Re: [PATCH v2] netfilter: ipset: Fix sleeping memory allocation in atomic context

[Nikolay Borisov <kernel@kyup.com>]

On 10/15/2015 05:32 PM, Eric Dumazet wrote:
> On Thu, 2015-10-15 at 16:41 +0300, Nikolay Borisov wrote:
>>
>> On 10/15/2015 04:32 PM, Eric Dumazet wrote:
>>> On Thu, 2015-10-15 at 13:56 +0300, Nikolay Borisov wrote:
>>>> Commit 00590fdd5be0 introduced RCU locking in list type and in
>>>> doing so introduced a memory allocation in list_set_add, which
>>>> results in the following splat:
>>>>
>>>> BUG: sleeping function called from invalid context at mm/page_alloc.c:2759
>>>> in_atomic(): 1, irqs_disabled(): 0, pid: 9664, name: ipset
>>>> CPU: 18 PID: 9664 Comm: ipset Tainted: G           O 3.12.47-clouder3 #1
>>>> Hardware name: Supermicro X10DRi/X10DRi, BIOS 1.1 04/14/2015
>>>>  0000000000000002 ffff881fd14273c8 ffffffff8163d891 ffff881fcb4264b0
>>>>  ffff881fcb4260c0 ffff881fd14273e8 ffffffff810ba5bf ffff881fd1427558
>>>>  0000000000000000 ffff881fd1427568 ffffffff81142b33 ffff881f00000000
>>>> Call Trace:
>>>>  [<ffffffff8163d891>] dump_stack+0x58/0x7f
>>>>  [<ffffffff810ba5bf>] __might_sleep+0xdf/0x110
>>>>  [<ffffffff81142b33>] __alloc_pages_nodemask+0x243/0xc20
>>>>  [<ffffffff81181c6e>] alloc_pages_current+0xbe/0x170
>>>>  [<ffffffff81188315>] new_slab+0x295/0x340
>>>>  [<ffffffff81189a40>] __slab_alloc+0x2c0/0x5a0
>>>>  [<ffffffff8164000c>] ? __schedule+0x2dc/0x760
>>>>  [<ffffffff8118a71b>] __kmalloc+0x11b/0x230
>>>>  [<ffffffffa02bd0ac>] ? ip_set_get_byname+0xec/0x100 [ip_set]
>>>>  [<ffffffffa02d23fb>] list_set_uadd+0x16b/0x314 [ip_set_list_set]
>>>>  [<ffffffff81642148>] ? _raw_write_unlock_bh+0x28/0x30
>>>>  [<ffffffffa02d1cfc>] list_set_uadt+0x21c/0x320 [ip_set_list_set]
>>>>  [<ffffffffa02d2290>] ? list_set_create+0x1a0/0x1a0 [ip_set_list_set]
>>>>  [<ffffffffa02be242>] call_ad+0x82/0x200 [ip_set]
>>>>  [<ffffffffa02bb171>] ? find_set_type+0x51/0xa0 [ip_set]
>>>>  [<ffffffff8133f275>] ? nla_parse+0xf5/0x130
>>>>  [<ffffffffa02be8ae>] ip_set_uadd+0x20e/0x2d0 [ip_set]
>>>>  [<ffffffffa02be013>] ? ip_set_create+0x2a3/0x450 [ip_set]
>>>>  [<ffffffffa02be6a0>] ? ip_set_udel+0x2e0/0x2e0 [ip_set]
>>>>  [<ffffffff815b316e>] nfnetlink_rcv_msg+0x31e/0x330
>>>>  [<ffffffff815b2e91>] ? nfnetlink_rcv_msg+0x41/0x330
>>>>  [<ffffffff815b2e50>] ? nfnl_lock+0x30/0x30
>>>>  [<ffffffff815ae179>] netlink_rcv_skb+0xa9/0xd0
>>>>  [<ffffffff815b2d45>] nfnetlink_rcv+0x15/0x20
>>>>  [<ffffffff815ade5f>] netlink_unicast+0x10f/0x190
>>>>  [<ffffffff815aedb0>] netlink_sendmsg+0x2c0/0x660
>>>>  [<ffffffff81567f00>] sock_sendmsg+0x90/0xc0
>>>>  [<ffffffff81565b03>] ? move_addr_to_user+0xa3/0xc0
>>>>  [<ffffffff81568552>] ? ___sys_recvmsg+0x182/0x300
>>>>  [<ffffffff81568064>] SYSC_sendto+0x134/0x180
>>>>  [<ffffffff811c4e01>] ? mntput+0x21/0x30
>>>>  [<ffffffff81572d2f>] ? __kfree_skb+0x3f/0xa0
>>>>  [<ffffffff815680be>] SyS_sendto+0xe/0x10
>>>>  [<ffffffff816434b2>] system_call_fastpath+0x16/0x1b
>>>>
>>>> The call chain leading to this is as follow:
>>>> call_ad -> list_set_uadt -> list_set_uadd -> kzalloc(, GFP_KERNEL).
>>>> And since GFP_KERNEL allows initiating direct reclaim thus
