This reverts commit b9cd20619e359d199b755543474c3d853c8e3415.
That patch causes much fewer node segments (which can be used for SSR)
than before, and in the corner case (e.g. create and delete *.txt files in
one same directory, there will be very few node segments but many data
segments), if the reserved free segments are all used up during gc, then
the write_checkpoint can still flush dentry pages to data ssr segments,
but will probably fail to flush node pages to node ssr segments, since
there are not enough node ssr segments left (the left ones are all
full).
So revert this patch to give a fair chance to let node segments remain
for SSR, which provides more robustness for corner cases.
Conflicts:
fs/f2fs/gc.c
---
fs/f2fs/gc.c | 12 +-----------
1 file changed, 1 insertion(+), 11 deletions(-)
diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
index bfe6a8c..f777e07 100644
--- a/fs/f2fs/gc.c
+++ b/fs/f2fs/gc.c
@@ -267,16 +267,6 @@ static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
}
-static unsigned int get_greedy_cost(struct f2fs_sb_info *sbi,
- unsigned int segno)
-{
- unsigned int valid_blocks =
- get_valid_blocks(sbi, segno, true);
-
- return IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
- valid_blocks * 2 : valid_blocks;
-}
-
static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
unsigned int segno, struct victim_sel_policy *p)
{
@@ -285,7 +275,7 @@ static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
/* alloc_mode == LFS */
if (p->gc_mode == GC_GREEDY)
- return get_greedy_cost(sbi, segno);
+ return get_valid_blocks(sbi, segno, true);
else
return get_cb_cost(sbi, segno);
}
--
1.8.5.2
On 2017/9/23 17:02, Yunlong Song wrote:
> This reverts commit b9cd20619e359d199b755543474c3d853c8e3415.
>
> That patch causes much fewer node segments (which can be used for SSR)
> than before, and in the corner case (e.g. create and delete *.txt files in
> one same directory, there will be very few node segments but many data
> segments), if the reserved free segments are all used up during gc, then
> the write_checkpoint can still flush dentry pages to data ssr segments,
> but will probably fail to flush node pages to node ssr segments, since
> there are not enough node ssr segments left (the left ones are all
> full).
IMO, greedy algorithm wants to minimize price of moving one dirty segment, our
behavior is accord with semantics of our algorithm to select victim with least
valid blocks. Pengyang's patch tries to adjust greedy algorithm to consider
minimizing total number of valid blocks in all selected victim segments during
whole FGGC cycle, but its algorithm is corrupted, since if all valid data blocks
in current victim segment is not belong to different dnode block, our selection
may be incorrect.
Anyway, I agree to revert Pengyang's patch first before we got a entire scheme.
BTW, for SSR or LFS selection, there is a trade-off in between: a) SSR-write
costs less free segment and move less data/node blocks, but it triggers random
write which results in bad performance. b) LFS-write costs more free segment and
move more data/node blocks, but it triggers sequential write which results in
good performance. So I don't think more SSR we trigger, lower latency our FGGC
faces.
Thanks,
>
> So revert this patch to give a fair chance to let node segments remain
> for SSR, which provides more robustness for corner cases.
>
> Conflicts:
> fs/f2fs/gc.c
> ---
> fs/f2fs/gc.c | 12 +-----------
> 1 file changed, 1 insertion(+), 11 deletions(-)
>
> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> index bfe6a8c..f777e07 100644
> --- a/fs/f2fs/gc.c
> +++ b/fs/f2fs/gc.c
> @@ -267,16 +267,6 @@ static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
> return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
> }
>
> -static unsigned int get_greedy_cost(struct f2fs_sb_info *sbi,
> - unsigned int segno)
> -{
> - unsigned int valid_blocks =
> - get_valid_blocks(sbi, segno, true);
> -
> - return IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
> - valid_blocks * 2 : valid_blocks;
> -}
> -
> static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
> unsigned int segno, struct victim_sel_policy *p)
> {
> @@ -285,7 +275,7 @@ static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
>
> /* alloc_mode == LFS */
> if (p->gc_mode == GC_GREEDY)
> - return get_greedy_cost(sbi, segno);
> + return get_valid_blocks(sbi, segno, true);
> else
> return get_cb_cost(sbi, segno);
> }
>
On 2017/9/25 10:52, Chao Yu wrote:
> On 2017/9/23 17:02, Yunlong Song wrote:
>> This reverts commit b9cd20619e359d199b755543474c3d853c8e3415.
>>
>> That patch causes much fewer node segments (which can be used for SSR)
>> than before, and in the corner case (e.g. create and delete *.txt files in
>> one same directory, there will be very few node segments but many data
>> segments), if the reserved free segments are all used up during gc, then
>> the write_checkpoint can still flush dentry pages to data ssr segments,
>> but will probably fail to flush node pages to node ssr segments, since
>> there are not enough node ssr segments left (the left ones are all
>> full).
>
> IMO, greedy algorithm wants to minimize price of moving one dirty segment, our
> behavior is accord with semantics of our algorithm to select victim with least
> valid blocks. Pengyang's patch tries to adjust greedy algorithm to consider
> minimizing total number of valid blocks in all selected victim segments during
> whole FGGC cycle, but its algorithm is corrupted, since if all valid data blocks
> in current victim segment is not belong to different dnode block, our selection
> may be incorrect.
>
> Anyway, I agree to revert Pengyang's patch first before we got a entire scheme.
Please replace old commit message with above reason. :)
Thanks,
>
> BTW, for SSR or LFS selection, there is a trade-off in between: a) SSR-write
> costs less free segment and move less data/node blocks, but it triggers random
> write which results in bad performance. b) LFS-write costs more free segment and
> move more data/node blocks, but it triggers sequential write which results in
> good performance. So I don't think more SSR we trigger, lower latency our FGGC
> faces.
>
> Thanks,
>
>>
>> So revert this patch to give a fair chance to let node segments remain
>> for SSR, which provides more robustness for corner cases.
>>
>> Conflicts:
>> fs/f2fs/gc.c
>> ---
>> fs/f2fs/gc.c | 12 +-----------
>> 1 file changed, 1 insertion(+), 11 deletions(-)
>>
>> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
>> index bfe6a8c..f777e07 100644
>> --- a/fs/f2fs/gc.c
>> +++ b/fs/f2fs/gc.c
>> @@ -267,16 +267,6 @@ static unsigned int get_cb_cost(struct f2fs_sb_info *sbi, unsigned int segno)
>> return UINT_MAX - ((100 * (100 - u) * age) / (100 + u));
>> }
>>
>> -static unsigned int get_greedy_cost(struct f2fs_sb_info *sbi,
>> - unsigned int segno)
>> -{
>> - unsigned int valid_blocks =
>> - get_valid_blocks(sbi, segno, true);
>> -
>> - return IS_DATASEG(get_seg_entry(sbi, segno)->type) ?
>> - valid_blocks * 2 : valid_blocks;
>> -}
>> -
>> static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
>> unsigned int segno, struct victim_sel_policy *p)
>> {
>> @@ -285,7 +275,7 @@ static inline unsigned int get_gc_cost(struct f2fs_sb_info *sbi,
>>
>> /* alloc_mode == LFS */
>> if (p->gc_mode == GC_GREEDY)
>> - return get_greedy_cost(sbi, segno);
>> + return get_valid_blocks(sbi, segno, true);
>> else
>> return get_cb_cost(sbi, segno);
>> }
>>