lkml.org 
[lkml]   [2019]   [Sep]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [f2fs-dev] [PATCH 1/2] f2fs: do not select same victim right again
On 09/19, Chao Yu wrote:
> On 2019/9/19 0:47, Jaegeuk Kim wrote:
> > On 09/18, Chao Yu wrote:
> >> On 2019/9/18 11:12, Jaegeuk Kim wrote:
> >>> On 09/18, Chao Yu wrote:
> >>>> On 2019/9/18 4:55, Jaegeuk Kim wrote:
> >>>>> On 09/17, Chao Yu wrote:
> >>>>>> On 2019/9/16 23:37, Jaegeuk Kim wrote:
> >>>>>>> On 09/16, Chao Yu wrote:
> >>>>>>>> On 2019/9/9 20:04, Jaegeuk Kim wrote:
> >>>>>>>>> On 09/09, Chao Yu wrote:
> >>>>>>>>>> On 2019/9/9 16:06, Jaegeuk Kim wrote:
> >>>>>>>>>>> On 09/09, Chao Yu wrote:
> >>>>>>>>>>>> On 2019/9/9 9:25, Jaegeuk Kim wrote:
> >>>>>>>>>>>>> GC must avoid select the same victim again.
> >>>>>>>>>>>>
> >>>>>>>>>>>> Blocks in previous victim will occupy addition free segment, I doubt after this
> >>>>>>>>>>>> change, FGGC may encounter out-of-free space issue more frequently.
> >>>>>>>>>>>
> >>>>>>>>>>> Hmm, actually this change seems wrong by sec_usage_check().
> >>>>>>>>>>> We may be able to avoid this only in the suspicious loop?
> >>>>>>>>>>>
> >>>>>>>>>>> ---
> >>>>>>>>>>> fs/f2fs/gc.c | 2 +-
> >>>>>>>>>>> 1 file changed, 1 insertion(+), 1 deletion(-)
> >>>>>>>>>>>
> >>>>>>>>>>> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> >>>>>>>>>>> index e88f98ddf396..5877bd729689 100644
> >>>>>>>>>>> --- a/fs/f2fs/gc.c
> >>>>>>>>>>> +++ b/fs/f2fs/gc.c
> >>>>>>>>>>> @@ -1326,7 +1326,7 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync,
> >>>>>>>>>>> round++;
> >>>>>>>>>>> }
> >>>>>>>>>>>
> >>>>>>>>>>> - if (gc_type == FG_GC)
> >>>>>>>>>>> + if (gc_type == FG_GC && seg_freed)
> >>>>>>>>>>
> >>>>>>>>>> That's original solution Sahitya provided to avoid infinite loop of GC, but I
> >>>>>>>>>> suggest to find the root cause first, then we added .invalid_segmap for that
> >>>>>>>>>> purpose.
> >>>>>>>>>
> >>>>>>>>> I've checked the Sahitya's patch. So, it seems the problem can happen due to
> >>>>>>>>> is_alive or atomic_file.
> >>>>>>>>
> >>>>>>>> For some conditions, this doesn't help, for example, two sections contain the
> >>>>>>>> same fewest valid blocks, it will cause to loop selecting them if it fails to
> >>>>>>>> migrate blocks.
> >>>>>>>>
> >>>>>>>> How about keeping it as it is to find potential bug.
> >>>>>>>
> >>>>>>> I think it'd be fine to merge this. Could you check the above scenario in more
> >>>>>>> detail?
> >>>>>>
> >>>>>> I haven't saw this in real scenario yet.
> >>>>>>
> >>>>>> What I mean is if there is a bug (maybe in is_alive()) failing us to GC on one
> >>>>>> section, when that bug happens in two candidates, there could be the same
> >>>>>> condition that GC will run into loop (select A, fail to migrate; select B, fail
> >>>>>> to migrate, select A...).
> >>>>>>
> >>>>>> But I guess the benefit of this change is, if FGGC fails to migrate block due to
> >>>>>> i_gc_rwsem race, selecting another section and later retrying previous one may
> >>>>>> avoid lock race, right?
> >>>>>
> >>>>> In any case, I think this can avoid potenial GC loop. At least to me, it'd be
> >>>>> quite risky, if we remain this just for debugging purpose only.
> >>>>
> >>>> Yup,
> >>>>
> >>>> One more concern is would this cur_victim_sec remain after FGGC? then BGGC/SSR
> >>>> will always skip the section cur_victim_sec points to.
> >>>
> >>> Then, we can get another loop before using it by BGGC/SSR.
> >>
> >> I guess I didn't catch your point, do you mean, if we reset it in the end of
> >> FGGC, we may encounter the loop during BGGC/SSR?
> >
> > FGGC failed in a loop and last victim was remained in cur_victim_sec.
>
> It won't run into a loop because we keep below condition?

The following FGGC will be likely to select this victim again, which doesn't
mean "this loop" but "loop of f2fs_gc".

>
> + if (gc_type == FG_GC && seg_freed)
> + sbi->cur_victim_sec = NULL_SEGNO;
>
> if (sync)
> goto stop;
>
> I meant add below logic in addition:
>
> + if (gc_type == FG_GC)
> + sbi->cur_victim_sec = NULL_SEGNO;
>
> mutex_unlock(&sbi->gc_mutex);
>
> Thanks,
>
> > Next FGGC kicked in and did the same thing again. I don't expect BGGC/SSR
> > wants to select this victim much, since it will have CB policy.
> >
> >>
> >> I meant:
> >>
> >> f2fs_gc()
> >> ...
> >>
> >> + if (gc_type == FG_GC)
> >> + sbi->cur_victim_sec = NULL_SEGNO;
> >>
> >> mutex_unlock(&sbi->gc_mutex);
> >>
> >> put_gc_inode(&gc_list);
> >> ...
> >>
> >> Thanks,
> >>
> >>>
> >>>>
> >>>> So could we reset cur_victim_sec in the end of FGGC?
> >>>>
> >>>> Thanks,
> >>>>
> >>>>>
> >>>>>>
> >>>>>> Thanks,
> >>>>>>
> >>>>>>>
> >>>>>>> Thanks,
> >>>>>>>
> >>>>>>>>
> >>>>>>>> Thanks,
> >>>>>>>>
> >>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> Thanks,
> >>>>>>>>>>
> >>>>>>>>>>> sbi->cur_victim_sec = NULL_SEGNO;
> >>>>>>>>>>>
> >>>>>>>>>>> if (sync)
> >>>>>>>>>>>
> >>>>>>>>> .
> >>>>>>>>>
> >>>>>>> .
> >>>>>>>
> >>>>> .
> >>>>>
> >>> .
> >>>
> > .
> >

\
 
 \ /
  Last update: 2019-09-19 19:11    [W:0.094 / U:0.476 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site