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

[Nikolay Borisov <kernel@kyup.com>]

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?


> 
> 
> 
>