Re: [PATCH] mm: Fix kswapd livelock on single core, no preempt kernel

[Mike Waychison]

On Tue, Dec 13, 2011 at 6:24 PM, Shaohua Li <shaohua.li@xxxxxxxxx> wrote:
> On Wed, 2011-12-14 at 01:44 +0800, Mike Waychison wrote:
>> On a single core system with kernel preemption disabled, it is possible
>> for the memory system to be so taxed that kswapd cannot make any forward
>> progress.  This can happen when most of system memory is tied up as
>> anonymous memory without swap enabled, causing kswapd to consistently
>> fail to achieve its watermark goals.  In turn, sleeping_prematurely()
>> will consistently return true and kswapd_try_to_sleep() to never invoke
>> schedule().  This causes the kswapd thread to stay on the CPU in
>> perpetuity and keeps other threads from processing oom-kills to reclaim
>> memory.
>>
>> The cond_resched() instance in balance_pgdat() is never called as the
>> loop that iterates from DEF_PRIORITY down to 0 will always set
>> all_zones_ok to true, and not set it to false once we've passed
>> DEF_PRIORITY as zones that are marked ->all_unreclaimable are not
>> considered in the "all_zones_ok" evaluation.
>>
>> This change modifies kswapd_try_to_sleep to ensure that we enter
>> scheduler at least once per invocation if needed.  This allows kswapd to
>> get off the CPU and allows other threads to die off from the OOM killer
>> (freeing memory that is otherwise unavailable in the process).
> your description suggests zones with all_unreclaimable set. but in this
> case sleeping_prematurely() will return false instead of true, kswapd
> will do sleep then. is there anything I missed?

Debugging this, I didn't get a dump from oom-kill as it never ran
(until I binary patched in a cond_resched() into live hung machines --
this reproduced in a VM).

I was however able to capture the following data while it was hung:


/cloud/vmm/host/backend/perfmetric/node0/zone0/active_anon : long long = 773
/cloud/vmm/host/backend/perfmetric/node0/zone0/active_file : long long = 6
/cloud/vmm/host/backend/perfmetric/node0/zone0/anon_pages : long long = 1,329
/cloud/vmm/host/backend/perfmetric/node0/zone0/bounce : long long = 0
/cloud/vmm/host/backend/perfmetric/node0/zone0/dirtied : long long = 4,425
/cloud/vmm/host/backend/perfmetric/node0/zone0/file_dirty : long long = 0
/cloud/vmm/host/backend/perfmetric/node0/zone0/file_mapped : long long = 5
/cloud/vmm/host/backend/perfmetric/node0/zone0/file_pages : long long = 330
/cloud/vmm/host/backend/perfmetric/node0/zone0/free_pages : long long = 2,018
/cloud/vmm/host/backend/perfmetric/node0/zone0/inactive_anon : long long = 865
/cloud/vmm/host/backend/perfmetric/node0/zone0/inactive_file : long long = 13
/cloud/vmm/host/backend/perfmetric/node0/zone0/kernel_stack : long long = 10
/cloud/vmm/host/backend/perfmetric/node0/zone0/mlock : long long = 0
/cloud/vmm/host/backend/perfmetric/node0/zone0/pagetable : long long = 74
/cloud/vmm/host/backend/perfmetric/node0/zone0/shmem : long long = 0
/cloud/vmm/host/backend/perfmetric/node0/zone0/slab_reclaimable : long long = 54
/cloud/vmm/host/backend/perfmetric/node0/zone0/slab_unreclaimable :
long long = 130
/cloud/vmm/host/backend/perfmetric/node0/zone0/unevictable : long long = 0
/cloud/vmm/host/backend/perfmetric/node0/zone0/writeback : long long = 0
/cloud/vmm/host/backend/perfmetric/node0/zone0/written : long long = 47,184

/cloud/vmm/host/backend/perfmetric/node0/zone1/active_anon : long long = 359,251
/cloud/vmm/host/backend/perfmetric/node0/zone1/active_file : long long = 67
/cloud/vmm/host/backend/perfmetric/node0/zone1/anon_pages : long long = 441,180
/cloud/vmm/host/backend/perfmetric/node0/zone1/bounce : long long = 0
/cloud/vmm/host/backend/perfmetric/node0/zone1/dirtied : long long = 6,457,125
/cloud/vmm/host/backend/perfmetric/node0/zone1/file_dirty : long long = 0
/cloud/vmm/host/backend/perfmetric/node0/zone1/file_mapped : long long = 134
/cloud/vmm/host/backend/perfmetric/node0/zone1/file_pages : long long = 38,090
/cloud/vmm/host/backend/perfmetric/node0/zone1/free_pages : long long = 1,630
/cloud/vmm/host/backend/perfmetric/node0/zone1/inactive_anon : long
long = 119,779
/cloud/vmm/host/backend/perfmetric/node0/zone1/inactive_file : long long = 81
/cloud/vmm/host/backend/perfmetric/node0/zone1/kernel_stack : long long = 173
/cloud/vmm/host/backend/perfmetric/node0/zone1/mlock : long long = 0
/cloud/vmm/host/backend/perfmetric/node0/zone1/pagetable : long long = 15,222
/cloud/vmm/host/backend/perfmetric/node0/zone1/shmem : long long = 1
/cloud/vmm/host/backend/perfmetric/node0/zone1/slab_reclaimable : long
long = 1,677
/cloud/vmm/host/backend/perfmetric/node0/zone1/slab_unreclaimable :
long long = 7,152
/cloud/vmm/host/backend/perfmetric/node0/zone1/unevictable : long long = 0
/cloud/vmm/host/backend/perfmetric/node0/zone1/writeback : long long = 8
/cloud/vmm/host/backend/perfmetric/node0/zone1/written : long long = 16,639,708

These value were static while the machine was hung up in kswapd.  I
unfortunately don't have the low/min/max or lowmem watermarks handy.

>From stepping through with gdb, I was able to determine that
ZONE_DMA32 would fail zone_watermark_ok_safe(), causing a scan   up to
end_zone == 1.  If memory serves, it would not get the
->all_unreclaimable flag.  I didn't get the chance to root cause this
internal inconsistency though.

FYI, this was seen with a 2.6.39-based kernel with no-numa, no-memcg
and swap-enabled.

If I get the chance, I can reproduce and look at this closer to try
and root cause why zone_reclaimable() would return true, but I won't
be able to do that until after the holidays -- sometime in January.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/