have to wrap it
under one of those config options anyway...
>> +static LIST_HEAD(dangling_memcgs);
>> +static DEFINE_MUTEX(dangling_memcgs_mutex);
>> +
>> +static inline void memcg_dangling_free(struct mem_cgroup *memcg)
>> +{
>> + mutex_lock(&dangling_memcgs_mutex);
>> + list_del(&memcg->dead);
>> + mutex_unlock(&dangling_memcgs_mutex);
>> + kfree(memcg->memcg_name);
>> +}
>> +
>> +static inline void memcg_dangling_add(struct mem_cgroup *memcg)
>> +{
>> +
>> + memcg->memcg_name = kstrdup(cgroup_name(memcg->css.cgroup), GFP_KERNEL);
>
> Who gets charged for this allocation? What if the allocation fails (not
> that it would be probable but still...)?
>
Well, yeah, the lack of test is my bad - sorry.
As for charging, This will be automatically charged to whoever calls
mem_cgroup_destroy(). It is certainly not in the cgroup being destroyed,
otherwise it would have a task and destruction would not be possible.
But now that you mention, maybe it would be better to get it to the root
cgroup every time? This way this can't itself hold anybody in memory.
>> +
>> + INIT_LIST_HEAD(&memcg->dead);
>> + mutex_lock(&dangling_memcgs_mutex);
>> + list_add(&memcg->dead, &dangling_memcgs);
>> + mutex_unlock(&dangling_memcgs_mutex);
>> +}
>> +#else
>> +static inline void memcg_dangling_free(struct mem_cgroup *memcg) {}
>> +static inline void memcg_dangling_add(struct mem_cgroup *memcg) {}
>> +#endif
>> +
>> /* internal only representation about the status of kmem accounting. */
>> enum {
>> KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
>> @@ -4868,6 +4912,92 @@ static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
>> return simple_read_from_buffer(buf, nbytes, ppos, str, len);
>> }
>>
>> +#if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MEMCG_SWAP)
>> +static void
>> +mem_cgroup_dangling_swap(struct mem_cgroup *memcg, struct seq_file *m)
>> +{
>> +#ifdef CONFIG_MEMCG_SWAP
>> + u64 kmem;
>> + u64 memsw;
>> +
>> + /*
>> + * kmem will also propagate here, so we are only interested in the
>> + * difference. See comment in mem_cgroup_reparent_charges for details.
>> + *
>> + * We could save this value for later consumption by kmem reports, but
>> + * there is not a lot of problem if the figures differ slightly.
>> + */
>> + kmem = res_counter_read_u64(&memcg->kmem, RES_USAGE);
>> + memsw = res_counter_read_u64(&memcg->memsw, RES_USAGE) - kmem;
>> + seq_printf(m, "\t%llu swap bytes\n", memsw);
>> +#endif
>> +}
>> +
>> +static void
>> +mem_cgroup_dangling_kmem(struct mem_cgroup *memcg, struct seq_file *m)
>> +{
>> +#ifdef CONFIG_MEMCG_KMEM
>> + u64 kmem;
>> + struct memcg_cache_params *params;
>> +
>> +#ifdef CONFIG_INET
>> + struct tcp_memcontrol *tcp = &memcg->tcp_mem;
>> + s64 tcp_socks;
>> + u64 tcp_bytes;
>> +
>> + tcp_socks = percpu_counter_sum_positive(&tcp->tcp_sockets_allocated);
>> + tcp_bytes = res_counter_read_u64(&tcp->tcp_memory_allocated, RES_USAGE);
>> + seq_printf(m, "\t%llu tcp bytes, in %lld sockets\n",
>> + tcp_bytes, tcp_socks);
>> +
>> +#endif
>
> Looks like this deserves its own function rather than this ifdef games
> inside functions.
>
ok...
>> +
>> + kmem = res_counter_read_u64(&memcg->kmem, RES_USAGE);
>> + seq_printf(m, "\t%llu kmem bytes", kmem);
>> +
>> + /* list below may not be initialized, so not even try */
>> + if (!kmem)
>> + return;
>> +
>> + seq_printf(m, " in caches");
>> + mutex_lock(&memcg->slab_caches_mutex);
>> + list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
>> + struct kmem_cache *s = memcg_params_to_cache(params);
>> +
>> + seq_printf(m, " %s", s->name);
>> + }
>> + mutex_unlock(&memcg->slab_caches_mutex);
>> + seq_printf(m, "\n");
>> +#endif
>> +}
>> +
>> +/*
>> + * After a memcg is destroyed, it may still be kept around in memory.
>> + * Currently, the two main reasons for it are swap entries, and kernel memory.
>> + * Because they will be freed assynchronously, they will pin the memcg structure
>> + * and its resources until the last reference goes away.
>> + *
>> + * This root-only file will show information about which users
>> + */
>> +static int mem_cgroup_dangling_read(struct cgroup *cont, struct cftype *cft,
>> + struct seq_file *m)
>> +{
>> + struct mem_cgroup *memcg;
>> +
>> + mutex_lock(&dangling_memcgs_mutex);
>> +
>> + list_for_each_entry(memcg, &dangling_memcgs, dead) {
>> + seq_printf(m, "%s:\n", memcg->memcg_name);
>
> Hmm, we have lost the cgroup path so know there is something called A
> but we do not know whether it was A/A A/B/A A/......../A (aka we have
> lost the hierarchy information and a group with the same name might
> exist which can be really confusing).
>
Since all I store in the now-dead cgroup structure is a pointer, I guess
I could store the whole path...
> That being said I would prefer if this was covered by a debugging
> option, off by default.
Ok.
> It would be better if we could preserve the whole group name (something
> like cgroup_path does) but I guess this would break caches names, right?
I can't see how it would influence the cache names either way. I mean -
the effect of that would be that patches 1 and 2 here would be totally
independent, since we would be using cgroup_path instead of cgroup_name
in this facility.
> And finally it would be really nice if you described what is the
> exported information good for. Can I somehow change the current state
> (e.g. force freeing those objects so that the memcg can finally pass out
> in piece)?
>
I am open, but I would personally like to have this as a view-only
interface, just so we suspect a leak occurs, we can easily see what is
the dead memcg contribution to it. It shows you where the data come
from, and if you want to free it, you go search for subsystem-specific
ways to force a free should you want.
I really can't see anything good coming from being able to force changes
to the kernel from this interface.
--
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>