Re: [f2fs-dev] [PATCH 3/3] f2fs: skip migration only when BG_GC is called

[Chao Yu]

On 2020/2/19 11:04, Jaegeuk Kim wrote:
> On 02/19, Chao Yu wrote:
>> On 2020/2/19 7:27, Jaegeuk Kim wrote:
>>> On 02/17, Chao Yu wrote:
>>>> On 2020/2/15 2:58, Jaegeuk Kim wrote:
>>>>> FG_GC needs to move entire section more quickly.
>>>>>
>>>>> Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
>>>>> ---
>>>>>  fs/f2fs/gc.c | 2 +-
>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
>>>>> index bbf4db3f6bb4..1676eebc8c8b 100644
>>>>> --- a/fs/f2fs/gc.c
>>>>> +++ b/fs/f2fs/gc.c
>>>>> @@ -1203,7 +1203,7 @@ static int do_garbage_collect(struct f2fs_sb_info *sbi,
>>>>>  
>>>>>  		if (get_valid_blocks(sbi, segno, false) == 0)
>>>>>  			goto freed;
>>>>> -		if (__is_large_section(sbi) &&
>>>>> +		if (gc_type == BG_GC && __is_large_section(sbi) &&
>>>>>  				migrated >= sbi->migration_granularity)
>>>>
>>>> I knew migrating one large section is a more efficient way, but this can
>>>> increase long-tail latency of f2fs_balance_fs() occasionally, especially in
>>>> extreme fragmented space.
>>>
>>> FG_GC requires to wait for whole section migration which shows the entire
>>> latency.
>>
>> That will cause long-tail latency for single f2fs_balance_fs() procedure,
>> which it looks a very long hang when userspace call f2fs syscall, so why
>> not splitting total elapsed time into several f2fs_balance_fs() to avoid that.
> 
> Then, other ops can easily make more dirty segments. The intention of FG_GC is

Yup, that's a problem, if there are more dirty datas being made, reserved segments
may be ran out during FG_GC.

> to block everything and make min. free segments as a best shot.

I just try to simulate write GC's logic in FTL to mitigate single op's max latency,
otherwise benchmark looks hang during FG_GC (in a 500mb+ section).

Thanks,

> 
>>
>> Thanks,
>>
>>>
>>>>
>>>> Thanks,
>>>>
>>>>>  			goto skip;
>>>>>  		if (!PageUptodate(sum_page) || unlikely(f2fs_cp_error(sbi)))
>>>>>
>>> .
>>>
> .
>