Re: [PATCH] f2fs: skip checkpoint if having a dirty segment but no prefree at BG_GC -> FG_GC

[Jaegeuk Kim <jaegeuk@kernel.org>]

On 02/27, Jaegeuk Kim wrote:
> On 02/27, Chao Yu wrote:
> > On 2017/2/26 3:56, Jaegeuk Kim wrote:
> > > On 02/25, guoweichao wrote:
> > >> Hi Jaegeuk,
> > >>
> > >> I regard no enough free sections as a precondition when talking about
> > >> BG_GC -> FG_GC. I mean that for both case a) and b) I mentioned has no enough
> > >> free sections implicitly. 
> > >>
> > >> On 2017/2/25 2:49, Jaegeuk Kim wrote:
> > >>> Hi Weichao,
> > >>>
> > >>> On 02/25, Weichao Guo wrote:
> > >>>> When turning to FG_GC from BG_GC, we need to write checkpoint in 2 cases:
> > >>>> * a) BG_GC have made some progress, e.g.: some prefree segments.
> > >>>> * b) There is no victim and no prefree segment.
> > >>>
> > >>> You missed
> > >>>   * c) has_not_enough_free_secs() introduced by
> > >>>       6e17bfbc75a5cb ("f2fs: fix to overcome inline_data floods")
> > >> As we have enabled SSR for warm node(5b6c6be2d8 ("f2fs: use SSR for warm node as well")),
> > >> I think inline data floods should not be a problem in most cases.
> > >>>
> > >>> And, Yunlong pointed that we can't find a case to avoid write_checkpoint()
> > >>> mostly due to c) condition.
> > >> As inline data floods is an extreme case, and there is little possibility caused panic
> > >> for inline data floods now, there should be lots of chance to skip checkpoint. I mean
> > >> that we can make some accurate inline data floods checking before writing checkpoint.
> > > 
> > > For now, the safest way is our first option. So, I decided to start with doing
> > > checkpoint due to previous inline_data flooding issue even though it's an
> > > extreme case under SSR.
> > > 
> > > Anyway, I agree that we need to find a way to detect when to avoid checkpoint.
> > 
> > Hi all,
> > 
> > I proposed a approach before, can you please check that one?
> > 
> > https://www.mail-archive.com/linux-f2fs-devel@lists.sourceforge.net/msg03632.html
> 
> Oh, right, let's take a look at this. ;)

Hmm, I just read this patch again, and realized it doesn't quite address the
current issue. This patch flushes inline_data inodes in background, which does
not guarantee this worst case. The key idea would be how to measure the space
we can do SSR and use it in has_not_enough_free_secs().

Thanks,

> 
> Thanks,
> 
> > 
> > Thanks,
> > 
> > > 
> > > Thanks,
> > > 
> > >>>
> > >>> Thanks,
> > >>>
> > >>>>
> > >>>> For case a), previously, we also check if there is a dirty segment for
> > >>>> infering blocks moving in last BG_GC. But dirty segments do not always
> > >>>> indicate that, BG_GC may just start and do not move any blocks at all.
> > >>>> Futhermore, skipping checkpoint if there are some dirty segments but no
> > >>>> prefree segments is OK.
> > >>>
> > >>>
> > >>>>
> > >>>> Signed-off-by: Weichao Guo <guoweichao@huawei.com>
> > >>>> ---
> > >>>>  fs/f2fs/gc.c | 7 ++++++-
> > >>>>  1 file changed, 6 insertions(+), 1 deletion(-)
> > >>>>
> > >>>> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> > >>>> index 6c996e3..30d206a 100644
> > >>>> --- a/fs/f2fs/gc.c
> > >>>> +++ b/fs/f2fs/gc.c
> > >>>> @@ -958,7 +958,12 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync, bool background)
> > >>>>  		 * enough free sections, we should flush dent/node blocks and do
> > >>>>  		 * garbage collections.
> > >>>>  		 */
> > >>>> -		ret = write_checkpoint(sbi, &cpc);
> > >>>> +		if (prefree_segments(sbi))
> > >>>> +			ret = write_checkpoint(sbi, &cpc);
> > >>>> +		else if (!__get_victim(sbi, &segno, gc_type) {
> > >>>> +			segno = NULL_SEGNO;
> > >>>> +			ret = write_checkpoint(sbi, &cpc);
> > >>>> +		}
> > >>>>  		if (ret)
> > >>>>  			goto stop;
> > >>>>  	} else if (gc_type == BG_GC && !background) {
> > >>>> -- 
> > >>>> 2.10.1
> > >>>
> > >>> .
> > >>>
> > >>
> > >> Thanks,
> > >> Weichao
> > > 
> > > .
> > > 
> 
> ------------------------------------------------------------------------------
> Check out the vibrant tech community on one of the world's most
> engaging tech sites, SlashDot.org! http://sdm.link/slashdot
> _______________________________________________
> Linux-f2fs-devel mailing list
> Linux-f2fs-devel@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, SlashDot.org! http://sdm.link/slashdot