Hi,
during analysis of a customer's problem on a 4.12 based kernel
(deadlock due to a blocking mmu notifier in a Xen driver) I came
across upstream patches 93065ac753e4 ("mm, oom: distinguish
blockable mode for mmu notifiers") et al.
The backtrace of the blocked tasks was typically something like:
#0 [ffffc9004222f228] __schedule at ffffffff817223e2
#1 [ffffc9004222f2b8] schedule at ffffffff81722a02
#2 [ffffc9004222f2c8] schedule_preempt_disabled at ffffffff81722d0a
#3 [ffffc9004222f2d0] __mutex_lock at ffffffff81724104
#4 [ffffc9004222f360] mn_invl_range_start at ffffffffc01fd398 [xen_gntdev]
#5 [ffffc9004222f398] __mmu_notifier_invalidate_page at ffffffff8123375a
#6 [ffffc9004222f3c0] try_to_unmap_one at ffffffff812112cb
#7 [ffffc9004222f478] rmap_walk_file at ffffffff812105cd
#8 [ffffc9004222f4d0] try_to_unmap at ffffffff81212450
#9 [ffffc9004222f508] shrink_page_list at ffffffff811e0755
#10 [ffffc9004222f5c8] shrink_inactive_list at ffffffff811e13cf
#11 [ffffc9004222f6a8] shrink_node_memcg at ffffffff811e241f
#12 [ffffc9004222f790] shrink_node at ffffffff811e29c5
#13 [ffffc9004222f808] do_try_to_free_pages at ffffffff811e2ee1
#14 [ffffc9004222f868] try_to_free_pages at ffffffff811e3248
#15 [ffffc9004222f8e8] __alloc_pages_slowpath at ffffffff81262c37
#16 [ffffc9004222f9f0] __alloc_pages_nodemask at ffffffff8121afc1
#17 [ffffc9004222fa48] alloc_pages_current at ffffffff8122f350
#18 [ffffc9004222fa78] __get_free_pages at ffffffff8121685a
#19 [ffffc9004222fa80] __pollwait at ffffffff8127e795
#20 [ffffc9004222faa8] evtchn_poll at ffffffffc00e802b [xen_evtchn]
#21 [ffffc9004222fab8] do_sys_poll at ffffffff8127f953
#22 [ffffc9004222fec8] sys_ppoll at ffffffff81280478
#23 [ffffc9004222ff30] do_syscall_64 at ffffffff81004954
#24 [ffffc9004222ff50] entry_SYSCALL_64_after_hwframe at ffffffff818000b6
It was found that the notifier of the Xen gntdev driver was using a
mutex resulting in the deadlock.
Michal Hocko suggested that backporting above mentioned patch might
help, as the mmu notifier call is happening in atomic context.
Looking into that I was wondering whether try_to_unmap_one() shouldn't
call mmu_notifier_invalidate_range_start_nonblock() instead of
mmu_notifier_invalidate_range_start() if this is true. Otherwise I
can't see how this deadlock could be avoided.
Any thoughts?
Juergen
Attachment:
OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key
Attachment:
OpenPGP_signature
Description: OpenPGP digital signature
