Hi Gang,
On 2017/11/28 16:32, Gang He wrote:
> Hi Alex,
>
>
>>>>
>> Hi Gang,
>>
>> On 2017/11/28 15:38, Gang He wrote:
>>> Hi Alex,
>>>
>>>
>>>>>>
>>>> Hi Gang,
>>>>
>>>> On 2017/11/28 13:33, Gang He wrote:
>>>>> Hello Alex,
>>>>>
>>>>>
>>>>>>>>
>>>>>> Hi Gang,
>>>>>>
>>>>>> On 2017/11/27 17:46, Gang He wrote:
>>>>>>> Add ocfs2_overwrite_io function, which is used to judge if
>>>>>>> overwrite allocated blocks, otherwise, the write will bring extra
>>>>>>> block allocation overhead.
>>>>>>>
>>>>>>> Signed-off-by: Gang He <[email protected]>
>>>>>>> ---
>>>>>>> fs/ocfs2/extent_map.c | 67
>>>>>> +++++++++++++++++++++++++++++++++++++++++++++++++++
>>>>>>> fs/ocfs2/extent_map.h | 3 +++
>>>>>>> 2 files changed, 70 insertions(+)
>>>>>>>
>>>>>>> diff --git a/fs/ocfs2/extent_map.c b/fs/ocfs2/extent_map.c
>>>>>>> index e4719e0..98bf325 100644
>>>>>>> --- a/fs/ocfs2/extent_map.c
>>>>>>> +++ b/fs/ocfs2/extent_map.c
>>>>>>> @@ -832,6 +832,73 @@ int ocfs2_fiemap(struct inode *inode, struct
>>>>>> fiemap_extent_info *fieinfo,
>>>>>>> return ret;
>>>>>>> }
>>>>>>>
>>>>>>> +/* Is IO overwriting allocated blocks? */
>>>>>>> +int ocfs2_overwrite_io(struct inode *inode, u64 map_start, u64 map_len,
>>>>>>> + int wait)
>>>>>>> +{
>>>>>>> + int ret = 0, is_last;
>>>>>>> + u32 mapping_end, cpos;
>>>>>>> + struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
>>>>>>> + struct buffer_head *di_bh = NULL;
>>>>>>> + struct ocfs2_extent_rec rec;
>>>>>>> +
>>>>>>> + if (wait)
>>>>>>> + ret = ocfs2_inode_lock(inode, &di_bh, 0);
>>>>>>> + else
>>>>>>> + ret = ocfs2_try_inode_lock(inode, &di_bh, 0);
>>>>>>> + if (ret)
>>>>>>> + goto out;
>>>>>>> +
>>>>>>> + if (wait)
>>>>>>> + down_read(&OCFS2_I(inode)->ip_alloc_sem);
>>>>>>> + else {
>>>>>>> + if (!down_read_trylock(&OCFS2_I(inode)->ip_alloc_sem)) {
>>>>>>> + ret = -EAGAIN;
>>>>>>> + goto out_unlock1;
>>>>>>> + }
>>>>>>> + }
>>>>>>> +
>>>>>>> + if ((OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) &&
>>>>>>> + ((map_start + map_len) <= i_size_read(inode)))
>>>>>>> + goto out_unlock2;
>>>>>>> +
>>>>>>> + cpos = map_start >> osb->s_clustersize_bits;
>>>>>>> + mapping_end = ocfs2_clusters_for_bytes(inode->i_sb,
>>>>>>> + map_start + map_len);
>>>>>>> + is_last = 0;
>>>>>>> + while (cpos < mapping_end && !is_last) {
>>>>>>> + ret = ocfs2_get_clusters_nocache(inode, di_bh, cpos,
>>>>>>> + NULL, &rec, &is_last);
>>>>>>> + if (ret) {
>>>>>>> + mlog_errno(ret);
>>>>>>> + goto out_unlock2;
>>>>>>> + }
>>>>>>> +
>>>>>>> + if (rec.e_blkno == 0ULL)
>>>>>>> + break;
>>>>>> I think here the blocks is not overwrite, because the hold is found and the
>>>>>> blocks
>>>>>> should be allocated.
>>>>> If the rec.e_blkno == NULL, this means there is a hole.
>>>>> The file hole means that these blocks are not allocated, it does not like
>>>> unwritten block.
>>>>> The unwritten blocks means that these blocks are allocated, but still have
>>>> not been unwritten.
