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, 12 Jul 2023 08:55:20 -0700
From:   Jaegeuk Kim <jaegeuk@kernel.org>
To:     Chao Yu <chao@kernel.org>
Cc:     Daejun Park <daejun7.park@samsung.com>,
        linux-f2fs-devel@lists.sourceforge.net,
        linux-kernel@vger.kernel.org
Subject: Re: [f2fs-dev] [PATCH v2] f2fs: do not issue small discard commands
 during checkpoint
Message-ID: <ZK7M6EkLkV8UbqUl@google.com>
References: <20230613203947.2745943-1-jaegeuk@kernel.org>
 <ZInmkgjDnAUD5Nk0@google.com>
 <50d5fa8c-4fe9-8a03-be78-0b5383e55b62@kernel.org>
 <ZKP6EJ5dZ4f4wScp@google.com>
 <65143701-4c19-ab66-1500-abd1162639cd@kernel.org>
 <ZKWovWZDiHjMavtB@google.com>
 <cadfb8d7-f5d0-a3ec-cafb-a0c06ad7d290@kernel.org>
 <ZK2FT9CUjxXvQ2K5@google.com>
 <330c96f7-fbad-dd17-6368-f1378b3b5375@kernel.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <330c96f7-fbad-dd17-6368-f1378b3b5375@kernel.org>
Precedence: bulk

On 07/12, Chao Yu wrote:
> On 2023/7/12 0:37, Jaegeuk Kim wrote:
> > On 07/06, Chao Yu wrote:
> > > On 2023/7/6 1:30, Jaegeuk Kim wrote:
> > > > On 07/04, Chao Yu wrote:
> > > > > On 2023/7/4 18:53, Jaegeuk Kim wrote:
> > > > > > On 07/03, Chao Yu wrote:
> > > > > > > On 2023/6/15 0:10, Jaegeuk Kim wrote:
> > > > > > > > If there're huge # of small discards, this will increase checkpoint latency
> > > > > > > > insanely. Let's issue small discards only by trim.
> > > > > > > > 
> > > > > > > > Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
> > > > > > > > ---
> > > > > > > > 
> > > > > > > >      Change log from v1:
> > > > > > > >       - move the skip logic to avoid dangling objects
> > > > > > > > 
> > > > > > > >      fs/f2fs/segment.c | 2 +-
> > > > > > > >      1 file changed, 1 insertion(+), 1 deletion(-)
> > > > > > > > 
> > > > > > > > diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> > > > > > > > index 8c7af8b4fc47..0457d620011f 100644
> > > > > > > > --- a/fs/f2fs/segment.c
> > > > > > > > +++ b/fs/f2fs/segment.c
> > > > > > > > @@ -2193,7 +2193,7 @@ void f2fs_clear_prefree_segments(struct f2fs_sb_info *sbi,
> > > > > > > >      			len = next_pos - cur_pos;
> > > > > > > >      			if (f2fs_sb_has_blkzoned(sbi) ||
> > > > > > > > -			    (force && len < cpc->trim_minlen))
> > > > > > > > +					!force || len < cpc->trim_minlen)
> > > > > > > >      				goto skip;
> > > > > > > 
> > > > > > > Sorry for late reply.
> > > > > > > 
> > > > > > > We have a configuration for such case, what do you think of setting
> > > > > > > max_small_discards to zero? otherwise, w/ above change, max_small_discards
> > > > > > > logic may be broken?
> > > > > > > 
> > > > > > > What:           /sys/fs/f2fs/<disk>/max_small_discards
> > > > > > > Date:           November 2013
> > > > > > > Contact:        "Jaegeuk Kim" <jaegeuk.kim@samsung.com>
> > > > > > > Description:    Controls the issue rate of discard commands that consist of small
> > > > > > >                    blocks less than 2MB. The candidates to be discarded are cached until
> > > > > > >                    checkpoint is triggered, and issued during the checkpoint.
> > > > > > >                    By default, it is disabled with 0.
> > > > > > > 
> > > > > > > Or, if we prefer to disable small_discards by default, what about below change:
> > > > > > 
> > > > > > I think small_discards is fine, but need to avoid long checkpoint latency only.
> > > > > 
> > > > > I didn't get you, do you mean we can still issue small discard by
> > > > > fstrim, so small_discards functionality is fine?
> > > > 
> > > > You got the point.
> > > 
> > > Well, actually, what I mean is max_small_discards sysfs entry's functionality
> > > is broken. Now, the entry can not be used to control number of small discards
> > > committed by checkpoint.
> > 
> > Could you descrbie this problem first?
> 
> Oh, alright, actually, I've described this problem literally, but maybe it's not
> clear, let me give some examples as below:
> 
> echo 0 > /sys/fs/f2fs/vdb/max_small_discards
> xfs_io -f /mnt/f2fs/file -c "pwrite 0 2m" -c "fsync"
> xfs_io /mnt/f2fs/file -c "fpunch 0 4k"
> sync
> cat /proc/fs/f2fs/vdb/discard_plist_info |head -2
> 
> echo 100 > /sys/fs/f2fs/vdb/max_small_discards
> rm /mnt/f2fs/file
> xfs_io -f /mnt/f2fs/file -c "pwrite 0 2m" -c "fsync"
> xfs_io /mnt/f2fs/file -c "fpunch 0 4k"
> sync
> cat /proc/fs/f2fs/vdb/discard_plist_info |head -2
> 
> Before the patch:
> 
> Discard pend list(Show diacrd_cmd count on each entry, .:not exist):
>   0         .       .       .       .       .       .       .       .
> 
> Discard pend list(Show diacrd_cmd count on each entry, .:not exist):
>   0         3       1       .       .       .       .       .       .
> 
> After the patch:
> Discard pend list(Show diacrd_cmd count on each entry, .:not exist):
>   0         .       .       .       .       .       .       .       .
> 
> Discard pend list(Show diacrd_cmd count on each entry, .:not exist):
>   0         .       .       .       .       .       .       .       .
> 
> So, now max_small_discards can not be used to control small discard number
> cached by checkpoint.

