lkml.org 
[lkml]   [2017]   [Jul]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] f2fs: let __get_victim successfully get a segno in corner case
Hi Yunlong,

On 07/20, Yunlong Song wrote:
> Hi, Jay,
> The distribution is like this, unit is segment counts:
> cnt_free: 0 (free blocks)
> cnt_full: 25182 (segment which has 512 blocks)
> cnt_over: 25192 (segment which valid blocks over fggc_threshold)
> cnt_below: 870 (segment which valid blocks below fggc_threshold)

What are the values of fggc_threshold and ovp ratio or reserved_segments?
How was the proc/segment_info?

Thanks,

>
> The test is fragment test, add and delete small files over and over.
>
> On 2017/7/18 0:56, Jaegeuk Kim wrote:
> > On 07/16, sylinux wrote:
> >> In fact,this is not "suppose" case yet, we have already met this problem several times in some test suits for corner case, or I cannot notice that this could happen.
> > So, have you taken a look at distribution of valid blocks in that case?
> > I'm wondering what distribution can cause this.
> >
> > Thanks,
> >
> >>
> >>
> >>
> >> 发自网易邮箱大师
> >> On 07/16/2017 09:09, Jaegeuk Kim wrote:
> >> Hi Yunlong,
> >>
> >> On 07/14, Yunlong Song wrote:
> >>> Suppose that the valid blocks of each section are all over sbi->fggc_threshold,
> >> How about adding a kernel message first to detect your *supposed* sceanrio?
> >> If that happens, it'll be a sort of bug which we haven't assumed.
> >>
> >> Thanks,
> >>
> >>> and even has_not_enough_free_secs is true, f2fs_gc cannot do its job since the
> >>> no_fggc_candidate always returns true. As a result, the reserved segments can be
> >>> used up, and finally there is no free segment at all, and get_new_segment cannot
> >>> get a free segment, filesystem will trap into a wrong status.
> >>>
> >>> To fix this, we record the segno which has a rough minimum cost and return it to
> >>> __get_victim to continue f2fs_gc's job.
> >>>
> >>> Signed-off-by: Yunlong Song <yunlong.song@huawei.com>
> >>> ---
> >>> fs/f2fs/gc.c | 19 ++++++++++++++-----
> >>> fs/f2fs/segment.h | 17 ++++++++++++++---
> >>> 2 files changed, 28 insertions(+), 8 deletions(-)
> >>>
> >>> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> >>> index fa3d2e2..965e783 100644
> >>> --- a/fs/f2fs/gc.c
> >>> +++ b/fs/f2fs/gc.c
> >>> @@ -178,6 +178,8 @@ static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
> >>> p->offset = 0;
> >>> else
> >>> p->offset = SIT_I(sbi)->last_victim[p->gc_mode];
> >>> +
> >>> + p->min_cost_r = UINT_MAX;
> >>> }
> >>>
> >>> static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
> >>> @@ -194,7 +196,7 @@ static unsigned int get_max_cost(struct f2fs_sb_info *sbi,
> >>> return 0;
> >>> }
> >>>
> >>> -static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
> >>> +static unsigned int check_bg_victims(struct f2fs_sb_info *sbi, struct victim_sel_policy *p)
> >>> {
> >>> struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
> >>> unsigned int secno;
> >>> @@ -208,11 +210,12 @@ static unsigned int check_bg_victims(struct f2fs_sb_info *sbi)
> >>> if (sec_usage_check(sbi, secno))
> >>> continue;
> >>>
> >>> - if (no_fggc_candidate(sbi, secno))
