On Wed, 17 Jan 2007, Benjamin LaHaise wrote:
On Mon, Jan 15, 2007 at 08:25:15PM -0800, Nate Diller wrote:
the right thing to do from a design perspective. Hopefully it enables
a new architecture that can reduce context switches in I/O completion,
and reduce overhead. That's the real motive ;)
And it's a broken motive. Context switches per se are not bad, as they
make it possible to properly schedule code in a busy system (which is
*very* important when realtime concerns come into play). Have a look
at how things were done in the 2.4 aio code to see how completion would
get done with a non-retry method, typically in interrupt context. I had
code that did direct I/O rather differently by sharing code with the
read/write code paths at some point, the catch being that it was pretty
invasive, which meant that it never got merged with the changes to handle
writeback pressure and other work that happened during 2.5.
I'm having some trouble understanding your concern. From my perspective,
any unnecessary context switch represents not only performance loss, but
extra complexity in the code. In this case, I'm not suggesting that the
aio.c code causes problems, quite the opposite. The code I'd like to change
is FS and md levels, where context switches happen because of timers,
workqueues, and worker threads. For sync I/O, these layers could be doing
their completion work in process context, but because waiting on sync I/O is
done in layers above, they must resort to other means, even for the common
case. The dm-crypt module is the most straightforward example.
I took a look at some 2.4.18 aio patches in kernel.org/.../bcrl/aio/, and if
I understand what you did, you were basically operating at the aops level
rather than f_ops. I actually like that idea, it's nicer than having the
direct-io code do its work seperately from the aio code. Part of where I'm
going with this patch is a better integration between the block layer
(make_request), page layer (aops), and FS layer (f_ops), particularly in the
completion paths. The direct-io code is an improvement over the common code
on that point, do_readahead() and friends all wait on individual pages to
become uptodate. I'd like to bring some improvements from the directIO
architecture into use in the common case, which I hope will help
performance.
I know that might seem somewhat unrelated, but I don't think it is. This
change goes hand in hand with using completion handlers in the aops. That
will link together the completion callback in the bio with the aio callback,
so that the whole stack can finish its work in one context.
That said, you can't make kiocb private without completely removing the
ability of the rest of the kernel to complete an aio sanely from irq context.
You need some form of i/o descriptor, and a kiocb is just that. Adding more
layering is just going to make things messier and slower for no real gain.
This patchset does not change how or when I/O completion happens,
aio_complete() will still get called from direct-io.c, nfs-direct.c, et al.
The iocb structure is still passed to aio_complete, just like before. The
only difference is that the lower level code doesn't know that it's got an
iocb, all it sees is an opaque cookie. It's more like enforcing a layer
that's already in place, and I think things got simpler rather than messier.
Whether things are slower or not remains to be seen, but I expect no
measurable changes either way with this patch.
I'm releasing a new version of the patch soon, it will use a new iodesc
structure to keep track of iovec state, which simplifies things further. It
also will have a new version of the usb gadget code, and some general
cleanups. I hope you'll take a look at it.
NATE
-
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