>>>> potentially sleeping in the allocation path, this leads to the
>>>> aforementioned splat.
>>>>
>>>> To fix it change the allocation type to GFP_ATOMIC, to
>>>> correctly reflect that it is occuring in an atomic context.
>>>>
>>>> Fixes: 00590fdd5be0 ("netfilter: ipset: Introduce RCU locking in list type")
>>>>
>>>> Acked-by: Jozsef Kadlecsik <kadlec@blackhole.kfki.hu>
>>>> Signed-off-by: Nikolay Borisov <kernel@kyup.com>
>>>> ---
>>>>
>>>> Changes since V1: 
>>>>  * Added acked-by 
>>>>  * Fixed patch header 
>>>>
>>>>  net/netfilter/ipset/ip_set_list_set.c | 2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/net/netfilter/ipset/ip_set_list_set.c b/net/netfilter/ipset/ip_set_list_set.c
>>>> index a1fe537..5a30ce6 100644
>>>> --- a/net/netfilter/ipset/ip_set_list_set.c
>>>> +++ b/net/netfilter/ipset/ip_set_list_set.c
>>>> @@ -297,7 +297,7 @@ list_set_uadd(struct ip_set *set, void *value, const struct ip_set_ext *ext,
>>>>  	      ip_set_timeout_expired(ext_timeout(n, set))))
>>>>  		n =  NULL;
>>>>  
>>>> -	e = kzalloc(set->dsize, GFP_KERNEL);
>>>> +	e = kzalloc(set->dsize, GFP_ATOMIC);
>>>>  	if (!e)
>>>>  		return -ENOMEM;
>>>>  	e->id = d->id;
>>>
>>> This patch looks very bogus to me.
>>>
>>> Could we fix the root cause please ?
>>>
>>> Root cause is that somewhere in this controlling path, an erroneous
>>> rcu_read_lock() is used, while it is very probably not needed, as
>>> controlling path should be protected by a mutex, which definitely is
>>> sane, because it allows us to perform GFP_KERNEL allocations and being
>>> preempted.
>>>
>>> Why are we using rcu_read_lock() in list_set_list() ?
>>>
>>> This looks as yet another bit of 'let us throw
>>> rcu_read_lock()/rcu_read_unlock() pairs' all over the places because it
>>> feels so good.
>>
>> I did check the call paths and there isn't an rcu_read_lock called in
>> list_set_uadt/list_set_uadd. On the contrary, this "write" operation to
>> the list is being serialised in call_ad() via set->lock spin_lock.
>>
>> What am I missing here?
> 
> I was not complaining to you, but to Jozsef ;)
> 
> Looking at commit 00590fdd5be0d7 terse changelog, we have to look at
> whole commit, and find suspicious rcu_read_lock()
> 
> Apparently, before this commit, allocations were safe, in process
> context (and using GFP_KERNEL), but after the commit, we have the
> unfortunate side effect of having potential burst of allocations
> done from bh context, using the special reserves dedicated to GFP_ATOMIC
> true users (processing packets from softirq handlers)

I guess the reason why a spinlock is used is due to the gc code which is
responsible for reaping entries whose timeout has elapsed. All of this
is happening in set_cleanup_entries. Given the state of things I don't
think using a mutex is a feasible solution unless the gc mechanism is
reworked.

> 
> I hate when we add more GFP_ATOMIC allocations all over the places,
> as this increases the risk of depleting memory reserves.
> 
> Can the spinlock be converted to a mutex ? If not, then instead of
> putting a stack trace in your changelog, a good explanation would be
> more useful.
> 
> Thanks !
> 
> 
>