Re: [PATCH V2] mlx4_core: allocate ICM memory in page size chunks

[Qing Huang <qing.huang@oracle.com>]

On 5/15/2018 12:08 PM, Eric Dumazet wrote:
>
> On 05/15/2018 11:53 AM, Qing Huang wrote:
>>> This is control path so it is less latency-sensitive.
>>> Let's not produce unnecessary degradation here, please call kvzalloc so we maintain a similar behavior when contiguous memory is available, and a fallback for resiliency.
>> No sure what exactly degradation is caused by vzalloc here. I think it's better to keep physically contiguous pages
>> to other requests which really need them. Besides slow path/mem compacting can be really expensive.
>>
> Just use kvzalloc(), and you get the benefit of having contiguous memory if available,
> without expensive compact phase.
>
> This thing _automatically_ falls back to vmalloc(), thus your problem will be solved.
>
> If you are not sure, trust others.

Thanks for the review. There are many places in kernel and applications 
where physically contiguous pages are needed.
We saw quite a few issues when there were not enough contiguous phy mem 
available. My main concern here is that why
using physically contiguous pages when they are not really needed?