Re: [PATCH] ipc/sem: do not sleep with a spin lock held

From: Vasily Averin <vvs@virtuozzo.com>
To: Manfred Spraul <manfred@colorfullife.com>,
	cgel.zte@gmail.com, akpm@linux-foundation.org
Cc: shakeelb@google.com, rdunlap@infradead.org, dbueso@suse.de,
	unixbhaskar@gmail.com, chi.minghao@zte.com.cn, arnd@arndb.de,
	linux-kernel@vger.kernel.org, Zeal Robot <zealci@zte.com.cn>
Subject: Re: [PATCH] ipc/sem: do not sleep with a spin lock held
Date: Wed, 22 Dec 2021 22:08:39 +0300	[thread overview]
Message-ID: <bb788afc-b267-987c-278f-b844184f4767@virtuozzo.com> (raw)
In-Reply-To: <d89356b2-c7d8-a3d0-5d3f-1ef37b631193@virtuozzo.com>

On 22.12.2021 20:38, Vasily Averin wrote:
> On 22.12.2021 20:06, Manfred Spraul wrote:
>> Hi Vasily,
>>
>> On 12/22/21 16:50, Vasily Averin wrote:
>>> On 22.12.2021 18:31, Vasily Averin wrote:
>>>> On 22.12.2021 14:45, Manfred Spraul wrote:
>>>>> Hi Minghao,
>>>>>
>>>>> On 12/22/21 09:10, cgel.zte@gmail.com wrote:
>>>>>> From: Minghao Chi <chi.minghao@zte.com.cn>
>>>>>>
>>>>>> We can't call kvfree() with a spin lock held, so defer it.
>>>> I'm sorry, but I do not understand why exactly we cannot use kvfree?
>>>> Could you explain it in more details?
>>> Got it,
>>> there is cond_resched() called in __vfree() -> __vunmap()
>>>
>>> However I'm still not sure that in_interrupt() is used correctly here.
>>
>> I see three different topics:
>>
>> - is the current code violating the API? I think yes, thus there is a bug that needs to be fixed.
> 
> I'm agree. Found issue is a bug and it should be fixed ASAP,
> I'm sorry for a mistake in my patch.
> 
>> - Where is __vunmap() sleeping? Would it be possible to make __vunmap() safe to be called when owning a spinlock?
> 
> I think it is possible, and we should do it to prevent similar incidents in future.
> vfree() should check preempt count to detect this situation (i.e. execution under taken spinlock)
> generate WARN_ON and then call __vfree_deferred() to avoid sleep.
> 
>> - should kvfree() use vfree() [i.e. unsafe when owning a spinlock] or vfree_atomic [i.e. a bit slower, but safe]
> 
> I think it's better to change vfree.

I mean something like this:

--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -2674,7 +2674,7 @@ void vfree_atomic(const void *addr)
 
 static void __vfree(const void *addr)
 {
-       if (unlikely(in_interrupt()))
+       if (unlikely(in_atomic()))    <<<< VvS: do not sleep in atomic ...
                __vfree_deferred(addr);
        else
                __vunmap(addr, 1);
@@ -2703,7 +2703,7 @@ void vfree(const void *addr)
 
        kmemleak_free(addr);
 
-       might_sleep_if(!in_interrupt());
+       might_sleep_if(in_task());  <<<<< VvS: ... but generate warning if vfree was called
                                    <<<<<      in task context with taken spin_lock or spin_lock_bh
 
        if (!addr)
                return;


>> As we did quite many s/kfree/kvfree/ changes, perhaps just switching to vfree_atomic() is the best solution.
>>
>> @Andrew: What would you prefer?
>>
>> In addition, if we do not use vfree_atomic(): Then I would propose to copy the might_sleep_if() from vfree() into kvfree()
> I think it does not help, as far as I understand we are in task context, just under taken spinlock.
> 
> Thank you,
> 	vasily Averin
> 