>>>>>
>>>> If we break the loop when we find the hold, out of this function we will
>>>> allocate the blocks in
>>>> ocfs2_file_write_iter()->..->ocfs2_direct_IO->__blockdev_direct_IO->..->ocfs2_dio_wr_g
>>>> et_block()
>>>> ->ocfs2_write_begin_nolock. Does this violate the semantics of 'IOCB_NOWAIT';
>>> Yes, then we need to check if this is a overwrite before doing direct-io.
>>>
>>
>> I mean here we should return 0 instead of break and we should immediately
>> return -EAGAIN
>> to upper apps, otherwise, some block allocation will be happen, which
>> violates the
>> semantics of 'IOCB_NOWAIT'.
> Before we do a direct-io, I need to check if this is a overwrite allocated blocks IO.
> If not, we will return -EAGAIN in 'IOCB_NOWAIT' mode. this should not trigger any block allocation.
> I am not sure if we understand your concern totally.
>
Yes, your description is correct.
So we should return 0 instead of break when we find the hold in ocfs2_overwrite_io();
> Thanks
> Gang
>
>>
>> Thanks,
>> Alex
>>
>>>>
>>>> BTW, should we consider the down_write() and ocfs2_inode_lock() in
>>>> ocfs2_dio_wr_get_block() when
>>>> the flag 'IOCB_NOWAIT' is set;
>>> I think that we should not consider that layer lock, otherwise, the code
>> change will become more and more complex and big.
>>> I also refer to ext4 file system code change for this
>> feature(728fbc0e10b7f3ce2ee043b32e3453fd5201c055), they did not do any change
>> in that layer.
>>>
>>
>> OK.
>>
>>> Thanks
>>> Gang
>>>
>>>>
>>>>>>> +
>>>>>>> + if (rec.e_flags & OCFS2_EXT_REFCOUNTED)
>>>>>>> + break;
>>>>>>> +
>>>>>>> + cpos = le32_to_cpu(rec.e_cpos) +
>>>>>>> + le16_to_cpu(rec.e_leaf_clusters);
>>>>>>> + }
>>>>>>> +
>>>>>>> + if (cpos < mapping_end)
>>>>>>> + ret = 1;
>>>>>>> +
>>>>>>> +out_unlock2:
>>>>>>
>>>>>> I think the 'out_up_read' is more readable than the 'out_unlock2' .
>>>>> Ok, I will use more readable tag here.
>>>>>>
>>>>>>> + brelse(di_bh);
>>>>>>> +
>>>>>>> + up_read(&OCFS2_I(inode)->ip_alloc_sem);
>>>>>>> +
>>>>>>> +out_unlock1:
>>>>>>
>>>>>> We should release buffer head here.
>>>>>>
>>>>>>> + ocfs2_inode_unlock(inode, 0);
>>>>>>> +
>>>>>>> +out:
>>>>>>> + return (ret ? 0 : 1);
>>>>>>> +}
>>>>>>> +
>>>>>>> int ocfs2_seek_data_hole_offset(struct file *file, loff_t *offset, int
>>>>>> whence)
>>>>>>> {
>>>>>>> struct inode *inode = file->f_mapping->host;
>>>>>>> diff --git a/fs/ocfs2/extent_map.h b/fs/ocfs2/extent_map.h
>>>>>>> index 67ea57d..fd9e86a 100644
>>>>>>> --- a/fs/ocfs2/extent_map.h
>>>>>>> +++ b/fs/ocfs2/extent_map.h
>>>>>>> @@ -53,6 +53,9 @@ int ocfs2_extent_map_get_blocks(struct inode *inode, u64
>>>>>> v_blkno, u64 *p_blkno,
>>>>>>> int ocfs2_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
>>>>>>> u64 map_start, u64 map_len);
>>>>>>>
>>>>>>> +int ocfs2_overwrite_io(struct inode *inode, u64 map_start, u64 map_len,
>>>>>>> + int wait);
>>>>>>> +
>>>>>>> int ocfs2_seek_data_hole_offset(struct file *file, loff_t *offset, int
>>>>>> origin);
>>>>>>>
>>>>>>> int ocfs2_xattr_get_clusters(struct inode *inode, u32 v_cluster,
>>>>>>>
>>>>>
>>>>>
>>>>> .
>>>>>
>>>
>>> .
>>>
>
> .
>
From 1585300066650915501@xxx Tue Nov 28 09:05:57 +0000 2017
X-GM-THRID: 1585212205986077742
X-Gmail-Labels: Inbox,Category Forums,HistoricalUnread