Re: [Qemu-devel] [PATCH] Align file accesses with cache=off (O_DIRECT)

[Laurent Vivier <Laurent.Vivier@bull.net>]

Le mardi 29 avril 2008 à 18:21 +0200, Kevin Wolf a écrit :
> Laurent Vivier schrieb:
> >> Disabling O_DIRECT for a single aio request is impossible (after all, 
> >> aio is asynchronous), and disabling it for at least one aio request is 
> > 
> > Perhaps I'm wrong, but I think it is possible: the only consequence is
> > the asynchronous I/O becomes synchronous...
> 
> Hm, yes. We could call raw_pread in raw_aio_read when O_DIRECT is used 
> and the request is not properly aligned. Is this what you meant?

No, it was just a (stupid) comment. I think we must not convert
asynchronous I/O to synchronous I/O.

> >> going to be ugly. So maybe we better turn O_DIRECT off for snapsnot 
> >> saving/loading, even if it's not the generic fix I wanted to have when I 
> >> started.
> > 
> > I don't think it is a good idea:
> > 
> > In linux world, there are three reasons to use O_DIRECT:
> > 
> > 1- to use linux AIO (not POSIX AIO).
> > 
> > 2- to avoid a buffer copy between user- and kernel- space
> > (performance ?)
> > 
> > 3- to increase reliability: by using O_DIRECT you are sure your data are
> > on the disk when the write is over and your system can now crash (if it
> > wants).
> > 
> > And I think reliability is better when the snapshot is being saved...
> 
> I think we agree that it's mostly item 3 why one would use O_DIRECT with 
> qemu. In terms of reliability, it is important that the data really is 
> written to the disk when the guest OS thinks so. But when for example 
> qemu crashes, I don't think it's too important if 40% or 50% of a 
> snapshot have already been written - it's unusable anyway. A sync 
> afterwards could be enough there.

I don't speak about "qemu crashes" but about "host crashes".

> 
> >> I'm still undecided, though. What do you think?
> > 
> > Is it possible to align the last AIO ?
> 
> I have to admit that I neither know how to recognize the "last AIO" in 
> the generic code nor do I understand what you want to achieve with it. 
> The main problem are unaligned buffers and these occur on any request.

I'm not in the spirit "my patch is better than yours" (and I don't think
so); but could you try to test my patch ? Because if I remember
correctly I think It manages all cases and this can help you to find a
solution (or perhaps you can add to your patch the part of my patch
about block-qcow2.c)

> > And see comments below
> 
> Good suggestion, will change the patch accordingly.
> 
> Kevin
> 

Regards,
Laurent
-- 
------------- Laurent.Vivier@bull.net ---------------
"The best way to predict the future is to invent it."
- Alan Kay