Since we do not submit small discards anymore during checkpoint. Why not relying
on the discard thread to issue them?

> 
> Thanks,
> 
> > 
> > > 
> > > I think there is another way to achieve "avoid long checkpoint latency caused
> > > by committing huge # of small discards", the way is we can set max_small_discards
> > > to small value or zero, w/ such configuration, it will take checkpoint much less
> > > time or no time to committing small discard due to below control logic:
> > > 
> > > f2fs_flush_sit_entries()
> > > {
> > > ...
> > > 			if (!(cpc->reason & CP_DISCARD)) {
> > > 				cpc->trim_start = segno;
> > > 				add_discard_addrs(sbi, cpc, false);
> > > 			}
> > > ...
> > > }
> > > 
> > > add_discard_addrs()
> > > {
> > > ...
> > > 	while (force || SM_I(sbi)->dcc_info->nr_discards <=
> > > 				SM_I(sbi)->dcc_info->max_discards) {
> > > 
> > > It will break the loop once nr_discards is larger than max_discards, if
> > > max_discards is set to zero, checkpoint won't take time to handle small discards.
> > > 
> > > ...
> > > 		if (!de) {
> > > 			de = f2fs_kmem_cache_alloc(discard_entry_slab,
> > > 						GFP_F2FS_ZERO, true, NULL);
> > > 			de->start_blkaddr = START_BLOCK(sbi, cpc->trim_start);
> > > 			list_add_tail(&de->list, head);
> > > 		}
> > > ...
> > > 	}
> > > ...
> > > 
> > > Thanks,
> > > 
> > > > 
> > > > > 
> > > > > Thanks,
> > > > > 
> > > > > > 
> > > > > > > 
> > > > > > >    From eb89d9b56e817e3046d7fa17165b12416f09d456 Mon Sep 17 00:00:00 2001
> > > > > > > From: Chao Yu <chao@kernel.org>
> > > > > > > Date: Mon, 3 Jul 2023 09:06:53 +0800
> > > > > > > Subject: [PATCH] Revert "f2fs: enable small discard by default"
> > > > > > > 
> > > > > > > This reverts commit d618ebaf0aa83d175658aea5291e0c459d471d39 in order
> > > > > > > to disable small discard by default, so that if there're huge number of
> > > > > > > small discards, it will decrease checkpoint's latency obviously.
> > > > > > > 
> > > > > > > Also, this patch reverts 9ac00e7cef10 ("f2fs: do not issue small discard
> > > > > > > commands during checkpoint"), due to it breaks small discard feature which
> > > > > > > may be configured via sysfs entry max_small_discards.
> > > > > > > 
> > > > > > > Fixes: 9ac00e7cef10 ("f2fs: do not issue small discard commands during checkpoint")
> > > > > > > Signed-off-by: Chao Yu <chao@kernel.org>
> > > > > > > ---
> > > > > > >     fs/f2fs/segment.c | 4 ++--
> > > > > > >     1 file changed, 2 insertions(+), 2 deletions(-)
> > > > > > > 
> > > > > > > diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> > > > > > > index 14c822e5c9c9..0a313368f18b 100644
> > > > > > > --- a/fs/f2fs/segment.c
> > > > > > > +++ b/fs/f2fs/segment.c
> > > > > > > @@ -2193,7 +2193,7 @@ void f2fs_clear_prefree_segments(struct f2fs_sb_info *sbi,
> > > > > > >     			len = next_pos - cur_pos;
> > > > > > > 
> > > > > > >     			if (f2fs_sb_has_blkzoned(sbi) ||
> > > > > > > -					!force || len < cpc->trim_minlen)
> > > > > > > +			    (force && len < cpc->trim_minlen))
> > > > > > >     				goto skip;
> > > > > > > 
> > > > > > >     			f2fs_issue_discard(sbi, entry->start_blkaddr + cur_pos,
> > > > > > > @@ -2269,7 +2269,7 @@ static int create_discard_cmd_control(struct f2fs_sb_info *sbi)
> > > > > > >     	atomic_set(&dcc->queued_discard, 0);
> > > > > > >     	atomic_set(&dcc->discard_cmd_cnt, 0);
> > > > > > >     	dcc->nr_discards = 0;
> > > > > > > -	dcc->max_discards = MAIN_SEGS(sbi) << sbi->log_blocks_per_seg;
> > > > > > > +	dcc->max_discards = 0;
> > > > > > >     	dcc->max_discard_request = DEF_MAX_DISCARD_REQUEST;
> > > > > > >     	dcc->min_discard_issue_time = DEF_MIN_DISCARD_ISSUE_TIME;
> > > > > > >     	dcc->mid_discard_issue_time = DEF_MID_DISCARD_ISSUE_TIME;
> > > > > > > -- 
> > > > > > > 2.40.1
> > > > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > >      			f2fs_issue_discard(sbi, entry->start_blkaddr + cur_pos,