Re: [Xen-devel] [patch 0/6] xenblk: Add O_DIRECT and O_SYNC support.

[Keir Fraser <keir.fraser@xxxxxxxxxxxxx>]

On 5/11/08 09:37, "Jens Axboe" <jens.axboe@xxxxxxxxxx> wrote:

>>> However, frontend driver could not got the write's flag at all.
>>> if we could know the request with O_SYNC flag should be easy to handle,
>>> need not touch filesystem layer.
>> 
>> So what does O_SYNC mean to Linux then? If it's not passed down to the
>> blkdev layer then it can only mean that requests must be synchronously
>> committed as far as that layer. I would therefore imagine you could lose as
>> much data natively as you do running on Xen. Where are these megabytes of
>> data sitting when they get lost, and why does it not happen when running
>> natively? If O_SYNC is not waiting for completion responses from the block
>> layer, that's a limitation of Linux's generic block subsystem, isn't it?
> 
> The block layer doesn't care, it's async by nature. What makes O_SYNC
> work is that callers will wait on the submitted writes afterwards, not
> returning a result until they have reached the drive (that's as far as
> that guarentee goes, beyond that you need flushes or barriers).
> 
> So it is definitely NOT a limitation of the block layer, that's very
> much how it is designed.

The issue at hand is whether Xen's block drivers need to handle O_SYNC
specially in any way. It sounds like it is actually handled properly at a
higher level (e.g. individual filesystems?) and hence the patches proposed
by Oracle make not much sense.

 -- Keir



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel