[RFC] add ovp valid_blocks check for bg gc victim to fg_gc

[RFC] add ovp valid_blocks check for bg gc victim to fg_gc

[Hou Pengyang]

For foreground gc, greedy algorithm should be adapted, which makes
this formula work well:

	(2 * (100 / config.overprovision + 1) + 6)

But currently, we fg_gc have a prior to select bg_gc victim segments to gc first,
these victims are selected by cost-benefit algorithm, we can't guarantee such segments
have the small valid blocks, which may destroy the f2fs rule, on the worstest case, would 
consume all the free segments.

This patch fix this by add a filter in check_bg_victims, if segment's has # of valid blocks
over overprovision ratio, skip such segments.

Signed-off-by: Hou Pengyang <houpengyang@huawei.com>
---
 fs/f2fs/f2fs.h    |  1 +
 fs/f2fs/gc.c      |  4 ++++
 fs/f2fs/segment.h | 12 ++++++++++++
 fs/f2fs/super.c   | 10 ++++++++++
 4 files changed, 27 insertions(+)

diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
index c9a97c3..0afe94f 100644
--- a/fs/f2fs/f2fs.h
+++ b/fs/f2fs/f2fs.h
@@ -881,6 +881,7 @@ struct f2fs_sb_info {
 	struct mutex gc_mutex;			/* mutex for GC */
 	struct f2fs_gc_kthread	*gc_thread;	/* GC thread */
 	unsigned int cur_victim_sec;		/* current victim section num */
+	unsigned int fggc_threshold;	/* threshold for converting bg victims for fg */
 
 	/* maximum # of trials to find a victim segment for SSR and GC */
 	unsigned int max_victim_search;
diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
index 88e5e7b..18f3efa 100644
--- a/fs/f2fs/gc.c
+++ b/fs/f2fs/gc.c
@@ -199,6 +199,10 @@ static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
 	for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
 		if (sec_usage_check(sbi, secno))
 			continue;
+
+		if (valid_blocks_ovp_check(sbi, secno))
+			continue;
+
 		clear_bit(secno, dirty_i->victim_secmap);
 		return secno * sbi->segs_per_sec;
 	}
diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
index 5cb5755..2a6a1a9 100644
--- a/fs/f2fs/segment.h
+++ b/fs/f2fs/segment.h
@@ -716,6 +716,18 @@ static inline block_t sum_blk_addr(struct f2fs_sb_info *sbi, int base, int type)
 				- (base + 1) + type;
 }
 
+static inline bool valid_blocks_ovp_check(struct f2fs_sb_info *sbi, unsigned int secno)
+{
+	struct sit_info *sit_i = SIT_I(sbi);
+	unsigned int valid_blocks;
+
+	valid_blocks = sit_i->sec_entries[secno].valid_blocks;
+	if (valid_blocks >= sbi->fggc_threshold)
+		return true;
+	return false;
+
+}
+
 static inline bool sec_usage_check(struct f2fs_sb_info *sbi, unsigned int secno)
 {
 	if (IS_CURSEC(sbi, secno) || (sbi->cur_victim_sec == secno))
diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
index 68a555b..54c4dac 100644
--- a/fs/f2fs/super.c
+++ b/fs/f2fs/super.c
@@ -2090,6 +2090,16 @@ static int f2fs_fill_super(struct super_block *sb, void *data, int silent)
 		if (err)
 			goto free_kobj;
 	}
+
+	sbi->fggc_threshold = (le32_to_cpu(sbi->ckpt->overprov_segment_count) -
+					le32_to_cpu(sbi->ckpt->rsvd_segment_count)) * 100 /
+					(le32_to_cpu(sbi->raw_super->segment_count_main) -
+					le32_to_cpu(sbi->ckpt->rsvd_segment_count)) *
+					sbi->blocks_per_seg * sbi->segs_per_sec / 100;
+
+	sbi->fggc_threshold = sbi->blocks_per_seg * sbi->segs_per_sec -
+					sbi->fggc_threshold;
+
 	kfree(options);
 
 	/* recover broken superblock */
-- 
2.10.1


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

Re: [RFC] add ovp valid_blocks check for bg gc victim to fg_gc

[Jaegeuk Kim]

Hi Pengyang,

Nice catch!

I think fggc_threshold needs to be revised, and we need to consider about
general victim selection as well.

Could you take a look at this?

