Re: [PATCH 3/3] f2fs: provide more chance for node and data to get ssr segment

[Jaegeuk Kim]

On 02/25, Yunlong Song wrote:
> HiïJaegeuk and Chao,
>     SorryïI misunderstood CURSEG_HOT_DATA is greater than CURSEG_WARM_DATA and use "i--" in the codes, so just forget that.
>     But there is still a issue, when type is CURSEG_WARM_DATA, chao's patch will select cold SSR segment first, but my patch will select hot SSR segment first, since I think the probability of having hot SSR segment is bigger than having cold SSR segment due to the temperature itself.

One thing that I'm seeing is, theoretically hot segments will get dirty more
frequently. So if we select dirty cold segments for warm data, we can gain
fully valid cold segments more, which can mitigate log thrashing problem.

Thanks,

> 
> 
> On 02/25/2017 01:58, Jaegeuk Kim wrote:
> On 02/24, Yunlong Song wrote:
> > Hi, Chao,
> >
> > Not looks good to me, since there is some case your code does not include:
> > if type is CURSEG_HOT_DATA, and if get_victim also returns 0 for both CURSEG_HOT_DATA and
> > CURSEG_WARM_DATA, then i will be -1 and pass to get_victim in your code.  So I still suggest
> > my original patch attached below.
> 
> Why does i become -1?
> 
> >
> > On 2017/2/24 18:47, Chao Yu wrote:
> > > On 2017/2/24 17:19, Yunlong Song wrote:
> > >> Hi Jaegeuk and Chao,
> > >>
> > >>     How about the question I pointed out in last mail:
> > >>     Why not take "neighboring temperature" for ssr? For example, if type == CURSEG_COLD_DATA,
> > >>     the new patch selects CURSEG_HOT_DATA first, why not select CURSEG_WARM_DATA first?
> > >>     The patch I sent ensure this "neighboring temperature" for ssr. This is to reduce the influence of
> > >>     mixing different levels of hot/code node types.
> > > Agreed, I sent one patch for changing the policy of SSR, how do you think of it?
> > >
> > > Thanks,
> > >
> > >> On 2017/2/24 17:05, Chao Yu wrote:
> > >>> Hi Jaegeuk,
> > >>>
> > >>> Reviewed-by: Chao Yu <yuchao0@xxxxxxxxxx>
> > >>>
> > >>> For your attached two patches.
> > >>>
> > >>> Thanks,
> > >>>
> > >>> On 2017/2/23 9:17, Jaegeuk Kim wrote:
> > >>>> Hi Yunlong,
> > >>>>
> > >>>> I've been testing the similar patches as I attached.
> > >>>>
> > >>>> Thanks,
> > >>>>
> > >>>> On 02/22, Yunlong Song wrote:
> > >>>>> Signed-off-by: Yunlong Song <yunlong.song@xxxxxxxxxx>
> > >>>>> ---
> > >>>>>  fs/f2fs/segment.c | 9 +++++++++
> > >>>>>  1 file changed, 9 insertions(+)
> > >>>>>
> > >>>>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> > >>>>> index 9d13743..5fe71b9 100644
> > >>>>> --- a/fs/f2fs/segment.c
> > >>>>> +++ b/fs/f2fs/segment.c
> > >>>>> @@ -1540,12 +1540,17 @@ static int get_ssr_segment(struct f2fs_sb_info *sbi, int type)
> > >>>>>  {
> > >>>>>      struct curseg_info *curseg = CURSEG_I(sbi, type);
> > >>>>>      const struct victim_selection *v_ops = DIRTY_I(sbi)->v_ops;
> > >>>>> +    int old_type = type;
> > >>>>>  
> > >>>>>      if (IS_NODESEG(type)) {
> > >>>>>          for (; type >= CURSEG_HOT_NODE; type--)
> > >>>>>              if (v_ops->get_victim(sbi, &(curseg)->next_segno,
> > >>>>>                              BG_GC, type, SSR))
> > >>>>>                  return 1;
> > >>>>> +        for (type = old_type + 1; type <= CURSEG_COLD_NODE; type++)
> > >>>>> +            if (v_ops->get_victim(sbi, &(curseg)->next_segno,
> > >>>>> +                            BG_GC, type, SSR))
> > >>>>> +                return 1;
> > >>>>>          return 0;
> > >>>>>      }
> > >>>>>  
> > >>>>> @@ -1554,6 +1559,10 @@ static int get_ssr_segment(struct f2fs_sb_info *sbi, int type)
> > >>>>>          if (v_ops->get_victim(sbi, &(curseg)->next_segno,
> > >>>>>                          BG_GC, type, SSR))
> > >>>>>              return 1;
> > >>>>> +    for (type = old_type + 1; type <= CURSEG_COLD_DATA; type++)
> > >>>>> +        if (v_ops->get_victim(sbi, &(curseg)->next_segno,
> > >>>>> +                        BG_GC, type, SSR))
> > >>>>> +            return 1;
> > >>>>>      return 0;
> > >>>>>  }
> > >>>>>  
> > >>>>> --
> > >>>>> 1.8.5.2
> > >>> .
> > >>>
> > >>
> > >
> > > .
> > >
> >
> >
> > --
> > Thanks,
> > Yunlong Song
> >