Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Wed, 5 Apr 2023 08:55:47 -0700
From:   Jaegeuk Kim <jaegeuk@kernel.org>
To:     Chao Yu <chao@kernel.org>
Cc:     linux-f2fs-devel@lists.sourceforge.net,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH] f2fs: fix to trigger a checkpoint in the end of
 foreground garbage collection
Message-ID: <ZC2aA+i5+HpdJ6M2@google.com>
References: <20230324071028.336982-1-chao@kernel.org>
 <ZCyZGgf4RSEjyHTF@google.com>
 <a4e49177-3959-eb2b-996c-5d07b7390495@kernel.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <a4e49177-3959-eb2b-996c-5d07b7390495@kernel.org>
Precedence: bulk

On 04/05, Chao Yu wrote:
> On 2023/4/5 5:39, Jaegeuk Kim wrote:
> > Can we do like this?
> > 
> >  From 9a58f0e59364241aa31b555cfe793d278e39b0dc Mon Sep 17 00:00:00 2001
> > From: Jaegeuk Kim <jaegeuk@kernel.org>
> > Date: Tue, 4 Apr 2023 14:36:00 -0700
> > Subject: [PATCH] f2fs: do checkpoint when there's not enough free sections
> > 
> > We didn't do checkpoint in FG_GC case, which may cause losing to reclaim prefree
> > sctions in time.
> > 
> > Fixes: 6f8d4455060d ("f2fs: avoid fi->i_gc_rwsem[WRITE] lock in f2fs_gc")
> > Signed-off-by: Chao Yu <chao@kernel.org>
> > Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
> > ---
> >   fs/f2fs/gc.c | 24 +++++++++++-------------
> >   1 file changed, 11 insertions(+), 13 deletions(-)
> > 
> > diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> > index 56c53dbe05c9..f1d0dd9c5a6c 100644
> > --- a/fs/f2fs/gc.c
> > +++ b/fs/f2fs/gc.c
> > @@ -1806,6 +1806,7 @@ int f2fs_gc(struct f2fs_sb_info *sbi, struct f2fs_gc_control *gc_control)
> >   	};
> >   	unsigned int skipped_round = 0, round = 0;
> >   	unsigned int upper_secs;
> > +	bool stop_gc = false;
> >   	trace_f2fs_gc_begin(sbi->sb, gc_type, gc_control->no_bg_gc,
> >   				gc_control->nr_free_secs,
> > @@ -1876,19 +1877,15 @@ int f2fs_gc(struct f2fs_sb_info *sbi, struct f2fs_gc_control *gc_control)
> >   				(gc_type == FG_GC) ? sec_freed : 0, 0)) {
> >   		if (gc_type == FG_GC && sec_freed < gc_control->nr_free_secs)
> >   			goto go_gc_more;
> > -		goto stop;
> > -	}
> > -
> > -	/* FG_GC stops GC by skip_count */
> > -	if (gc_type == FG_GC) {
> > +		stop_gc = true;
> 
> I guess below condition is for emergency recycle of prefree segments during
> foreground GC, in order to avoid exhausting free sections due to to many
> metadata allocation during CP.
> 
> 	if (free_sections(sbi) <= upper_secs + NR_GC_CHECKPOINT_SECS &&
> 				prefree_segments(sbi)) {
> 
> But for common case, free_sections() is close to reserved_segments(), and
> upper_secs + NR_GC_CHECKPOINT_SECS value may be far smaller than free_sections(),
> so checkpoint may not be trggered as expected, IIUC.
> 
> So it's fine to just trigger CP in the end of foreground garbage collection?

My major concern is to avoid unnecessary checkpointing given multiple FG_GC
requests were pending in parallel. And, I don't want to add so many combination
which gives so many corner cases, and feel f2fs_gc() needs to call checkpoint
automatically in the worst case scenario only.

By the way, do we just need to call checkpoint here including FG_GC as well?

1832
1833         if (gc_type == BG_GC && has_not_enough_free_secs(sbi, 0, 0)) {
1834                 /*
1835                  * For example, if there are many prefree_segments below given
1836                  * threshold, we can make them free by checkpoint. Then, we
1837                  * secure free segments which doesn't need fggc any more.
1838                  */
1839                 if (prefree_segments(sbi)) {
1840                         ret = f2fs_write_checkpoint(sbi, &cpc);
1841                         if (ret)
1842                                 goto stop;
1843                 }
1844                 if (has_not_enough_free_secs(sbi, 0, 0))
1845                         gc_type = FG_GC;
1846         }

> 
> One other concern is for those path as below:
> - disable_checkpoint
> - ioc_gc
> - ioc_gc_range
> - ioc_resize
> ...

I think the upper caller should decide to call checkpoint, if they want to
reclaim the prefree likewise f2fs_disable_checkpoint.

> 
> We've passed gc_type as FG_GC, but the demand here is to migrate block in time,
> rather than dirtying blocks, and callers don't expect checkpoint in f2fs_gc(),
> instead the callers will do the checkpoit as it needs.
> 
> That means it's better to decouple FG_GC and write_checkpoint behavior, so I
> added another parameter .reclaim_space to just let f2fs_balance_fs() to trigger
> checkpoit in the end of f2fs_gc().

> 
> Thanks,
> 
> > +	} else if (gc_type == FG_GC) {
> > +		/* FG_GC stops GC by skip_count */
> >   		if (sbi->skipped_gc_rwsem)
> >   			skipped_round++;
> >   		round++;
> >   		if (skipped_round > MAX_SKIP_GC_COUNT &&
> > -				skipped_round * 2 >= round) {
> > -			ret = f2fs_write_checkpoint(sbi, &cpc);
> > -			goto stop;
> > -		}
> > +				skipped_round * 2 >= round)
> > +			stop_gc = true;
> >   	}
> >   	__get_secs_required(sbi, NULL, &upper_secs, NULL);
> > @@ -1901,12 +1898,13 @@ int f2fs_gc(struct f2fs_sb_info *sbi, struct f2fs_gc_control *gc_control)
> >   				prefree_segments(sbi)) {
> >   		ret = f2fs_write_checkpoint(sbi, &cpc);
> >   		if (ret)
> > -			goto stop;
> > +			stop_gc = true;
> >   	}
> >   go_gc_more:
> > -	segno = NULL_SEGNO;
> > -	goto gc_more;
> > -
> > +	if (!stop_gc) {
> > +		segno = NULL_SEGNO;
> > +		goto gc_more;
> > +	}
> >   stop:
> >   	SIT_I(sbi)->last_victim[ALLOC_NEXT] = 0;
> >   	SIT_I(sbi)->last_victim[FLUSH_DEVICE] = gc_control->victim_segno;