Re: [dpdk-stable] [PATCH] mem: fix allocation failure on non-NUMA kernel

[Nick Connolly <nick.connolly@mayadata.io>]

On 05/08/2020 15:36, Nicolas Dichtel wrote:
> Le 05/08/2020 à 16:20, Nick Connolly a écrit :
> [snip]
>>>> Fixes: 2a96c88be83e ("mem: ease init in a docker container")
>>> I'm wondering if the bug existed before this commit.
>>>
>>> Before this commit, it was:
>>>          move_pages(getpid(), 1, &addr, NULL, &cur_socket_id, 0);
>>>          if (cur_socket_id != socket_id) {
>>>                  /* error */
>>>
>>> Isn't it possible to hit this error case if CONFIG_NUMA is unset in the kernel?
>> I've just run the previous code to test this out and you are right that
>> move_pages does indeed return -1 with errno set to ENOSYS, but nothing checks
>> this so execution carries on and compares cur_socket_id (which will be unchanged
>> from the zero initialization) with socket_id (which is presumably also zero),
>> thus allowing the allocation to succeed!
> I came to this conclusion, but I didn't check if socket_id could be != from 0.
>
>>> [snip]
>>>> +    if (check_numa()) {
>>>> +        ret = get_mempolicy(&cur_socket_id, NULL, 0, addr,
>>>> +                    MPOL_F_NODE | MPOL_F_ADDR);
>>>> +        if (ret < 0) {
>>>> +            RTE_LOG(DEBUG, EAL, "%s(): get_mempolicy: %s\n",
>>>> +                __func__, strerror(errno));
>>>> +            goto mapped;
>>>> +        } else if (cur_socket_id != socket_id) {
>>>> +            RTE_LOG(DEBUG, EAL,
>>>> +                    "%s(): allocation happened on wrong socket (wanted %d,
>>>> got %d)\n",
>>>> +                __func__, socket_id, cur_socket_id);
>>>> +            goto mapped;
>>>> +        }
>>>> +    } else {
>>>> +        if (rte_socket_count() > 1)
>>>> +            RTE_LOG(DEBUG, EAL, "%s(): not checking socket for allocation
>>>> (wanted %d)\n",
>>>> +                    __func__, socket_id);
>>> nit: maybe an higher log level like WARNING?
>> Open to guidance here - my concern was that this is going to be generated for
>> every call to alloc_seg() and I'm not sure what the frequency will be - I'm
>> cautious about flooding the log with warnings under 'normal running'.  Are the
>> implications of running on a multi socket system with NUMA support disabled in
>> the kernel purely performance related for the DPDK or is there a functional
>> correctness issue as well?
> Is it really a 'normal running' to have CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES in
> dpdk and not CONFIG_NUMA in the kernel?

I'm not an expert of DPDK, but I think it needs to be treated as 'normal 
running', for the following reasons:

 1. The existing code in eal_memalloc_alloc_seg_bulk() is designed to
    work even if check_numa() indicates that NUMA support is not enabled:

    #ifdef RTE_EAL_NUMA_AWARE_HUGEPAGES
    if (check_numa()) {
             oldmask = numa_allocate_nodemask();
             prepare_numa(&oldpolicy, oldmask, socket);
             have_numa = true;
         }
    #endif
 2. The DPDK application could be built with
    CONFIG_RTE_EAL_NUMA_AWARE_HUGE_PAGES and then the binary run on
    different systems with and without NUMA support.

Regards,
Nick