On 02/25, Yunlong Song wrote: > The benefit is much, let me give an example to make the point more clear, the reserved_sections for a 64G image is > about 500M, and if the current free_sections is 600M, and the IO pattern is like this: > > Before this patch: > time 1: node & dent * imeta 20M to write, 100M data to write > free_sections(sbi) <= node_secs + 2 * dent_secs + imeta_secs + reserved_sections(sbi); > then has_not_enough_free_secs returns false, since 600M > 20M + 500M > > time 2: node & dent * imeta 20M*2 to write, 100M*2 data to write > free_sections(sbi) <= node_secs + 2 * dent_secs + imeta_secs + reserved_sections(sbi); > then has_not_enough_free_secs returns false, since 600M > 20M*2 + 500M > > time 3: node & dent * imeta 20M*3 to write, 100M*3 data to write > free_sections(sbi) <= node_secs + 2 * dent_secs + imeta_secs + reserved_sections(sbi); > then has_not_enough_free_secs returns false, since 600M > 20M*3 + 500M > > time 4: node & dent * imeta 20M*4 to write, 100M*4 data to write > free_sections(sbi) <= node_secs + 2 * dent_secs + imeta_secs + reserved_sections(sbi); > then has_not_enough_free_secs returns false, since 600M > 20M*4 + 500M > > then here comes a sync, and wait for all the node & dent * imeta and data to flush to the flash device > what will happen after this sync? > the free_sections will decrease to 600M-20M*4(node & dent * imeta)-100M*4(data) = 120M > next time in f2fs_balance_fs: > free_sections(sbi) <= node_secs + 2 * dent_secs + imeta_secs + reserved_sections(sbi); > 120M <= 0 + 500M This is wrong, and should be 500M <= 0 + 500M, since gc will be done during sync. Besides, you're missing normal checkpoint which doesn't flush user data. Thanks, > then f2fs_gc will gc_more times and times again until free_sections increases from 120M to 500M...... > It will cost a lot of time! > > After this patch: > time 1: node & dent * imeta 20M to write, 100M data to write > free_sections(sbi) <= node_secs + 2 * dent_secs + imeta_secs + *data* + reserved_sections(sbi); > then has_not_enough_free_secs returns true, since 600M < 20M + *100M* + 500M > this time f2fs_gc will only gc_more for the gap 20M + *100M* + 500M - 600M = 20M > > time 2: node & dent * imeta 20M*2 to write, 100M*2 data to write > free_sections(sbi) <= node_secs + 2 * dent_secs + imeta_secs + *data* + reserved_sections(sbi); > then has_not_enough_free_secs returns true, since 600M + 20M (gc_more from time 1) < 20M*2 + *100M*2* + 500M > this time f2fs_gc will only gc_more for the gap 20M*2 + *100M* *2 + 500M - (600M + 20M) = 120M > > time 3: node & dent * imeta 20M*3 to write, 100M*3 data to write > free_sections(sbi) <= node_secs + 2 * dent_secs + imeta_secs + *data* + reserved_sections(sbi); > then has_not_enough_free_secs returns true, since 600M + 20M (gc_more from time 1) + 120M (gc_more from time 2) < 20M*3 + *100M*3* + 500M > this time f2fs_gc will only gc_more for the gap 20M*3 + *100M* *3 + 500M - (600M + 20M + 120M) = 120M > > time 4: node & dent * imeta 20M*4 to write, 100M*4 data to write > free_sections(sbi) <= node_secs + 2 * dent_secs + imeta_secs + *data* + reserved_sections(sbi); > then has_not_enough_free_secs returns true, since 600M + 20M (gc_more from time 1) + 120M (gc_more from time 2) + 120M (gc_more from time 3) < 20M*4 + *100M*4* + 500M > this time f2fs_gc will only gc_more for the gap 20M*4 + *100M* *4 + 500M - (600M + 20M + 120M + 120M) = 120M > > then here comes a sync, and wait for all the node & dent * imeta and data to flush to the flash device > what will happen after this sync? > the free_sections will decrease to 600M + 20M (gc_more from time 1) + 120M (gc_more from time 2) + 120M (gc_more from time 3) +120M (gc_more from time 4) - 20M*4(node & dent * imeta)-100M*4(data) = 500M > next time in f2fs_balance_fs: > free_sections(sbi) <= node_secs + 2 * dent_secs + imeta_secs + reserved_sections(sbi); > 500M <= 0 + 500M > this time f2fs_gc will only gc for 1 free segment, comparing with the current design without the patch, which