Re: kernel panic on null pointer on page->mem_cgroup

[Brad Bolen]

On Tue, Aug 8, 2017 at 1:37 PM, Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> On Tue, Aug 08, 2017 at 09:56:01AM -0700, Jaegeuk Kim wrote:
>> On 08/08, Johannes Weiner wrote:
>> > Hi Jaegeuk and Bradley,
>> >
>> > On Mon, Aug 07, 2017 at 09:01:50PM -0400, Bradley Bolen wrote:
>> > > I am getting a very similar error on v4.11 with an arm64 board.
>> > >
>> > > I, too, also see page->mem_cgroup checked to make sure that it is not
>> > > NULL and then several instructions later it is NULL.  It does appear
>> > > that someone is changing that member without taking the lock.  In my
>> > > setup, I see
>> > >
>> > > crash> bt
>> > > PID: 72     TASK: e1f48640  CPU: 0   COMMAND: "mmcqd/1"
>> > >  #0 [<c00ad35c>] (__crash_kexec) from [<c0101080>]
>> > >  #1 [<c0101080>] (panic) from [<c028cd6c>]
>> > >  #2 [<c028cd6c>] (svcerr_panic) from [<c028cdc4>]
>> > >  #3 [<c028cdc4>] (_SvcErr_) from [<c001474c>]
>> > >  #4 [<c001474c>] (die) from [<c00241f8>]
>> > >  #5 [<c00241f8>] (__do_kernel_fault) from [<c0560600>]
>> > >  #6 [<c0560600>] (do_page_fault) from [<c00092e8>]
>> > >  #7 [<c00092e8>] (do_DataAbort) from [<c055f9f0>]
>> > >     pc : [<c0112540>]    lr : [<c0112518>]    psr: a0000193
>> > >     sp : c1a19cc8  ip : 00000000  fp : c1a19d04
>> > >     r10: 0006ae29  r9 : 00000000  r8 : dfbf1800
>> > >     r7 : dfbf1800  r6 : 00000001  r5 : f3c1107c  r4 : e2fb6424
>> > >     r3 : 00000000  r2 : 00040228  r1 : 221e3000  r0 : a0000113
>> > >     Flags: NzCv  IRQs off  FIQs on  Mode SVC_32  ISA ARM
>> > >  #8 [<c055f9f0>] (__dabt_svc) from [<c0112518>]
>> > >  #9 [<c0112540>] (test_clear_page_writeback) from [<c01046d4>]
>> > > #10 [<c01046d4>] (end_page_writeback) from [<c0149bcc>]
>> > > #11 [<c0149bcc>] (end_swap_bio_write) from [<c0261460>]
>> > > #12 [<c0261460>] (bio_endio) from [<c042c800>]
>> > > #13 [<c042c800>] (dec_pending) from [<c042e648>]
>> > > #14 [<c042e648>] (clone_endio) from [<c0261460>]
>> > > #15 [<c0261460>] (bio_endio) from [<bf60aa00>]
>> > > #16 [<bf60aa00>] (crypt_dec_pending [dm_crypt]) from [<bf60c1e8>]
>> > > #17 [<bf60c1e8>] (crypt_endio [dm_crypt]) from [<c0261460>]
>> > > #18 [<c0261460>] (bio_endio) from [<c0269e34>]
>> > > #19 [<c0269e34>] (blk_update_request) from [<c026a058>]
>> > > #20 [<c026a058>] (blk_update_bidi_request) from [<c026a444>]
>> > > #21 [<c026a444>] (blk_end_bidi_request) from [<c026a494>]
>> > > #22 [<c026a494>] (blk_end_request) from [<c0458dbc>]
>> > > #23 [<c0458dbc>] (mmc_blk_issue_rw_rq) from [<c0459e24>]
>> > > #24 [<c0459e24>] (mmc_blk_issue_rq) from [<c045a018>]
>> > > #25 [<c045a018>] (mmc_queue_thread) from [<c0048890>]
>> > > #26 [<c0048890>] (kthread) from [<c0010388>]
>> > > crash> sym c0112540
>> > > c0112540 (T) test_clear_page_writeback+512
>> > >  /kernel-source/include/linux/memcontrol.h: 518
>> > >
>> > > crash> bt 35
>> > > PID: 35     TASK: e1d45dc0  CPU: 1   COMMAND: "kswapd0"
>> > >  #0 [<c0559ab8>] (__schedule) from [<c0559edc>]
>> > >  #1 [<c0559edc>] (schedule) from [<c055e54c>]
>> > >  #2 [<c055e54c>] (schedule_timeout) from [<c055a3a4>]
>> > >  #3 [<c055a3a4>] (io_schedule_timeout) from [<c0106cb0>]
>> > >  #4 [<c0106cb0>] (mempool_alloc) from [<c0261668>]
>> > >  #5 [<c0261668>] (bio_alloc_bioset) from [<c0149d68>]
