Ocfs2 devel - Aug 2015 - [PATCH 0/9 v6] ocfs2: support append O_DIRECT write

On 08/05/2015 02:40 PM, Joseph Qi wrote:
> On 2015/8/5 12:40, Ryan Ding wrote:
>> Hi Joseph,
>>
>>
>> On 08/04/2015 05:03 PM, Joseph Qi wrote:
>>> Hi Ryan,
>>>
>>> On 2015/8/4 14:16, Ryan Ding wrote:
>>>> Hi Joseph,
>>>>
>>>> Sorry for bothering you with the old patches. But I really need
to know what this patch is for.
>>>>
>>>>
https://oss.oracle.com/pipermail/ocfs2-devel/2015-January/010496.html
>>>>
>>>>   From above email archive, you mentioned those patches aim to
reduce the host page cache consumption. But in my opinion, after append direct
io, the page used for buffer is clean. System can realloc those cached pages. We
can even call invalidate_mapping_pages to fast that process. Maybe more pages
will be needed during direct io. But direct io size can not be too large, right?
>>>>
>>> We introduced the append direct io because originally ocfs2 would
fall
>>> back to buffer io in case of thin provision, which was not the
actual
>>> behavior that user expect.
>> direct io has 2 semantics:
>> 1. io is performed synchronously, data is guaranteed to be transferred
after write syscall return.
>> 2. File I/O is done directly to/from user space buffers. No page buffer
involved.
>> But I think #2 is invisible to user space, #1 is the only thing that
user space is really interested in.
>> We should balance the benefit and disadvantage to determine whether #2
should be supported.
>> The disadvantage is: bring too much complexity to the code, bugs will
come along. And involved a incompatible feature.
>> For example, I did a single node sparse file test, and it failed.
> What do you mean by "failed"? Could you please send out the test
case
> and the actual output?
> And which version did you test? Because some bug fixes were submitted
later.
> Currently doing direct io with hole is not support.
I use linux 4.0 latest commit 39a8804455fb23f09157341d3ba7db6d7ae6ee76
A simplified test case is:
dd if=/dev/zero of=/mnt/hello bs=512 count=1 oflag=direct && truncate 
/mnt/hello -s 2097152
file 'hello' is not exist before test. After this command, file
'hello'
should be all zero. But 512~4096 is some random data.
>
>> The original way of ocfs2 handling direct io(turn to buffer io when
it's append write or write to a file hole) has 2 consideration:
>> 1. easier to support cluster wide coherence.
>> 2. easier to support sparse file.
>> But it seems that your patch handle #2 not very well.
>> There may be more issues that I have not found.
>>> I didn't get you that more pages would be needed during direct
io. Could
>>> you please explain it more clearly?
>> I mean the original way of handle append-dio will consume some page
cache. The page cache size it consume depend on the direct io size. For example,
1MB direct io will consume 1MB page cache.But since direct io size can not be
too large, the page cache it consume can not be too large also. And those pages
can be freed after direct io finished by calling invalidate_mapping_pages().
> I've got your point. Please consider the following user scenario.
> 1. A node mounted several ocfs2 volumes, for example, 10.
> 2. For each ocfs2 volume, there are several thin provision VMs.
Is there many direct io in parallelthat had been tested out?
About o2net_wq will block reclaim cache issue you mentioned in another 
mail. invalidate_mapping_pages() only free the page cache pages that 
stored data. It will not affect meta data cache. So that will not wait 
unlock. Is that right?
>
>>> Thanks,
>>> Joseph
>>>
>>>> Thanks,
>>>> Ryan
>>>>
>>>>
>>
>> .
>>
>

On 2015/8/5 16:07, Ryan Ding wrote:
> 
> On 08/05/2015 02:40 PM, Joseph Qi wrote:
>> On 2015/8/5 12:40, Ryan Ding wrote:
>>> Hi Joseph,
>>>
>>>
>>> On 08/04/2015 05:03 PM, Joseph Qi wrote:
>>>> Hi Ryan,
>>>>
>>>> On 2015/8/4 14:16, Ryan Ding wrote:
>>>>> Hi Joseph,
>>>>>
>>>>> Sorry for bothering you with the old patches. But I really
need to know what this patch is for.