>From 23b265f5ca6405032d092e240c94a827f743e42b Mon Sep 17 00:00:00 2001
From: Hou Pengyang <houpengyang@huawei.com>
Date: Thu, 16 Feb 2017 12:34:31 +0000
Subject: [PATCH] f2fs: add ovp valid_blocks check for bg gc victim to fg_gc

For foreground gc, greedy algorithm should be adapted, which makes
this formula work well:

	(2 * (100 / config.overprovision + 1) + 6)

But currently, we fg_gc have a prior to select bg_gc victim segments to gc first,
these victims are selected by cost-benefit algorithm, we can't guarantee such segments
have the small valid blocks, which may destroy the f2fs rule, on the worstest case, would
consume all the free segments.

This patch fix this by add a filter in check_bg_victims, if segment's has # of valid blocks
over overprovision ratio, skip such segments.

Cc: <stable@vger.kernel.org>
Signed-off-by: Hou Pengyang <houpengyang@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
---
 fs/f2fs/f2fs.h    |  3 +++
 fs/f2fs/gc.c      | 22 ++++++++++++++++++++--
 fs/f2fs/segment.h |  9 +++++++++
 3 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
index cc22dc458896..1c9f0cc8f027 100644
--- a/fs/f2fs/f2fs.h
+++ b/fs/f2fs/f2fs.h
@@ -888,6 +888,9 @@ struct f2fs_sb_info {
 	struct f2fs_gc_kthread	*gc_thread;	/* GC thread */
 	unsigned int cur_victim_sec;		/* current victim section num */
 
+	/* threshold for converting bg victims for fg */
+	u64 fggc_threshold;
+
 	/* maximum # of trials to find a victim segment for SSR and GC */
 	unsigned int max_victim_search;
 
diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
index 88e5e7b10ab6..fd4e479820e6 100644
--- a/fs/f2fs/gc.c
+++ b/fs/f2fs/gc.c
@@ -166,7 +166,8 @@ static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
 		p->ofs_unit = sbi->segs_per_sec;
 	}
 
-	if (p->max_search > sbi->max_victim_search)
+	/* we need to check every dirty segments in the FG_GC case */
+	if (gc_type != FG_GC && p->max_search > sbi->max_victim_search)
 		p->max_search = sbi->max_victim_search;
 
 	p->offset = sbi->last_victim[p->gc_mode];
@@ -199,6 +200,10 @@ static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
 	for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
 		if (sec_usage_check(sbi, secno))
 			continue;
+
+		if (no_fggc_candidate(sbi, secno))
+			continue;
+
 		clear_bit(secno, dirty_i->victim_secmap);
 		return secno * sbi->segs_per_sec;
 	}
@@ -322,13 +327,15 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
 			nsearched++;
 		}
 
-
 		secno = GET_SECNO(sbi, segno);
 
 		if (sec_usage_check(sbi, secno))
 			goto next;
 		if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
 			goto next;
+		if (gc_type == FG_GC && p.alloc_mode == LFS &&
+					no_fggc_candidate(sbi, secno))
+			goto next;
 
 		cost = get_gc_cost(sbi, segno, &p);
 
@@ -989,5 +996,16 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync, bool background)
 
 void build_gc_manager(struct f2fs_sb_info *sbi)
 {
+	u64 user_block_count, ovp_count, blocks_per_sec, th;
+
 	DIRTY_I(sbi)->v_ops = &default_v_ops;
+
+	/* threshold of # of valid blocks in a section for victims of FG_GC */
+	user_block_count = sbi->user_block_count;
+	ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
+	blocks_per_sec = sbi->blocks_per_seg * sbi->segs_per_sec;
+
+	th = user_block_count * 100 * blocks_per_sec /
+			((user_block_count + ovp_count) * 100);
+	sbi->fggc_threshold = th;
 }
diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
index 5cb5755c75d9..f4020f141d83 100644
--- a/fs/f2fs/segment.h
+++ b/fs/f2fs/segment.h
@@ -716,6 +716,15 @@ static inline block_t sum_blk_addr(struct f2fs_sb_info *sbi, int base, int type)
 				- (base + 1) + type;
 }
 
+static inline bool no_fggc_candidate(struct f2fs_sb_info *sbi,
+						unsigned int secno)
+{
+	if (get_valid_blocks(sbi, secno, sbi->segs_per_sec) >=
+						sbi->fggc_threshold)
+		return true;
+	return false;
+}
+
 static inline bool sec_usage_check(struct f2fs_sb_info *sbi, unsigned int secno)
 {
 	if (IS_CURSEC(sbi, secno) || (sbi->cur_victim_sec == secno))
-- 
2.11.0


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

Re: [RFC] add ovp valid_blocks check for bg gc victim to fg_gc

[Hou Pengyang]