> >>> + p->cur_segno_r = GET_SEG_FROM_SEC(sbi, secno);
> >>> + if (no_fggc_candidate(sbi, secno, p))
> >>> continue;
> >>>
> >>> clear_bit(secno, dirty_i->victim_secmap);
> >>> - return GET_SEG_FROM_SEC(sbi, secno);
> >>> + return p->cur_segno_r;
> >>> }
> >>> return NULL_SEGNO;
> >>> }
> >>> @@ -332,7 +335,7 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> >>>
> >>> last_victim = sm->last_victim[p.gc_mode];
> >>> if (p.alloc_mode == LFS && gc_type == FG_GC) {
> >>> - p.min_segno = check_bg_victims(sbi);
> >>> + p.min_segno = check_bg_victims(sbi, &p);
> >>> if (p.min_segno != NULL_SEGNO)
> >>> goto got_it;
> >>> }
> >>> @@ -369,8 +372,9 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> >>> goto next;
> >>> if (gc_type == BG_GC && test_bit(secno, dirty_i->victim_secmap))
> >>> goto next;
> >>> + p.cur_segno_r = segno;
> >>> if (gc_type == FG_GC && p.alloc_mode == LFS &&
> >>> - no_fggc_candidate(sbi, secno))
> >>> + no_fggc_candidate(sbi, secno, &p))
> >>> goto next;
> >>>
> >>> cost = get_gc_cost(sbi, segno, &p);
> >>> @@ -403,6 +407,11 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> >>> trace_f2fs_get_victim(sbi->sb, type, gc_type, &p,
> >>> sbi->cur_victim_sec,
> >>> prefree_segments(sbi), free_segments(sbi));
> >>> + } else if (has_not_enough_free_secs(sbi, 0, 0)) {
> >>> + p.min_segno = p.min_segno_r;
> >>> + if (p.alloc_mode == LFS && gc_type == FG_GC)
> >>> + clear_bit(GET_SEC_FROM_SEG(sbi, p.min_segno), dirty_i->victim_secmap);
> >>> + goto got_it;
> >>> }
> >>> out:
> >>> mutex_unlock(&dirty_i->seglist_lock);
> >>> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
> >>> index 6b871b4..7d2d0f3 100644
> >>> --- a/fs/f2fs/segment.h
> >>> +++ b/fs/f2fs/segment.h
> >>> @@ -169,6 +169,9 @@ struct victim_sel_policy {
> >>> unsigned int ofs_unit; /* bitmap search unit */
> >>> unsigned int min_cost; /* minimum cost */
> >>> unsigned int min_segno; /* segment # having min. cost */
> >>> + unsigned int min_cost_r; /* rough minimum cost */
> >>> + unsigned int min_segno_r; /* segment # having rough min. cost */
> >>> + unsigned int cur_segno_r; /* segment # rough process is handling */
> >>> };
> >>>
> >>> struct seg_entry {
> >>> @@ -743,11 +746,19 @@ static inline block_t sum_blk_addr(struct f2fs_sb_info *sbi, int base, int type)
> >>> }
> >>>
> >>> static inline bool no_fggc_candidate(struct f2fs_sb_info *sbi,
> >>> - unsigned int secno)
> >>> + unsigned int secno, struct victim_sel_policy *p)
> >>> {
> >>> - if (get_valid_blocks(sbi, GET_SEG_FROM_SEC(sbi, secno), true) >=
> >>> - sbi->fggc_threshold)
> >>> + unsigned int cur_cost;
> >>> +
> >>> + cur_cost = get_valid_blocks(sbi, GET_SEG_FROM_SEC(sbi, secno), true);
> >>> + if (cur_cost >= sbi->fggc_threshold) {
> >>> + if (p->min_cost_r > cur_cost) {
> >>> + p->min_cost_r = cur_cost;
> >>> + p->min_segno_r = p->cur_segno_r;
> >>> + }
> >>> return true;
> >>> + }
> >>> +
> >>> return false;
> >>> }
> >>>
> >>> --
> >>> 1.8.5.2
> > .
> >
>
>
> --
> Thanks,
> Yunlong Song
>

\
 
 \ /
  Last update: 2017-07-21 23:08    [W:0.076 / U:0.260 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site