Re: [PATCH] mm: fix hanging shrinker management on long do_shrink_slab

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




On 12/4/19 11:35 AM, Michal Hocko wrote:
> On Sat 30-11-19 00:45:41, Pavel Tikhomirov wrote:
>> We have a problem that shrinker_rwsem can be held for a long time for
>> read in shrink_slab, at the same time any process which is trying to
>> manage shrinkers hangs.
>>
>> The shrinker_rwsem is taken in shrink_slab while traversing shrinker_list.
>> It tries to shrink something on nfs (hard) but nfs server is dead at
>> these moment already and rpc will never succeed. Generally any shrinker
>> can take significant time to do_shrink_slab, so it's a bad idea to hold
>> the list lock here.
> 
> Yes, this is a known problem and people have already tried to address it
> in the past. Have you checked previous attempts? SRCU based one
> http://lkml.kernel.org/r/153365347929.19074.12509495712735843805.stgit@localhost.localdomain
> but I believe there were others (I only had this one in my notes).
> Please make sure to Cc Dave Chinner when posting a next version because
> he had some concerns about the change of the behavior.

The approach of the patch you are referencing is quiet different, it 
will still hold global srcu_read_lock(&srcu) when diving in 
do_shrink_slab and hanging nfs will still block all [un]register_shrinker.

> 
>> We have a similar problem in shrink_slab_memcg, except that we are
>> traversing shrinker_map+shrinker_idr there.
>>
>> The idea of the patch is to inc a refcount to the chosen shrinker so it
>> won't disappear and release shrinker_rwsem while we are in
>> do_shrink_slab, after that we will reacquire shrinker_rwsem, dec
>> the refcount and continue the traversal.
> 
> The reference count part makes sense to me. RCU role needs a better
> explanation.

I have 2 rcu's in patch, 1-st is to protect shrinker_map same as it was 
before in memcg_set_shrinker_bit, 2-nd is to protect shrinker struct in 
put_shrinker from being freed, as unregister_shrinker can see refcnt==0 
without actually going to schedule().

> Also do you have any reason to not use completion for
> the final step? Openconding essentially the same concept sounds a bit
> awkward to me.

Thanks for a good hint, from the first glance we can rework wait_event 
part to wait_for_completion.

> 
>> We also need a wait_queue so that unregister_shrinker can wait for the
>> refcnt to become zero. Only after these we can safely remove the
>> shrinker from list and idr, and free the shrinker.
> [...]
>>    crash> bt ...
>>    PID: 18739  TASK: ...  CPU: 3   COMMAND: "bash"
>>     #0 [...] __schedule at ...
>>     #1 [...] schedule at ...
>>     #2 [...] rpc_wait_bit_killable at ... [sunrpc]
>>     #3 [...] __wait_on_bit at ...
>>     #4 [...] out_of_line_wait_on_bit at ...
>>     #5 [...] _nfs4_proc_delegreturn at ... [nfsv4]
>>     #6 [...] nfs4_proc_delegreturn at ... [nfsv4]
>>     #7 [...] nfs_do_return_delegation at ... [nfsv4]
>>     #8 [...] nfs4_evict_inode at ... [nfsv4]
>>     #9 [...] evict at ...
>>    #10 [...] dispose_list at ...
>>    #11 [...] prune_icache_sb at ...
>>    #12 [...] super_cache_scan at ...
>>    #13 [...] do_shrink_slab at ...
> 
> Are NFS people aware of this? Because this is simply not acceptable
> behavior. Memory reclaim cannot be block indefinitely or for a long
> time. There must be a way to simply give up if the underlying inode
> cannot be reclaimed.

Sorry that I didn't cc nfs people from the begining.

> 
> I still have to think about the proposed solution. It sounds a bit over
> complicated to me.
> 

-- 
Best regards, Tikhomirov Pavel
Software Developer, Virtuozzo.




[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux