Re: [f2fs-dev] [External Mail] Re: [PATCH] f2fs: prevent the current section from being selected as a victim during garbage collection

[Chao Yu via Linux-f2fs-devel <linux-f2fs-devel@lists.sourceforge.net>]

On 3/27/25 14:43, yohan.joung@sk.com wrote:
>> From: Chao Yu <chao@kernel.org>
>> Sent: Thursday, March 27, 2025 3:02 PM
>> To: Yohan Joung <jyh429@gmail.com>; jaegeuk@kernel.org; daeho43@gmail.com
>> Cc: chao@kernel.org; linux-f2fs-devel@lists.sourceforge.net; linux-
>> kernel@vger.kernel.org; 정요한(JOUNG YOHAN) Mobile AE <yohan.joung@sk.com>
>> Subject: [External Mail] Re: [PATCH] f2fs: prevent the current section
>> from being selected as a victim during garbage collection
>>
>> On 3/26/25 22:14, Yohan Joung wrote:
>>> When selecting a victim using next_victim_seg in a large section, the
>>> selected section might already have been cleared and designated as the
>>> new current section, making it actively in use.
>>> This behavior causes inconsistency between the SIT and SSA.
>>
>> Hi, does this fix your issue?
> 
> This is an issue that arises when dividing 
> a large section into segments for garbage collection.
> caused by the background GC (garbage collection) thread in large section
> f2fs_gc(victim_section) -> f2fs_clear_prefree_segments(victim_section)-> 
> cursec(victim_section) -> f2fs_gc(victim_section by next_victim_seg)

I didn't get it, why f2fs_get_victim() will return section which is used
by curseg? It should be avoided by checking w/ sec_usage_check().

Or we missed to check gcing section which next_victim_seg points to
during get_new_segment()?

Can this happen?

e.g.
- bggc selects sec #0
- next_victim_seg: seg #0
- migrate seg #0 and stop
- next_victim_seg: seg #1
- checkpoint, set sec #0 free if sec #0 has no valid blocks
- allocate seg #0 in sec #0 for curseg
- curseg moves to seg #1 after allocation
- bggc tries to migrate seg #1

Thanks,

> 
> Because the call stack is different, 
> I think that in order to handle everything at once, 
> we need to address it within do_garbage_collect, 
> or otherwise include it on both sides. What do you think?
> 
> [30146.337471][ T1300] F2FS-fs (dm-54): Inconsistent segment (70961) type [0, 1] in SSA and SIT
> [30146.346151][ T1300] Call trace:
> [30146.346152][ T1300]  dump_backtrace+0xe8/0x10c
> [30146.346157][ T1300]  show_stack+0x18/0x28
> [30146.346158][ T1300]  dump_stack_lvl+0x50/0x6c
> [30146.346161][ T1300]  dump_stack+0x18/0x28
> [30146.346162][ T1300]  f2fs_stop_checkpoint+0x1c/0x3c
> [30146.346165][ T1300]  do_garbage_collect+0x41c/0x271c
> [30146.346167][ T1300]  f2fs_gc+0x27c/0x828
> [30146.346168][ T1300]  gc_thread_func+0x290/0x88c
> [30146.346169][ T1300]  kthread+0x11c/0x164
> [30146.346172][ T1300]  ret_from_fork+0x10/0x20
> 
> struct curseg_info : 0xffffff803f95e800 {
> 	segno        : 0x11531 : 70961
> }
> 
> struct f2fs_sb_info : 0xffffff8811d12000 {
> 	next_victim_seg[0] : 0x11531 : 70961
> }
> 
>>
>> https://lore.kernel.org/linux-f2fs-devel/20250325080646.3291947-2-
>> chao@kernel.org
>>
>> Thanks,
>>
>>>
>>> Signed-off-by: Yohan Joung <yohan.joung@sk.com>
>>> ---
>>>  fs/f2fs/gc.c | 4 ++++
>>>  1 file changed, 4 insertions(+)
>>>
>>> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c index
>>> 2b8f9239bede..4b5d18e395eb 100644
>>> --- a/fs/f2fs/gc.c
>>> +++ b/fs/f2fs/gc.c
>>> @@ -1926,6 +1926,10 @@ int f2fs_gc(struct f2fs_sb_info *sbi, struct
>> f2fs_gc_control *gc_control)
>>>  		goto stop;
>>>  	}
>>>
>>> +	if (__is_large_section(sbi) &&
>>> +			IS_CURSEC(sbi, GET_SEC_FROM_SEG(sbi, segno)))
>>> +		goto stop;
>>> +
>>>  	seg_freed = do_garbage_collect(sbi, segno, &gc_list, gc_type,
>>>  				gc_control->should_migrate_blocks,
>>>  				gc_control->one_time);
> 



_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel