On 2016/6/3 13:17, Jaegeuk Kim wrote: > On Fri, Jun 03, 2016 at 01:13:21PM +0800, Chao Yu wrote: >> On 2016/6/3 13:08, Jaegeuk Kim wrote: >>> On Tue, May 31, 2016 at 02:10:50PM +0800, Chao Yu wrote: >>>> Hi Jaegeuk, >>>> >>>> On 2016/5/30 10:37, Jaegeuk Kim wrote: >>>>> Hi Chao, >>>>> >>>>> On Sat, May 21, 2016 at 01:19:11PM +0800, Chao Yu wrote: >>>>>> From: Chao Yu <yuch...@huawei.com> >>>>>> >>>>>> If we fail to move data page during foreground GC, we should give another >>>>>> chance to writeback that page which was set dirty previously by writer. >>>>>> >>>>>> Signed-off-by: Chao Yu <yuch...@huawei.com> >>>>>> --- >>>>>> fs/f2fs/gc.c | 5 ++++- >>>>>> 1 file changed, 4 insertions(+), 1 deletion(-) >>>>>> >>>>>> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c >>>>>> index 38d56f6..ee213a8 100644 >>>>>> --- a/fs/f2fs/gc.c >>>>>> +++ b/fs/f2fs/gc.c >>>>>> @@ -653,12 +653,15 @@ static void move_data_page(struct inode *inode, >>>>>> block_t bidx, int gc_type) >>>>>> .page = page, >>>>>> .encrypted_page = NULL, >>>>>> }; >>>>>> + bool is_dirty = PageDirty(page); >>>>>> + >>>>>> set_page_dirty(page); >>>>>> f2fs_wait_on_page_writeback(page, DATA, true); >>>>>> if (clear_page_dirty_for_io(page)) >>>>>> inode_dec_dirty_pages(inode); >>>>>> set_cold_data(page); >>>>>> - do_write_data_page(&fio); >>>>>> + if (do_write_data_page(&fio) && is_dirty) >>>>>> + set_page_dirty(page); >>>>> >>>>> If this page is truncated with -ENOENT, we don't need to set it dirty >>>>> again. >>>> >>>> Agree >>>> >>>>> I expect that, if we get an error here, do_garbage_collect() would retry >>>>> FG_GC >>>> >>>> IIRC, you have reworked the FG_GC flows changed from an infinite loop to >>>> trying >>>> do the movement just one time. Here, I think if there are just few of >>>> blocks are >>>> failed to be moved, we can give one more time for retrying. How do you >>>> think? >>> >>> Mostly I expected here -ENOENT caused by race condition. >> >> If we hit ENOENT case, we can pass get_valid_blocks check, so we don't need >> to >> worry about this case, right? >> >>> Do we have another expectation? >> >> ENOMEM or EIO? > > EIO will stop everything. > ENOMEM would be better to wait for a while from page reclaim?
Agree, but for ioctl path, IMO, we don't need to let user waiting for ENOMEM case looping. > >> >> Thanks, >> >>> >>> Thanks, >>> >>>> >>>>> again. >>>>> >>>>> Thanks, >>>>> >>>>>> clear_cold_data(page); >>>>>> } >>>>>> out: >>>>>> -- >>>>>> 2.7.2 >>>>> . >>>>> >>> . >>> > . >
