Messages in this thread | | | Date | Wed, 18 Sep 2019 09:47:54 -0700 | From | Jaegeuk Kim <> | Subject | Re: [f2fs-dev] [PATCH 1/2] f2fs: do not select same victim right again |
| |
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. 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) > >>>>>>>>> > >>>>>>> . > >>>>>>> > >>>>> . > >>>>> > >>> . > >>> > > . > >
| |