Benjamin LaHaise
2007-Oct-11 12:40 UTC
[Ocfs2-devel] Re: [PATCH -mm 0/10][RFC] aio: make struct kiocb private
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. 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. -ben -- "Time is of no importance, Mr. President, only life is important." Don't Email: <dont@kvack.org>.