On 2017/2/17 7:48, Jaegeuk Kim wrote:
> Hi Pengyang,
>
> Nice
Hi Jaegeuk,
catch!
>
> I think fggc_threshold needs to be revised, and we need to consider about
> general victim selection as well.
>
> Could you take a look at this?
>
>>From 23b265f5ca6405032d092e240c94a827f743e42b Mon Sep 17 00:00:00 2001
> From: Hou Pengyang <houpengyang@huawei.com>
> Date: Thu, 16 Feb 2017 12:34:31 +0000
> Subject: [PATCH] f2fs: add ovp valid_blocks check for bg gc victim to fg_gc
>
> For foreground gc, greedy algorithm should be adapted, which makes
> this formula work well:
>
> 	(2 * (100 / config.overprovision + 1) + 6)
>
> But currently, we fg_gc have a prior to select bg_gc victim segments to gc first,
> these victims are selected by cost-benefit algorithm, we can't guarantee such segments
> have the small valid blocks, which may destroy the f2fs rule, on the worstest case, would
> consume all the free segments.
>
> This patch fix this by add a filter in check_bg_victims, if segment's has # of valid blocks
> over overprovision ratio, skip such segments.
>
> Cc: <stable@vger.kernel.org>
> Signed-off-by: Hou Pengyang <houpengyang@huawei.com>
> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
> ---
>   fs/f2fs/f2fs.h    |  3 +++
>   fs/f2fs/gc.c      | 22 ++++++++++++++++++++--
>   fs/f2fs/segment.h |  9 +++++++++
>   3 files changed, 32 insertions(+), 2 deletions(-)
>
> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> index cc22dc458896..1c9f0cc8f027 100644
> --- a/fs/f2fs/f2fs.h
> +++ b/fs/f2fs/f2fs.h
> @@ -888,6 +888,9 @@ struct f2fs_sb_info {
>   	struct f2fs_gc_kthread	*gc_thread;	/* GC thread */
>   	unsigned int cur_victim_sec;		/* current victim section num */
>
> +	/* threshold for converting bg victims for fg */
> +	u64 fggc_threshold;
> +
>   	/* maximum # of trials to find a victim segment for SSR and GC */
>   	unsigned int max_victim_search;
>
> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> index 88e5e7b10ab6..fd4e479820e6 100644
> --- a/fs/f2fs/gc.c
> +++ b/fs/f2fs/gc.c
> @@ -166,7 +166,8 @@ static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
>   		p->ofs_unit = sbi->segs_per_sec;
>   	}
>
> -	if (p->max_search > sbi->max_victim_search)
> +	/* we need to check every dirty segments in the FG_GC case */
> +	if (gc_type != FG_GC && p->max_search > sbi->max_victim_search)
>   		p->max_search = sbi->max_victim_search;
>
>   	p->offset = sbi->last_victim[p->gc_mode];
> @@ -199,6 +200,10 @@ static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
>   	for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
>   		if (sec_usage_check(sbi, secno))
>   			continue;
> +
> +		if (no_fggc_candidate(sbi, secno))
> +			continue;
> +
>   		clear_bit(secno, dirty_i->victim_secmap);
>   		return secno * sbi->segs_per_sec;
>   	}
> @@ -322,13 +327,15 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>   			nsearched++;
>   		}
>
> -
>   		secno = GET_SECNO(sbi, segno);
>
>   		if (sec_usage_check(sbi, secno))
>   			goto next;
>   		if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
>   			goto next;
> +		if (gc_type == FG_GC && p.alloc_mode == LFS &&
> +					no_fggc_candidate(sbi, secno))
> +			goto next;
>
>   		cost = get_gc_cost(sbi, segno, &p);
>
> @@ -989,5 +996,16 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync, bool background)
>
>   void build_gc_manager(struct f2fs_sb_info *sbi)
>   {
> +	u64 user_block_count, ovp_count, blocks_per_sec, th;
> +
>   	DIRTY_I(sbi)->v_ops = &default_v_ops;
> +
> +	/* threshold of # of valid blocks in a section for victims of FG_GC */
> +	user_block_count = sbi->user_block_count;
> +	ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;

About the ovp_count calculation,

in mkfs.f2fs, we get ovp_segment by

     set_cp(overprov_segment_count, (get_sb(segment_count_main) -
            get_cp(rsvd_segment_count)) *
             config.overprovision / 100);

     set_cp(overprov_segment_count, get_cp(overprov_segment_count) +
             get_cp(rsvd_segment_count));

where the overprov calculation is based on the space excluding the
rsvd segment, and the final overprov_segment is sum of the REAL
overprov segments and the rsvd ones.

