Message-ID: <cd2f027c-b20e-4ea1-87f5-c0fa41180595@redhat.com>
Date: Tue, 26 Mar 2024 09:17:38 +0800
From: Xiubo Li <xiubli@...hat.com>
To: Ilya Dryomov <idryomov@...il.com>, David Hildenbrand <david@...hat.com>
Cc: linux-mm@...ck.org, linux-kernel@...r.kernel.org,
 Ceph Development <ceph-devel@...r.kernel.org>,
 linux-fsdevel@...r.kernel.org, Vlastimil Babka <vbabka@...e.cz>
Subject: Re: kernel BUG at mm/usercopy.c:102 -- pc : usercopy_abort


On 3/26/24 03:37, Ilya Dryomov wrote:
> On Mon, Mar 25, 2024 at 6:39 PM David Hildenbrand <david@...hat.com> wrote:
>> On 25.03.24 13:06, Xiubo Li wrote:
>>> On 3/25/24 18:14, David Hildenbrand wrote:
>>>> On 25.03.24 08:45, Xiubo Li wrote:
>>>>> Hi guys,
>>>>>
>>>>> We are hitting the same crash frequently recently with the latest kernel
>>>>> when testing kceph, and the call trace will be something likes:
>>>>>
>>>>> [ 1580.034891] usercopy: Kernel memory exposure attempt detected from
>>>>> SLUB object 'kmalloc-192' (offset 82, size 499712)!^M
>>>>> [ 1580.045866] ------------[ cut here ]------------^M
>>>>> [ 1580.050551] kernel BUG at mm/usercopy.c:102!^M
>>>>> ^M
>>>>> Entering kdb (current=0xffff8881211f5500, pid 172901) on processor 4
>>>>> Oops: (null)^M
>>>>> due to oops @ 0xffffffff8138cabd^M
>>>>> CPU: 4 PID: 172901 Comm: fsstress Tainted: G S 6.6.0-g623393c9d50c #1^M
>>>>> Hardware name: Supermicro SYS-5018R-WR/X10SRW-F, BIOS 1.0c 09/07/2015^M
>>>>> RIP: 0010:usercopy_abort+0x6d/0x80^M
>>>>> Code: 4c 0f 44 d0 41 53 48 c7 c0 1c e9 13 82 48 c7 c6 71 62 13 82 48 0f
>>>>> 45 f0 48 89 f9 48 c7 c7 f0 6b 1b 82 4c 89 d2 e8 63 2b df ff <0f> 0b 49
>>>>> c7 c1 44 c8 14 82 4d 89 cb 4d 89 c8 eb a5 66 90 f3 0f 1e^M
>>>>> RSP: 0018:ffffc90006dfba88 EFLAGS: 00010246^M
>>>>> RAX: 000000000000006a RBX: 000000000007a000 RCX: 0000000000000000^M
>>>>> RDX: 0000000000000000 RSI: ffff88885fd1d880 RDI: ffff88885fd1d880^M
>>>>> RBP: 000000000007a000 R08: 0000000000000000 R09: c0000000ffffdfff^M
>>>>> R10: 0000000000000001 R11: ffffc90006dfb930 R12: 0000000000000001^M
>>>>> R13: ffff8882b7bbed12 R14: ffff88827a375830 R15: ffff8882b7b44d12^M
>>>>> FS:  00007fb24c859500(0000) GS:ffff88885fd00000(0000)
>>>>> knlGS:0000000000000000^M
>>>>> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033^M
>>>>> CR2: 000055c2bcf9eb00 CR3: 000000028956c005 CR4: 00000000001706e0^M
>>>>> Call Trace:^M
>>>>>      <TASK>^M
>>>>>      ? kdb_main_loop+0x32c/0xa10^M
>>>>>      ? kdb_stub+0x216/0x420^M
>>>>> more>
>>>>>
>>>>> You can see more detail in ceph tracker
>>>>> https://tracker.ceph.com/issues/64471.
>>>> Where is the full backtrace? Above contains only the backtrace of kdb.
>>>>
>>> Hi David,
>>>
>>> The bad news is that there is no more backtrace. All the failures we hit
>>> are similar with the following logs:
>>>
>> That's unfortunate :/
>>
>> "exposure" in the message means we are in copy_to_user().
>>
>> SLUB object 'kmalloc-192' means that we come from __check_heap_object()
>> ... we have 192 bytes, but the length we want to access is 499712 ...
>> 488 KiB.
>>
>> So we ended  up somehow in
>>
>> __copy_to_user()->check_object_size()->__check_object_size()->
