Hi,
Imagine all of the QUARANTINE_BATCHES elements of
global_quarantine array is of size 4MB + 1MB, now a new call
to quarantine_put is invoked, one of the element will be of size 4MB +
1MB + 1MB, so on and on.
On Tue, Nov 28, 2017 at 4:58 PM, Dmitry Vyukov <[email protected]> wrote:
> On Tue, Nov 28, 2017 at 9:33 AM, Zhouyi Zhou <[email protected]> wrote:
>> Hi,
>> Please take a look at function quarantine_put, I don't think following
>> code will limit the batch size below quarantine_batch_size. It only advance
>> quarantine_tail after qlist_move_all.
>>
>> qlist_move_all(q, &temp);
>>
>> spin_lock(&quarantine_lock);
>> WRITE_ONCE(quarantine_size, quarantine_size + temp.bytes);
>> qlist_move_all(&temp, &global_quarantine[quarantine_tail]);
>> if (global_quarantine[quarantine_tail].bytes >=
>> READ_ONCE(quarantine_batch_size)) {
>> int new_tail;
>>
>> new_tail = quarantine_tail + 1;
>> if (new_tail == QUARANTINE_BATCHES)
>> new_tail = 0;
>> if (new_tail != quarantine_head)
>> quarantine_tail = new_tail;
>
>
> As far as I see this code can exceed global quarantine batch size by
> at most 1 per-cpu batch. Per-cpu batch is caped at 1MB. So max global
> batch size will be 4MB+1MB. Which does not radically change situation.
>
>
>> On Tue, Nov 28, 2017 at 4:12 PM, Dmitry Vyukov <[email protected]> wrote:
>>> On Tue, Nov 28, 2017 at 9:00 AM, Zhouyi Zhou <[email protected]> wrote:
>>>> Thanks for reviewing
>>>> My machine has 128G of RAM, and runs many KVM virtual machines.
>>>> libvirtd always
>>>> report "internal error: received hangup / error event on socket" under
>>>> heavy memory load.
>>>> Then I use perf top -g, qlist_move_cache consumes 100% cpu for
>>>> several minutes.
>>>
>>> For 128GB of RAM, batch size is 4MB. Processing such batch should not
>>> take more than few ms. So I am still struggling to understand how/why
>>> your change helps and why there are issues in the first place...
>>>
>>>
>>>
>>>> On Tue, Nov 28, 2017 at 3:45 PM, Dmitry Vyukov <[email protected]> wrote:
>>>>> On Tue, Nov 28, 2017 at 5:05 AM, Zhouyi Zhou <[email protected]> wrote:
>>>>>> When there are huge amount of quarantined cache allocates in system,
>>>>>> number of entries in global_quarantine[i] will be great. Meanwhile,
>>>>>> there is no relax in while loop in function qlist_move_cache which
>>>>>> hold quarantine_lock. As a result, some userspace programs for example
>>>>>> libvirt will complain.
>>>>>
>>>>> Hi,
>>>>>
>>>>> The QUARANTINE_BATCHES thing was supposed to fix this problem, see
>>>>> quarantine_remove_cache() function.
>>>>> What is the amount of RAM and number of CPUs in your system?
>>>>> If system has 4GB of RAM, quarantine size is 128MB and that's split
>>>>> into 1024 batches. Batch size is 128KB. Even if that's filled with the
>>>>> smallest objects of size 32, that's only 4K objects. And there is a
>>>>> cond_resched() between processing of every batch.
>>>>> I don't understand why it causes problems in your setup. We use KASAN
>>>>> extremely heavily on hundreds of machines 24x7 and we have not seen
>>>>> any single report from this code...
>>>>>
>>>>>
>>>>>> On Tue, Nov 28, 2017 at 12:04 PM, <[email protected]> wrote:
>>>>>>> From: Zhouyi Zhou <[email protected]>
>>>>>>>
>>>>>>> This patch fix livelock by conditionally release cpu to let others
>>>>>>> has a chance to run.
>>>>>>>
>>>>>>> Tested on x86_64.
>>>>>>> Signed-off-by: Zhouyi Zhou <[email protected]>
>>>>>>> ---
>>>>>>> mm/kasan/quarantine.c | 12 +++++++++++-
>>>>>>> 1 file changed, 11 insertions(+), 1 deletion(-)
>>>>>>>
>>>>>>> diff --git a/mm/kasan/quarantine.c b/mm/kasan/quarantine.c
>>>>>>> index 3a8ddf8..33eeff4 100644
>>>>>>> --- a/mm/kasan/quarantine.c
>>>>>>> +++ b/mm/kasan/quarantine.c
>>>>>>> @@ -265,10 +265,13 @@ static void qlist_move_cache(struct qlist_head *from,
>>>>>>> struct kmem_cache *cache)
>>>>>>> {
>>>>>>> struct qlist_node *curr;
>>>>>>> + struct qlist_head tmp_head;
>>>>>>> + unsigned long flags;
>>>>>>>
>>>>>>> if (unlikely(qlist_empty(from)))
>>>>>>> return;
>>>>>>>
>>>>>>> + qlist_init(&tmp_head);
>>>>>>> curr = from->head;
>>>>>>> qlist_init(from);
>>>>>>> while (curr) {
>>>>>>> @@ -278,10 +281,17 @@ static void qlist_move_cache(struct qlist_head *from,
>>>>>>> if (obj_cache == cache)
>>>>>>> qlist_put(to, curr, obj_cache->size);
>>>>>>> else
>>>>>>> - qlist_put(from, curr, obj_cache->size);
>>>>>>> + qlist_put(&tmp_head, curr, obj_cache->size);
>>>>>>>
>>>>>>> curr = next;
>>>>>>> +
>>>>>>> + if (need_resched()) {
>>>>>>> + spin_unlock_irqrestore(&quarantine_lock, flags);
>>>>>>> + cond_resched();
>>>>>>> + spin_lock_irqsave(&quarantine_lock, flags);
>>>>>>> + }
>>>>>>> }
>>>>>>> + qlist_move_all(&tmp_head, from);
>>>>>>> }
>>>>>>>
>>>>>>> static void per_cpu_remove_cache(void *arg)
>>>>>>> --
>>>>>>> 2.1.4
>>>>>>>
>>>>>>
>>>>>> --
>>>>>> You received this message because you are subscribed to the Google Groups "kasan-dev" group.
>>>>>> To unsubscribe from this group and stop receiving emails from it, send an email to [email protected].
>>>>>> To post to this group, send email to [email protected].
>>>>>> To view this discussion on the web visit https://groups.google.com/d/msgid/kasan-dev/CAABZP2zEup53ZcNKOEUEMx_aRMLONZdYCLd7s5J4DLTccPxC-A%40mail.gmail.com.
>>>>>> For more options, visit https://groups.google.com/d/optout.
From 1585299694835864432@xxx Tue Nov 28 09:00:02 +0000 2017
X-GM-THRID: 1585281151451498932
X-Gmail-Labels: Inbox,Category Forums,HistoricalUnread