So, when to calculate the overprov ratio, the rsvd segments should
be subtracted from the ckpt->overprov_semgents?

Thanks,

> +	blocks_per_sec = sbi->blocks_per_seg * sbi->segs_per_sec;
> +
> +	th = user_block_count * 100 * blocks_per_sec /
> +			((user_block_count + ovp_count) * 100);
> +	sbi->fggc_threshold = th;
>   }
> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
> index 5cb5755c75d9..f4020f141d83 100644
> --- a/fs/f2fs/segment.h
> +++ b/fs/f2fs/segment.h
> @@ -716,6 +716,15 @@ static inline block_t sum_blk_addr(struct f2fs_sb_info *sbi, int base, int type)
>   				- (base + 1) + type;
>   }
>
> +static inline bool no_fggc_candidate(struct f2fs_sb_info *sbi,
> +						unsigned int secno)
> +{
> +	if (get_valid_blocks(sbi, secno, sbi->segs_per_sec) >=
> +						sbi->fggc_threshold)
> +		return true;
> +	return false;
> +}
> +
>   static inline bool sec_usage_check(struct f2fs_sb_info *sbi, unsigned int secno)
>   {
>   	if (IS_CURSEC(sbi, secno) || (sbi->cur_victim_sec == secno))
>



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

Re: [RFC] add ovp valid_blocks check for bg gc victim to fg_gc

[Jaegeuk Kim]

On 02/17, Hou Pengyang wrote:
> On 2017/2/17 7:48, Jaegeuk Kim wrote:
> > Hi Pengyang,
> > 
> > Nice
> Hi Jaegeuk,
> catch!
> > 
> > I think fggc_threshold needs to be revised, and we need to consider about
> > general victim selection as well.
> > 
> > Could you take a look at this?
> > 
> > > From 23b265f5ca6405032d092e240c94a827f743e42b Mon Sep 17 00:00:00 2001
> > From: Hou Pengyang <houpengyang@huawei.com>
> > Date: Thu, 16 Feb 2017 12:34:31 +0000
> > Subject: [PATCH] f2fs: add ovp valid_blocks check for bg gc victim to fg_gc
> > 
> > For foreground gc, greedy algorithm should be adapted, which makes
> > this formula work well:
> > 
> > 	(2 * (100 / config.overprovision + 1) + 6)
> > 
> > But currently, we fg_gc have a prior to select bg_gc victim segments to gc first,
> > these victims are selected by cost-benefit algorithm, we can't guarantee such segments
> > have the small valid blocks, which may destroy the f2fs rule, on the worstest case, would
> > consume all the free segments.
> > 
> > This patch fix this by add a filter in check_bg_victims, if segment's has # of valid blocks
> > over overprovision ratio, skip such segments.
> > 
> > Cc: <stable@vger.kernel.org>
> > Signed-off-by: Hou Pengyang <houpengyang@huawei.com>
> > Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
> > ---
> >   fs/f2fs/f2fs.h    |  3 +++
> >   fs/f2fs/gc.c      | 22 ++++++++++++++++++++--
> >   fs/f2fs/segment.h |  9 +++++++++
> >   3 files changed, 32 insertions(+), 2 deletions(-)
> > 
> > diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> > index cc22dc458896..1c9f0cc8f027 100644
> > --- a/fs/f2fs/f2fs.h
> > +++ b/fs/f2fs/f2fs.h
> > @@ -888,6 +888,9 @@ struct f2fs_sb_info {
> >   	struct f2fs_gc_kthread	*gc_thread;	/* GC thread */
> >   	unsigned int cur_victim_sec;		/* current victim section num */
> > 
> > +	/* threshold for converting bg victims for fg */
> > +	u64 fggc_threshold;
> > +
> >   	/* maximum # of trials to find a victim segment for SSR and GC */
> >   	unsigned int max_victim_search;
> > 
> > diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> > index 88e5e7b10ab6..fd4e479820e6 100644
> > --- a/fs/f2fs/gc.c
> > +++ b/fs/f2fs/gc.c
> > @@ -166,7 +166,8 @@ static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
> >   		p->ofs_unit = sbi->segs_per_sec;
> >   	}
> > 
> > -	if (p->max_search > sbi->max_victim_search)
> > +	/* we need to check every dirty segments in the FG_GC case */
> > +	if (gc_type != FG_GC && p->max_search > sbi->max_victim_search)
> >   		p->max_search = sbi->max_victim_search;
> > 
> >   	p->offset = sbi->last_victim[p->gc_mode];
> > @@ -199,6 +200,10 @@ static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
> >   	for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
> >   		if (sec_usage_check(sbi, secno))
> >   			continue;
> > +
> > +		if (no_fggc_candidate(sbi, secno))
> > +			continue;
> > +
> >   		clear_bit(secno, dirty_i->victim_secmap);
> >   		return secno * sbi->segs_per_sec;
> >   	}
> > @@ -322,13 +327,15 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> >   			nsearched++;
> >   		}
> > 
> > -
> >   		secno = GET_SECNO(sbi, segno);
> > 
> >   		if (sec_usage_check(sbi, secno))
> >   			goto next;
> >   		if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
> >   			goto next;
> > +		if (gc_type == FG_GC && p.alloc_mode == LFS &&
> > +					no_fggc_candidate(sbi, secno))
> > +			goto next;
> > 
> >   		cost = get_gc_cost(sbi, segno, &p);
> > 
> > @@ -989,5 +996,16 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync, bool background)
> > 
> >   void build_gc_manager(struct f2fs_sb_info *sbi)
> >   {
> > +	u64 user_block_count, ovp_count, blocks_per_sec, th;
> > +
> >   	DIRTY_I(sbi)->v_ops = &default_v_ops;
> > +
> > +	/* threshold of # of valid blocks in a section for victims of FG_GC */
> > +	user_block_count = sbi->user_block_count;
> > +	ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
> 
> About the ovp_count calculation,
> 
> in mkfs.f2fs, we get ovp_segment by
> 
>     set_cp(overprov_segment_count, (get_sb(segment_count_main) -
>            get_cp(rsvd_segment_count)) *
>             config.overprovision / 100);
> 
>     set_cp(overprov_segment_count, get_cp(overprov_segment_count) +
>             get_cp(rsvd_segment_count));
> 
> where the overprov calculation is based on the space excluding the
> rsvd segment, and the final overprov_segment is sum of the REAL
> overprov segments and the rsvd ones.
> 
> So, when to calculate the overprov ratio, the rsvd segments should
> be subtracted from the ckpt->overprov_semgents?