>>>>>
>>>>>
https://oss.oracle.com/pipermail/ocfs2-devel/2015-January/010496.html
>>>>>
>>>>>   From above email archive, you mentioned those patches aim
to reduce the host page cache consumption. But in my opinion, after append
direct io, the page used for buffer is clean. System can realloc those cached
pages. We can even call invalidate_mapping_pages to fast that process. Maybe
more pages will be needed during direct io. But direct io size can not be too
large, right?
>>>>>
>>>> We introduced the append direct io because originally ocfs2
would fall
>>>> back to buffer io in case of thin provision, which was not the
actual
>>>> behavior that user expect.
>>> direct io has 2 semantics:
>>> 1. io is performed synchronously, data is guaranteed to be
transferred after write syscall return.
>>> 2. File I/O is done directly to/from user space buffers. No page
buffer involved.
>>> But I think #2 is invisible to user space, #1 is the only thing
that user space is really interested in.
>>> We should balance the benefit and disadvantage to determine whether
#2 should be supported.
>>> The disadvantage is: bring too much complexity to the code, bugs
will come along. And involved a incompatible feature.
>>> For example, I did a single node sparse file test, and it failed.
>> What do you mean by "failed"? Could you please send out the
test case
>> and the actual output?
>> And which version did you test? Because some bug fixes were submitted
later.
>> Currently doing direct io with hole is not support.
> I use linux 4.0 latest commit 39a8804455fb23f09157341d3ba7db6d7ae6ee76
> A simplified test case is:
> dd if=/dev/zero of=/mnt/hello bs=512 count=1 oflag=direct &&
truncate /mnt/hello -s 2097152
> file 'hello' is not exist before test. After this command, file
'hello' should be all zero. But 512~4096 is some random data.
I've got the issue.
dd if=/dev/zero of=/mnt/hello bs=512 count=1 oflag=direct
The above dd command will allocate a cluster, but only write 512B, left
512B to cluster size uninitialized.
truncate /mnt/hello -s 2097152
The above truncate is just block aligned, so it will only zero out 4k to
2M.
In my design, I have only considered zeroing out the current allocated
cluster head, and left the tail zeroing out to the next direct io for
performance consideration (we have no need to zero out first and then
write).
So to fix this issue, we should at least zero out the block aligned pad.
But this may be unnecessary in case of continuous direct io. Do you have
any suggestions?
>>
>>> The original way of ocfs2 handling direct io(turn to buffer io when
it's append write or write to a file hole) has 2 consideration:
>>> 1. easier to support cluster wide coherence.
>>> 2. easier to support sparse file.
>>> But it seems that your patch handle #2 not very well.
>>> There may be more issues that I have not found.
>>>> I didn't get you that more pages would be needed during
direct io. Could
>>>> you please explain it more clearly?
>>> I mean the original way of handle append-dio will consume some page
cache. The page cache size it consume depend on the direct io size. For example,
1MB direct io will consume 1MB page cache.But since direct io size can not be
too large, the page cache it consume can not be too large also. And those pages
can be freed after direct io finished by calling invalidate_mapping_pages().
>> I've got your point. Please consider the following user scenario.
>> 1. A node mounted several ocfs2 volumes, for example, 10.
>> 2. For each ocfs2 volume, there are several thin provision VMs.
> Is there many direct io in parallelthat had been tested out?
> About o2net_wq will block reclaim cache issue you mentioned in another
mail. invalidate_mapping_pages() only free the page cache pages that stored
data. It will not affect meta data cache. So that will not wait unlock. Is that
right?
>>
>>>> Thanks,
>>>> Joseph
>>>>
>>>>> Thanks,
>>>>> Ryan
>>>>>
>>>>>
>>>
>>> .
>>>
>>
> 
> 
> .
>