Re: [Qemu-devel] [RFC 1/2] pci-dma-api-v1

[Andrea Arcangeli]

On Sun, Nov 30, 2008 at 10:27:22PM +0200, Avi Kivity wrote:
> Oh okay.  In that case it should be committed with DEBUG_BOUNCE enabled, 
> and that removed when we have proper aio *v.

Exactly.

But before this can be merged the emulation must be solved. Currently
bdrv_aio_cancel can't work safe (with last patch it works safe only
for drivers using the dma api). I guess one way is to move the
emulation to the block-raw-posix.c layer. But in general I'm not
really sure how to best handle this bdrv readv/writev in short-term
merging terms. Perhaps we should just get rid of posix aio and add the
real thing that will handle readv/writev as well even if that will
surely take a bit more than some emulation form in block-raw-posix.c
that could handle a safe cancellation for all read/write/readv/writev.

> Overkill IMO (glibc likely caches mallocs), but can't hurt.

Yes.