I just got calculation from fresh mounted image. What I could confirm was that
user can see (main_segments - ovp_segments).
BTW, it's worth to verify reserved_segments with that.

> 
> Thanks,
> 
> > +	blocks_per_sec = sbi->blocks_per_seg * sbi->segs_per_sec;
> > +
> > +	th = user_block_count * 100 * blocks_per_sec /
> > +			((user_block_count + ovp_count) * 100);
> > +	sbi->fggc_threshold = th;
> >   }
> > diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
> > index 5cb5755c75d9..f4020f141d83 100644
> > --- a/fs/f2fs/segment.h
> > +++ b/fs/f2fs/segment.h
> > @@ -716,6 +716,15 @@ static inline block_t sum_blk_addr(struct f2fs_sb_info *sbi, int base, int type)
> >   				- (base + 1) + type;
> >   }
> > 
> > +static inline bool no_fggc_candidate(struct f2fs_sb_info *sbi,
> > +						unsigned int secno)
> > +{
> > +	if (get_valid_blocks(sbi, secno, sbi->segs_per_sec) >=
> > +						sbi->fggc_threshold)
> > +		return true;
> > +	return false;
> > +}
> > +
> >   static inline bool sec_usage_check(struct f2fs_sb_info *sbi, unsigned int secno)
> >   {
> >   	if (IS_CURSEC(sbi, secno) || (sbi->cur_victim_sec == secno))
> > 
> 

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

Re: [RFC] add ovp valid_blocks check for bg gc victim to fg_gc

[Hou Pengyang]

