Re: [PATCH 4/5] aio: vmap ringbuffer

[Kent Overstreet <koverstreet@google.com>]

On Tue, Oct 09, 2012 at 03:32:10PM -0700, Zach Brown wrote:
> > If it is measurable I'll take another stab at using memory from
> > __get_free_pages() for the ringbuffer. That really would be the ideal
> > solution.
> 
> No, then you'll run into high order allocation failures with rings that
> don't fit in a single page.

Not if we decouple the ringbuffer size from max_requests.

This would be useful to do anyways because right now, allocating a kiocb
has to take a global refcount and check head and tail in the ringbuffer
just so it can avoid overflowing the ringbuffer.

If we change aio_complete() so that if the ringbuffer is full then the
kiocb just goes on a linked list - we can size the ringbuffer so this
doesn't happen normally and avoid the global synchronization in the fast
path.

> > The other reason I wanted to do this was for the aio attributes stuff -
> > for return values, I think the only sane way is for the return values to
> > go in the ringbuffer, which means records are no longer fixed size so
> > dealing with pages is even more of a pain.
> 
> Then let's see that, please.

I was starting on that, but then I got sidetracked with refactoring...
:P

> And can we please stop calling them attributes?  They're inputs and
> outputs that change behaviour -- they're interfaces.

Attributes isn't a good name but neither is interfaces, because they
don't exist on their own; they're always attached to some other
interface.

I dunno.

> And no, just for the record, I don't think generic packed variable size
> structs are worth the trouble.
> 
> If we're going to do a generic interface extension mechanism then we
> should put it in its own well thought out system calls, not staple it on
> to the side of aio because it's there.  It's a really crummy base to
> work from.

Not arguing with you about aio, but most of the use cases I have for it
want aio.

So unless we're going to deprecate the existing aio interfaces and make
something better (I wouldn't complain about that!) I do need to make it
work with aio.

Not that I'm opposed to new syscalls passing attributes to sync versions
of read/write/etc, I just haven't started that yet or really thought
about it.