On 2017/10/14 20:53, Yunlong Song wrote:
> Oh, yes it is. I found that problem in a kernel tree which does not have
> commit
> c6f82fe90d7458e5fa190a6820bfc24f96b0de4e (Revert "f2fs: put allocate_segment
> after refresh_sit_entry"). In that kernel, the allocate_segment is still
> behind
> refresh_sit_entry. Now I understand the commit message:
> "This makes a leak to register dirty segments. I reproduced the issue by
> modified postmark which injects a lot of file create/delete/update and
> finally triggers huge number of SSR allocations."
>
> The reason is that if refresh_sit_entry is before allocate_segment, then the
> dirty status of CURSEG is not updated, as a result, the count of dirty
> segments
> is wrong, which is much smaller than its real value. Then the f2fs_gc
> can not
> do its work since it can not even get one victim, then the free segments are
> used up and then triggers much SSR. So Jay reverts the patch.
>
> It seems there are two options:
> (1) keep this patch ([PATCH v2] f2fs: update dirty status for CURSEG as
> well)
> and we can recover commit 3436c4bdb30de421d46f58c9174669fbcfd40ce0
> (f2fs: put allocate_segment after refresh_sit_entry)
> (2) remove this patch at all
>
> It seems (1) is robust, but (2) avoids unnecessary check.
What about reverting 5e443818fa0b ("f2fs: handle dirty segments inside
refresh_sit_entry") to keep the original order:
1. update sit info
2. allocate new segment
3. update dirty status of segment
Thanks,
>
> On 2017/10/14 8:14, Chao Yu wrote:
>> On 2017/10/13 21:21, Yunlong Song wrote:
>>> Without this patch, it will cause all the free segments using up in some
>>> corner case. For example, there are 100 segments, and 20 of them are
>>> reserved for ovp. If 79 segments are full of data, segment 80 becomes
>>> CURSEG segment, write 512 blocks and then delete 511 blocks. Since it is
>>> CURSEG segment, the __locate_dirty_segment will not update its dirty
>>> status. Then the dirty_segments(sbi) is 0, f2fs_gc will fail to
>>> get_victim, and f2fs_balance_fs will fail to trigger gc action. After
>>> f2fs_balance_fs returns, f2fs can continue to write data to segment 81.
>>> Again, segment 81 becomes CURSEG segment, write 512 blocks and delete
>>> 511 blocks, the dirty_segments(sbi) is 0 and f2fs_gc fail again. This
>>> can finally use up all the free segments and cause panic.
>> Look into this patch again, I found refresh_sit_entry is called after
>> ->allocate_segment, so if all 512 blocks were allocated, log header should
>> have been moved to another segment, so locate_dirty_segment in
>> refresh_sit_entry should update dirty status of previous segment correctly,
>> anything I'm missing?
>>
>> Thanks,
>>
>>> Signed-off-by: Yunlong Song <yunlong.song@xxxxxxxxxx>
>>> ---
>>> fs/f2fs/segment.c | 4 ++--
>>> 1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
>>> index bfbcff8..0fce076 100644
>>> --- a/fs/f2fs/segment.c
>>> +++ b/fs/f2fs/segment.c
>>> @@ -687,7 +687,7 @@ static void __locate_dirty_segment(struct f2fs_sb_info *sbi, unsigned int segno,
>>> struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
>>>
>>> /* need not be added */
>>> - if (IS_CURSEG(sbi, segno))
>>> + if (IS_CURSEG(sbi, segno) && dirty_type == PRE)
>>> return;
>>>
>>> if (!test_and_set_bit(segno, dirty_i->dirty_segmap[dirty_type]))
>>> @@ -737,7 +737,7 @@ static void locate_dirty_segment(struct f2fs_sb_info *sbi, unsigned int segno)
>>> struct dirty_seglist_info *dirty_i = DIRTY_I(sbi);
>>> unsigned short valid_blocks;
>>>
>>> - if (segno == NULL_SEGNO || IS_CURSEG(sbi, segno))
>>> + if (segno == NULL_SEGNO)
>>> return;
>>>
>>> mutex_lock(&dirty_i->seglist_lock);
>>>
>> .
>>
>