On 2017/2/17 10:54, Jaegeuk Kim wrote:
> On 02/17, Hou Pengyang wrote:
>> On 2017/2/17 7:48, Jaegeuk Kim wrote:
>>> Hi Pengyang,
>>>
>>> Nice
>> Hi Jaegeuk,
>> catch!
>>>
>>> I think fggc_threshold needs to be revised, and we need to consider about
>>> general victim selection as well.
>>>
>>> Could you take a look at this?
>>>
>>>>  From 23b265f5ca6405032d092e240c94a827f743e42b Mon Sep 17 00:00:00 2001
>>> From: Hou Pengyang <houpengyang@huawei.com>
>>> Date: Thu, 16 Feb 2017 12:34:31 +0000
>>> Subject: [PATCH] f2fs: add ovp valid_blocks check for bg gc victim to fg_gc
>>>
>>> For foreground gc, greedy algorithm should be adapted, which makes
>>> this formula work well:
>>>
>>> 	(2 * (100 / config.overprovision + 1) + 6)
>>>
>>> But currently, we fg_gc have a prior to select bg_gc victim segments to gc first,
>>> these victims are selected by cost-benefit algorithm, we can't guarantee such segments
>>> have the small valid blocks, which may destroy the f2fs rule, on the worstest case, would
>>> consume all the free segments.
>>>
>>> This patch fix this by add a filter in check_bg_victims, if segment's has # of valid blocks
>>> over overprovision ratio, skip such segments.
>>>
>>> Cc: <stable@vger.kernel.org>
>>> Signed-off-by: Hou Pengyang <houpengyang@huawei.com>
>>> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
>>> ---
>>>    fs/f2fs/f2fs.h    |  3 +++
>>>    fs/f2fs/gc.c      | 22 ++++++++++++++++++++--
>>>    fs/f2fs/segment.h |  9 +++++++++
>>>    3 files changed, 32 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
>>> index cc22dc458896..1c9f0cc8f027 100644
>>> --- a/fs/f2fs/f2fs.h
>>> +++ b/fs/f2fs/f2fs.h
>>> @@ -888,6 +888,9 @@ struct f2fs_sb_info {
>>>    	struct f2fs_gc_kthread	*gc_thread;	/* GC thread */
>>>    	unsigned int cur_victim_sec;		/* current victim section num */
>>>
>>> +	/* threshold for converting bg victims for fg */
>>> +	u64 fggc_threshold;
>>> +
>>>    	/* maximum # of trials to find a victim segment for SSR and GC */
>>>    	unsigned int max_victim_search;
>>>
>>> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
>>> index 88e5e7b10ab6..fd4e479820e6 100644
>>> --- a/fs/f2fs/gc.c
>>> +++ b/fs/f2fs/gc.c
>>> @@ -166,7 +166,8 @@ static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
>>>    		p->ofs_unit = sbi->segs_per_sec;
>>>    	}
>>>
>>> -	if (p->max_search > sbi->max_victim_search)
>>> +	/* we need to check every dirty segments in the FG_GC case */
>>> +	if (gc_type != FG_GC && p->max_search > sbi->max_victim_search)
>>>    		p->max_search = sbi->max_victim_search;
>>>
>>>    	p->offset = sbi->last_victim[p->gc_mode];
>>> @@ -199,6 +200,10 @@ static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
>>>    	for_each_set_bit(secno, dirty_i->victim_secmap, MAIN_SECS(sbi)) {
>>>    		if (sec_usage_check(sbi, secno))
>>>    			continue;
>>> +
>>> +		if (no_fggc_candidate(sbi, secno))
>>> +			continue;
>>> +
>>>    		clear_bit(secno, dirty_i->victim_secmap);
>>>    		return secno * sbi->segs_per_sec;
>>>    	}
>>> @@ -322,13 +327,15 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>>>    			nsearched++;
>>>    		}
>>>
>>> -
>>>    		secno = GET_SECNO(sbi, segno);
>>>
>>>    		if (sec_usage_check(sbi, secno))
>>>    			goto next;
>>>    		if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
>>>    			goto next;
>>> +		if (gc_type == FG_GC && p.alloc_mode == LFS &&
>>> +					no_fggc_candidate(sbi, secno))
>>> +			goto next;
>>>
>>>    		cost = get_gc_cost(sbi, segno, &p);
>>>
>>> @@ -989,5 +996,16 @@ int f2fs_gc(struct f2fs_sb_info *sbi, bool sync, bool background)
>>>
>>>    void build_gc_manager(struct f2fs_sb_info *sbi)
>>>    {
>>> +	u64 user_block_count, ovp_count, blocks_per_sec, th;
>>> +
>>>    	DIRTY_I(sbi)->v_ops = &default_v_ops;
>>> +
>>> +	/* threshold of # of valid blocks in a section for victims of FG_GC */
>>> +	user_block_count = sbi->user_block_count;
>>> +	ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
>>
>> About the ovp_count calculation,
>>
>> in mkfs.f2fs, we get ovp_segment by
>>
>>      set_cp(overprov_segment_count, (get_sb(segment_count_main) -
>>             get_cp(rsvd_segment_count)) *
>>              config.overprovision / 100);
>>
>>      set_cp(overprov_segment_count, get_cp(overprov_segment_count) +
>>              get_cp(rsvd_segment_count));
>>
>> where the overprov calculation is based on the space excluding the
>> rsvd segment, and the final overprov_segment is sum of the REAL
>> overprov segments and the rsvd ones.
>>
>> So, when to calculate the overprov ratio, the rsvd segments should
>> be subtracted from the ckpt->overprov_semgents?
>
> I just got calculation from fresh mounted image. What I could confirm was that
> user can see (main_segments - ovp_segments).

