On 2016/7/20 15:47, Michal Hocko wrote:
> On Wed 20-07-16 09:33:30, Yisheng Xie wrote:
>>
>>
>> On 2016/7/19 22:14, Vlastimil Babka wrote:
>>> On 07/19/2016 03:48 PM, Xishi Qiu wrote:
> [...]
>>>> mode:0x2000d1 means it expects to alloc from zone_dma, (on arm64 zone_dma is 0-4G)
>>>
>>> Yes, but I don't see where the __GFP_DMA comes from. The backtrace
>>> suggests it's alloc_thread_info_node() which uses THREADINFO_GFP
>>> which is GFP_KERNEL | __GFP_NOTRACK. There shouldn't be __GFP_DMA,
>>> even on arm64. Are there some local modifications to the kernel
>>> source?
>>>
>>>> The page cache is very small(active_file:292kB inactive_file:240kB),
>>>> so did_some_progress may be zero, and will not retry, right?
>>>
>>> Could be, and then __alloc_pages_may_oom() has this:
>>>
>>> /* The OOM killer does not needlessly kill tasks for lowmem */
>>> if (ac->high_zoneidx < ZONE_NORMAL)
>>> goto out;
>>>
>>> So no oom and no faking progress for non-costly order that would
>>> result in retry, because of that mysterious __GFP_DMA...
>>
>> hi Vlastimil,
>> We do make change and add __GFP_DMA flag here for our platform driver's problem.
>
> Why would you want to force thread_info to the DMA zone?
>
Hi Michal,
Because of our platform driver's problem, so we change the code(add GFP_DMA) to let
it alloc from zone_dma. (on arm64 zone_dma is 0-4G)
Thanks,
Xishi Qiu
>> Another question is why it will do retry here, for it will goto out
>> with did_some_progress=0 ?
>>
>> if (!did_some_progress)
>> goto nopage;
>
> Do you mean:
> /*
> * If we fail to make progress by freeing individual
> * pages, but the allocation wants us to keep going,
> * start OOM killing tasks.
> */
> if (!did_some_progress) {
> page = __alloc_pages_may_oom(gfp_mask, order, ac,
> &did_some_progress);
> if (page)
> goto got_pg;
> if (!did_some_progress)
> goto nopage;
> }
>
> If yes then this code simply tells that if even oom path didn't make any
> progress then we should fail. As DMA request doesn't invoke OOM killer
> because it is effectively a lowmem request (see above check pointed
> by Vlastimil) then the OOM path couldn't make any progress and we are
> failing. If invoked the OOM killer then we would consider this as a
> forward progress and retry the allocation request.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>