Re: [PATCH bpf-next] bpf: use raw_spin_trylock() for pcpu_freelist_push/pop in NMI

[Song Liu <songliubraving@fb.com>]

> On Oct 2, 2020, at 4:09 PM, Daniel Borkmann <daniel@iogearbox.net> wrote:
> 
> On 9/26/20 2:07 AM, Song Liu wrote:
>> Recent improvements in LOCKDEP highlighted a potential A-A deadlock with
>> pcpu_freelist in NMI:
>> ./tools/testing/selftests/bpf/test_progs -t stacktrace_build_id_nmi
>> [   18.984807] ================================
>> [   18.984807] WARNING: inconsistent lock state
>> [   18.984808] 5.9.0-rc6-01771-g1466de1330e1 #2967 Not tainted
>> [   18.984809] --------------------------------
>> [   18.984809] inconsistent {INITIAL USE} -> {IN-NMI} usage.
>> [   18.984810] test_progs/1990 [HC2[2]:SC0[0]:HE0:SE1] takes:
>> [   18.984810] ffffe8ffffc219c0 (&head->lock){....}-{2:2}, at:
>> __pcpu_freelist_pop+0xe3/0x180
>> [   18.984813] {INITIAL USE} state was registered at:
>> [   18.984814]   lock_acquire+0x175/0x7c0
>> [   18.984814]   _raw_spin_lock+0x2c/0x40
>> [   18.984815]   __pcpu_freelist_pop+0xe3/0x180
>> [   18.984815]   pcpu_freelist_pop+0x31/0x40
>> [   18.984816]   htab_map_alloc+0xbbf/0xf40
>> [   18.984816]   __do_sys_bpf+0x5aa/0x3ed0
>> [   18.984817]   do_syscall_64+0x2d/0x40
>> [   18.984818]   entry_SYSCALL_64_after_hwframe+0x44/0xa9
>> [   18.984818] irq event stamp: 12
>> [ ... ]
>> [   18.984822] other info that might help us debug this:
>> [   18.984823]  Possible unsafe locking scenario:
>> [   18.984823]
>> [   18.984824]        CPU0
>> [   18.984824]        ----
>> [   18.984824]   lock(&head->lock);
>> [   18.984826]   <Interrupt>
>> [   18.984826]     lock(&head->lock);
>> [   18.984827]
>> [   18.984828]  *** DEADLOCK ***
>> [   18.984828]
>> [   18.984829] 2 locks held by test_progs/1990:
>> [ ... ]
>> [   18.984838]  <NMI>
>> [   18.984838]  dump_stack+0x9a/0xd0
>> [   18.984839]  lock_acquire+0x5c9/0x7c0
>> [   18.984839]  ? lock_release+0x6f0/0x6f0
>> [   18.984840]  ? __pcpu_freelist_pop+0xe3/0x180
>> [   18.984840]  _raw_spin_lock+0x2c/0x40
>> [   18.984841]  ? __pcpu_freelist_pop+0xe3/0x180
>> [   18.984841]  __pcpu_freelist_pop+0xe3/0x180
>> [   18.984842]  pcpu_freelist_pop+0x17/0x40
>> [   18.984842]  ? lock_release+0x6f0/0x6f0
>> [   18.984843]  __bpf_get_stackid+0x534/0xaf0
>> [   18.984843]  bpf_prog_1fd9e30e1438d3c5_oncpu+0x73/0x350
>> [   18.984844]  bpf_overflow_handler+0x12f/0x3f0
>> This is because pcpu_freelist_head.lock is accessed in both NMI and
>> non-NMI context. Fix this issue by using raw_spin_trylock() in NMI.
>> For systems with only one cpu, there is a trickier scenario with
>> pcpu_freelist_push(): if the only pcpu_freelist_head.lock is already
>> locked before NMI, raw_spin_trylock() will never succeed. Unlike,
>> _pop(), where we can failover and return NULL, failing _push() will leak
>> memory. Fix this issue with an extra list, pcpu_freelist.extralist. The
>> extralist is primarily used to take _push() when raw_spin_trylock()
>> failed on all the per cpu lists. It should be empty most of the time.
>> The following table summarizes the behavior of pcpu_freelist in NMI
>> and non-NMI:
>> non-NMI pop(): 	use _lock(); check per cpu lists first;
>>                 if all per cpu lists are empty, check extralist;
>>                 if extralist is empty, return NULL.
>> non-NMI push(): use _lock(); only push to per cpu lists.
>> NMI pop():    use _trylock(); check per cpu lists first;
>>               if all per cpu lists are locked or empty, check extralist;
>>               if extralist is locked or empty, return NULL.
>> NMI push():   use _trylock(); check per cpu lists first;
>>               if all per cpu lists are locked; try push to extralist;
>>               if extralist is also locked, keep trying on per cpu lists.
> 
> Code looks reasonable to me, is there any practical benefit to keep the
> extra list around for >1 CPU case (and not just compile it out)? For
> example, we could choose a different back end *_freelist_push/pop()
> implementation depending on CONFIG_SMP like ...
> 
> ifeq ($(CONFIG_SMP),y)
> obj-$(CONFIG_BPF_SYSCALL) += percpu_freelist.o
> else
> obj-$(CONFIG_BPF_SYSCALL) += onecpu_freelist.o
> endif
> 
> ... and keep the CONFIG_SMP simplified in that we'd only do the trylock
> iteration over CPUs under NMI with pop aborting with NULL in worst case
> and push keep looping, whereas for the single CPU case, all the logic
> resides in onecpu_freelist.c and it has a simpler two list implementation?

Technically, it is possible to have similar deadlock in SMP. For N cpus, 
there could be N NMI at the same time, and they may block N non-NMI raw
spinlock, and then all these NMI push() would spin forever. Of course, 
this is almost impossible to trigger with a decent N. 

On the other hand, I feel current code doesn't add too much complexity 
to SMP case. Maintaining two copies may require more work down the road. 
If we found current version too complex for SMP, we can do the split in 
the future. 

Does this make sense?

Thanks,
Song