[dpdk-dev] [PATCH v4 0/4] fix race condition in enqueue/dequeue because of cpu reorder

[Jia He]

Hi Bruce


On 11/8/2017 8:15 PM, Bruce Richardson Wrote:
> On Wed, Nov 08, 2017 at 09:54:37AM +0000, Jia He wrote:
>> We watched a rte panic of mbuf_autotest in our qualcomm arm64 server
>> due to a possible race condition.
>>
>> To fix this race, there are 2 options as suggested by Jerin: 1. use
>> rte_smp_rmb 2. use load_acquire/store_release(refer to [2]).
>> CONFIG_RTE_RING_USE_C11_MEM_MODEL is provided, and by default it is
>> "y" only on arm64 so far.
>>
>> The reason why providing 2 options is due to the performance benchmark
>> difference in different arm machines.
>>
>> Already fuctionally tested on the machines as follows: - on X86 - on
>> arm64 with CONFIG_RTE_RING_USE_C11_MEM_MODEL=y - on arm64 with
>> CONFIG_RTE_RING_USE_C11_MEM_MODEL=n
>>
>> --- Changelog: V4: split into small patches V3: arch specific
>> implementation for enqueue/dequeue barrier V2: let users choose
>> whether using load_acquire/store_release V1: rte_smp_rmb() between 2
>> loads
>>
>> Jia He (4): eal/arm64: remove the braces {} for dmb() and dsb() ring:
>> guarantee load/load order in enqueue and dequeue ring: introduce new
>> header file to include common functions ring: introduce new header
>> file to support C11 memory model
>>
> I'm wondering what the merge plans are for this set, given we are now
> past RC3 in 17.11? As the rings are broken on ARM machines we need to
> merge in some fix, but I'm a little concerned about the scope of the
> changes from the 3rd and 4th patches. Would it be acceptable to just
> merge in patches 1 & 2 in 17.11 and leave the rework and C11 memory
> model additions in patches 3 & 4 to 18.02 release?
As far as I'm concerned, it is ok.

Cheers,
Jia