lkml.org 
[lkml]   [2018]   [Oct]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH] kasan: convert kasan/quarantine_lock to raw_spinlock
On Wed, Oct 10, 2018 at 11:29 AM, Sebastian Andrzej Siewior
<bigeasy@linutronix.de> wrote:
> On 2018-10-10 10:25:42 [+0200], Dmitry Vyukov wrote:
>> > That loop should behave like your on_each_cpu() except it does not
>> > involve the remote CPU.
>>
>>
>> The problem is that it can squeeze in between:
>>
>> + spin_unlock(&q->lock);
>>
>> spin_lock(&quarantine_lock);
>>
>> as far as I see. And then some objects can be left in the quarantine.
>
> Okay. But then once you are at CPU10 (in the on_each_cpu() loop) there
> can be objects which are added to CPU0, right? So based on that, I
> assumed that this would be okay to drop the lock here.

What happens here is a bit tricky.

When a kmem cache is freed, all objects from the cache must be already
free. That is kmem_cache_free has returned for all objects (otherwise
it's a user bug), these calls could have have happened on other CPUs.
So is we are freeing a cache on CPU10, it is not possible that CPU0
frees an object from this cache right now/concurrently, the objects
were already freed but they still can be sitting in per-cpu quarantine
caches.
Now a free on an unrelated object on CPU0 can decide to drain CPU
cache concurrently, it grabs whole CPU cache, unlocks the cache
spinlock (with your patch), and now proceeds to pushing the batch to
global quarantine list. At this point CPU10 drains quarantine to evict
all objects related to the cache, but it misses the batch that CPU0
transfers from cpu cache to global quarantine, because at this point
it's referenced from just local stack variable temp in quarantine_put.
Wrapping the whole transfer sequence with irq disable/enable, ensures
that the transfer happens atomically wrt quarantine_remove_cache. That
is, that quarantine_remove_cache will see the object either in cpu
cache or in global quarantine, but not somewhere in the flight.


>> > But this is debug code anyway, right? And it is highly complex imho.
>> > Well, maybe only for me after I looked at it for the first time…
>>
>> It is debug code - yes.
>> Nothing about its performance matters - no.
>>
>> That's the way to produce unusable debug tools.
>> With too much overhead timeouts start to fire and code behaves not the
>> way it behaves in production.
>> The tool is used in continuous integration and developers wait for
>> test results before merging code.
>> The tool is used on canary devices and directly contributes to usage experience.
>
> Completely understood. What I meant is that debug code in general (from
> RT perspective) increases latency to a level where the device can not
> operate. Take lockdep for instance. Debug facility which is required
> for RT as it spots locking problems early. It increases the latency
> (depending on the workload) by 50ms+ and can't be used in production.
> Same goes for SLUB debug and most other.
>
>> We of course don't want to trade a page of assembly code for cutting
>> few cycles here (something that could make sense for some networking
>> code maybe). But otherwise let's not introduce spinlocks on fast paths
>> just for refactoring reasons.
>
> Sure. As I said. I'm fine with patch Clark initially proposed. I assumed
> the refactoring would make things simpler and avoiding the cross-CPU
> call a good thing.
>
>> > Can you take it as-is or should I repost it with an acked-by?
>>
>> Perhaps it's the problem with the way RT kernel changes things then?
>> This is not specific to quarantine, right?
>
> We got rid of _a lot_ of local_irq_disable/save() + spin_lock() combos
> which were there for reasons which are no longer true or due to lack of
> the API. And this kasan thing is just something Clark stumbled upon
> recently. And I try to negotiate something where everyone can agree on.
>
>> Should that mutex detect
>> that IRQs are disabled and not try to schedule? If this would happen
>> in some networking code, what would we do?
>
> It is not only that it is supposed not to schedule. Assuming the "mutex"
> is not owned you could acquire it right away. No scheduling. However,
> you would record current() as the owner of the lock which is wrong and
> you get into other trouble later on. The list goes on :)
> However, networking. If there is something that breaks then it will be
> addressed. It will be forwarded upstream if this something where it
> is likely to assume that RT won't change. So networking isn't special.
>
> Should I repost Clark's patch?


I am much more comfortable with just changing the type of the lock.
What are the bad implications of using the raw spinlock? Will it help
to do something along the following lines:

// Because of ...
#if CONFIG_RT
#define quarantine_spinlock_t raw_spinlock_t
#else
#define quarantine_spinlock_t spinlock_t
#endif

?

\
 
 \ /
  Last update: 2018-10-10 11:46    [W:0.141 / U:1.152 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site