>> check_heap_object()->__check_heap_object()->usercopy_abort()
>>
>>
>> ... but the big question is which code tried to copy way too much memory
>> out of a slab folio to user space.
>>
>>>> That link also contains:
>>>>
>>>> Entering kdb (current=0xffff9115d14fb980, pid 61925) on processor 5
>>>> Oops: (null)^M
>>>> due to oops @ 0xfffffffface3a1d2^M
>>>> CPU: 5 PID: 61925 Comm: ld Kdump: loaded Not tainted
>>>> 5.14.0-421.el9.x86_64 #1^M
>>>> Hardware name: Supermicro SYS-5018R-WR/X10SRW-F, BIOS 2.0 12/17/2015^M
>>>> RIP: 0010:usercopy_abort+0x74/0x76^M
>>>> Code: 14 74 ad 51 48 0f 44 d6 49 c7 c3 cb 9f 73 ad 4c 89 d1 57 48 c7
>>>> c6 60 83 75 ad 48 c7 c7 00 83 75 ad 49 0f 44 f3 e8 1b 3b ff ff <0f> 0b
>>>> 0f b6 d3 4d 89 e0 48 89 e9 31 f6 48 c7 c7 7f 83 75 ad e8 73^M
>>>> RSP: 0018:ffffbb97c16af8d0 EFLAGS: 00010246^M
>>>> RAX: 0000000000000072 RBX: 0000000000000112 RCX: 0000000000000000^M
>>>> RDX: 0000000000000000 RSI: ffff911d1fd60840 RDI: ffff911d1fd60840^M
>>>> RBP: 0000000000004000 R08: 80000000ffff84b4 R09: 0000000000ffff0a^M
>>>> R10: 0000000000000004 R11: 0000000000000076 R12: ffff9115c0be8b00^M
>>>> R13: 0000000000000001 R14: ffff911665df9f68 R15: ffff9115d16be112^M
>>>> FS:  00007ff20442eb80(0000) GS:ffff911d1fd40000(0000)
>>>> knlGS:0000000000000000^M
>>>> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033^M
>>>> CR2: 00007ff20446142d CR3: 00000001215ec003 CR4: 00000000003706e0^M
>>>> DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000^M
>>>> DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400^M
>>>> Call Trace:^M
>>>>    <TASK>^M
>>>>    ? show_trace_log_lvl+0x1c4/0x2df^M
>>
>> ... are we stuck in show_trace_log_lvl(), probably deadlocked not being
>> able to print the actuall callstack? If so, that's nasty.
> Hi David,
>
> I don't think so.  This appears to be a cut-and-paste from what is
> essentially a non-interactive serial console.  Stack trace entries
> prefixed with ? aren't exact and kdb prompt
>
>      more>
>
> is there in all cases which is what hides the rest of the stack.
>
> There are four ways to get the entire stack trace here:
>
> a) try to attach to the serial console and interact with kdb -- this
>     is very much hit or miss due to general IPMI/BMC unreliability and
>     the fact that it would be already attached to for logging
> b) disable kdb by passing "kdb: false" in the job definition -- this
>     should result in /sys/module/kgdboc/parameters/kgdboc cleared after
>     booting into the kernel under test (or just hack teuthology to not
>     pass "kdb: true" which it does by default if "-k <kernel>" is given
>     when scheduling)
> c) if b) fails, rebuild the kernel with kdb disabled in Kconfig
> d) configure kdump and grab a vmcore -- these is no teuthology support
>     for this, so it would be challenging but would provide the most data
>     to chew on
>
> Xiubo, I'd recommend going with b), but take your pick ;)

Thanks Ilya

Let's try the b) first and to see could we find something new.

Thanks

- Xiubo


> Thanks,
>
>                  Ilya
>

Powered by blists - more mailing lists