On 2018/2/1 6:28, Jaegeuk Kim wrote:
> On 01/31, Chao Yu wrote:
>> On 2018/1/31 14:39, Gaoxiang (OS) wrote:
>>> Previously, we attempt to flush the whole cp pack in a single bio,
>>> however, when suddenly powering off at this time, we could get into
>>> an extreme scenario that cp pack 1 page and cp pack 2 page are updated
>>> and latest, but payload or current summaries are still partially
>>> outdated. (see reliable write in the UFS specification)
>>>
>>> This patch submits the whole cp pack except cp pack 2 page at first,
>>> and then writes the cp pack 2 page with an extra independent
>>> bio with pre-io barrier.
>>>
>>> Signed-off-by: Gao Xiang <gaoxiang25@xxxxxxxxxx>
>>> Reviewed-by: Chao Yu <yuchao0@xxxxxxxxxx>
>>> ---
>>> Change log from v3:
>>> - further review comments are applied from Jaegeuk and Chao
>>> - Tested on this patch (without multiple-device): mount, boot Android with f2fs userdata and make fragment
>>> - If any problem with this patch or I miss something, please kindly share your comments, thanks :)
>>> Change log from v2:
>>> - Apply the review comments from Chao
>>> Change log from v1:
>>> - Apply the review comments from Chao
>>> - time data from "finish block_ops" to " finish checkpoint" (tested on ARM64 with TOSHIBA 128GB UFS):
>>> Before patch: 0.002273 0.001973 0.002789 0.005159 0.002050
>>> After patch: 0.002502 0.001624 0.002487 0.003049 0.002696
>>> fs/f2fs/checkpoint.c | 67 ++++++++++++++++++++++++++++++++++++----------------
>>> 1 file changed, 46 insertions(+), 21 deletions(-)
>>>
>>> diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c
>>> index 14d2fed..916dc72 100644
>>> --- a/fs/f2fs/checkpoint.c
>>> +++ b/fs/f2fs/checkpoint.c
>>> @@ -1158,6 +1158,39 @@ static void update_ckpt_flags(struct f2fs_sb_info *sbi, struct cp_control *cpc)
>>> spin_unlock_irqrestore(&sbi->cp_lock, flags);
>>> }
>>>
>>> +static void commit_checkpoint(struct f2fs_sb_info *sbi,
>>> + void *src, block_t blk_addr)
>>> +{
>>> + struct writeback_control wbc = {
>>> + .for_reclaim = 0,
>>> + };
>>> +
>>> + /*
>>> + * pagevec_lookup_tag and lock_page again will take
>>> + * some extra time. Therefore, update_meta_pages and
>>> + * sync_meta_pages are combined in this function.
>>> + */
>>> + struct page *page = grab_meta_page(sbi, blk_addr);
>>> + int err;
>>> +
>>> + memcpy(page_address(page), src, PAGE_SIZE);
>>> + set_page_dirty(page);
>>> +
>>> + f2fs_wait_on_page_writeback(page, META, true);
>>> + f2fs_bug_on(sbi, PageWriteback(page));
>>> + if (unlikely(!clear_page_dirty_for_io(page)))
>>> + f2fs_bug_on(sbi, 1);
>>> +
>>> + /* writeout cp pack 2 page */
>>> + err = __f2fs_write_meta_page(page, &wbc, FS_CP_META_IO);
>>> + f2fs_bug_on(sbi, err);
>>> +
>>> + f2fs_put_page(page, 0);
>>> +
>>> + /* submit checkpoint (with barrier if NOBARRIER is not set) */
>>> + f2fs_submit_merged_write(sbi, META_FLUSH);
>>> +}
>>> +
>>> static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
>>> {
>>> struct f2fs_checkpoint *ckpt = F2FS_CKPT(sbi);
>>> @@ -1260,16 +1293,6 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
>>> }
>>> }
>>>
>>> - /* need to wait for end_io results */
>>> - wait_on_all_pages_writeback(sbi);
>>> - if (unlikely(f2fs_cp_error(sbi)))
>>> - return -EIO;
>>> -
>>> - /* flush all device cache */
>>> - err = f2fs_flush_device_cache(sbi);
>>> - if (err)
>>> - return err;
>>> -
>>> /* write out checkpoint buffer at block 0 */
>>> update_meta_page(sbi, ckpt, start_blk++);
>>>
>>> @@ -1297,15 +1320,6 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
>>> start_blk += NR_CURSEG_NODE_TYPE;
>>> }
>>>
>>> - /* writeout checkpoint block */
>>> - update_meta_page(sbi, ckpt, start_blk);
>>> -
>>> - /* wait for previous submitted node/meta pages writeback */
>>> - wait_on_all_pages_writeback(sbi);
>>> -
>>> - if (unlikely(f2fs_cp_error(sbi)))
>>> - return -EIO;
>>> -
>>> filemap_fdatawait_range(NODE_MAPPING(sbi), 0, LLONG_MAX);
>>> filemap_fdatawait_range(META_MAPPING(sbi), 0, LLONG_MAX);
>
> - remove
Agreed.
>
>>>
>>> @@ -1313,12 +1327,23 @@ static int do_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
>>> sbi->last_valid_block_count = sbi->total_valid_block_count;
>>> percpu_counter_set(&sbi->alloc_valid_block_count, 0);
>>>
>>> - /* Here, we only have one bio having CP pack */
>>> - sync_meta_pages(sbi, META_FLUSH, LONG_MAX, FS_CP_META_IO);
>>> + /* Here, we have one bio having CP pack except cp pack 2 page */
>>> + sync_meta_pages(sbi, META, LONG_MAX, FS_CP_META_IO);
>>> +
>>> + /* flush all device cache */
>>> + err = f2fs_flush_device_cache(sbi);
>>> + if (err)
>>> + return err;
>>>
>>> /* wait for previous submitted meta pages writeback */
>>> wait_on_all_pages_writeback(sbi);
>>
>> Move f2fs_flush_device_cache here? since meta area can cross the multiple
>> devices, we should make sure all metadata were in device cache at least, and
>> then trigger the flush.
>
> Agreed, and need to flush, only if we have multiple devices.
f2fs_flush_device_cache is designed as only flushing devices except the first
one when f2fs enables multiple device. Calling it directly will be OK. :)
>
>>
>>>
>>> + if (unlikely(f2fs_cp_error(sbi)))
>>> + return -EIO;
>>> +
>>> + /* barrier and flush checkpoint cp pack 2 page if it can */
>>> + commit_checkpoint(sbi, ckpt, start_blk);
>>
>> Jaegeuk, are we really allow to make critical do_checkpoint which is on path of
>> fsync()/sync() be asynchronous?
>
> Yeah, so we need to wait end_io on synchronous paths like f2fs_sync_fs(1).
I think we should consider the case Xiang mentioned:
1. write async checkpoint #1 from gc;
2. write sync checkpoint #2 from sync_fs, end_io finished w/ error; -> cp #2 becomes corrupted
3. checkpoint #1's end_io finished w/ error; -> cp #1 becomes corrupted
Then, entire filesystem becomes corrupted now.
Thanks,
>
>>
>> Thanks,
>>
>>> +
>>> release_ino_entry(sbi, false);
>>>
>>> if (unlikely(f2fs_cp_error(sbi)))
>>>