I add some log into mkfs.f2fs, and get the config.overprovision,
rsvd/ovrp segments, user_block_count, and use the mkfs.f2fs to 
formatting a 512M disk, where:

config.overprovision: 10

ovrp : 50 segments(25600 blocks)
rsvd : 28 segments(14336 blocks)
user_block_count: 101376 blocks

Use ovrp/rsvd/user_block_count to get the ovrp ratio info reversely :

1) user_block_count/(user_block_count + ovrp blocks)

   101376/(101376+25600) = 79%

2) user_block_count/(user_block_count + ovrp_blocks - rsvd_blocks)

   101376/(101376+25600-14336) = 90% = 1-10%

By 2), 1-90%, we can get the original config.overprovion when mkfs.f2fs
So I think we should use 2) to calculate ovrp ratio. Anything I miss
in 2)?

Thanks,

> BTW, it's worth to verify reserved_segments with that.
>


>>
>> Thanks,
>>
>>> +	blocks_per_sec = sbi->blocks_per_seg * sbi->segs_per_sec;
>>> +
>>> +	th = user_block_count * 100 * blocks_per_sec /
>>> +			((user_block_count + ovp_count) * 100);
>>> +	sbi->fggc_threshold = th;
>>>    }
>>> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
>>> index 5cb5755c75d9..f4020f141d83 100644
>>> --- a/fs/f2fs/segment.h
>>> +++ b/fs/f2fs/segment.h
>>> @@ -716,6 +716,15 @@ static inline block_t sum_blk_addr(struct f2fs_sb_info *sbi, int base, int type)
>>>    				- (base + 1) + type;
>>>    }
>>>
>>> +static inline bool no_fggc_candidate(struct f2fs_sb_info *sbi,
>>> +						unsigned int secno)
>>> +{
>>> +	if (get_valid_blocks(sbi, secno, sbi->segs_per_sec) >=
>>> +						sbi->fggc_threshold)
>>> +		return true;
>>> +	return false;
>>> +}
>>> +
>>>    static inline bool sec_usage_check(struct f2fs_sb_info *sbi, unsigned int secno)
>>>    {
>>>    	if (IS_CURSEC(sbi, secno) || (sbi->cur_victim_sec == secno))
>>>
>>
>
> .
>



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

Re: [RFC] add ovp valid_blocks check for bg gc victim to fg_gc

[Jaegeuk Kim]

On 02/17, Hou Pengyang wrote:

...

> > > On 2017/2/17 7:48, Jaegeuk Kim wrote:
> > > > +	ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
> > > 
> > > About the ovp_count calculation,
> > > 
> > > in mkfs.f2fs, we get ovp_segment by
> > > 
> > >      set_cp(overprov_segment_count, (get_sb(segment_count_main) -
> > >             get_cp(rsvd_segment_count)) *
> > >              config.overprovision / 100);
> > > 
> > >      set_cp(overprov_segment_count, get_cp(overprov_segment_count) +
> > >              get_cp(rsvd_segment_count));
> > > 
> > > where the overprov calculation is based on the space excluding the
> > > rsvd segment, and the final overprov_segment is sum of the REAL
> > > overprov segments and the rsvd ones.
> > > 
> > > So, when to calculate the overprov ratio, the rsvd segments should
> > > be subtracted from the ckpt->overprov_semgents?
> > 
> > I just got calculation from fresh mounted image. What I could confirm was that
> > user can see (main_segments - ovp_segments).
> 
> I add some log into mkfs.f2fs, and get the config.overprovision,
> rsvd/ovrp segments, user_block_count, and use the mkfs.f2fs to formatting a
> 512M disk, where:
> 
> config.overprovision: 10
> 
> ovrp : 50 segments(25600 blocks)
> rsvd : 28 segments(14336 blocks)
> user_block_count: 101376 blocks
> 
> Use ovrp/rsvd/user_block_count to get the ovrp ratio info reversely :
> 
> 1) user_block_count/(user_block_count + ovrp blocks)
> 
>   101376/(101376+25600) = 79%
> 
> 2) user_block_count/(user_block_count + ovrp_blocks - rsvd_blocks)
> 
>   101376/(101376+25600-14336) = 90% = 1-10%
> 
> By 2), 1-90%, we can get the original config.overprovion when mkfs.f2fs
> So I think we should use 2) to calculate ovrp ratio. Anything I miss
> in 2)?

