On 8/27/20 1:53 AM, Li,Rongqing wrote:
>
>
>> -----Original Message-----
>> From: Eric Dumazet [mailto:eric.duma...@gmail.com]
>> Sent: Thursday, August 27, 2020 4:26 PM
>> To: Li,Rongqing <lirongq...@baidu.com>; netdev@vger.kernel.org;
>> intel-wired-...@lists.osuosl.org
>> Subject: Re: [PATCH] iavf: use kvzalloc instead of kzalloc for rx/tx_bi
>> buffer
>>
>>
>>
>> On 8/27/20 12:53 AM, Li RongQing wrote:
>>> when changes the rx/tx ring to 4096, kzalloc may fail due to a
>>> temporary shortage on slab entries.
>>>
>>> kvmalloc is used to allocate this memory as there is no need to have
>>> this memory area physical continuously.
>>>
>>> Signed-off-by: Li RongQing <lirongq...@baidu.com>
>>> ---
>>
>>
>> Well, fallback to vmalloc() overhead because order-1 pages are not readily
>> available when the NIC is setup (usually one time per boot) is adding TLB
>> cost
>> at run time, for billions of packets to come, maybe for months.
>>
>> Surely trying a bit harder to get order-1 pages is desirable.
>>
>> __GFP_RETRY_MAYFAIL is supposed to help here.
>
> Could we add __GFP_RETRY_MAYFAIL to kvmalloc, to ensure the allocation
> success ?
__GFP_RETRY_MAYFAIL does not _ensure_ the allocation success.
The idea here is that for large allocations (bigger than PAGE_SIZE),
kvmalloc_node() will not force __GFP_NORETRY, meaning that page allocator
will not bailout immediately in case of memory pressure.
This gives a chance for page reclaims to happen, and eventually the high order
page
allocation will succeed under normal circumstances.
It is a trade-off, and only worth it for long living allocations.
