Re: [PATCH 1/1] mm: add swapiness= arg to memory.reclaim

[Huan Yang <11133793@vivo.com>]

在 2023/12/1 10:17, Yosry Ahmed 写道:
> On Thu, Nov 30, 2023 at 6:14 PM Huan Yang <11133793@vivo.com> wrote:
>>
>> 在 2023/12/1 10:05, Yosry Ahmed 写道:
>>>> @@ -2327,7 +2330,8 @@ static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
>>>>           struct pglist_data *pgdat = lruvec_pgdat(lruvec);
>>>>           struct mem_cgroup *memcg = lruvec_memcg(lruvec);
>>>>           unsigned long anon_cost, file_cost, total_cost;
>>>> -       int swappiness = mem_cgroup_swappiness(memcg);
>>>> +       int swappiness = sc->swappiness ?
>>>> +               *sc->swappiness : mem_cgroup_swappiness(memcg);
>>>>
>>>> Should we use "unlikely" here to indicate that sc->swappiness is an unexpected behavior?
>>>> Due to current use case only apply in proactive reclaim.
>>> On a system that is not under memory pressure, the rate of proactive
>>> reclaim could be higher than reactive reclaim. We should only use
>>> likely/unlikely when it's obvious a scenario will happen most of the
>>> time. I don't believe that's the case here.
>> Not all vendors will use proactive interfaces, and reactive reclaim are
>> a normal
>> system behavior. In this regard, I think it is appropriate to add
>> "unlikely".
> The general guidance is not to use likely/unlikely when it's not
> certain, which I believe is the case here. I think the CPU will make
OK, I will remember this part.
> better decisions on its own than if we give it hints that's wrong in
> some situations. Others please correct me if I am wrong.
No, you're right. CPU is good to do this.
>
>>>>           u64 fraction[ANON_AND_FILE];
>>>>           u64 denominator = 0;    /* gcc */
>>>>           enum scan_balance scan_balance;
>>>> @@ -2608,6 +2612,9 @@ static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
>>>>               mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
>>>>                   return 0;
>>>>
>>>> +       if (sc->swappiness)
>>>> +               return *sc->swappiness;
>>>>
>>>> Also there.
>>>>
>>>> +
>>>>           return mem_cgroup_swappiness(memcg);
>>>>    }
>>>>
>>>> @@ -6433,7 +6440,8 @@ unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
>>>>    unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
>>>>                                              unsigned long nr_pages,
>>>>                                              gfp_t gfp_mask,
>>>> -                                          unsigned int reclaim_options)
>>>> +                                          unsigned int reclaim_options,
>>>> +                                          int *swappiness)
>>>>    {
>>>>           unsigned long nr_reclaimed;
>>>>           unsigned int noreclaim_flag;
>>>> @@ -6448,6 +6456,7 @@ unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
>>>>                   .may_unmap = 1,
>>>>                   .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
>>>>                   .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
>>>> +               .swappiness = swappiness,
>>>>           };
>>>>           /*
>>>>            * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
>>>> --
>>>> 2.34.1
>>>>
>>>> My previous patch attempted to ensure fully deterministic semantics under extreme swappiness.
>>>> For example, when swappiness is set to 200, only anonymous pages will be reclaimed.
>>>> Due to code in MGLRU isolate_folios will try scan anon if no scanned, will try other type.(We do not want
>>>> it to attempt this behavior.)
>>>> How do you think about extreme swappiness scenarios?
>>> I think having different semantics between swappiness passed to
>>> proactive reclaim and global swappiness can be confusing. If it's
>>> needed to have a swappiness value that says "anon only no matter
>>> what", perhaps we should introduce such a new value and make it
>>> supported by both global and proactive reclaim swappiness? We could
>>> support writing "max" or something similar instead of a special value
>>> to mean that.
>> Yes, use other hint more suitable for this scenario.
>>
>> However, from this patch, it seems that this feature is not supported.
>> Do you have a demand for this scenario?
> We do anonymous-only proactive reclaim in some setups, so it would be
> nice to have. I am not sure if it's absolutely needed vs. just using
> swappiness=200 and living with the possibility of reclaiming some file
> pages.
Right now, the scenario where swappiness=200 is sufficient for us, but 
having the
tendency to only reclaim anonymous pages has a clear semantics that is
suitable for upper-level strategy scenarios, rather than relying solely on
the functionality of swappiness.