has to gc_more from 120M to 500M.... > > You see, after the patch, gc_more are separated to 4 times but the old design make gc_more to one time, which will cost much performance. > > Besides, after the patch, we can make sure the reserved_sections are remained unused and free all the time, which can avoid the segment using up case! > > So now we can use the patch I sent last time, which changes the mkfs.f2fs to reduce the reserved_segments a lot and use SSR, I remember you pointed out an > issue that if there are not enough SSR segments then there is a problem, now I can solve your issue with this patch, since we can make sure the reserved_sections > are free all the time, then we are always able to make gc to have more free segments (or just take up the reserved_segments for temporary use when SSR segments > are not enough for performance). > > Finally, the performance can be improved and the reserved_segments can be reduced a lot. > > However, if this path is not allowed finally, I still suggest to reduce the reserved_segments to a smaller number, this is to avoid the amount of gc_more segments > after the sync in my example, because you have to gc_more until free_sections equals to reserved_segments, so the smaller the reserved_segment is, the smaller > times f2f2_gc will gc_more. > > Anyway, I will test the patch in device for stability. > > On 2017/2/25 4:07, Jaegeuk Kim wrote: > > On 02/24, Yunlong Song wrote: > >> Currently, it miss the part of F2FS_DIRTY_DATA to check whether there is enough > >> free segments for the "reserved_sections" originally set in the mkfs.f2fs. As a > >> result, it will use the reserved_sections part to write dirty data, and has to > >> do gc_more to free a lot of sections together next time. This will cost much > >> time to do so many fggc. So let's add the F2FS_DIRTY_DATA part and do a few gc > >> gradually each time, which will avoid to do a large number of gc at the same time. > >> > >> And this will also make sure the pre-set "reserved_sections" is not used all the > >> time and can be used anytime for gc when ssr segments are not enough. > > Not sure any side effect of this patch. Let's investigate more severely how to > > trigger SSR more eagerly in order to avoid write_checkpoint. > > > > Thanks, > > > >> Signed-off-by: Yunlong Song <yunlong.song@xxxxxxxxxx> > >> --- > >> fs/f2fs/segment.h | 6 ++++-- > >> 1 file changed, 4 insertions(+), 2 deletions(-) > >> > >> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h > >> index f4020f1..44f2a46 100644 > >> --- a/fs/f2fs/segment.h > >> +++ b/fs/f2fs/segment.h > >> @@ -490,12 +490,13 @@ static inline bool need_SSR(struct f2fs_sb_info *sbi) > >> int node_secs = get_blocktype_secs(sbi, F2FS_DIRTY_NODES); > >> int dent_secs = get_blocktype_secs(sbi, F2FS_DIRTY_DENTS); > >> int imeta_secs = get_blocktype_secs(sbi, F2FS_DIRTY_IMETA); > >> + int data_secs = get_blocktype_secs(sbi, F2FS_DIRTY_DATA); > >> > >> if (test_opt(sbi, LFS)) > >> return false; > >> > >> return free_sections(sbi) <= (node_secs + 2 * dent_secs + imeta_secs + > >> - reserved_sections(sbi) + 1); > >> + data_secs + reserved_sections(sbi) + 1); > >> } > >> > >> static inline bool has_not_enough_free_secs(struct f2fs_sb_info *sbi, > >> @@ -504,13 +505,14 @@ static inline bool has_not_enough_free_secs(struct f2fs_sb_info *sbi, > >> int node_secs = get_blocktype_secs(sbi, F2FS_DIRTY_NODES); > >> int dent_secs = get_blocktype_secs(sbi, F2FS_DIRTY_DENTS); > >> int imeta_secs = get_blocktype_secs(sbi, F2FS_DIRTY_IMETA); > >> + int data_secs = get_blocktype_secs(sbi, F2FS_DIRTY_DATA); > >> > >> if (unlikely(is_sbi_flag_set(sbi, SBI_POR_DOING))) > >> return false; > >> > >> return (free_sections(sbi) + freed) <= > >> (node_secs + 2 * dent_secs + imeta_secs + > >> - reserved_sections(sbi) + needed); > >> + data_secs + reserved_sections(sbi) + needed); > >> } > >> > >> static inline bool excess_prefree_segs(struct f2fs_sb_info *sbi) > >> -- > >> 1.8.5.2 > > . > > > > > -- > Thanks, > Yunlong Song >
