Re: [Qemu-devel] [RFC] Replace posix-aio with custom thread pool

[Anthony Liguori <anthony@codemonkey.ws>]

Andrea Arcangeli wrote:
> On Tue, Dec 09, 2008 at 10:01:34AM -0600, Anthony Liguori wrote:
>   
>> Yes, that's what I'm going to do before committing it.
>>     
>
> I've been hacking on this too, just to push out a full tested patchset
> without the _em badness... problem is if we use more than one thread,
> there's a thread race between lseek and writev, pread/pwrite don't
> have the issue as they store the pos on the kernel stack, and they
> don't pass through the shared file->f_pos. We'd really need
> preadv/pwritev...
>
> To solve this in userland without kernel aio we'd need to open (not
> just dup)

Why not just dup?  I've implemented this and it seems to work.

>  the file in each thread, then the file->f_pos will become
> thread local and we can cache the last lseek value and avoid the lseek
> syscall for contiguous I/O. Or we need to reduce the number of threads
> to 1 per fd (screwing seeking I/O). kernel aio wouldn't have this
> trouble and a single fd/file would be enough, but that would only work
> on linux.
>   

I've also done a preadv/pwritev implementation in userspace using linux-aio.

My current plan is to finish the refactoring, then test out each 
implementation (dup() + lseek + readv vs. linux-aio) to see which one 
performs better.  If they're equal, dup() wins out because it's more 
portable.

Regards,

Anthony Liguori

> --
> To unsubscribe from this list: send the line "unsubscribe kvm" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>