On 02/14/2018 01:02 PM, Jason Wang wrote:
> On 2018年02月14日 19:51, Michal Hocko wrote:
>> On Wed 14-02-18 19:47:30, Jason Wang wrote:
>>> On 2018年02月14日 17:28, Daniel Borkmann wrote:
>>>> [ +Jason, +Jesper ]
>>>> On 02/14/2018 09:43 AM, Michal Hocko wrote:
>>>>> On Tue 13-02-18 18:55:33, Matthew Wilcox wrote:
>>>>>> On Tue, Feb 13, 2018 at 03:59:01PM -0800, syzbot wrote:
>>>>> [...]
>>>>>>>    kvmalloc include/linux/mm.h:541 [inline]
>>>>>>>    kvmalloc_array include/linux/mm.h:557 [inline]
>>>>>>>    __ptr_ring_init_queue_alloc include/linux/ptr_ring.h:474 [inline]
>>>>>>>    ptr_ring_init include/linux/ptr_ring.h:492 [inline]
>>>>>>>    __cpu_map_entry_alloc kernel/bpf/cpumap.c:359 [inline]
>>>>>>>    cpu_map_update_elem+0x3c3/0x8e0 kernel/bpf/cpumap.c:490
>>>>>>>    map_update_elem kernel/bpf/syscall.c:698 [inline]
>>>>>> Blame the BPF people, not the MM people ;-)
>>>> Heh, not really. ;-)
>>>>> Yes. kvmalloc (the vmalloc part) doesn't support GFP_ATOMIC semantic.
>>>> Agree, that doesn't work.
>>>> Bug was added in commit 0bf7800f1799 ("ptr_ring: try vmalloc() when 
>>>> kmalloc() fails").
>>>> Jason, please take a look at fixing this, thanks!
>>> It looks to me the only solution is to revert that commit.
>> Do you really need this to be GFP_ATOMIC? I can see some callers are
>> under RCU read lock but can we perhaps do the allocation outside of this
>> section?
> If I understand the code correctly, the code would be called by XDP program 
> (usually run inside a bh) which makes it hard to do this.
> Rethink of this, we can probably test gfp and not call kvmalloc if GFP_ATOMIC 
> is set in __ptr_ring_init_queue_alloc().

That would be one option indeed (probably useful in any case to make the API
more robust). Another one is to just not use GFP_ATOMIC in cpumap. Looking at
it, update can neither be called out of a BPF prog since prevented by verifier
nor under RCU reader side when updating this type of map from syscall path.
Jesper, any concrete reason we still need GFP_ATOMIC here?

Reply via email to