Yup, indeed 2) is correct, which I found my calcuation was wrong, and tried
to verify with:

threshold = (main_segments - ovp_segments) / (main_segments - reserved_segments)
            ------------------------------    ---------------------------------
                User-visible space                  F2FS-writable space
              (user_block_count) / (user_block_count + ovp_blocks - rsvd_blocks)

For example,

1. 8GB
 ovp reported by mkfs = 2.22%
 main_segments = 4065
 ovp_segments = 186
 reserved_segments = 98
 threshold = 500 (mine) vs. 502 (yours)

 -> Actual ovp = (512-500) / 512 = 2.34%
 -> reserved segments = 98 / 2 = 49 can cover 2.34%

2. 64GB
 ovp reported by mkfs = 0.78%
 main_segments = 32581
 ovp_segments = 516
 reserved_segments = 264
 threshold = 508 (mine) vs. 512 (yours)

 -> Actual ovp = (512-508) / 512 = 0.78%
 -> reserved segments = 264 / 2 = 132 can cover 0.78%

Thanks,

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

Re: [RFC] add ovp valid_blocks check for bg gc victim to fg_gc

[Chao Yu]

On 2017/2/18 4:17, Jaegeuk Kim wrote:
> On 02/17, Hou Pengyang wrote:
> 
> ...
> 
>>>> On 2017/2/17 7:48, Jaegeuk Kim wrote:
>>>>> +	ovp_count = SM_I(sbi)->ovp_segments << sbi->log_blocks_per_seg;
>>>>
>>>> About the ovp_count calculation,
>>>>
>>>> in mkfs.f2fs, we get ovp_segment by
>>>>
>>>>      set_cp(overprov_segment_count, (get_sb(segment_count_main) -
>>>>             get_cp(rsvd_segment_count)) *
>>>>              config.overprovision / 100);
>>>>
>>>>      set_cp(overprov_segment_count, get_cp(overprov_segment_count) +
>>>>              get_cp(rsvd_segment_count));
>>>>
>>>> where the overprov calculation is based on the space excluding the
>>>> rsvd segment, and the final overprov_segment is sum of the REAL
>>>> overprov segments and the rsvd ones.
>>>>
>>>> So, when to calculate the overprov ratio, the rsvd segments should
>>>> be subtracted from the ckpt->overprov_semgents?
>>>
>>> I just got calculation from fresh mounted image. What I could confirm was that
>>> user can see (main_segments - ovp_segments).
>>
>> I add some log into mkfs.f2fs, and get the config.overprovision,
>> rsvd/ovrp segments, user_block_count, and use the mkfs.f2fs to formatting a
>> 512M disk, where:
>>
>> config.overprovision: 10
>>
>> ovrp : 50 segments(25600 blocks)
>> rsvd : 28 segments(14336 blocks)
>> user_block_count: 101376 blocks
>>
>> Use ovrp/rsvd/user_block_count to get the ovrp ratio info reversely :
>>
>> 1) user_block_count/(user_block_count + ovrp blocks)
>>
>>   101376/(101376+25600) = 79%
>>
>> 2) user_block_count/(user_block_count + ovrp_blocks - rsvd_blocks)
>>
>>   101376/(101376+25600-14336) = 90% = 1-10%
>>
>> By 2), 1-90%, we can get the original config.overprovion when mkfs.f2fs
>> So I think we should use 2) to calculate ovrp ratio. Anything I miss
>> in 2)?
> 
> Yup, indeed 2) is correct, which I found my calcuation was wrong, and tried
> to verify with:
> 
> threshold = (main_segments - ovp_segments) / (main_segments - reserved_segments)
>             ------------------------------    ---------------------------------
>                 User-visible space                  F2FS-writable space
>               (user_block_count) / (user_block_count + ovp_blocks - rsvd_blocks)
> 
> For example,
> 
> 1. 8GB
>  ovp reported by mkfs = 2.22%
>  main_segments = 4065
>  ovp_segments = 186
>  reserved_segments = 98
>  threshold = 500 (mine) vs. 502 (yours)
> 
>  -> Actual ovp = (512-500) / 512 = 2.34%
>  -> reserved segments = 98 / 2 = 49 can cover 2.34%
> 
> 2. 64GB
>  ovp reported by mkfs = 0.78%
>  main_segments = 32581
>  ovp_segments = 516
>  reserved_segments = 264
>  threshold = 508 (mine) vs. 512 (yours)
> 
>  -> Actual ovp = (512-508) / 512 = 0.78%
>  -> reserved segments = 264 / 2 = 132 can cover 0.78%

The last updated patch wasn't been sent out, so I reply in this thread:

Reviewed-by: Chao Yu <yuchao0@huawei.com>

> 
> Thanks,
> 
> .
> 


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