Re: [kmemleak] unreferenced object 0xcd9c1a80 (size 192):

[Yonghong Song <yhs@fb.com>]

On 2/12/18 7:55 AM, Alexei Starovoitov wrote:
> On Mon, Feb 12, 2018 at 09:28:33AM +0100, Daniel Borkmann wrote:
>> On 02/12/2018 06:47 AM, Yonghong Song wrote:
>>> On 2/11/18 11:18 AM, Mathieu Malaterre wrote:
>>>> On Sun, Feb 11, 2018 at 5:54 PM, Alexei Starovoitov
>>>> <alexei.starovoitov@gmail.com> wrote:
>>>>> On Sun, Feb 11, 2018 at 7:24 AM, Mathieu Malaterre <malat@debian.org> wrote:
>>>>>> Alexei,
>>>>>>
>>>>>> Could you please comment on why I am seeing those memleaks being
>>>>>> reported on my ppc32 system ? Should they be marked as false positive
>>>>>> ?
>>>>>>
>>>>>> System is Mac Mini G4, git/master (4.15.0+), ppc.
>>>>>>
>>>>>> Thanks for your time
>>>>>>
>>>>>> $ dmesg
>>>>>> ...
>>>>>> [ 1281.504173] kmemleak: 36 new suspected memory leaks (see
>>>>>> /sys/kernel/debug/kmemleak)
>>>>>>
>>>>>> Where:
>>>>>>
>>>>>> # cat /sys/kernel/debug/kmemleak
>>>>>> unreferenced object 0xdee25000 (size 192):
>>>>>>     comm "systemd", pid 1, jiffies 4294894348 (age 1438.580s)
>>>>>>     hex dump (first 32 bytes):
>>>>>>       c0 56 2f 88 00 00 00 00 00 00 00 0b 00 00 00 0c  .V/.............
>>>>>>       00 00 00 08 00 00 00 01 00 00 00 01 00 00 00 01  ................
>>>>>>     backtrace:
>>>>>>       [<6c69baf5>] trie_alloc+0xb0/0x150
>>>>>>       [<fa093284>] SyS_bpf+0x288/0x1458
>>>>>>       [<82182f53>] ret_from_syscall+0x0/0x38
>>>>>> unreferenced object 0xdee25900 (size 192):
>>>>>>     comm "systemd", pid 1, jiffies 4294894540 (age 1437.812s)
>>>>>>     hex dump (first 32 bytes):
>>>>>>       c0 56 2f 88 00 00 00 00 00 00 00 0b 00 00 00 08  .V/.............
>>>>>>       00 00 00 08 00 00 00 01 00 00 00 01 00 00 00 01  ................
>>>>>>     backtrace:
>>>>>>       [<6c69baf5>] trie_alloc+0xb0/0x150
>>>>>>       [<fa093284>] SyS_bpf+0x288/0x1458
>>>>>>       [<82182f53>] ret_from_syscall+0x0/0x38
>>>>>
>>>>> hmm. looks real. Is there a reproducer?
>>>>> Yonghong, lpm map not cleaning after itself?
>>>>
>>>> Not really. I simply boot up my machine and wait for the first kmemleak scan.
>>>
>>> I am not able to reproduce the issue. Tried with latest net-next on FC26 with kmemleak on. I only got this one after bootup,
>>> 'cat /sys/kernel/debug/kmemleak' or
>>> 'echo scan > /sys/kernel/debug/kmemleak
>>>   cat /sys/kernel/debug/kmemleak':
>>>
>>> unreferenced object 0xffff99701a7386e0 (size 32):
>>>    comm "mount", pid 1856, jiffies 4294669263 (age 98.440s)
>>>    hex dump (first 32 bytes):
>>>      00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
>>>      00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
>>>    backtrace:
>>>      [<000000004668ec00>] security_sb_parse_opts_str+0x36/0x50
>>>      [<00000000a9807d2b>] parse_security_options+0x3d/0x60
>>>      [<00000000cc1e1d58>] btrfs_mount_root+0x139/0x720
>>>      [<00000000bdc4f1a3>] mount_fs+0x30/0x150
>>>      [<00000000f189f1bd>] vfs_kern_mount.part.26+0x54/0x100
>>>      [<0000000093ae5db7>] btrfs_mount+0x184/0x914
>>>      [<00000000bdc4f1a3>] mount_fs+0x30/0x150
>>>      [<00000000f189f1bd>] vfs_kern_mount.part.26+0x54/0x100
>>>      [<000000003b67b9fc>] do_mount+0x5b9/0xc70
>>>      [<00000000de4073a0>] SyS_mount+0x80/0xd0
>>>      [<00000000fc5a968a>] do_syscall_64+0x5d/0x110
>>>      [<000000003d61f5fc>] entry_SYSCALL_64_after_hwframe+0x21/0x86
>>>      [<00000000458a6ffa>] 0xffffffffffffffff
>>>
>>> Not sure whether the above is a true issue or not.
>>>
>>> However, by inspecting the code, I do find the trie_free in lpm_trie.c
>>> may have missed freeing the trie memory.
>>>
>>> The change likes below should work:
>>> -bash-4.2$ git diff
>>> diff --git a/kernel/bpf/lpm_trie.c b/kernel/bpf/lpm_trie.c
>>> index 7b469d1..cecb259 100644
>>> --- a/kernel/bpf/lpm_trie.c
>>> +++ b/kernel/bpf/lpm_trie.c
>>> @@ -589,6 +589,7 @@ static void trie_free(struct bpf_map *map)
>>>
>>>   unlock:
>>>          raw_spin_unlock(&trie->lock);
>>> +       kfree(trie);
> 
> also looks like trie_free() is missing
> synchronize_rcu() + rcu_barrier()
> it doesn't wait for parallel lookup/update/delete to complete
> before freeing the elements.

Thanks, Alexei. I will address this in the patch as well.