Re: [RFC PATCH 00/22] loop: Issue O_DIRECT aio with pages

[Dave Kleikamp <dave.kleikamp@oracle.com>]

On 02/27/2012 04:27 PM, Zach Brown wrote:
> On 02/27/2012 04:19 PM, Dave Kleikamp wrote:
>> This patchset was begun by Zach Brown and was originally submitted for
>> review in October, 2009.
> 
> Man, it's been a while.  I remembered almost none of the details of this
> work when I read your introductory message.  As I read the patches it
> all came flooding back, though.  Yikes :).
> 
> My biggest fear about this patch series is the sheer amount of very
> fiddly code motion.  I remember spending a lot of time verifying that
> the patches didn't accidentally lose new changes as the patches were
> ported to newer kernels.
>
> Has someone gone through these most recent patches with an absurdly fine
> toothed comb?  The patches that touch fs/direct-io.c and mm/filemap.c
> are the most risky, by far.

I was pretty careful porting these patches, trying not to lose the
effects of any newer changes to the affected functions. It wouldn't hurt
to go through them again, but I am putting it out as an RFC since I
don't think these are ready to merge quite yet.

>> This series was written to reduce the current overhead loop imposes by
>> performing synchronus buffered file system IO from a kernel thread. 
>> These
>> patches turn loop into a light weight layer that translates bios into
>> iocbs.
> 
> It'd be worth including some simple benchmark results, I think.  I
> remember testing with concurrent O_DIRECT aio with fio on a loopback
> device on top of files in each underlying file system.

Actually, some preliminary tests I ran showed a significant slowdown. I
suspect it may have to do with the unconditional setting of O_DIRECT,
but I haven't verified that yet. I'll do further performance analysis,
but I wanted to get some eyes on this in the mean time.

Thanks,
Shaggy