>> > >  #6 [<c0149d68>] (get_swap_bio) from [<c014a280>]
>> > >  #7 [<c014a280>] (__swap_writepage) from [<c014a3bc>]
>> > >  #8 [<c014a3bc>] (swap_writepage) from [<c011e5c8>]
>> > >  #9 [<c011e5c8>] (shmem_writepage) from [<c011a9b8>]
>> > > #10 [<c011a9b8>] (shrink_page_list) from [<c011b528>]
>> > > #11 [<c011b528>] (shrink_inactive_list) from [<c011c160>]
>> > > #12 [<c011c160>] (shrink_node_memcg) from [<c011c400>]
>> > > #13 [<c011c400>] (shrink_node) from [<c011d7dc>]
>> > > #14 [<c011d7dc>] (kswapd) from [<c0048890>]
>> > > #15 [<c0048890>] (kthread) from [<c0010388>]
>> > >
>> > > It appears that uncharge_list() in mm/memcontrol.c is not taking the
>> > > page lock when it sets mem_cgroup to NULL.  I am not familiar with the
>> > > mm code so I do not know if this is on purpose or not.  There is a
>> > > comment in uncharge_list that makes me believe that the crashing code
>> > > should not have been running:
>> > > /*
>> > >  * Nobody should be changing or seriously looking at
>> > >  * page->mem_cgroup at this point, we have fully
>> > >  * exclusive access to the page.
>> > >  */
>> > > However, I am new to looking at this area of the kernel so I am not
>> > > sure.
>> >
>> > The lock is for pages that are actively being used, whereas the free
>> > path requires the page refcount to be 0; nobody else should be having
>> > access to the page at that time.
>>
>> Given various trials for a while, using __mod_memcg_state() instead of
>> mod_memcg_state() ssems somehow blowing the panic away. It might be caused
>> by kernel preemption?
>
> That's puzzling. Is that reliably the case? Because on x86-64,
> __this_cpu_add and this_cpu_add should result in the same code:
>
> #define raw_cpu_add_8(pcp, val)                 percpu_add_op((pcp), val)
> #define this_cpu_add_8(pcp, val)                percpu_add_op((pcp), val)
>
> which boils down to single instructions - incq, decq, addq - and so
> never needs explicit disabling of scheduler or interrupt preemption.
>
>> > > I was able to create a reproducible scenario by using a udelay to
>> > > increase the time between the if (page->mem_cgroup) check and the later
>> > > dereference of it to increase the race window.  I then mounted an empty
>> > > ext4 partition and ran the following no more than twice before it
>> > > crashed.
>> > > dd if=/dev/zero of=/tmp/ext4disk/test bs=1M count=100
>> >
>> > Thanks, that's useful. I'm going to try to reproduce this also.
>> >
>> > There is a
>> >
>> >     VM_BUG_ON_PAGE(!PageHWPoison(page) && page_count(page), page);
>> >
>> > inside uncharge_list() that verifies that there shouldn't in fact be
>> > any pages ending writeback when they get into that function. Can you
>> > build your kernel with CONFIG_DEBUG_VM to enable that test?
>>
>> I'll test this as well. ;)
>
> Thanks. I'm trying various udelays in between the NULL-check and the
> dereference, but I cannot seem to make it happen with the ext4-dd on
> my local machine.
I am using a fairly fat udelay of 100.

I turned on CONFIG_DEBUG_VM but it didn't bug for me, although the original
problem still reproduced.

I can insert some tracing to try to get some more information, but unfortunately
I am stuck with v4.11 at the moment and can't try v4.13-rc4 on this setup.